You are on page 1of 41

Bifurcations in Maps

Juan Alejandro Valdivia Hepp


jvaldivi@uniandes.edu.co
. . . in dedication to Carolina y Nicolas
Contents
1 Concepts in MAPS 3
1.1 Bifurcation diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The attractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Lebesgue measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Basin of attraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.1 invariant measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Path integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Periodic Orbits and Period doubling bifurcation . . . . . . . . . . . . . . . . . . . . 9
1.5 Renormalization group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.1 General Approach to RGT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 Path integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Other situations of interest 17
2.1 Symbolic dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 More than one attractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Higher Dimensions 18
3.1 Map examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.1 Hennon Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.2 prey-predator model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.3 Bakers Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.4 Horse shoe map and non-attracting chaotic sets . . . . . . . . . . . . . . . . 23
3.1.5 The cat map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.6 Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.7 City Trac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Stable and unstable Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1
4 Generic Bifurcations 31
4.1 Generic Bifurcations 1-D and normal forms . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Hysteresis as a non-local bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Non-generic local bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 Generic local Bifurcations N-D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.5 Global bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.6 Discontinuous transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.6.1 Intermittent transitions to chaos . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.6.2 Interior and exterior crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2
1 Concepts in MAPS
There are a lot of models in physics that are based in discrete maps,
x
n+1
= M[x
n
, p]
with p as a controlling parameter. Similarly, there is a straight forward way to describe a set of
dierential equations with discrete maps by using the Poincare surface of section.
For a given initial condition x
0
the limit of x
n
as n is described as the -set of x. In
many situations characterizing the behavior of this system involves determining the possible limit
sets (or asymptotic sets) to which the dynamics can evolve.
1.1 Bifurcation diagram
By changing the parameters p we can observe dierent behavior through a Bifurcation diagram
initiated at x
0
that display the behavior of the limit set of a given initial condition as a function
of the parameter. In general, since in many situations this approach is not uniformly continuous,
interesting things can happen such as discontinuous transitions in the attractor features (size,
disappearance, etc).
One of the most studied equations is the logistic equation that represents the reproduction of a
certain specie for a given xed carrying capacity of the food source
x
n+1
= rx
n
(1 x
n
)
with r as the controlling parameter. As we will see this system shows chaotic behavior with
sensitivity to initial conditions. First, notice that the points in (, 0)U(1, +) map to the negative
values for r > 0 and to for (r > 1). The set [0, 1] maps to itself for (r > 0). We will see that
the range of interest (for 1 r 4) is for x in [0, 1].
The asymptotic state of the system can be computed numerically by starting with an initial
condition x
0
= 0.1 and after removing the transient (say n < 1000) we plot for a given value of
r, the following iterations of the map (say n < 2000). The bifurcation diagram is constructed by
repeating the same procedure for dierent r. For the case in Fig. 1a-b we can see that the limit
set for the same initial conditions converges to a xed point of period one and two respectively.
Figures 1e-f show the bifurcation diagram, and we see in Fig. 1d that the trajectory seems to
become chaotic for r 3.8. This is called a period doubling bifurcation route to chaos.
1.2 The attractor
An attractor A intuitively is a set to which all neighborhood trajectories converge. In general for
the case of maps these sets are (a) xed points, and (b) strange attractors. In the case of dierential
equations, we will see that we can also have limit cycles.
Mathematically, A is the minimal (no smaller set satisfy these conditions) closed invariant set
(if x
0
A x
n
A n > 0) that attracts an open set U of initial conditions. The largest set U
is the basin of attraction of A.
3
Figure 1: (a) The logistic map for r=2.5 and the zig-zig line represent the trajectory of the x
0
= 0.1.
(b) The same but for r = 3.2 representing a orbit of period 2. (c) The trajectory of the map f
2
for
r = 3.2. (d) The logistic map for r = 3.5 a chaotic case. (e) The bifurcation diagram. (e)A zoom
of the bifurcation diagram.
4
Figure 2: (a) The bifurcation diagram
5
Numerically, we start with a set of initial conditions and see numerically whether they converges
to a xed point for a given number of iterations. If there is a set of points that do not converge
to the already observed xed points, we initially assume that we have a strange attractor SA. Of
course this is not necessarily true, for we may have a nonatracting set and those points just take a
long time to go to one of the previous attractors, so we must be sure that the number of points that
stay in the SA does not goes down in time. Dierent attractors have dierent basin of attractors.
Strange attractors are attractors in which there is sensitive dependence on initial conditions.
Sometimes are referred as fractal attractors or chaotic attractors.
In Fig. 1 the attractor is obtained from the initial condition x
0
= 0.1, but in principle there
could be more attractors than the one described in the bifurcation diagram, since in principle we
started with one initial condition, and other initial conditions may converge to other attractors.
In the bifurcation diagram displayed above, we see two possible attractors, the one at and
another one in [0, 1] for r > 0.
Numerically, in the case of multiple attractors we may have to take an ensemble of initial
conditions and construct a bifurcation diagram for each. If the attractors are well separated, then
the bifurcations diagram can be superposed, but if the attractors are somehow interwoven, then it
may be more dicult to separate.
1.2.1 Lebesgue measure
For the case shown in Fig. 1, almost all (in the Lebesgue sense) initial conditions converge to
the attractor shown. The concept of the Lebesgue is important and necessary, since it is obvious
that if we start with an initial condition at an unstable periodic orbit, then the trajectory does not
evolve towards the attractor. But this set is countable and has Lebesgue measure zero, and does
not aect the statement above.
HOMEWORK PROBLEM: show numerically that almost all initial conditions con-
verge to the same attractor
1.3 Basin of attraction
The closure (see comment on Lebesgue measure below) of the set of initial conditions that converge
to a given attractor is dened as its basin of attraction. In the case of the logistic map, we have
the trivial attractor at with basin [, 0]U[1, +]. The other attractor is in [0,1] with basing
of attraction [0, 1].
Numerically we start with an ensemble of initial conditions and follow their evolution. In
principle some initial conditions will converge to a given set of xed points (i.e., trajectories that
repeat themselves after some period. If after a certain number of iterations, the trajectory does
not converge to a xed point, we dene it initially as part of a strange attractor. The set of initial
conditions that dont converge to a xed point form the strange attractor. First, there may be
more that one, and second, care must be taken that the set is not a non-attracting chaotic set by
observing if the number of points in the strange attractors does not go down in time.
6
1.3.1 invariant measure
It is possible to assign a weight, or density (in the Lebesgue sense), to the points in phase space.
Suppose we assign a certain initial density
0
(x) to each point in phase space, with
_

n
(x)dx = 1
then we can evolve that density using

n+1
(x) =
_

n
(y)(x M[y])dy =

n
(y
(i)
)
|M

(y
(i)
)|
(1)
where M(y
(i)
) = x. An invariant density is one that satises
n+1
(x) =
n
(x) = (x). In
practice (x) is usually a very discontinuous function, and in many situations is very dicult to
dene. For that it is more customary to dene a measure . The measure of a set S is described
by (s) =
_
S
(x)dx and the average of a function by
_
f(x)(x)dx =
_
f(x)d(x)
The measure is smoother that the density , so it is more manageable. It is possible to
estimate the relevance of each attractor in phase space by measuring the measure of its basin of
attraction (starting with
0
(x) = 1 in phase space).
0 0.2 0.4 0.6 0.8 1
x
0.0001
0.001
0.01
0.1
1

