PHYSICAI

ELSEVIER Physica D 85 (1995) 165-224

Characterization of dispersive chaos and related states of

binary-fluid convection
P a u l K o l o d n e r a, S a i d S l i m a n i b, N a d i n e A u b r y b, R i c a r d o Lima c
aAT& T Bell Laboratories, Murray Hill, NJ 07974-0636, USA
bBenjamin Levich Institute for Physico-Chemical Hydrodynamics, City University of New York, New York, N Y 10031, USA
CCentre de Physique Thdorique de Luminy, CNRS, Marseille, France
Received 15 August 1994; revised 2 February 1995; accepted 2 February 1995
Communicated by A.C. Newell

Abstract

We describe an extensive series of experiments on traveling-wave convection in an ethanol/water mixture with

separation ratio ~0= - 0 . 0 2 0 in a long, narrow, annular container. As the stress parameter is quasi-statically
increased through the range 0-2% above onset, this system evolves through a series of complex spatiotemporal
states before settling into an extended pattern of steady, spatially-uniform convective rolls. Much of the dynamical
behavior is influenced by strong nonlinear dispersion, which causes the repeated and often irregular appearance
and sudden collapse of spatially-localized bursts of traveling waves. Small-amplitude states seen very close to onset
are analyzed quantitatively using a complex Ginzburg-Landau equation. Complex dynamical states encountered
further above onset are analyzed using a number of classical statistical techniques, focusing mainly on correlation
functions and Fourier spectra. In addition, we use the technique of biorthogonal decomposition in an effort to
extract the basic structures of the patterns and to characterize the complexity of their dynamics.

1. I n t r o d u c t i o n system of size L [3]. If ~ ~ L, then the dynamics

are correlated throughout the system, and chaos
T h e last two decades have b o r n e witness to an can exist only in the t e m p o r a l domain. If, on the
explosion of interest in the behavior of nonlinear other hand, £ ~ L, then the dynamics of differ-
dynamical systems. In the 1970's and 1980's, ent spatial regions are uncorrelated, and
great progress was m a d e in the understanding of spatiotemporal chaos is possible. O n e of our
chaos: the erratic but deterministic temporal goals in studying such systems is to develop a
b e h a v i o r of non-extended or spatially-uniform general understanding of complex spatiotempor-
systems [1]. Increasingly, the attention of work- al behavior that is analogous t o - and borrows
ers in this field has b r o a d e n e d to include extend- tools f r o m - our understanding of non-extended,
ed systems which exhibit complex behavior in chaotic systems.
space as well as in time [2]. This work has given Theoretical progress in this field has always
rise to the notion of s p a t i o t e m p o r a l chaos based advanced hand-in-hand with well-characterized
on the definition of a coherence length ~ in a experimental observations. T h e r e are indeed

1995 Elsevier Science B.V.

S S D I 0167-2789(95)00061-5
166 P. Kolodner et al. / Physica D 85 (1995) 165-224

