Lecture notes for 12.006J/18.353J/2.

050J, Nonlinear Dynamics: Chaos

D. H. Rothman, MIT
November 14, 2022

Contents
1 Fractals 1
1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Correlation dimension ν . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Computation . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Relationship of ν to D . . . . . . . . . . . . . . . . . . . . . . 8

1 Fractals

References: [1–4]

We now proceed to quantify the “strangeness” of strange attractors. There

are two processes of interest, each associated with a measurable quantity:

• sensitivity to initial conditions, quantified by Lyaponov exponents.

• repetitive folding of attractors, quantified by the fractal dimension.

Now we consider fractals, and defer Lyaponov exponents to the next lecture.

We shall see that the fractal dimension can be associated with the effective
number of degrees of freedom that are “excited” by the dynamics, e.g.,

• the number of independent variables;

• the number of oscillatory modes; or
• the number of peaks in the power spectrum

1
1.1 Definition

Consider an attractor A formed by a set of points in a p-dimensional space:

ε

etc

We contain each point within a (hyper)-cube of linear dimension ε.

Let N (ε) = smallest number of cubes of size ε needed to cover A.

Then if
N (ε) = Cε−D , as ε → 0, C = const.
then D is called the fractal (or Hausdorf ) dimension.

Solve for D (in the limit ε → 0):

ln N (ε) − ln C
D= .
ln(1/ε)

Since ln C/ ln(1/ε) → 0 as ε → 0, we obtain the formal definition

ln N (ε)
D = lim .
ε→0 ln(1/ε)

1.2 Examples

Suppose A is a line segment of length L:

Then the “boxes” that cover A are just line segments of length ε, and it is
obvious that
N (ε) = Lε−1 =⇒ D = 1.

2
Next suppose A is a surface or area S. Then

N (ε) = Sε−2 =⇒ D = 2.

But we have yet to learn anything from D.

Consider instead the Cantor set. Start with a unit line segment:

0 1

The successively remove the middle third:

0 1/3 2/3 1

1/9 2/9
etc etc

Note that the structure is scale-invariant: from far away, you see the middle
1/3 missing; closer up, you see a different middle 1/3 missing.

The effect is visually similar to that seen in the Lorenz, Hénon, and Rössler
attractors.

The fractal dimension of the Cantor set is easily calculated from the definition
of D:
 
1
Obviously, N ε= = 2
3
 
1
Then N ε= = 4
9
 
1
N = 8...
27
Thus  
1
N = 2m .
3m
Taking ε = 1/3 and using the definition of D,
ln 2m ln 2
D = lim = ' 0.63
m→∞ ln 3m ln 3
3
1.3 Correlation dimension ν

We proceed now to an alternative procedure for the calculation of the fractal

dimension, which offers additional (physical) insight.

Rather than calculating the fractal dimension via its definition, we calculate
the correlation dimension ν.

We shall show that ν ≤ D. But first we define it.

1.3.1 Definition

Consider a set of points distributed on a plane.

Let N (r) = number of points located inside a circle of radius r.

Assume the points are uniformly distributed on a curve like

For r sufficiently small compared to the curvature of the curve, we have

N (r) ∝ r

or
N (r) ∝ rν , ν = 1.
Now assume the points are uniformly distributed along a surface in two di-
mensions:

4
Now
N (r) ∝ r2 =⇒ ν = 2.
Next, reconsider the Cantor set:

We expect that N (r) will grow more slowly than r.

Indeed, calculations show that ν ' 0.63 = D, just as before.

1.3.2 Computation

Our implicit definition of ν is clearly generalized by considering

• an attractor in a p-dimensional space, and

• N (r) = number of points in a p-dimensional hypersphere of radius r.

For a time series x(t), we reconstruct a p-dimensional phase space with the
coordinates

x(t), x(t + τ ), x(t + 2τ ), . . . x(t + (p − 1)τ ) = ~x(t).

Suppose there are m points on the attractor. We quantify the spatial corre-
lation of these points by defining
1
C(r) = lim (number of pairs i, j for which |~xi − ~xj | < r) .
m→∞ m2

More formally,
m m
1 XX
C(r) = lim 2 H(r − |~xi − ~xj |)
m→∞ m
i j

5
where 
1 x>0
H(x) =
0 else.
The summation is performed by centering hyperspheres on each of the m
points.