0 0.2 0.4 0.6 0.8 1


x
0.0001
0.001
0.01
0.1
1

Figure 3: The analytical density for r=3.25 and r=3.8 using Eq. 1 for n = 1 (red line), n = 5 (blue
line), and n = 10 (green line), assuming
0
= 1. The black line corresponds to the the density
constructed in time after n = 10
6
iterations using x
0
= 0.1. It is possible to show that this is indeed
the natural invariant density.
Notice, that we have looked from the point of view of phase space, we will now look from the
point of view of time. Lets start with an initial condition x
0
= 0.1 in the basin of attraction of
an attractor, and compute the fraction of time the trajectory spends in an interval S, denoted by
(S, x
o
). In Fig. 3 we display the density (x, x
0
) that is generated by the fraction of time the
trajectory spends around x. If (S, x
o
) is the same for almost all the points in x
0
in the basin of
attraction then we denote it by the natural measure (S). For smooth 1-D maps natural measures
usually exists, but in higher dimensional systems is an open question. This implies that the measure
is ergodic since it satises (only in the basin of attraction B)
7
< f >
t
= lim
N
1
N
N

n=0
f(M
n
[x
0
]) =
_
B
f(x)(x)dx =< f >
x
for any function f(x) and almost any point x
0
B.
1.3.2 Path integrals
Suppose we start with an initial density
0
(x
0
) in phase space (essentially it accounts for the fact
that we cannot determine our initial conditions with innite precision). Then starting from
s
n+1
= M(s
n
) +
n
with
n
as a random noise, then the probability of a particular sequence [x
0
, x
1
, x
2
, . . . , x
n1
] can
be written as
P[x
0
, x
1
, . . . , x
n
] = [x
1
s
1
] . . . [x
n
s
n
]
from an initial condition x
0
, where the average is over the noise sequence
n
. Using
[x] =
1
2
_
dke
ikx
we can rewrite
P[x
0
, x
1
, . . . , x
n
] =
_
1
2
_
n
_
[
n
m=1
dk
m
]

e
i
P
n
m=1
km(xmsm)
_
=
_
1
2
_
n
_
[
n
m=1
dk
m
]

e
i
P
n
m=1
km(xmM[x
m1
]
m1
)
_
given that x
m
= M(x
m1
) +
m1
and x
m1
= s
m1
.
Usually the dierential DK =
n
m=1
dk
m
describe all possible paths connecting this specic
sequence from x
0
and x
n
. It is interesting to compare this formalism with the propagator in
quantum mechanics. We see that
n
(x
n
) is just the phase space average of this probability over all
intermediate sequences [x
0
, x
1
, x
2
, . . . , x
n1
] with initial probability
0
(x
0
).
The probability that connects an initial condition x
0
with a nal condition x
n
is
P[x
n
] =
_

n1
m=0

0
(x
0
)P[x
0
, x
1
, . . . , x
n
]
and the phase space density at the nth step is then

n
(x
n
) =
1
Z
n
_
dx
0

0
(x
0
)
_

n1
m=1
dx
m
dk
m

dk
n
_
e
i
P
n
m=1
km(xmM[x
m1
]
m1
)
_
where we have dened the partition function at the nth step as
Z
n
=
_
[
n
m=1
dx
m
dk
m
]
0
(x
0
)
_
e
i
P
n
m=1
km(xmM[x
m1
]
m1
)
_
This partition function can be use to describe any system, even away from equilibrium. An equi-
librium is reached if an invariant density is obtained in some basin. From the partition function,
8
using functional derivation, dynamical variables for the system can be obtained. But that is a topic
for another book.
The probability of a phase space variable of function as a function of time can be calculated
now as
< f >
n
=
_
f(x)
n
(x)dx
Lets note that this equations can be derived from Eq. 1 using
(x M[y
n
]) =
_
e
ikn(xM[yn])
dk
n
This path integral formalism is useful when the noise distribution is know, so that

e
ik
n+1
(x
n+1
M(xn)n)
_
can be evaluated. For example
P[] =
1

2
e

e
b+a
_
= e
b

a
2
4

HOMEWORK PROBLEM: For the logistic map for r = 3.8 start with
0
(x) = 1, what
happens to
n
as n ?. Is this an invariant measure? Is it equal to the invariant
natural measure?
HOMEWORK PROBLEM: Evolve
n
with
0
(x) = 1 using Eq. 1 analytically
HOMEWORK PROBLEM: For the map z
n+1
= z
2
n
+(0.341, 0.341) construct the basin of
attraction of innity in the complex plane. The basin boundary should look fractal.
Estimate the dimension of the basin boundary using boxes of size . Help: the number
of boxes that cover the basin boundary diverges as N
D
as 0 (see below).
HOMEWORK PROBLEM: For the identity map, and a Gaussian independent noise,
calculate the partition function and the probability density for a few n. Repeat the
same analysis for the logistic map for a few values of r and n
1.4 Periodic Orbits and Period doubling bifurcation
As seen in the gures above, the dynamics seems to be controlled to a certain extent by the stability
properties of the periodic orbits of the map. The periodic orbits of period m can be calculated
from the xed points of M
m
given by
M
m
(x

) = x

9
Lets take the logistic map and study the details of the bifurcation diagram. The period doubling
bifurcation to chaos seems to be a continuous bifurcation. In the bifurcation diagram we observe
windows of periodicity inside the chaotic range, where the transition to the periodic windows seems
to occur discontinuously. These are the discontinuous transitions mentioned above and that will
be studied in detail below.
0.5 1 1.5 2 2.5 3 3.5 4
r
0.2
0.4
0.6
0.8
1
x
0.5 1 1.5 2 2.5 3 3.5 4
r
0.2
0.4
0.6
0.8
1
x
3.2 3.4 3.6 3.8 4
r
0.2
0.4
0.6
0.8
1
x
3.82 3.84 3.86 3.88 3.9
r
0.2
0.4
0.6
0.8
1
x
Figure 4: The stability of the period (a) m=1, (b) m=4, (c) zoom of m=4 and (d) m=3 orbits.
The period m = 1 orbits are dened by the solutions of
M(x

) = x


_
x
(1)
0
= 0
x
(2)
0
= 1
1
r
and we still need to clarify the stability of them. The stability of a periodic m = 1 orbit x

can be
constructed from all possible perturbations x
n
= x

+
n
, then
x
n+1
= x

+
n+1
= M(x

+
n
) M(x

) +
dM(x

)
dx

n

n+1

dM(x

)
dx

n
therefore, the periodic orbit is attractive if |
n+1
| < |
n
|. Furthermore, the trajectory will converge
at a rate

n
e
n

0
with
ln|
d(M(x

)
dx
|
10
In this particular case, we have

0
=
_

(1)
0
= ln|r|

(2)
0
= ln|2 r|
respectively. Therefore, the x
1
is an attractive xed point for r < 1 while x
2
is an attractive xed
point for 1 < r < 3. Otherwise they are repulsive.
The stability of a periodic m orbit x
0
, . . . , x
m1
can be constructed from all possible perturba-
tions x
n
= x

+
n
, then
x
n+m
= x

+
n+m
= M
m
(x

+
n
) M
m
(x

) +
d(M
m
)
dx

n

n+m

d(M
m
)
dx

n
=
d(M(x
0
)
dx
...
d(M(x
m1
)
dx

n
therefore, the periodic orbit is attractive if the product in absolute value is less than one. Further-
more, the trajectory will converge at a rate

n
e
n

0
with

1
m
m1

n=0
ln|
d(M(x
n
)
dx
|
Figure 4 shows the behavior of a few of these orbits. We see that in the period doubling
bifurcation at the point at which the periodic m-orbit looses stability, an stable and unstable period
2m-orbit appears. Notice that the other 2m-orbits are still there but are unstable. Obviously the
M
2m
also describe the M
m
orbits as it should. This is a very important point. Take notice.
It is interesting to note that we could dene the stability criteria of a m periodic orbit, i.e.,
something that is non-periodic. In essence all the periodic orbits are there, but are unstable. The
chaotic attractor include the set of all unstable periodic orbits. We dene the Lyapunov exponent
as the innite limit
Lim
N
1
N
N

n=0
ln|
d(M(x
n
)
dx
|
and due to the ergodicity of the system, it can also be computed from

