Volume 110A, number 7,8 PHYSICS LETTERS 12 August 1985

T H E C H A O T I C B E H A V I O R O F T H E LEAKY F A U C E T

P. M A R T I E N , S.C. POPE, P.L. S C O T T and R.S. S H A W

Board of Studies in Physicw, University of California, Santa Cruz, CA 95064, USA

Received 28 August 1984; revised manuscript received 3 June 1985; accepted for publication 4 June 1985

A variety of phenomena associated with the chaotic behavior of a leaky faucet are displayed and discussed. Interpretations
are suggested, both in terms of a simple one-dimensional analog simulation, and in terms of a Shannon-based information
theory, in which a useful descriptive function, the stored information, is calculated from observed streams of data.

Several years ago R6ssler suggested that the drips gate pulse. The sequence of drips thus produces a
falling from a leaky faucet might provide a familiar stream of uniformly shaped pulses, one pulse per drip,
example of a dynamical system capable of exhibiting which is fed to a Z80-based microcomputer ,2 for tim-
chaotic behavior [ 1 ]. In the following, we report a ing, storage and subsequent analysis. The microcom-
summary of our experimental investigations to date puter also is used to set the needle valve opening,
on such a system. We find substantial evidence for a through the use of a stepper motor coupled to the
broad range of dynamical behavior, including period valve shaft, allowing the drip rate to be varied between
doubling, a transition to chaos, hysteresis, and a varie- about 30 drips per minute (the stable period 1 regime)
ty of complex behavior that remains as yet incomplete- and about 1200 drips per minute (just prior to the
ly understood. We find that a mathematical model of transition to smooth laminar flow). A typical data set,
a simple one-dimensional nonlinear oscillator may be obtained at a single setting of the needle valve, consists
used to simulate some of the simpler behavior of the of a few thousand pulses.
leaky faucet, with good qualitative agreement. We also Although a complete description of the detailed
analyze the dynamical behavior in terms of a function behavior of the drips falling from an orifice requires a
we call the stored information, which quantifies our very large number of variables, we have chosen to focus
ability to predict future system states. Preliminary our attention on a single, easily accessible variable, the
reports of these investigations have been presented time interval T n between successive drips - the "drip
previously [2]. Aspects of this work have also been interval". A particularly revealing way to display the
described by two of us in undergraduate senior theses data from a single data set is to plot a map of Tn÷1
[3]. A lengthy paper providing additional description versus T n. Fig. 2 shows a sequence of such plots, as the
and analysis is available [4]. valve opening is set to increasingly larger values.
The basic apparatus is shown in fig. 1 * 1. Distilled Simple periodic regimes such as the period-1 and
water, kept at a constant pressure head by means of a period-2 attractors (figs. 2a, 2b) dominate the behavior
float valve (a modified Model A Ford carburator) is of the leaky faucet at flow rates below about 200
allowed to flow through a needle valve to an orifice. drips/min. A period-doubling sequence leading to
Drips from the orifice interrupt the beam of a h e l i u m - chaos appears to exist; the system is quiet enough to
neon laser directed onto a photocell; each of the re- observe doubling up to period 4, or, with some imagi-
sulting voltage pulses from the photocell is made to nation, period 8.
trigger the sweep of an oscilloscope, which provides a At drip rates ranging from about 200 drips/min to

~1 Figure drawn by Chris Shaw. ,2 The microcomputer was constructed by J.P. Crutchfield.

0.375-9601/85/$ 03.30 © Elsevier Science Publishers B.V. 399

(North-Holland Physics Publishing Division)
Volume l l0A, number 7,8 PHYSICS LETTERS 12 August 1985

i;iiill ii

Fig. 1. A depiction of the experimental apparatus.

about 750 drips/min, "fuzzy humps" and "worms" string-like structures to attractors with apparent di-
predominate (figs. 2d-2f). These string-like, one- mension greater than one, and a sharp drop in the fre-
dimensional structures are often interspersed with sim- quency of appearance of simple periodic attractors.
ple periodic structures; fig. 2f shows such a string well Two types of stability can be qualitatively distin-
on its way to becoming a period-3 attractor. guished. Stability against mechanical shock and randon
A wide variety of more complex structures are perturbations appears to be equally prevalent over the
found at drip rates above 750 drips/min (figs. 2g-2i). entire range o f drip rates. However, the stability of
At this point, the relative scatter of the measured given attractor structures in valve parameter space (the
periods, on the order of a few per cent for the previous range o f valve settings over which an attractor remains
attractors, increases dramatically to the order of fifty essentially unchanged) is much greater at the high flow
per cent. Accompanying this increase is a shift from rates.