many experimental systems which undergo a onset of convection causes the system to evolve
transition from a trivial state to a one exhibiting through a series of dynamical TW states of
complex spatiotemporal behavior as some pa- varying spatiotemporal complexity. In the largest
rameter is increased through a threshold. These range of e, the dynamics consist of the repeated
systems typically exhibit increasingly complex appearance and abrupt collapse of spatially-
spatiotemporal dynamics as the parameter is localized bursts of TW. Because this behavior is
further increased; this behavior is often char- associated with strong nonlinear dispersion, this
acterized by a correlation length ~ which de- dynamical state has been called d i s p e r s i v e c h a o s
creases as the stress parameter is increased. One [131.
system which exhibits such behavior is the Fara- This experiment has a number of advantages
day experiment, in which the free surface of a over other pattern-forming systems. First, it is
liquid becomes unstable to the appearance of an one-dimensional. As noted above, data-manage-
erratic pattern of capillary waves as the container ment problems are much easier to handle in one
is vibrated vertically with increasing amplitude dimension than in two. Theoretical progress is
[4]. Another example is "spiral chaos", seen in also much less difficult in one-dimensional sys-
convection in thin, horizontal layers of gas at tems. Second, this particular experiment has
high pressure [5]. been developed into an exceptionally clean sys-
These two examples of experiments in which tem: drifts, nonuniformities, and distortions have
spatiotemporal chaos can b e seen are three-di- been reduced to extremely low levels [14,15]. A
mensional systems which exhibit complex, two- disadvantage is that the dynamics have a very
dimensional patterns. Even at this late date in specific character, consisting of dispersive burst-
the evolution of computer technology, the stor- ing, rather than, say, the "phase turbulence"
age and analysis of the data representing such seen in two-dimensional convection patterns in
large-scale patterns present great practical dif- pure fluids [5,16]. This kind of specificity, how-
ficulties: they simply contain too much infor- ever, appears to be endemic to experiments in
mation. A more tractable situation is presented this field: many experimental systems exhibit
by experiments on one-dimensional patterns, for unique dynamics, and it often appears that
which large domains of space-time can be more general theoretical progress must be based large-
conveniently sampled and analyzed. Of the many ly on understanding of such individual behaviors
recent examples of one-dimensional pattern- on a case-by-case basis.
forming systems, we would particularly like to Convection in binary fluid mixtures is char-
bring the reader's attention to experiments on acterized by four dimensionless parameters: the
the "printer's instability" [6], directional order- Rayleigh number R, which quantifies the stress
ing of plastic crystals [7] and of nematics [7,8], applied to the system, the separation ratio 0,
and convection in pure fluids in narrow channels which characterizes the extent to which the Soret
[9,10]. effect influences density stratification [17], and
This paper presents observations of spatiotem- the Prandtl and Lewis numbers P and L, which
poral dynamics in yet another one-dimensional express the dissipation due to diffusion of
experimental system: traveling-wave convection momentum and solute, respectively. The
in a narrow, annular container [11]. In this Rayleigh number is proportional to the tempera-
system, convection in the form of traveling ture difference AT applied across the fluid layer;
waves (TW) is triggered immediately above in this paper, we express R in terms of a
onset by a subcritical, oscillatory instability [12]. fractional distance above the convective onset
For the fluid parameters in the present series of ATco : e = (AT-- A T c o ) / A T c o . In ethanol/water
experiments, increasing the stress parameter e mixtures, the separation ratio q, can be tuned
through a narrow range of values just above the over a wide range of adjusting the ethanol
 P. Kolodner et al. / Physica D 85 (1995) 165-224 167