_
ln|
d(M(x)
dx
|d(x) =
_
ln|r(1 2x)|d(x)
If the orbit is periodic, we naturally obtain the previous result for the stability of the orbit. Figure
5 shows the Lyapunov exponent for the logistic map.
Note that in the period m=4, the other period 4 orbit becomes stable at r a little less that r,
but through a bifurcation that has a dierent nature from the period doubling. We will study these
bifurcations below.
11
HOMEWORK PROBLEM: repeat the whole analysis for the tent map x
n+1
= r (0.5
|x 0.5|), including attractors, bifurcation diagrams, Lyapunov exponents.
HOMEWORK PROBLEM: repeat the whole analysis for the tent map x
n+1
= r(1x
2
),
including attractors, bifurcation diagrams, Lyapunov exponents.
HOMEWORK PROBLEM:For the logistic map show that the two methods for com-
puting the Lyapunov exponents give the same results, hence the system is ergodic.
1.5 Renormalization group
From the map of the logistic map we can characterize the universal behavior of the period doubling
bifurcation. This universal behavior occurs for all maps with a single quadratic maximum. Lets
do the logistic map and then we can generalize. From the Fig 6 we can observe that the m and
(m + 1) period doubling bifurcations look the same except for a zooming factor and a shifting.
Lets dene the m-super-stable xed point at r
m
when dM
2
m
(x

, r
m
)/dx = 0. It is found that

m
=
r
m
r
m1
r
m+1
r
m


= 4.669201 . . .
is universal for all maps with a single quadratic maximum. This implies that |r
m
r
c
|

.
Numerically it is found that r
c

= 3.57 . . .. There are other universal quantities. When the stable
xed point of M
m
(x

) become super-stable (dM


m
(x

, r
m
)/dx = 0), we can estimate the value

m
= M
2
m1
_
1
2
, r
m
_

1
2

m
=

m

m+1


= 2.50280 . . .
We will use a renormalization group transformation (RGT) to estimate these values (see
Fig. 7). The RGT is a coarse graining transformation that in essence transform the map M
2
m
(x)
to the map M
2
m+1
(x). Notice from Fig. 6 that the map close to the supercritical xed points of M
2
at r
2
looks exactly the same as the map M at r
1
, except for a coordinate transformation (actually
a ip and a zooming) which can be described by
y
1
2
=
_
x
1
2
_
and we expect to be larger than one and negative. We therefore, expect
x
n+2
= M
2
(x
n
, r
m+1
)
y
n+1
= M(y
n
, r
m
)
12
Figure 5: (a) The Lyapunov exponent of the logistic map.
0.2 0.4 0.6 0.8 1
x
0.2
0.4
0.6
0.8
1
f
r1=2
0.2 0.4 0.6 0.8 1
x
0.2
0.4
0.6
0.8
1
f
r1=2
Figure 6: (a) The super-stable xed points for M and M
2

1
Figure 7: (a) Gracal description of the parameters
13
which can be written in the general case as
M [y, r
n
]
1
2
=
_
M
2
_
1
2
+
_
y
1
2
_

, r
n+1
_

1
2
_
The can construct the transformation r
m
= R(r
m+1
) and = G(r
m+1
) which are called the RGT
which do the zooming and shifting, and permits to transform from the r
m
to the r
m+1
super-stable
points down the period doubling bifurcation. This transform from the m period doubling to the
m + 1 period doubling, and following this way we approach to chaos at m . This coarse
graining transformation is enough to determine the behavior of the period doubling bifurcation.
In general is usually very dicult to carry a RGT exactly, which means that we have to resort
to some approximation. Notice that this transformation only makes sense close to the super-stable
xed points which permits us to do the analysis close to x 1/2. We rewrite the logistic map as
M(x, r) = r
_
1
4

_
x
1
2
_
2
_
(2)
It is left to the reader to work out the transformation of other similar maps proving that the results
are in essence universal. We then have using Eq. 2 that
M
2
(x, r) =
r
4

_
r
4

1
2
_
2
+ 2r
2
_
r
4

1
2
__
x
1
2
_
2
+O(
_
x
1
2
_
4
)
Therefore, following the above prescription we have obtained 3 solutions for r
m
= R(r
m+1
) and
= G(r
m
). These solutions give a nontrivial xed point (of the R
r
transformation) only in one
case, namely
R(r) = 1 +
1
8
_
64 + 8r
2
(r 2)(8 4r
2
+r
2
)
where the xed point of this new transformation corresponds to the saturation point r
c
= 1 +
_
3 + 2

3

= 3.54 . . . which is close to the numerical value. This values is not universal, and
depend on the particular system. At the same time, = G(r
c
) 2.732.. corresponds to the
Feigenbaum number.
Now, lets estimate . First notice that close to the xed point of the RGT (i.e., close to the
saturation point of the period doubling bifurcation) the variable
m
= r
m
r
c
represent the only
relevant variable, and we can conclude that

m
= R(r
m+1
) r
c
= R(
m+1
+r
c
) r
c

= (4 +

3)
m+1
+O(
2
m+1
)
hence
m
(4+

3) = 5.73, which is ok due to the approximation we made initially. Furthermore,


lets dene N() = N(r r
c
) as the number of period doubling at r (notice that N depends only
on r r
c
). Clearly, we have
N(
m
) = 2N(
m1
) = 2N(
m
) = 2
n
N(
n

m
)
14
from this relationship we estimate N since it applies for all n. Then chose an n, such that
n

m
1
(n = ln
2

m
/ln
2
), then
N(
m
) = 2
n
N(1) = 2
ln
2
2/ln
2

N(r)
1
|r r
c
|

with = 1/ln
2
= 0.396 . . ..
1.5.1 General Approach to RGT
We discuss rapidly the general RGT, which has a much more general applicability.
The renormalization group (RG) approach is a method to describe the behavior of scaling in
a system that has some inherent symmetry. Furthermore, within the RG approach the presence
of scaling usually represents some kind of critical behavior as we will see below. The RG method
usually starts with a set of equations, that can include dynamical equations,
f(x, a)
described by a set of variables x and a set of parameters a. We make explicit use of the symmetry
to make a scaling transformation that transform the variables and parameters as
x


= R
x
(x)
a


= R
a
(a)
so that the equation is left (almost) invariant
f(x, a) f(x

, a

)
The interesting point, is that the scaling properties are completely described by the transfor-
mation equation of the parameters, as it should be, and the critical behavior can be completely
characterized.
Suppose that the transformation a

= R
a
(a) is repeated many times, then the system will
approach a xed point in the parameter space dened by
a
c
= R
a
(a
c
)
If the initial system start with those parameters it becomes invariant under those transforma-
tions and we say that the system is at the critical state. Of course there are trivial xed points
like at innity or zero. But in general if there is critical behavior there will be nite critical points
(or xed points). Furthermore, such a xed point in general have stable and unstable manifolds.
The stable manifold described the criticality of such a system, which means that a system on
the stable manifold approaches the critical point under the RG transformation. In general, many
critical points will describe dierent critical or scaling behaviors of the system.
Around a xed point a
c
we can linearize the transformation R
a
L
a
to characterize the
stable (
i
< 0) and unstable (
i
> 0) manifolds by the respective eigenvectors V
i
with eigenvalue

i
. Since any a =

a
i
V
i
then we note that there are only a few relevant directions described by
the dimension of the unstable manifold.
15
Any physical macroscopic parameter of the system P(a) can now be described in this scaling
region by the behavior of the relevant directions, i.e., (like generalized coordinates)
P(a)

