Volume 99A, number 1 PHYSICS LETTERS 14 November 1983

THE FRACTAL DIMENSION OF THE LORENZ ATTRACTOR

Mark J. McGUlNNESS l p2
Applied Mathematics, CaliforniaInstitute of Technology, Pasadena, CA 91125, USA

Received 21 July 1983

Taken’s boxcounting algorithm for computing the fractal dimension of a strange attractor is applied to the Lorenz equa-
tions. A convergence problem is discussed, and an approximate dimension is computed.

Lorenz’s dicovery [l] of random motion on a The rate at which this probability tends to zero is found
“strange attractor” in phase space for a simple set of to be algebraic, in the region of the attractor which
nonlinear ordinary differential equations has prompted yields a one-dimensional cuspshaped return map. If
much subsequent interest in this type of behaviour, the algebraic decay rate holds on the whole attractor,
in one-dimensional maps, ordinary and partial differ- it implies an algebraically slow rate of convergence in
ential equations, and turbulent flow experiments. A time of the number of boxes visited by a solution tra-
distinctive feature of the “strange attractor” is its frac- jectory (over some intermediate time range). This sug-
tal (noninteger) dimension [2] - in the Lorenz equa- gests an extrapolation in time to iind the total number
tions the noninteger part allows the motion to be ape- of boxes covering the attractor.
riodic. In particular, the measurement of a fractal di- Lorenz obtained the equations
mension in a fluid experiment would suggest the pres-
?=u(Y-X),
ence of a strange attractor, and would indicate the min-
imum size phase space needed to reproduce the attrac-
tor (e.g. with a set of ordinary differential equations).
i=xy-bz, (1)
A step in this direction is to be able to measure the
fractal dimension of strange attractors in multi-dimen- by truncating a Fourier series expansion of a convect-
sional maps and ordinary and partial differential equa- ing fluid model. They also appear at the onset of tur-
tions. Takens [3] has suggested a box-counting algo- bulence in baroclinic flow [8,9] and alaser model [8],
rithm, which has been succesfully applied to 2D maps and they model a convecting fluid annulus and a disc
[4,5] a set of three ordinary differential equations [4], dynamo [lO].Wheno= 16,r=40andb=4,the
and a delay differential equation [6]. (numerical) solution trajectories in phase space are
However, the algorithm fails to converge for the aperiodic [ 111, and they all appear to approach and
Lorenz equations. This convergence problem has been remain close to a “double-screw” shaped surface.
encountered by others [5], for the Lorenz equations These are the parameter values used in this paper.
and for the Curry modeL As they suggest, a major rea- The firactal dimension or limit capacity of an at-
son for the problem is that some parts of the attractor tractor is given by [2]
are visited with a probability that tends to zero [7].
D=eyo -htN(e)/lne, (2)
where N(e) is the number of boxes of side E in phase
’ This work supported by Control Data Dorporation, through space needed to cover the attractor. This definition,
access to Cyber 205 computing time, and by the Office of
Naval Research.
analogous to the Hausdorff dimension, may be arrived
2 Current Address: Applied Mathematics Division, DSIR, at intuitively by realising that if the attractor has vol-
P.O. Box 1335, Wellington, New Zeeland. ume V, then
5
0.031-9163/83/0000-0000/$03.00 0 1983 North-Holland
Volume 99A, number 1 PHYSICS LETTERS 14 November 1983