concentration, while the Prandtl and Lewis num- eigenfunctions or "hidden dynamical structures".
bers are relatively insensitive to concentration The second challenge presented by these data is
and m e a n temperature. It is well known that, for to model the observed dynamics. We know that
~b < 0, convection is triggered in this system by a the complex Ginzburg-Landau equation
subcritical H o p f bifurcation, leading to dynam- ( C G L E ) , which can be derived from first princi-
ical states of T W over a ~0-dependent range of ples for this system from the full N a v i e r - S t o k e s
Rayleigh number. In the range ~b~>-0.06, the equations [21-24], can in principle be a quantita-
T W dynamics seen just above onset in a very tive model for at t h e least t h e first dynamical
long container are dominated by dispersive states seen above onset [18] and can be consid-
bursting, leading to persistent dispersive chaos ered as a spatiotemporal normal form [2,25]. We
over a narrow range of e [13,18]. As e is also expect that an extension of this model which
increased beyond this range, the system under- incorporates the effects of the convective con-
goes a transition to a spatially-uniform state in centration field [26] should be able to account at
which steady convective rolls fill the experimen- least semi-quantitatively for the dynamics seen
tal cell [19]. further above onset. However, the theoretical
In the present series of experiments, we have program of extending the C G L E appropriately
acquired a wealth of data which document the and quantitatively comparing its predictions to
evolution of all the dynamical T W states seen in the data has not yet been completed. It is our
this system for ~b = -0.020. Approximately fifty hope that the data presented herein will stimu-
long, well-equilibrated data sets have been ac- late progress in this area.
quired, covering a narrow range of e at small The remainder of this paper is organised as
intervals. Each data set samples a wide domain follows: The apparatus and experimental tech-
of s p a c e - t i m e on a fine mesh. Noise, drifts, niques are described in Section 2. Section 3
nonuniformities, and distortions have been re- discusses the classical statistical analysis we have
duced to very low levels. These data present two applied to the d a t a . Section 4 describes the
serious challenges to our understanding of technique of biorthogonal decomposition. In
spatiotemporal chaos in one-dimensional sys- Section 5, we present the complex G i n z b u r g -
tems. The first challenge is to characterize the Landau model. As an aid to following the
observed dynamics in a natural way. We have complex and detailed data described in this
begun to answer this challenge by applying work, Section 6 is a short overview of the
classical statistical techniques to the data-correla- dynamical states that have been encountered.
tion functions, Fourier spectra, probability den- These states are then characterized in detail in
sity functions, moments and averages, etc. How- the subsections of Section 7. The subsections of
ever, it is qualitatively clear that these statistical Section 8 describe results of the statistical analy-
diagnostics are insufficient to fully characterize sis of the observed dynamics over the entire
dynamical behavior which is manifestly non- range of Rayleigh numbers explored in these
linear. As a complement to these analyses, we experiments. Section 9 follows with a summary
have also used the technique of biorthogonal and discussion.
decomposition ( B O D ) [20]. The first results of
this analysis are quantitative measures of the
complexity of the dynamics. The promise of this
program of research is that a projection of 2. Apparatus and experimental procedures
apparently high-dimensional dynamics onto an
abstract space of reduced dimension will allow a The apparatus and techniques used in these
simplified representation of complex behavior in experiments have been described in great detail
terms of a small n u m b e r of spatial and temporal in recent publications [14,15,27]. The cell is an
168 P. Kolodner et al. / Physica D 85 (1995) 165-224

annular channel of height d = 0.2737 cm, radial temporal oscillation frequencies than of spatial
width 1.677d, and mean circumference 82.47d, wavenumbers; the mesh covers the fastest ob-
formed by a concentric disk and ring machined served temporal oscillations with approximately
from inert, polyetherimide plastic. These are 5 points per cycle.
clamped between an electrically-heated, mirror- Approximately fifty data sets were recorded
polished silicon bottom plate and a transparent, over the course of six months in this series of
water-cooled sapphire top plate. The long- and experiments. In order to track and correct for
short-term fractional stability of the applied slow drifts in the apparatus [15], we adopted a
vertical temperature difference is 5 x 10 -2. By protocol in which data sets at constant Rayleigh
means of 24 adjustable trim heaters arranged in number were recorded in groups, separated by
a circle on the underside of the bottom plate of onset measurements. Each group began with a
the cell, and using the pulse-drift technique careful servo-controlled stabilization of small-
described in Ref. [14], the spatial profile of the amplitude, linear T W as described in Ref. [27].
applied temperature difference can be measured This gave an accurate measurement of the tem-
and made uniform to within ~Erm s ~ 4 x 10 -4. perature difference ATco, defining e = 0. Then,
T h e fluid used in these experiments was a AT was increased by a small amount to produce
0.4 wt-% solution of ethanol in water at a mean a nonlinear state. After an equilibration time
temperature of 27.26°C. This fluid has separation which varied from 18 to 24 hours, a data set of
ratio 0 = - 0 . 0 2 0 4 , Prandtl number P = 5 . 9 6 9 , length 37 hours was recorded. Then, AT was
and Lewis n u m b e r L = 0.0085 [17]. The vertical changed again, and, after equilibration, another
thermal diffusion time in this system is ~'v = data set was recorded. After several such acqui-
51.2 sec. Except where given explicit dimension- sitions, the system was returned to e = 0 for
al units, time and space measurements presented recalibration of ATco. 17 such groups of data
in this paper are scaled with this time and with runs were performed. Fitting a smooth function
the cell height d, respectively. of date to the measurements of ATco allowed us
The flow patterns studied consist of superposi- to correct the recorded value of e for drifts. The
tions of one-dimensional T W which propagate drifts encountered in the present series of experi-
azimuthaUy around the cell in directions that we ments were much smaller than those reported in
refer to as "left" and "right". In our shadow- Ref. [15]. Indeed, for most of the data sets, the
graphic flow-visualization system, a refractive total drift over 37 hours was smaller than our
image of this pattern is projected onto an annu- temperature resolution of 0.1 m K and could only
lar array of 720 photodiodes under the control of be inferred from the long-term trend of the onset
a small computer. The shadowgraph system is measurements. The largest drift correction ap-
adjusted to be in the linear contrast regime. The plied in this work was 8e = 4 x 10 -4, and the
photodiode signals are digitized by the control uncertainties contributed by this correction were
computer on a fine space-time mesh and stored in the range ( 0 - 5 ) x 10 -5, leading to total
for later analysis. We experimented with several uncertainties in e of (5 - 11) x 10 -5.
different meshes and settled on a mesh spacing We applied two corrections to the data sets
of 2 ° (0.458d) in space (ie, the signal from every before further quantitative analyses were per-
fourth photodiode was recorded) and 20sec formed. First, the dc background components of
(0.390~-v) in time. The mean wavenumber of the the signals were removed. In almost all cases,
patterns does not vary much in these experi- this was done by applying a normalizing tempo-
ments, so that the spatial mesh has approximate- ral hipass filter, as described in Ref. [27]. How-
ly 4 points per spatial wavelength for all data ever, some of our data exhibit regions of very
sets. The data exhibit a much stronger spread of slow T W or truly stationary convective rolls, and
 P. Kolodner et al. / Physica D 85 (1995) 165-224 169