= P(V
1
, V
2
, ...V
N
)
with N as the dimension of the unstable manifold. It is important to stress that this physical
variable transform under the RG transformation and it depends on the system under consideration
and the type of the transformation. In equilibrium statistical mechanics we are usually interested
rescaling the variables by a factor with the correlation length and the free energy F transforming
as
F(V
1
, V
2
, ...V
N
)
nd
F(V
1

n
1
, V
2

n
2
, ...V
N

n
N
)
(V
1
, V
2
, ...V
N
)
n
(V
1

n
1
, V
2

n
2
, ...V
N

n
N
)
(3)
since F is an extensive variable (Careful if we have non-extensive entropy) and clearly decreases as
we do the transformation of the d variables in x. Both F and gre invariant under the transformation
and the transformed variables x represent a smaller system. From this relation we can get rid of
one variable and obtain the scaling relations since n is arbitrary, i.e.,
V
1

n
1
= 1 n =
ln V
1
ln
1
F(V
1
, V
2
, ...V
N
) V
d
ln
ln
1
1
F(1,
V
2
V
ln
2
ln
1
1
, ...
V
N
V
ln
N
ln
1
1
)
(V
1
, V
2
, ...V
N
) V

ln
ln
1
1
(1,
V
2
V
ln
2
ln
1
1
, ...
V
N
V
ln
N
ln
1
1
)
which reduces to the scaling since what it inside the function are all constant but in general describe
macroscopic variables like temperature T and magnetic elds H as in the case of spin systems.
Figure 8: The RG transformation in parameter space
In Fig. 8 we show the RG transformation in parameter space, with multiple xed points. Usually
symmetry breaking occurs as we change from the neighborhood of one xed point to another.
16
1.6 Path integrals
It is interesting to note that this RG approach can be done directly using the path integral formalism
described above for P[x
0
, x
1
, . . . , x
n
] by
integrating every other term in the sequence
rewriting the result as it look before the summation, but with a renormalized function for r.
2 Other situations of interest
2.1 Symbolic dynamics
Take the map
x
n+1
= 2x
n
mod2
we can dene a point by its binary representation
x
0
= 0.a
1
a
2
a
3
. . . =

j=1
2
j
a
j
with a
i
= 0, 1 which applied to the map gives
x
1
= 0.a
2
a
3
. . . x
n
= 0.a
n+1
a
n+2
. . .
and it is called a binary shifting. It is then very easy to study periodic orbits in this system. Notice
that the xed points are all unstable since = ln2 and in fact they are dense in the interval [0, 1].
We then clearly can nd that the natural invariant density must be
(x) = 1 x[0, 1]
since the unstable periodic orbits have measure zero.
2.2 More than one attractor
In many situations we have more than one attractor. Lets study the map shown in Fig. 9c.
This is constructed in Mathematica by connecting points, namely connecting crossing points at
x = 0, 1/5, 2/5, 3/5, 4/5, 1 with lines. There are two clear attractors, namely x = 2 and x = 1,
which are attractive. The basin of attraction for each attractor is clearly fractal dening something
similar to a cantor set with nite structure at all scales. The basin of attraction is not a cantor set,
for it has dimension 1/2. But the basin boundary has a fractional dimension.
HOMEWORK PROBLEM: Estimate and and and r
c
numerically for the logistic
map.
17
-1 1 2
-1
-0.5
0.5
1
1.5
2
0.2 0.4 0.6 0.8 1
x
0.2
0.4
0.6
0.8
1
Figure 9: (a) The map and (b) and fractal basin boundary where the color determines which
attractor it belongs to.
HOMEWORK PROBLEM: Repeat the RGT for the logistic map, but by mapping
M
2
M
4
, and show you obtain similar results.
HOMEWORK PROBLEM: Estimate and and and r
c
numerically for the map
r(1 x
2
) and check the universality.
HOMEWORK PROBLEM: word out the RGT for the map M = r (1 x
2
) and nd
and . What are the values of for the map r (1 x
2p
)?
HOMEWORK PROBLEM: Try to do the RGT for the logistic map but using the path
integral approach
HOMEWORK PROBLEM: Estimate the dimension of the fractal basin boundary of
the map shown in Fig. 9
HOMEWORK PROBLEM: By mapping the logistic map for r = 4 to the map used
for symbolic dynamics, construct the density (x) in analytical form.
HOMEWORK PROBLEM: Repeat Fig 9 but including also the time it takes to ap-
proach the attractor. Construct the probability function for the time it takes to
approach the attractor. Calculate the dimension of the boundary of the basin.
3 Higher Dimensions
The stability of a xed point x

can be constructed from all possible perturbations x


n
= x

+
n
,
then
x
n+1
= x

+
n+1
= M(x

+
n
) M(x

) +D(M)
n

n+1
DM(x

)
n
18
therefore, the xed point is attractive if all the eigenvalues |w
i
| < 1. Notice that in this case
the eigenvalues can have an imaginary value, and therefore the stability criteria must be that the
absolute value must be less than one, in amplitude. For imaginary eigenvalues the orbits spiral in
or out of the xed point in the direction of the stable or unstable manifold perpendicular to the
two eigendirections corresponding to the complex eigenvalues. In the case of continuous systems,
it becomes extremely relevant the fact that we have complex eigenvalues, for it usually signal the
appearance of a quasiperiodic orbit when the xed point losses stability.
We can see that points that start close-by, begin separating at the rate of the largest eigenvalue.

n
w
n
i

0
= e
n

0
If we are interested in the looking at the stability of a xed point, we are only required to follow
any perturbation, which will eventually converge to the most unstable direction, or largest .
The stability of a periodic m-orbit x
0
, . . . , x
m1
can be constructed from all possible perturba-
tions x
n
= x

+
n
, then
x
n+m
= x

+
n+m
= M
m
(x

+
n
) M
m
(x

) +D(M
m
)
n

n+m
D(M
m
(x
n
))
n
= DM(x
m1
)..DM(x
0
)
n
therefore, the periodic orbit is attractive if all the eigenvalues |w
i
| < 1. Notice that includes the
factor m
1
.
It is interesting to note that we could dene the stability criteria of a m periodic orbit,
i.e., something that is non-periodic. Again we could start with a perturbation in any direction and
the evolution will take it towards the most unstable direction. But if we are interested in computing
the whole spectrum of Lyapunov exponents, as they are called, we have to resolve to a trick since
diagonalization of a large number of matrices is very unstable numerically. Also the stable and
unstable directions are changing as the trajectory progresses. Instead we do a QR decomposition
DM(x
n
)Q
n1
= Q
n
R
n
with Q
0
= I, Q as an orthogonal matrix, and R as an upper triangular matrix. This implies
numerically
D(M
n
) = Q
n
R
n
R
n1
. . . R
1
hence

i
= lim
n
1
N
N

n
ln|(R
n
)
i,i
|
because the R
n
are upper triangular matrices. In essence, the QR decomposition accounts for the
rotation of eigendirections and the numerical instability. Note that in the case of a periodic orbit,
this Lyapunov spectrum is exactly equal to the one computed above due to the periodicity in x
n
.
Note also that the convergence of the above spectrum is guaranteed by the Ergodic theorem for
very general systems.
19
It is very easy to discover that the QR decomposition is in general correcting for the rotation of
the manifolds. For example, lets take a 2-D system and we evolve two unitary vectors that are
initially perpendicular. Eventually these two vectors become v
1
and v
2
, which are not perpendicular,
and have grown in size if one of the Lyapunov exponents is positive. At this point it is convenient
to renormalize to two unitary vectors that are perpendicular to each other, namely,
v
1
=
v
1
|v
1
|
v
2
=
v
2
v
2
v
1
v
1
|v
2
v
2
v
1
v
1
|
or
v
1
= |v
1
| v
1
+ 0 v
2
v
2
=
v
2
v
1
|v
1
|
v
1
+

v
2

(v
2
v
1
)v
1
|v
1
|
2

v
2
The transformation can be rewritten in matrix form as
(v
1
, v
2
) = ( v
1
, v
2
)
T
_
|v
1
|
v
2
v
1
|v
1
|
0

v
2

(v
2
v
1
)v
1
|v
1
|
2

_
which is precisely what we have above DM(x
n
)Q
n1
= Q
n
R
n
. We see that |v
1
| corresponds to the
maximum Lyapunov exponents, and that

v
2

v
2
v
1
|v
1
|
2

corresponds to the 2nd Lyapunov exponents.