V+N&‘, (3)
for small enough e, that is
lnN=-Dlne+ln V. (4) N / ./

This is a more practical formula for fmdingD than (2),

since the error term is V/in e [4] which is slow to
vanish as e--f 0. Eq. (4) gives D as the slope of a least-
squares fit straight line to ln N versus - In E for small
enough E.
The numerical problem is to lind N(e), the number
of boxes entered by a trajectory on the attractor, given
the box side E. The variables (x, y, z) are normalised
so that the bounds of the attractor are 0, 1 in all direc-
tions. The coordinates of the current box entered are
given by dividing (x, y, z) by E and taking the integer
part *r. In the 8na.l versions of the program, the boxes
are represented by a bit matrix, a “1” bit being entered Fig. 1. The number N of boxes visited, against time T, where
when a box is visited. In one version 215 trajectories the box side is E = l/128.
are simultaneously followed, making the problem
ideal for vectorisation on the Control Data Corpora- o fthe z interval (normalised to 0,l). This decay means
tion CYBER 205 used for computations. Another ver- the probability of visiting boxes near these ends on the
sion follows a single trajectory, with no essential dif- attractor is small, and will give a convergence problem
ference in results. The equations are solved with a like that encountered above, especially noticeable at
fured-step fourth order Runge-Kutta method, and a smalI box sizes. The numerical results in ref. [7], where
stepsize of 0.01. For boxes ranging in size from l/8 the attractor is approximated by a branched two-di-
to l/128 the programs required about lo6 words of mensional manifold, suggest a similar decay on other
memory. Poincare sections through the attractor.
The convergence problem is illustrated in fig. 1 for The rate of decay of the “tails” of the probability
a box side of l/128 and the equivalent of 2 X lo6 density is determined by the local nature of the cusp
scalar iterations. A longer run shows N has not con- map near the cusp and near the ends z = 0, 1 [ 141.
verged to N(e) even after 20 X lo6 iterations. This Near the cusp the map is algebraic [ 15,161, that is, if
problem is apparent even at quite large box sizes. The the nth maximum of z normal&d between 0 and 1 is
smooth appearance at the curve ofN versus T for small denoted z, , and the cusp is located at zo, the map is
box size suggests extrapolating in time to find N(e). given near zc by
This requires some knowledge of whether the conver-
Zn+l - 1 --cyIz, -zcl”, z, -fzc 3 (5)
gence is, for example, algebraic or logarithmic.
An answer to this problem is suggested by a consid- where OL,Yare constants, and v may be given ln terms
eration of the cuspshaped return map that Lorenz ob- of the eigenvalues of the Lorenz equations linearised
tained by plotting successive values of maxima in z about the origin [ 161. For the parameter values chosen
against previous values of these maxima [ 11. This map, here, v is approximately l/4.5.
noninvertible with slope everywhere greater than one If the probability denisity g(z) is invariant under
in absolute value, has a unique integrable invariant the map zn+l = r(z,), it satisfies [11,17]
probability density [ 121, and hence is ergodic. The
probability density decays to zero [7,13] at the ends
&I=$ J g(x) dx * (6)
*l Originally suggested to the author by David Rand, Warwick 7-l l(O,z)l
University. This identity and eq. (5) imply that [ 141

6
Volume 99A, number 1 PHYSICS LETTERS 14 November 1983

Pl-J=l-P1, (13)
g(z) = [g(zc)/d/u] (1 - z)llV-l ) z + 1. (7)
and taking the limit At + 0,
Near the ends the map is linear, with a slope greater
than 1 in absolute value. This and eqs. (6) and (7) give W(n, t) = a(n + l)P+lP(n + 1, t)
g(z) - [&,)/~I [Y/W(Y - 1)] l” z~“-~ , z + 0, (8) - unp+1 P(n, t). (14)
where
This is a “deaths” equation in queuing theory [ 181 -
r(z) - -yz , 2+ 0 ) (9) intervals are “killed” by being visited. It holds for n G
M where M is the number of unvisited intervals at time
and
origin, andP(il4 + 1, t) is zero for all t.
7(z)-/J(l -z), z-+1. (10) Eq. (14) may be solved by taking a Laplace trans-
form in time and solving the resulting fust-order dif-
That is, the “tails” of the probability density g(z)
ference equation directly. The solution is
decay algebraically, at the rate (l/v - 1). Compare
the exact results in ref. [ 131. Assuming a similar decay
on other Poincare sections through the attractor (as
suggested by the results in ref. [7] ). the rate at which
P(n, t) = k t+lkfl+l E
I j=n
(exp(-@+l r)

boxes of side E are filled in phase space is given by the M

rate at which intervals of length Eare visited when iter- x kFn (kp+l +P+l)-l
ating on the return map. 1
. (15)
This rate may be shown to be algebraic for inter- k+j

mediate times by considering P(n, t), the probability The expected value of n(t), the number of unvisited
that exactly n intervals remain unvisited at time t *2. intervals remaining at time t, is
Then
P(n,t+At)=P(n,f)~~+P(n+ l,f)p,, (11) (16)
where p. is the probability that no new intervals are
visited in time At, pl is the probability that one new An approximation to A(t) for small t and large M fol-
interval is visited in time At, and At is small enough lows from a consideration of the Laplace transform
to neglect the probability of visiting two or more new for large values of the transform variable [ 141. This
intervals. To approximate po, assume counting of inter- Indicates that the dominant contribution to the sums
vals has proceeded long enough that the probability in (16) comes from large values of k, and
density has converged to the invariant density, and
P(k, t) - [l/u(M+ l)Pl] 6 (t + l/&W@- l/u/3k@),
that (mostly) only those intervals in the “tails” remain
unvisited. Since some of these intervals may have been
k+=. (17)
visited, assume further that the (discontinuous) prob-
ability density over unvisited intervals may still be ap Hence for large N,
proximated by an algebraic decay. Then given that n
t?(t) N (M-0 t’ f@t)-lI@ , t + 0, (18)
intervals remain unvisited at time t, the probability of
visiting one new interval in time At is approximated by or to leading order
p1 w a&+1 At (12) A(t) - (apt)-l/O , t+O,M+m,t.M+. (19)
for n large enough and At small enough. Here II is some The result suggests that N, the number of boxes
constant (dependent on e) and /3> l/u - 1. Then visited in phase space, converges algebraically in time
to N(e), provided enough time has elapsed that the
“tails” of the probability density are being entered.
*2 This approach was suggested by Jose-Luis Farah, IIMAS, Since p is not known, a nonlinear least-squares program
UNAM. is used to find p and M(e) by fitting