this procedure suppresses the low-temporal-fre- dc component in u(x, t) has an effect both on the
quency signals from these structures. Usually, we hipass filtering described in the previous section
are interested in TW dynamics only, so that this and on the complex demodulation mentioned
procedure is the one we adopted. However, in here. This situation is discussed further in Sec-
some of our analyses, we wished to retain these tions 7.4 and 8.5.
low-frequency structures. In these cases, we Complex demodulation is quantitatively useful
removed the background using a normalizing because the complex Ginzburg-Landau equation
spatial hipass f i l t e r - a less precise procedure. discussed in Section 5 below is in fact a model
Second, in order to correct for small distortions for the evolution of the slowly-varying complex
in the optical system, the filtered data were amplitudes extracted by the demodulation. It is
normalized by the spatial amplitude profile mea- also quite useful just for qualitative examination
sured in a state of spatially-uniform, slow TW, as of the data, because hidden-line graphs of the
described in Ref. [15]. As a diagnostic, we often spatially-varying amplitudes AL,R(x,t ) clearly
compute the amplitude and wave-number pro- illustrate the bursting and propagation that are
files of the left- and right-TW components of the the heart of dispersive chaos. This is easily
data using complex demodulation [27]; with the appreciated in Figs. 3 and 12 below. For this
correction just described, these computations are kind of qualitative examination, a simpler tech-
accurate to within 1-2%. nique is also useful: spatial demodulation of the
raw signal at the measured mean wavenumber at
each time step. As shown in the hiddenqine plots
3. Data analysis: classical statistical measures
in Figs. 8 and 13 below, this analysis also reveals
dynamical behavior without separating the left-
In this section and the next, we summarize and and right-TW components.
A particular application of the complex de-
describe the analyses that we have performed on
the data, so that the results can be presented modulation technique has been developed for
with minimal interruptions in Section 7 below. the quantitative analysis of the weakly-nonlinear
Figure numbers are given to illustrate each evolution of small-amplitude, spatially-uniform,
analysis, but a detailed discussion of these figures unidirectional TW. At sufficiently small am-
is deferred until Section 7. plitude, such a state is observed to exhibit no
Our first method of analysis is complex de- spatial or temporal variation in the
modulation [27]. This technique is applicable to wavenumber - this is the case in the beginning of
signals which can be written in the form the run shown below in Fig. 3. In this case, all of
the variation in the TW phase is due to the effect
u(x, t) = Re{AL(x, t) e i(wL(x't)t+kL(x't)x) of nonlinear dispersion on the temporal oscilla-
+ A R(x , t) e i('R(x' t),-kR(x,Ox)}. (1) tion frequency mR(X, t). We extract this fre-
quency by removing any stray signal component
Complex demodulation allows the real am- that might be due to weak, oppositely-propagat-
plitudes AL,R(x , t), wavenumbers kL,R(x, t) and ing TW, spatially averaging the phase of the
frequencies O)LI~(X,t) to be extracted from mea- remaining TW, and then differentiating this
surements of u(x, t). The reader is referred to phase in time. From the amplitude dependence
Ref. [27] for the details of this decomposition. of the resulting frequency, we can extract the
Here we merely point out that the conditions value of the corresponding coefficient in the
under which this procedure is valid are satisfied CGLE. This analysis is discussed further in
by most but not all of our data sets: at the Sections 5 and 7.1.
highest values of e, the presence of a significant Space-time diagrams are useful for qualitative
170 P. Kolodner et al. / Physica D 85 (1995) 165-224