For the general case, the QR decomposition is precisely this gram-Schmitt diagonalization in n
dimensions.
We could also construct the maximum Lyapunov exponents numerically, by following 2 trajec-
tories v
1
= x
n
and v
2
= x
n
+
n
separated by a small initial condition
0
. But we now that as
time increases, the separation of the two trajectories also increases up to the saturation size of the
attractor. Hence, every m steps we must re-normalize the separation of the two trajectories to
v
2
= x
m
+
0

m
|
m
|

i
=
|
m
|

0
and continue integrating v
1
and v
2
. The Lyapunov exponents is then the average of this quantity

1
=
1
Nm
N

i=1
log
i
If we want to compute the 2nd Lyapunov exponents this way, then we must follow 3 trajectories
v
1
, v
2
, v
3
such that the separation between v
1
and v
2
and the separation between v
1
and v
3
are
perpendicular. Every m steps, we 1st Lyapunov exponent is obtained the same manner as above
from v
1
and v
2
. The 2nd Lyapunov exponent is obtained by looking at the growth of the projection
of the vector v
3
v
1
in the direction perpendicular to v
2
v
1
. This is equivalent to looking at the
size of the area A
m
spanned by the vectors v
2
v
1
and v
3
v
1
(see Fig. 10) which is related to the
2nd Lyapunov exponent by
20

1
+
2
=
1
Nm
N

i=1
log A
m
But this is precisely the QR decomposition of the vectors v
2
v
1
and v
3
v
1
. The extension to the
other exponents and more dimensions is now straight forward. For example the renormalization of
volumes V
m
give

1
+
2
+
3
=
1
Nm
N

i=1
log V
m
and so on.
To do this in a general manner, the vector product and the determinant of a matrix may be useful
for calculations here.
Figure 10: Evolution of a small volume in 2-D space
Note that these formulations allow us to construct locally (at each point in phase space) stable
and unstable directions with local Lyapunov exponents. It is interesting to note that each local
Lyapunov exponent does not in principle has to converge to the global exponent for two reasons:
(a) the system may not be ergodic, (b) the orientation of the stable-unstable directions changes
over the attractor.
3.1 Map examples
Notice that the attractor usually has a volume (or a Lebesgue measure) smaller than its basin
of attraction, this gives rise to the concept of fractals, with dimension smaller that the original
volume.
3.1.1 Hennon Map
A typical 2-D map is the Hennon map, given by
x
n+1
= A x
2
n
+By
n
y
n+1
= x
n
21
Note that this map is in general dissipative for a range of parameters, hence there must be some
negative Lyapunov exponents. We can visualize this map with a bifurcation diagram as shown in
Fig. 11. Even though the attractor seems to be broken (composed of many parts), this is an eect
of the numerical procedure. We could take a square of initial conditions and iterate forward in
time, and because the map is continuous, the attractor should be connected. In essence, we can
see that the orientation of the stable and unstable directions changes along the attractor (in phase
space) forcing the system to have the shape shown in the picture (with one stable and one unstable
direction).
Figure 11: (a) A bifurcation diagram for the Hennon map for B = 0.3. (b) The attractor for
A = 1.4.
3.1.2 prey-predator model
Lets take a prey-predator model using a generalization of the logistic map
x
n+1
= r
a
x
n
(1 x
n
)(1 A
a
y
n
)
y
n+1
= r
b
y
n
(1 y
n
)(1 A
b
(1 x))
with the obvious parameters. This map will be useful later on.
A variant of the predator-pray model is to take the logistic map but with r changing from
period to period. For example, r
n
= A and r
n+1
= B (period two situation). The results are
shown in Fig. 12 for the Lyapunov exponent in the AB parameter space. We see that by giving
dynamics (period 2) to the parameter in a system we can change considerably the result. This is
why predator-pray models are so interesting.
3.1.3 Bakers Map
Another well know map is the Baker map, namely
22
3 3.2 3.4 3.6 3.8 4
a
3
3.2
3.4
3.6
3.8
4
b
Figure 12: The Lyapunov exponent of the variant of the logistic map in the A B space.
x
n+1
=
_

a
x
n
y
n
<
(1
b
) +
b
x
n
y
n
>
y
n+1
=
_
y
n
/ y
n
<
(y
n
)/ y
n
>
with + = 1 and
a
+
b
1. This is dissipative system that is very useful to study.
3.1.4 Horse shoe map and non-attracting chaotic sets
The complex horse shoe dynamics are useful to study the chaotic dynamics in non-attracting
sets, usually with the use of a symbolic description (see an example above). We show in Fig.
13 one iteration of the map, which corresponds to the set M(S) = V
1
V
2
. The inverse set
M
1
[M(V
i
] = H
i
is also shown. In the same manner we can construct the sets M
m
(S) (vertical
lines) and M
m
[M
m
(S)] (horizontal lines). If we continue the iterating the map, we can clearly
see that the set that stays forever has measure zero. But there is an invariant set that stays
forever, and this set is a chaotic non-attracting set. This set can be obtained by the successive
intersections of

m=1
M
m
[M
m
(S)]

m=1
M
m
(S)
Therefore, we can have chaos in sets that are not attractors.
3.1.5 The cat map
Sometimes it is useful to work with the following map
_
x
n+1
y
n+1
_
=
_
1 1
1 2
_
mod 1
23
V
1
V
2
H
2
H
1
H
2
H
1
V
1
V
2
Figure 13: (a) Horse shoe map, and its inverse. (b) The sequence to construct the invariant set.
(c) The equivalent dynamics by extending stable and unstable manifold in a Homoclinic Tangency.
24
3.1.6 Hamiltonian systems
In a conservative map we have volume preservation, which requires that the Jacobian
Det[DF[x]] = 1
If the map is volume preserving, it does not have an attractor since volume cannot decrease in
phase space. Hence the sum of the Lyapunov exponents must be zero,

i
= 0.
An example is a discretization of a Hamiltonian system with a Poincare surface of section
(or stroboscopic map). Hamiltonian systems have also the symplectic property that forces the
Lyapunov exponents to occur in pairs that cancel each other. A good example is the circle map
(Fig. 14) generated by a kicked pendulum without gravity generated by the Hamiltonian system.
H =
p
2
2
+

n
Cos(t nT)
p
n+1
= p
n
+Sin
n+1

n+1
=
n
+p
n
T
Figure 14: (a) The pendulum of the kicked rotator.
We can nd the xed points of the M map
p = 2n
= m

=
1
2
_
2 +(1)
m+2n

_
+ 4(1)
m+2n
_
hence in the range 0 < < 4 we have an elliptic point at = n and a hyperbolic xed point at
= 0. For > 4 the elliptic point becomes an hyperbolic point. Because of the elliptic nature of a
xed point, the orbits must spend a long time close to it, generating what is called an island. It is
like an attractor, but since there are not attractors, || = 1, the trajectory eventually must move
on. This is shown in Fig. 15b, where this particular island is clearly dened. Note that for = 2
it is still alive. Therefore, an orbit spends a long time close to these islands, with sporadic motion
between the islands. People have develop theories based on fractal integrals and derivatives for the
dynamics on these islands.
Hamiltonian systems will be studied in more detail later on.
25
3.1.7 City Trac
A single car moving through a sequence of N trac lights. The velocity dynamics between trac
lights is
dv
dt
=
_
a
+
(v
max
v) accelerate
a

(v) brake
The states between trac lights are:
accelerate with a
+
until reaching cruising speed v
max
constant cruising speed v
max
until decision point x
n
= L
n

v
2
max
2a

sin(
n
t
n

n
) =
_
> 0 green
< 0 red
a

braking while red


It generates a 2D map from light to light (T
n+1
, v
n+1
) = M[T
n
, v
n
]. Sometimes, it is possible to
normalize with respect to cruising time T
c
=
L
vmax
A
+
=
a
+
L
v
2
max
= 10 A

=
a

L
v
2
max
= 30 = T
c

= /2
3.2 Stable and unstable Manifolds
For a xed point, the eigendirections E
i
dened by DM(x