7
Volume 99A, number 1 PHYSICS LETTERS 14 November 1983

’ I
-i I ’ ’ I j

‘1, I i
I
‘I I’

Fig. 2. The residuals of a least-squares fit N = N(e) - ATtip

against @lo, with box side E = l/128. -InE

Fig. 3. ln N(E) against - In E for E = l/8 to l/128. N(E) ob-

tained by extrapolation in time for all except E = l/8.
N = N(E) -A T-l/p . (20)
Fig. 2 shows a typical plot of the residues of this fit
for E = l/128 and 0 = 0.28. The range of N is 25200 I am grateful to David Rand for suggesting the box-
- 27000. counting algorithm, and to Faust0 Milinazzo, Dan
These extrapolated values of N(e) were used (ex- Meiron, Javier Jimenez, Jose-Luis Farah and Tim
cept at the largest box size E = l/8) to obtain fig. 3, a Minzoni for many helpful discussions. The majority of
plot of In N(e) versus ln E. The slope gives D = 1.98 the computations were made possible by the support
f 0.02, where the uncertainty is taken as three times of Control Data Corporation, who made time available
the standard deviation of the least squares fit to fig. 3. on their CYBER 205. The office of Navel Research
This result was obtained both by following 215 tra- has also supported this work.
jectories and a single trajectory for 215 times as long.
Mori [ 191 and Kaplan and Yorke [20] have hypo- References
thesised different formulae for D in terms of the
Lyapunov exponents (length stretching and contrac- E.N. Lorenz, J. Atmos. Phys. 20 (1963) 130.
tion rates). Both formulae give the same answer, ;: Mandelbriit, Fractals: form, chance and dimension
(Freeman, San Francisco, 1977).
D i 2.06, [3 F. Takens, Detecting strange attractors in turbulence, to
be published in Proc. Warwick Symposium, eds. D. Rand
at the parameter values used here. The values of the and L.S. Young (1981); Distinguishing deterministic and
Lyapunov exponents are taken from Shimada and random systems, to be published.
Nagashima’s numerical study [ 2 1] . [4 D.A. Russel, J.D. Hanson and E. Ott, Phys. Rev. Lett.
4.5 (1980) 1175.
It is not surprising that this box-counting and ex-
15 H.S. Greenside, A. Wolf, J. Swift and T. Pignataro, Phys.
trapolation procedure has not given the expected frac- Rev. A25 (1982) 3453.
tional part of D, since the smallest box size used is of 16 J.D. Farmer, Order within chaos, Ph. D. thesis, Univer-
the order of the transverse thickness of the attractor sity of California Santa Cruz (June 1981), to be published.
0.01 [7]. Here the attractor appears to be a two-di- [71M. Diirfle and R. Graham, Phys. Rev. A27 (1983) 1096.
mensional manifold, with the fractional part unre- [81J.D. Gibbon and M.J. McGuinness, Physica 5D (1982)
108.
solved. To resolve this transverse structure would re-
[91 I.M. Moroz and J. Brindley, J. Fluid Mech., to be pub-
quire prohibitive computer time and storage. lished.

8
Volume 99A, number 1 PHYSICS LETTERS 14 November 1983

[lo] K.A. Robbins, Math. Proc. Cambridge Philos. Sot. 82 [16] J.A.YorkeandE.D.Yorke, J.Stat.Phys. 21(1979) 263.
(1977) 309. [ 171 A. Boyarsky and hi. Scarowsky, Trans. Am. Math. Sot.
[ll] T.-Y. Li and J.A. Yorke, Am. Math. Monthly 82 (1975) 255 (1979) 243.
985. [ 181 T. Saaty, Elements of queuing theory (McGraw-Hill,
[12] A. Lasota and J.A. Yorke, Trans. Am. Math. Sot. 186 New York, 1961).
(1973) 481. [19] H. Mori, Prog. Theor. Phys. 63 (1980) 1044.
[ 131 S. Grossman and S. Thomae, Z. Naturforsch. 32A [ 201 J.L. Kaplan and J.A. Yorke, in: Functional differential
(1977) 1353. equations and approximation of fixed points, eds. Dietten
[ 141 M.J. McGuinness, to be published. and Walther, Lecture Notes in Mathematics, Vol. 730
[15] A.C. Fowler and M.J. McGuinness, Physica SD (1982) (Springer, Berlin, 1979) p. 228.
149. [21] I. Shimada and T. Nagashima, Prog. Theor. Phys. 61
(1979) 1605.