examination of the data. In these diagrams, the 41b) have been computed for the raw signals
trajectories of all the individual convective rolls u(x, t), for their spatial gradients, and for the
(eg, the paths of constant TW phase) in a data demodulated TW amplitudes. We have also com-
set are plotted in space-time. An amplitude puted the moments of these distribution func-
threshold is applied, so that only regions of tions, as well as moments of the temporal spectra
significant TW amplitude are seen. Fig. 14 con- described previously.
tains nine examples of such space-time dia- PoincarO sections have been computed for
grams, and they clearly reveal the burst-like states of coexisting TW and steady rolls. As
nature of the dynamics in dispersive chaos and shown in Fig. 23, this diagnostic can reveal
its qualitative variation with e. subtle differences in the dynamics of bursts seen
Spectral analysis has usually been applied in in these states. The exact definitions of the
the time domain. We typically apply a windowed quantities being plotted will be given in Section
FFT algorithm to the time-series signal at each 7.4 below.
spatial point in a data set and then average all Finally, two recent publications have discussed
spatial points together to produce a space-aver- the observation that erratic, two-dimensional
aged temporal spectrum. It has also been useful patterns can exhibit ordered time averages [28].
to perform spectral analysis on the demodulated We have extended this concept by computing
amplitude profiles A L R ( x , t). Stripping off the comoving time averages- that is, by shifting the
carrier oscillations by demodulating often reveals spatial signal at time t to the right by an amount
the structure of modulation-induced sidebands Ax = Vmt before averaging; the results are ex-
that would otherwise be smeared out in the amined for a range of shift velocities v m. Exam-
broad spectrum of the raw data. ples appear in Figs. 43 and 44.
Correlation functions have been studied in a
number of ways. We have computed the equal-
time spatial correlation function at each time 4. Biorthogonal decomposition
step in a data set and then averaged all time
steps together to produce the time-averaged two- The statistical analyses described in the previ-
point spatial correlation matrix. The diagonal of ous section have a limited ability to fully char-
this matrix, denoted Cs(Ax), with Ax = x - x', is acterize the manifestly nonlinear dynamics ob-
referred to as the time-averaged spatial correla- served in these experiments. A general de-
tion function. An example is shown in Fig. 27 ficiency of such system-independent analyses is
below. Similarly, the space-averaged temporal that they are basically descriptive and take no
correlation function Ct(At) has been computed- particular advantage of the dynamical structure
see Fig. 32 for three examples. It has also been of the system under study. In contrast, the
useful to compute correlation functions from the theory of dynamical systems has provided an
demodulated amplitude profiles AL,R(x,t ) as array of dynamics-determined analytical tech-
well as from the raw signals u(x, t) (Figs. 34-36). niques for the study of single-channel time series
Finally, we have computed the time-averaged [1]. However, corresponding techniques for the
spatial cross-correlation function from the de- analysis of systems with both spatial and tempo-
modulated left- and right-TW amplitudes for ral degrees of freedom are much less advanced.
selected data s e t s - see Figs. 29-31. In the fig- One possibility for such an approach (pro-
ures in this paper, all correlation functions have posed in [20]; see also [29] for a review) is the
been normalized to unit peak amplitude. study of the integral operator whose kernel is
Probability distribution functions (Figs. 40a, precisely the spatiotemporal dynamics u(x, t)(x E
 P. Kolodner et al. / Physica D 85 (1995) 165-224 171