) dene a tangent space T


x
at x

. This
tangent space can be decomposed in stable E

(|w
i
| < 1), unstable E
+
(|w
i
| > 1) and centered
E
0
(|w
i
| = 1) subspaces or directions. If we continue the stable directions we obtain the stable
manifold W
s
(x

) of the xed point x

W
s
(x

) = {xU|Lim
n
M
n
(x) x

, with x
n
U nZ
+
}
Similarly if we continue the unstable directions we construct the unstable manifold of the periodic
orbit. Continuing the centered directions we obtained the center manifolds of the periodic orbit.
The manifold have the same dimension as the linear subspace, and are tangent to them. There
is a small issue related to the degeneracy of the eigenvalues, in which case the contraction (or
expansion) may not be exponential.
Notice that these sets are by construction invariant in the sense that M(W
s
(x

)) = W
s
(x

).
Numerically is easier to construct the unstable manifold of a xed point than stable manifolds,
specially in higher dimensions. Also, it may be dicult to construct the stable manifold if the map
is not invertible.
We can think of a number of ways to compute the manifolds. First, the unstable manifolds
can be trivially computed by iterating forward an initial condition close to the xed point but in
the unstable direction. If the map in invertible, we can compute the stable manifold by iterating
26
1 2 3 4 5 6
q
-6
-4
-2
2
4
6
P
1 2 3 4 5 6
q
-6
-4
-2
2
4
6
P
1 2 3 4 5 6
q
-6
-4
-2
2
4
6
P
1 2 3 4 5 6
q
-6
-4
-2
2
4
6
P
Figure 15: Trajectories in the phase space for (a) = 0, (b) = 0.5, (c) = 1, (d) = 2 using
many initial conditions.
Figure 16: Possible trajectories between trac lights
27
Figure 17: The bifurcation diagram, associated Lyapunov exponent.
backwards along the stable directions. If it is non-invertible, then we could iterate many initial
conditions forward in time. The initial conditions that stay in the initial volume should approximate
the stable manifold.
Mathematically, a n-dimensional manifold is a space which looks locally like R
n
, but it is not
necessarily R
n
. On each point in the manifold, there is a local coordinate neighborhood that can
be described by n coordinates (R
n
). This allow us to dene functions, dierential forms, etc. on
the manifold. The simpler example is the 2-sphere S
2
which is clearly not R
2
(non-Euclidean) but
that a neighborhood close to any point in the sphere can be considered as part of R
2
(Euclidean)
represented by two coordinates. Locally we cannot distinguish between this and a small domain of
R
2
.
Figure 18: (a) The stable and unstable manifold of a xed point. (b) Evolution of a square close
to the xed point.
The linear spaces, and the invariant (stable, unstable, center) manifolds of periodic m-orbits
can be dened similarly. In particular, the stable manifold applies to all the points in the period
28
m-orbit x
0
, x
1
, . . . , x
m1
. Visually and numerically, it may be easier to treat the periodic x-orbit
as a xed point of M
m
and study the local subspaces, with their invariant manifolds, of this xed
point with the understanding that the invariant manifold apply to all the elements of the xed
point.
We dene a hyperbolic xed point (or a periodic m-orbit) if it has no center subspace. The
tangent subspace can then be uniquely specied as the direct sum of T
x
= E
+
E

.
For example for a Hamiltonian map in 2D, we have that around a xed point
w
1
=
1
w
2

1
=
2
and we usually dene two situations
w > 0 Hyperbolic
w > e
i
Elliptic
Lets note that we cannot have an attractor (periodic orbit, etc) because they do not guarantee the
conservation of the volume.
The Baker map yields an example of an hyperbolic system, since
DM(x) =
_
w
x
(y) 0
0 w
y
(y)
_
with
x
(y) < 1 and
y
(y) > 1. The unstable manifold are vertical lines, and the stable manifolds
are horizontal lines. This is in essence one of the reasons for creating this map. The Horse shoe
map also generates horizontal and vertical manifolds. In fact they are cantor sets of horizontal and
vertical lines whose intersection is the invariant set.
The concept of hyperbolicity not only applies to xed points, but also to more general invariant
sets of a map, such as strange attractors, or nonattracting chaotic sets. We say that an invariant
set is hyperbolic if there is a direct sum decomposition of T
x
= E
+
E

for all x, and that


the splitting varies continuously with x and is invariant DM(E

x
) = E

M(x)
. We obtain that
_
yE

x
|DM
n
(x)y| < K
n
|y|
yE
+
x
|DM

n(x)y| < K
n
|y|
for K > 0 and 0 < < 1. The stable and unstable manifold can then be locally dened and
that points in the stable manifold approach each other exponentially, and points in the unstable
manifold separate exponentially (e.g., chaotic attractors).
Hyperbolic invariant sets are nice because a number of theorems can be proven, such as (a) they
are structurally stable, (b) there is shadowing, (c) a natural measure exists, (d) can be represented
by symbolic dynamics, etc. It seems that the logistic map is not structurally stable since a small
perturbation in the parameter can shift from chaotic to periodic.
In general, many of the interesting chaotic phenomena seems to occur in systems that are not
hyperbolic and not structurally stable. The Hennon map is non hyperbolic since there are points
on the attractor whose stable and unstable manifold at x are tangent, and therefore they cannot
span the whole space.
29
The manifolds generated by xed points and periodic orbits are extremely relevant in deter-
mining the dynamics of the system. There are situations in which the dynamics of the stable and
unstable manifolds of xed points are of interest in understanding chaos. For example, in the Hen-
non map, the strange attractor is the closure of the unstable manifold of a xed point in its basin
of attraction (see homework).
In the particular case of xed points of invertible maps (for example maps derived from
dierential equations) the stable manifold cannot intersect stable manifolds (in particular itself)
and similarly unstable manifold cannot intersect unstable manifolds (in particular itself) (WHY?).
But there can be intersections of stable with unstable manifolds (of itself or with other xed
points). In particular we have Homoclinic (intersections of the stable and unstable manifolds)
and Heteroclinic (intersections of the manifolds of two periodic orbits) tangencies. This is also
relevant in continuous systems as we will see later on.
In a Homoclinic intersection, chaotic behavior occurs because if there is one intersection, there
must be an innite number of them. Note that since W

are invariant, then intersections map


into intersections. This shows the complicated dynamics forming a chaotic or strange attractor.
The dynamics are horseshoe type for a high enough iteration of the map, and hence generate
non-attracting chaotic sets.
Figure 19: (a) Homoclinic Tangency, (b) Heteroclinic tangency
HOMEWORK PROBLEM:For the predator-pray model with A
a
, A
b
xed, draw the
maximum Lyapunov exponent in the r
a
r
b
plane.
HOMEWORK PROBLEM: For the Hennon Map in a chaotic regime, show that the
two Lyapunov exponents can be obtained by (a) using the QR decomposition of the
innitesimal vectors evolved by DM, and (b) by following trajectories with
0
= 10
10
HOMEWORK PROBLEM:Lets take the logistic map, but one in which r
n
= A and
r
n+1
= B. In the AB space paint the value of the Lyapunov exponent for that system.
Use a color function that makes the picture interesting
30
HOMEWORK PROBLEM: Sketch the dynamics of the Baker map (
a
+
b
= 1). Take
a square in [0, 1] [0, 1] and follow what happens to it. Is this map conservative? Is the
map chaotic? Calculate numerically the spectrum of Lyapunov exponents?
HOMEWORK PROBLEM: Calculate the stable and unstable manifolds for the xed
point of the Hennon map in the chaotic case by iterating the map forward and back-
ward. Show that the second method to compute the stable manifold also works.
Repeat for a period 2 orbit.
HOMEWORK PROBLEM: Show that the unstable manifold of a point in the attractor
is the chaotic attractor and that the stable manifold is tangent to it in many places.
This suggest that the system is not hyperbolic. Why?
HOMEWORK PROBLEM: Take the Horseshoe map. To visualize the complex dy-
namics that can happen, consider the invariant set . Follow the iterations of the 16
elements of M
2
[M
2
(S)] forward in time. Does it show chaos? Can you estimate the
Lyapunov exponent. Can you estimate the dimension of the invariant set?
HOMEWORK PROBLEM: For the city trac problem. Construct the map, and a
bifurcation diagram for A
+
= 10, A