400
Volume 110A, number 7,8 PHYSICS LETTERS 12 August 1985

4-

.~..

93 Tn 95 101 109 128 136

• ...
i~\. f

.(~< • ..:
" ' :;; 4"

300 Tn 210 80 90 81 91

5":
ii. . . . -".- ~::i

J~

"~.:~~ :~ • .~..
"i.., :i'f,~! y~:''':' ~': '"~: q.::' .

30 Tn 35 29 35 33 35
Fig. 2. Examples of Tn+ 1 versus T n return maps selected from the data. (a)- (c): Periodic behavior. (d)- (f): Low-dimensional
chaotic behavior. (g)- (i): More complex chaotic behavior, appearing at higher flow rates. All time values are in milliseconds.

A mathematical model which simulates some o f the time until the drop position exceeds a preset threshold.
behavior shown in fig. 2 may be developed as follows: At this point, its valve is suddenly reduced by an
A drop hanging from the orifice may oscillate, with a amount proportional to the speed of the drop at that
frequency that decreases as the drop mass increases. moment, thus simulating the breaking away of the
At a critical moment, which is sensitively dependent drop. Fig. 3 shows data, in the form of T n + 1 versus T n
on the drop size and position, the drop will break away, maps, produced by an analog computer on which this
setting the subsequent drop into a similar oscillation. equation is programmed. The behavior of the solutions
This one-dimensional "mass-on-a-spring" type of oscil- to eq. (1) depends on at least four independent param-
latory motion may be described by the following equa- eters (g, k, b and dm/dt), and we have not made a sys-
tion: tematicaUy exhaustive exploration of this behavior.
However it is apparent that there exist regions of the
d(mu)/dt =mg - ky - bu , (1)
parameter space for which this simple model produces
wherey is the position of the forming drop, u = d y / d t solutions whose return maps are remarkably similar,
is its velocity, m is its mass, and g, k and b are constant in a qualitative sense, to those produced by the faucet
parameters. The mass is made to increase linearly with when the flow rate is low enough (figs. 2 a - 2 0 . In

401
Volume IIOA, number 7,8 PHYSICS LETTERS 12 August 1985

V '\
/

\
\ \ /
/f
\, /
x\ .,77

Tn

Fig. 3. Tn+1 versus Tn return maps for the analog model, illustrating qualitative similarities with some of the attractor structures
shown in fig. 2. Units of time are arbitrary, and thus are not shown.

particular we observe period-doubling bifurcations observed sequence of drip intervals as a valid data
and low-dimensional chaotic motion for both systems. stream from which to estimate l(t).
Examination of the return maps of fig. 2 suggests To be more precise, consider a data set consisting
the presence of noisy but well-defined attractors of a few thousand drip interval values T1, T2, T 3 .....
underlying the behavior of the system dynamics. For The estimate of I ( t ) is somewhat complicated by the
some systems, a description of the chaotic motion is fact that these drip intervals are embedded in continu-
aided by the determination of a "dimension" for some ous time, with the result that there may be a signifi-
underlying attractor. For the leaky faucet, however, cant contribution to I(t) from the phase of the drip
calculations of dimension have proven unsatisfactory, pulse sequence. In what follows, to aid conceptual
owing to the limited dynamic range of the data, the clarity, we start by ignoring this "phase information",
noise scale being almost as large as that of the promi- to arrive at a function I 1 (t), the stored information
nent topological features. associated with only the abstracted set of drip interval
Hence we have developed a new technique, in values. Subsequently we shall modify the calculation
which we compute from a data set the function I(t), to include the phase variable.
the stored information. This function, which derives We denote by T a short sequence of a few adjacent
from Shannon's information theory [5], may in measured intervals (often only one); we call this the
general be estimated from an observed data stream history vector. T' represents a single future interval,
produced by any dynamical system using methods separated by k intervals from the history, where k = 1,
developed by one of the authors [4], and serves to 2, 3 . . . . . In this context, the stored information (in
provide a quantitative description of the system be- "bits") is given by
havior. It is a measure of predictability in that it speci- , [ ek(r, r')
fies how well we can predict the system's future state, Ii(k) = ffek(r, r') log 2 ~ - - ~ ( - ~ ) ~ dT d r ' . (2)
given knowledge of its past behavior. We expect that
l(t) will be a decreasing function of time, as we be- Here Pk(T, T') dTdT' represents the joint probability
come less able to predict behavior farther and farther that the history lies between T and T + dT and that
into the future. Furthermore, I(t) is an invariant the future interval lies between T' and T' + dT'.
under any arbitrary coordinate transformation of P(T) dT is the unconditioned probability that the his-
reasonable smoothness, a property allowing it to be tory vector lies between T and T + dT, while P(T') dT'
estimated from data provided by any set of conve- is the similar probability that the future interval lies
niently observed system variables. In particular, for between T' and T' + dT'. We obtain an estimate of
the dripping faucet system, we may treat the easily I 1 (k) from a data set through a binning procedure,