X, t E T), where X is a spatial domain and T a with

temporal interval. Such an operator maps a
AI~A2~'"~AN>O.
Hilbert space of spatial functions, H(X), to a
Hilbert space of temporal functions, H(T), such Thus, the original signal u(x, t) can be exactly
that reconstructed from the BOD. Here, N, the
number of non-zero eigenvalues, is known as the
V~b E H ( X ) , (U¢)(t) = f u(x, t) 49(x) dm(x), global dimension of u(x, t) [20].
x The one-to-one correspondence between the
(2) topos en and the chronos ~bn is given by the
operator U itself, namely Uen =AnO n. The
the adjoint operator U* mapping from H(T) to
square of the norm of each term in Eq. (5),
H(X) being defined as
equal to the square of the eigenvalue An, corre-
sponds to the spectral energy associated with
vOE H(T), (u*O)(x) = f u(x, t) O(t) drh(t), each pair (~bn, On). For pure data analysis pur-
T
poses, as in this paper, LZ-norms are used in the
(3) definitions of the operator V (Eqs. (2)-(4)),
where din(x) and drh(t) denote the measures although the choice of other norms may be
defining the scalar products in H(X) and H(T), necessary in other situations [30,31]. It is inter-
respectively, and the bar refers to complex esting to note that, with the L2-norm, the spec-
conjugation. Our method is based on the spec- tral decomposition of the operator U'U, whose
tral decomposition of U, or more precisely, that kernel is the two-point spatial correlation matrix
of the matrix operator V: Cs(x,x'), coincides with the probability theory
tool known variously as proper orthogonal de-
V=(0 U U*). (4)
composition (POD), Karhunen-Lo6ve expan-
Hereafter, this spectral decomposition is referred sion, or principal component analysis [32] when
to as the biorthogonal decomposition, or BOD, data at multiple positions are measured instanta-
of u(x, t). If U is a non-compact operator, one neously and the statistical average is chosen to be
can carry out the spectral decomposition by the time average. The application of the POD to
applying, for example, the theory of eigenfunc- turbulence was first proposed and developed by
tion expansion of self-adjoint operators [31]. Lumley [33] (see [34-35] for more recent discus-
This case .is not relevant for the data analysis sions), and it has been applied to a rather large
performed in this paper and will not be discussed number of flows (see e.g. [37]). It has also been
further here. If U is compact, then it possesses applied to the analysis of time series [38]. The
discrete, orthonormal sets of spatial eigenmodes difference between the spectral decompositions
~b,(x) (called topos) and of temporal eigenmodes of the operators U*U (whose kernel is Cs(x, x'))
qJ,(t) (chronos) such that (¢n, ~)m)H(X) = and UU* (whose kernel is Ct(t, t')), on the one
(l~n' ~ttm)H(T)=~n,m, where the parentheses de- hand, and that of the operator V, on the other
note scalar products in H(X) and H(T). The hand, has been discussed in Ref. [39]. The
space-time function u(x, t) can be written as a decomposition of V maintains the one-to-one
sum of products of chronos and topos which correspondence between space and time and
converges in norm: allows the original signal to be reconstructed
N
from its eigenfunctions, even in the presence of
degeneracies. This is not the case for the Kar-
U(X, t) = Z An~)n(X) ~n(t) , (5)
n=l hunen-Lo6ve expansion. This distinction is cru-
172 P. Kolodner et al. / PhysicaD 85 (1995) 165-224