= 30 with .
HOMEWORK PROBLEM: Take the kicked pendulum. Take many initial conditions
and calculate . Count how many have > 0 as a funcion of . This are called Arnold
tongues. Can you estimate the dimension of the chaotic set as a function of ?
4 Generic Bifurcations
One of the most important topics here is the issue of the type of allowed bifurcations. We have
observed a period doubling bifurcations in which an orbit of period 2
m
losses stability and an orbit
of period 2
m+1
appears and becomes stable.
4.1 Generic Bifurcations 1-D and normal forms
Dene a local generic bifurcation if its basic character cannot be altered by an arbitrary small
perturbation M(x, r) + g(x, r). There are 3 generic bifurcations in 1-D continuous maps. In fact
it is interesting to note that we can dene normal forms that describe (with the minimum detail
possible) the local bifurcation (not its global form), namely
Period doubling bifurcation,
f(x) = rx +x x
3
31
Tangent bifurcation
f(x) = r +x x
2
Inverse period doubling bifurcation
f(x) = rx +x +x
3
there are other bifurcations that are not generic, as we will see later. The 3 bifurcations are shown
in Fig. 2 can also be understood in terms of the x point solution and their stability.
S
S
U
S
Forward Backward
U
S
Inverse
Period doubling
Tangent
Doubling
Period
S
S
U
U
Figure 20: The generic bifurcations
In the case of the logistic map the period doubling bifurcation has been clearly described.
We can now understand that the periodic windows that appear in the chaotic see of the logistic
map occur because of a tangent bifurcation as described in Fig. 20. When we calculate the
xed point of a polynomial function, like the logistic map, in the tangent bifurcation 2 complex
conjugate xed points become real. In the case of the logistic map for the period 3 window, 3
pairs of complex conjugate become real at the same time. The other two solutions of the degree 8
polynomial corresponds to the 2 period one unstable xed points.
Since it can be shown that the logistic does not have backward bifurcations not inverse period
doubling, then we have explained the bifurcations diagram of the logistic map. There are a theorem
by Sarkovskii (1964) that describe the order of the windows.
Notice that all the unstable xed points created during the period doubling bifurcations (from
period one or from the periodic windows) are still present. In fact, they form what is called a
32
chaotic saddle (or a non-attracting chaotic set). Understanding these sets is one of the
crucial open problems in complex system theory.
It is important to notice that the by expanding the map M around a xed point, we could
investigate the type of bifurcation that will happen.
M(x) = f(x
o
) +f

(x
o
)(x x
o
) +f

(x
o
)(x x
o
)
2
. . .
-1 -0.5 0.5 1
x
-1
-0.5
0.5
1
f@xD
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75 1
r
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
x
-1 -0.5 0.5 1
x
-2.5
-2
-1.5
-1
-0.5
0.5
1
f@xD
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75 1
r
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
x
-1 -0.5 0.5 1
x
-2
-1
1
2
f@xD
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75 1
r
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
x
Figure 21: The generic bifurcations with the normal forms.
4.2 Hysteresis as a non-local bifurcation
The bifurcations studied above, are local bifurcations. But we can add a few of these bifurcations
to form a global type of bifurcation. For example we can take the map
M(x, , r) = x x
3
+r (4)
clearly it is not of the forms given above, and as such it will display nonlocal bifurcations. This
map could have 3 xed point solutions depending on r.
33
-1 -0.5 0.5 1
x
-1
-0.5
0.5
1
f@xD
-0.4 -0.2 0.2 0.4
r
-1
-0.5
0.5
1
x
Figure 22: The hysteresis generated by the collision of two tangent bifurcations, given by = 1.5.
(a) the function, (b) the bifurcation diagram. We should study the phase r diagram. But we
will leave it for a homework problem.
4.3 Non-generic local bifurcation
Notice that in the logistic map a non-generic bifurcation occurs at r = 1, which cannot be described
by the above 3 bifurcations. To see that it is non-generic, take the map
x
n+1
= r x
n
(1 x
n
) +
then the xed points no longer go through a bifurcation depending on the sign of .
0.25 0.5 0.75 1 1.25 1.5 1.75 2
r
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
x
0.25 0.5 0.75 1 1.25 1.5 1.75 2
r
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
x
Figure 23: Non generic bifurcations for > 0 and < 0.
4.4 Generic local Bifurcations N-D
In more dimension we can have the same type of transitions. There is the small complication that
the eigenvalues can be complex.
Lets take the predator-prey map and look at the bifurcation diagram, as shown in Fig. 24a,b,d.
We clearly see that we have a bifurcation (inverse) that we have not see before at around A
a
= 0.4.
If we study what is going on, we realize that we have complex eigenvalues

=
r
i
i
when the
eigenvalues go through|

| = 1, as shown in Fig. 24e. This is called a Hoft bifurcation and at the


transition we obtain a rotation that is quasiperiodic, which means that it rotates with frequency
that is an irrational number. A rational periodic orbit is equivalent to a m-periodic orbit. We can
estimate the frequency of the orbit by using
34
(r
a
)
1
N
N

i=1

x
i+1
2x
i
+x
i1
x
i

which is plotted if Fig. 24c, which clearly has to go to zero at the bifurcation because the xed point
is stable. We will study in more detail quasiperiodicity later on in the context of Hamiltonian
chaos.
But what is going on here? Well, to understand it better we should look at the following map
in polar coordinates (x = rcos,y = rsin), namely
r
n+1
= r
n
(1 r
n
)

n+1
=
n
+ +br
2
n
shown in Fig. 24f. Clearly, there is a period doubling bifurcation at (x = y = 0) for
c
= 0
in the radial variable with a stable solution at r(). On the other hand, rotates with angular
frequency which varies with the stable radius. This is called a supercritical Hoft bifurcation.
The eigenvalues are
w
i
=
1

2
_
1 +r

1 2r r
2
_
therefore, the eigenvalues, which are complex become in magnitude greater than one at
c
. If
+ br() = p/q is a rational number, then we have a periodic solution that repeats itself after q
iterations. On the other hand if it is an irrational number then we have a quasiperiodic orbit,
that never repeats itself. Since r() is continuous, we will get a countable set of periodic solutions
embedded in the aperiodic solutions. But note that even though we have no periodic orbits, this
is not chaos because the Lyapunov exponent is clearly zero, i.e., there is no sensitivity to initial
conditions.
There is the subcritical Hoft bifurcation in which we have a destabilizing cubic term, i.e.,
f(r) = r +r
2
r
5
.
The idea of a Hoft bifurcation makes also sense in a continuous dynamics system in which the
real part of the eigenvalue goes from negative to positive (we will see later). A xed point losses
stability and a limit cycle becomes stable.
4.5 Global bifurcations
There are other bifurcations that are a little more obscure and deserve some time.
First, we have the tangent bifurcations of cycles in which a tangent bifurcation occurs in
the radial variable and a stable radius appears.
r
n+1
= r
n
+r
3
n
r
5

n+1
=
n
+ +br
2
n
the r
5
is to make the origin stable.
Second, we have the innite period bifurcation of cycles in which a cycle losses stability
and a xed point appears at a particular angle,
35
0.1 0.2 0.3 0.4 0.5
r
0.2
0.4
0.6
0.8
1
x
0.1 0.2 0.3 0.4 0.5
r
0.001
0.1
10
1000