402
Volume 110A, number 7,8 PHYSICS LETTERS 12 August 1985

with the bin size less than the stochastic noise, to tion rate is simply I(1) -•(2), which may also be
allow the above integral to be approximated by a sum. thought of as the initial rate of loss of predictability
The relevant probabilities are estimated by counting as future times are considered. This definition is equiv-
occurrences of particular history vectors and future alent to the Kolmogorov-Sinai entropy of noiseless,
intervals. In our calculations, we include appropriate deterministic systems [7] in appropriate limits, but
correction factors to remove bias arising from the bin- avoids the difficulties which arise in the definition of
ning procedure, resulting in a determination o f I 1(k) the deterministic entropy when a noise element is
which is insensitive to bin size over a broad range. If present. As with the stored information itself, both
the history vector consists o f only a single interval T, the information storage capacity and the entropy
then Pk(T, T') is approximately equal to the density generation rate are geometric invariants.
of points on a Tn+k versus Tn return map, while P(T) Fig. 4 shows examples of both I 1 (k) and I(k), esti-
and P(T'), which must be identical functions, are sim- mated as described above, using the data of the "fuzzy
ply the projections of the return map point density h u m p " of fig. 2d, which occupies about 5 per cent of
onto each of the axes. the drip interval. The stored information I(k), initially
Now we proceed to include the additional contri- about 6.8 bits, is seen to decrease, at first more rapid-
bution to the stored information involving knowledge
of the phase. We denote by I(k) the stored information
which includes that associated with this additional
phase variable, so that I(k) serves to describe our abil-
ity to predict not only the magnitude of the future drip
intervals, but also the probable times at which future 6.0 0 (a)
drips will occur. If the drip intervals are fairly regular, []

that is, if the scatter of drip interval values is small []

compared with the average drip interval Tar , then the

appropriate expression for I(k) is given by O
4.0
~ ~ D D

~DDO D

Z(k) -- f f f Pk(r, S, T')

Xlog 2 \
[TavPg(T,S,T')]
~ , ~ ]dTOSdT', (3)
2.0
(b)
where T is the history vector, T' is the future interval []

Q
separated by k intervals from the history, and S is the o
[]
total intervening time between T and T'. Eq. (3) gives
0.0 [] [] ~-°~ .....
the stored information I(t) for t = kTav , where k = 1,
2, 3 . . . . . With the observation of only the drip interval 8 1'6 32
data, values of I ( t ) calculated using eq. (3) represent
time (mean drip interval units)
lower bounds of information stored by the actual sys-
tem. We do not give here complete derivations of eqs. Fig. 4. Plots of stored information versus time. The time,
(2) and (3). Further discussion of their applicability to which is expressed in multiplets of the mean drip interval, is
plotted on a logarithmic scale. Values of the stored informa-
our experimental results will be found in refs. [3,4]. tion, which are computed from the data set of fig. 2d using
There are two useful quantities associated with the eqs. (2) and (3), are indicated by the small squares. (a): The
stored information, which we call the information stored information I(k) includes that associated with the phase
storage capacity and the entropy generation rate. The variable. It decays from its initial value of 6.8 bits toward the
former is simply the maximum value of I(k), namely dashed line, whose slope is -0.5, the expected behavior when
only phase information is present. (b): Ii(k)includes only the
I(1). In the language of communication theory, this is
information associated with the sequence of drip interval val-
the "channel rate", or "mutual information" between ues, and decays toward zero as the drip intervals become un-
past and future system states [6]. The entropy genera- correlated.