cial in the study of bifurcations, since de- modes, then the entropy is maximal, equal to
generacies correspond to the presence of one. Likewise, if only a single mode is excited,
spatiotemporal symmetries [30,39], and bifurca- then the entropy is zero.
tions are signaled by the breaking of such While these global measures are necessarily
symmetries as parameters are varied [39-41]. characteristic of the dynamics as averaged over
It may happen that the global dimension N the entire space-time domain of the system, it is
appearing in Eq. (5) is less than the dimension of also possible to define spectral quantities which
the original subspaces H(X) and H(T); i.e., that integrate separately over space or over time,
the dynamics of the system evolves in a subspace allowing the system to be probed in a more
of reduced dimensionality. Defining Ker(U)-= "local" manner in space or in time. The spatial
{4) E H(X)/U(rb) = 0}, we can construct a vector energy is defined as:
~t(X) = [A101(t), A202(t ) . . . . . ANON(t)] in the
spatial characteristic space x(X) = Ker(U) ± . This
vector describes a time-dependent orbit in x(X)
Es(t ) =
f u(x, t) u(x, t)
X
am(x) = E A.10.(t)I2 2
n=l

which represents temporal aspect of the system's (8)

dynamics. Similarly, we can define a temporal
This quantity measures, as a function of time,
characteristic space x(T) = Ker(U*) ± and a vec-
the spectral energy corresponding to spatial
tor ~x(t) = [Alq~i(x), A2¢z(X ) . . . . , ANCN(X)] in
variations in the signal at each instant. Es(t ) is
this space. The spaces x(X) and x(T) are related
equal to the square of the norm of the vector
to each other via the isomorphism U, and it
~t(x). Similarly, a temporal energy equal to the
follows that they have the same dimension N.
square of the norm of ~?x(t) reflects the temporal
Standard phase-space methods can be used to
variations in the signal at each spatial point,
examine the orbits of ,?x(t) and ~,(x). Symmetry-
measured as a function of position:
breaking bifurcations, which correspond to the
splitting of degenerate eigenvalues, can be de-
tected via their effects on the topologies of these
ET(X) =
f U(X,t) u(x, t)
T
drh(t) = ~ An[C.(x)122.
n=l

orbits.
(9)
In addition to the global dimension N, two
other global quantities can be defined, namely In a similar manner, a spatial entropy measuring
the energy E and the entropy H [20]. The energy the degree of complexity in the spatial pattern at
is simply the sum of the squares of the eigen- each instant in time can be defined:
values
N Hs(t) = log1 N ,,~--1
= P4",(t) log p,,(t), (10)
E = Z A2, (6)
n=l where the temporal functions p,,(t) are defined
while the entropy (see [20]) is defined as from the eigenvalues A , and the chronos 0n(t):

1 N [Link].(t)[
H = logN 2 Pn log Pn, (7)
p,.(t)- u (11)
X Jnl0.(t)l
n=l
where pn is the normalized eigenvalue p , = A]/
Likewise, a temporal entropy measuring the
E. The entropy corresponds to the distribution
degree of complexity in the time series measured
of energy among the various eigenvalues and can
at each spatial position can be defined:
be understood as a quantitative measure of the
degree of space-time complexity of u(x, t). If the HT(X) = l o g1 N ~= pe,(x) log pe.(x), (12)
information is uniformly distributed among all
 P. Kolodner et al. / Physica D 85 (1995) 165-224 173

where the spatial functions p6,(x) are defined comparison will be taken up again in the Section
from the squared eigenvalues A n and the topos 7.1 and in the Discussion section.
4'n(x): We begin with the lowest-order nonlinear
C G L E that describes the convective amplifica-
Anl4,.(x)[ tion and nonlinear evolution of right-going TW:
P~n(x) = N (13)
N [Link]~.(x)[
n=l %(0, + s 0~)fi, R = e(1 + ic0)A R
The spectra energy and entropy defined in Eqs. + se2(1 + ic,) OxA
2 " R + g(l + iC2)]AR]2AR