0.2 0.3 0.4 0.5


r
0.2
0.4
0.6
0.8
1
l
-0.4 -0.2 0.2 0.4
r
-0.4
-0.2
0.2
0.4
x
Figure 24: (a) The bifurcation diagram for r
a
= 4, r
b
= 4, A
b
= 1.0. (b) The attractor r = 0.4.,
(c) The frequency as a function of r. (d) The attractor r = 0.35. (e) |
max
| as a function of r. (f)
The attractor for the polar map.
36
r
n+1
= (1 r
2
n
)

n+1
=
n
+ sin()
which make an angle stable when < 1.
There is also the possibility that a limit cycle can collide with the unstable manifold of periodic
orbit forming a Homoclinic orbit (at the bifurcation point). Beyond that the cycle will disappear.
This is called the Homoclinic bifurcation. Here the key is the unstable manifold of the saddle
which is capable of destroying the cycle. This is similar to a boundary crisis as we will see later.
4.6 Discontinuous transitions
We can see that the period-doubling bifurcations is in essence a continuous bifurcation, while the
tangent bifurcations is a discontinuous transition. In the same way, the logistic map grows abruptly
again at the end of the periodic window in another discontinuous transition of the attractor.
4.6.1 Intermittent transitions to chaos
See Fig. 26 for a zoom of the period 3 window. Observe that in the logistic map, the transition when
the periodic windows appear is clearly discontinuous and the attractor changes size abruptly (not
uniformly continuous) at r
3c
. We can appreciate from the gure that before the tangent bifurcations
there is an intermittent behavior, in which the trajectory spends a long time between burst close
to the xed point that is going to form (we could also nd this by computing the invariant natural
measure). This are called intermittent transition to a chaotic attractor or inverse as in the
case of the logistic map (period 3 window). We can dene P
T
as the mean time between burst, and
of course P
T
diverges at the bifurcation. We will estimate the scaling of P
T
.
In Fig. 25a we see what occurs close to r
3c
, and the trajectory spends a long time around the
tunnel. We can expand and form an equivalent map close to the xed point y
n
= x x
p
and
= r r
3c
,
yn + 1 = +y
n
+y
2
n
which is shown in Fig 25b (this is the normal form described above). It is interesting to note that
this is one of the normal forms for the transitions. Since we are very close to the transition, we can
assume that in the tunnel we have
dx
dn
x
n+1
x
n
= x
2
+ T
_

xo
dx
x
2
+
=
_

dx
x
2
+

1/2
hence P
T
(r r
T
)
1/2
.
Pomeau and Manneville (1980) distinguish, besides the tangent bifurcation, 3 generic types of
intermittent transitions (or inverse)
Tangent bifurcations to chaos, with P
T
(r r
T
)
1/2
, which we studied above.
Hopf bifurcations to chaos, with P
T
(r r
T
)
1
in which a quasiperiodic orbit (similar to a
tangent bifurcation) is formed (see latter).
37
0.52 0.54 0.56 0.58 0.6
0.52
0.54
0.56
0.58
0.6
r=2.5
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75 1
x
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
1
f@xD
r=2.5
Figure 25: (a) The iterated map f
3
close to the period 3 window and close to the formation of the
period 3 orbit. (b) the equivalent map.
Inverse period doubling, with P
T
(r r
T
)
1
in which a periodic, but unstable, orbit in the
attractor becomes stable by an inverse period doubling bifurcation.
where the other scalings can be obtained in a similar manner.
4.6.2 Interior and exterior crisis
In the same way, the logistic map grows abruptly again at the end of the periodic window, see Fig
26 in another discontinuous transition, called a crisis, of the attractor.
An interior crisis, in which the attractor collides with a unstable orbit inside the basin of
attraction of the attractor. Therefore, once every so often the trajectory will spend time away from
the initial attractor (before the transition) along the unstable manifold of the unstable periodic
orbit. But since this orbit is in the basin or attraction of the attractor it will eventually come back
to the place of the original attractor. of course now the attractor is the whole thing. In this way
the attractor growths in size at the crisis, showing an crisis induced intermittency as clearly in
Fig. 26 or Fig. 2.
It is possible to construct a scaling for the time between burst assuming that it has a distri-
bution which seems random. Hence,
P() e
/<>
This for is suggestive of the smoothness of the measure of the region of the attractor (before
crossing) that cross the unstable manifold, i.e, the measure should not be too singular (see also
below the exterior crisis situation).
The average time between burst can be derived in a manner similar to the one above (see Grebogi
et al., 1986, 1987). They determined that
< T > (r r
c
)

=
1
2
+
ln|
1
|
ln|
2
|
38
with
1
and
2
are the expanding and contracting eigenvalues of the map at A (in the case of a
period 3 orbit use the M
3
). The case of a 1-D map, we have that
2
= 0, i.e., only an expanding
direction. Then = 1/2.
3.82 3.83 3.84 3.85 3.86 3.87 3.88
r
0.2
0.4
0.6
0.8
1
x
3.82 3.83 3.84 3.85 3.86 3.87 3.88
r
0.35
0.4
0.45
0.5
0.55
0.6
0.65
x
Figure 26: (a) The periodic 3 window with the created periodic orbits at the tangent bifurcation.
(b) a zoom.
There are other types of crisis. In particular, there are boundary crisis in which the attractor
hits an unstable manifold on its basin boundary. This transition destroys the attractor leaving only
a chaotic transient. The number of points that stay decays as
P exp(/ < >)
Again, this for is suggestive of the smoothness of the measure of the region of the attractor (before
crossing) that cross the unstable manifold, i.e, the measure should not be too singular.
A clear example in the logistic map happens when r > 4 in which case the only attractor is .
Again with = (r 4) we can estimate that the set of points in which M > 1 is proportional to ,
and since this region has probability
1/2
, then < > (r 4)
1/2
. Remember that 1/ < > is the
probability that per iterate of falling in that area proportional to . The fact that this probability
density
1/2
can be found numerically, or can be determined from the fact that the invariant density
at r = 4 is smooth, and that the region that maps to |x 1/2| is of the order of
1/2
which gives
the above result. Notice the similarity with the above value.
In more dimensions we may distinguish between:
Homoclinic tangency crisis: the collision of the stable and unstable manifolds of a unstable
periodic orbit A become tangent. Grebogi et al., [1986, 1987] construct the scaling
< T > (r r
c
)

=
1
2
+
ln|
1
|
ln|
2
|
39
0 10 20 30 40 50
n
1
10
100
1000
x
0 0.2 0.4 0.6 0.8 1
xo
2
5
10
20
50
100
200
n
0 50 100 150 200
n
1
10
100
1000
10000
Pn
0.0001 0.001 0.01 0.1
4
1
5
10
50
100

0.524
Figure 27: (a) One trajectory for the logistic map with r = 4.02. (b) The time in takes for an
initial condition to leave the region [0, 1]. (c) The surviving probability. (d) The average lifetime
of as a function of 4.
40
with
1
and
2
are the expanding and contracting eigenvalues of the map at B.
Heteroclinic tangency crisis: The stable manifold of an unstable periodic orbit B becomes
tangent with the unstable manifold of an unstable periodic orbit A. Before the crisis A was on
the attractor, and B was on the boundary. Grebogi et al., [1986, 1987] construct the scaling
< T > (r r
c
)

=
ln|
2
|
ln|
1

2
|
2
Of course we have not discussed the other relevant issues, for example of a boundary crisis in
which an unstable manifold or a periodic orbit collides with a non attracting chaotic set.
Furthermore, there are other crisis of interest. For example, there are attractor merging
crisis in which the two attractor collides generating a large attractor showing clear intermittency.
Also, there are transitions in which the two states are for example low dimensional chaos to high
dimensional chaos. We will see more on this later.
HOMEWORK PROBLEM: Show numerically that indeed the mean time between
burst diverges as P
T
(r r
t
)
1/2
for the tangent bifurcation. Also check that for the
crisis we have P() exp(/ < >) with < > (r r
t
)
1/2
for the crisis of the periodic
3-window of the logistic map.
HOMEWORK PROBLEM: Repeat the analysis for the map M = r x
2
(1 x) with r
in [3.5, 6.74], and study the type of crisis that appear. Can you estimate when they
happen?
HOMEWORK PROBLEM: Describe the phase diagram r of the map given in Eq.
4.
41

You might also like