403
 Volume 110A, number 7,8 PHYSICS LETTERS 12 August 1985

ly, corresponding to the rapid loss of information as-

sociated with the chaotic behavior within the hump,
even drips :ii : odd drips
and then more slowly, as the phase information is also ;) :-i.:::
lost, eventually decreasing like - l o g ~ as expected
for a purely stochastic drip sequence, I 1 (k) shows the
stored information associated with the chaotic motion
only, decaying smoothly from about 1.6 bits to zero • .:- ~:

after about 10 drip intervals. Thus we may character-

ize this attractor as possessing an information storage 30 Tn 35 30 35
capacity of about 6.8 bits, of which about 5.2 bits are
Fig. 5. Tn+ 2 versus Tn return maps for the even and odd data
associated with the phase variable. Furthermore, the points of fig. 2g.
entropy generation rate is seen to be about 0.8 bits/
drip, about half of which is associated with the phase
variable. For the period-1 structure of fig. 2a, the in- system, perhaps taking additional system variables into
formation storage capacity, which is associated totally account, is under contemplation. We also contemplate
with the phase variable, is 8.3 bits, agreeing well with extending the application of information theory to
the observed (stochastic) scatter of 0.32 per cent of other dynamical systems.
the drip interval. Correspondingly, the entropy genera-
tion rate here is only slightly less than its theoretical This research was supported by Faculty Research
value of 0.5 bits/drip. As expected, I 1 (k) vanishes for funds granted by the University of California, Santa
this case. On the other hand, for the period-2 regime Cruz. One of us (PLS) also gratefully acknowledges
(fig. 2b), I 1(k) remains at about 1 bit, and the informa- the recent hospitality of Professor Carson Jeffries and
tion storage capacity is correspondingly greater. In his research group at the University of California,
general, for a period-n regime, I 1 (k) = log2(n ) bits. Berkeley.
An interesting type of behavior is illustrated in
figs. 2g, 2h. Attractors of this general form (the "eagle" References
in local parlance) appear over a fairly wide range of
flow rates. Here, although the faucet is clearly in a [1 ] O. R6ssler, in: Synergetics: a workshop, ed. H. Haken
chaotic regime, every other drip interval is nearly con- (Springer, Berlin, 1977) pp. 174-183.
stant. This is seen by plotting the Tn+ 2 versus T n map [2] R. Shaw, P. Maxtien and P. Scott, presented at the Order
in chaos Conf. (Los Alamos, 1982);
for the "even" and " o d d " drops, as shown in fig. 5 for R. Shaw, presented at the Dynamic days Conf. (La Jolla,
the data of fig. 2g. One set of points yields a nearly 1982).
periodic structure, while the other gives a chaotic at- [3 ] P. Martien, An example of a chaotic attractor: the dripping
tractor. faucet, UCSC Senior thesis in physics (1982);
A consequence of this behavior is that one bit of S. Pope, Information, chaos, and the kitchen sink, UCSC
Senior thesis in physics (1984).
information is propagated into the indefinite future, [4] R. Shaw, The dripping faucet as a model chaotic system
in the "phase" variable, indicating whether a particular (Aerial Press, Box 1360, Santa Cruz, CA 95061).
drop interval is "even" or "odd". This is to be con- [5] C.E. Shannon and W. Weaver, The mathematical theory
trasted with the data set of fig. 2h, which, though sim- of communication (Univ. of Illinois Press, 1962).
ilar in form, has a small cross-over region visible in t h e [6] D. Ruelle, Ann. N.Y. Acad. Sci. 316 (1978) 408;
R.S. Shaw, Z. Naturforsch. 36a (1981) 80;
lower-left corner of the figure. This results in a small J.P. Crutchfield and N.H. Packard, Physica 7D (1983) 201;
probability of "tunneling" between the even and odd K. Matsumoto and I. Tsuda, to be published.
phases, so that there is no long-term stored information. [7] A.N. Kolmogorov, Dokl. Akad. Nauk SSSR 124 (1959)
We do not yet completely understand the reasons under- 754;
lying this phenomenon. Y. Sinai, Dokl. Akad. Nauk SSSR 124 (1959) 768.
Further experimental work with the leaky faucet

404