In practice, one embeds the signal x(t) in a phase space of dimension p, for

p = 2, 3, 4, 5, . . .

p is called the embedding dimension.

For each p, we calculate C(r). Then, assuming

C(r) = rν

we plot log C vs. log r and estimating the slope ν:

log C(r)

slope = ν

log r

Consider the example of white noise. Then x(t) is a series of uncorrelated

random numbers, and we expect

C(r) ∝ rp , p = embedding dimension.

Graphically, one expect a series of plots like

p= 2 3 4 5
log C(r)

log r

6
Here
ν(p) = p,
a consequence of the fact that white noise has as many degrees of freedom
(i.e., independent “modes”) as there are data points.

Consider instead X(t) = periodic function, i.e., a limit cycle, with only one
fundamental frequency.

Then the attractor looks like

p=2 p=3

Provided that r is sufficiently smaller than the curvature of the limit cycle,
we expect
C(r) ∝ r1 , for p = 2, 3, 4, . . .
Graphically, we obtain
p= 2 3 4 5
log C(r)

log r

and therefore
ν(p) = 1, independent of p.

We conclude that ν measures something related to the “number of degrees

of freedom” needed to parameterize an attractor.

Specifically, suppose a dynamical regime has n oscillatory modes. The at-

7
tractor is then a torus T n , and we expect
C(r) ∝ rn .
Thus
p ≤ n =⇒ C(r) ∝ rp
and
p > n =⇒ C(r) ∝ rn , independent of p.

Conclusion: If, for embedding dimensions p ≥ p0 , ν is independent of p, then

ν is the number of degrees of freedom excited by the system.

This conclusion provides for an appealing conjecture: since white noise gives
ν(p) = p,
ν independent of p (and reasonably small) implies that the signal is deter-
ministic, and characterizable by ∼ν variables.

There are some practical limitations:

• r must be small compared to the attractor size.

• r and m must be large enough for reasonable statistics.
• Experimental noise, non-stationary time series, and difficulties extrapo-
lating r → 0 can also be a problem.

1.4 Relationship of ν to D

The correlation dimension is not strictly the same as the fractal dimension,
however it can be. We now derive their mathematical relation.

Suppose we cover an attractor A with N (r) hypercubes of size r.

If the points are uniformly distributed on A, the probability that a point falls
into the ith hypercube is
pi = 1/N (r).

8
By definition, for an attractor containing m points,
m m 
1 XX 1 x>0
C(r) = lim 2 H(r − |~xi − ~xj |), H(x) =
m→∞ m 0 else
i j

C(r) measures the number of pairs of points within a distance r of each other.
In a box of size r, there are on average mpi points, all within the range r.
Therefore, within a factor of O(1) (i.e., ignoring box boundaries and factors of two arising
from counting pairs twice),

N (r)
1 X
C(r) ' 2
(mpi )2
m i=1
N (r)
X
= p2i
i=1

Then, using angle brackets to represent mean quantities, we have, from

Schwartz’s inequality,
1
C(r) = N (r) hp2i i ≥ N (r) hpi i2 = .
N (r)

If the attractor has fractal dimension D, then

N (r) ∝ r−D , r → 0.

The definition of the correlation dimension ν, on the other hand, gives

C(r) ∝ rν .

Substituting these relations into both sides of the inequality, we find

rν ≥ rD

Thus as r → 0, we see that

ν≤D

The equality is obtained when hp2i i = hpi i2 .

Thus ν < D results from non-uniformity of points on the attractor.

9
References
1. Grassberger, P. & Procaccia, I. Measuring the strangeness of strange at-
tractors. Physica D 9, 189–208 (1983).
2. Bergé, P., Pomeau, Y. & Vidal, C. Order within Chaos: Towards a De-
terministic Approach to Turbulence (John Wiley and Sons, New York,
1984).
3. Grassberger, P. Grassberger-Procaccia algorithm. Scholarpedia 2. revi-
sion #91330, 3043 (2007).
4. Strogatz, S. Nonlinear Dynamics and Chaos: With Applications to Physics,
Biology, Chemistry, and Engineering (CRC Press, 2018).

10
MIT OpenCourseWare
https://ocw.mit.edu

12.006J/18.353J/2.050J Nonlinear Dynamics: Chaos

Fall 2022

For information about citing these materials or our Terms of Use, visit: https://ocw.mit.edu/terms.