(6) and (7) allow us to reduce the dynamics (14)

under study to simple, global quantities which
characterize its spatiotemporal complexity. The Here, r 0 is a characteristic time, s is the linear
spatial and temporal energies and entropies in group velocity, ~0 is a correlation length, g is a
Eqs. (8)-(10) and (12) additionally reflect the saturation parameter, and c i, i = 0, 1, 2 are dis-
spatial uniformity and temporal stability of both persion coefficients. The complex field A R(X, t)
the dynamics of the individual states and of the represents the slowly-varying deviation of the
experimental system itself. These are the issues amplitude of a right-going TW from that of a
explored with the B O D in this work, as dis- uniform TW of frequency % and wavenumber
cussed below in Section 8.5. The reduction of the k c. For the case of a convection pattern consist-
dynamics to trajectories in the spaces or sub- ing only of a right-TW component, the shadow-
spaces spanned by the topos and chronos is also graph signal defined in Eq. (1) is related to this
a powerful technique for understanding the be- amplitude by
havior of the system. We defer discussion of this u(x, t) = Re{AR(x, t) ei(~°"'-*cx)}. (15)
aspect of the problem to a later publication [41].
Thus, Aa(x, t) is related to the real amplitude
A R(X, t) obtained from complex demodulation of
5. The complex G i n z b u r g - L a n d a u equation the shadowgraph signal by

.,4R(X, t) = A R ( X , t) e i(A'°"'-A*R*) , (16)

In addition to characterizing the observed
dynamics using the tools described in the previ- where Ak a = kr~ - k c and Awg = o2R - w0.
ous two sections, we also interpret our results in Eq. (14) is a truncation of an expansion in
terms of the complex Ginzburg-Landau equa- powers of the amplitude "4R and its spatial
tion ( C G L E ) . This model was first applied to derivatives. The application of this kind of model
pattern-forming systems by Newell [21] and has to pattern-forming systems is often based solely
become one of the most intensively-studied theo- on qualitative grounds, with the signs of the
retical models of pattern evolution. The C G L E coefficients being deduced from symmetry con-
was derived for the Rayleigh-B6nard problem siderations. A recent example of this kind of
using a multiscale analysis (see for instance [42]), logic appears in Ref. [44]. In the case of binary-
and this equation can be considered a "normal fluid convection, however, the C G L E can be
f o r m " near a bifurcation for such partial dif- derived from first principles using a perturbation
ferential equations as the Swift-Hohenberg expansion of the full Navier-Stokes equations,
equation [2,25,43]. Our desire in this section is and the physically-correct values of the coeffi-
to motivate the generalized C G L E that is appro- cients in Eq. (14) for the fluids used in these
priate to our physical system and to give some experiments have been computed [22-24]. Thus,
indications of how this model can be quantita- this model has the potential to be a quantitative
tively compared to our experimental results. This description of the dynamics observed in these
174 P. Kolodner et al. / Physica D 85 (1995) 165-224

experiments, depending on the convergence o f The coefficients h and c 3 have been calculated
this expansion. And indeed, we show in Section for several fluids by Sch6pf and Zimmermann
7.1 below that, in the limit of small amplitude [23,24,49]. Qualitatively, we know from experi-
and very close to onset, the linear and nonlinear ments that h < 0 , since oppositely-propagating
behavior of unidirectional TW states in this TW quench each other [47,48]. In Section 7.2
system can be accurately fit to the predictions of below, we suggest that, by varying parameters in
Eq. (14), with the fitted coefficients corre- numerical simulations of Eqs. (17a,b) so as to
sponding well to the computed values. match the behavior of the regular, counter-
As the experimental system is driven further propagating wave packets seen in the present
above onset, the dynamical behavior of unidirec- experiments just above onset, quantitative values
tional TW is dominated by dispersive bursting of these coefficients might be deduced.
[13]. Qualitatively similar behavior is exhibit by The second physical effect absent from the
solutions of Eq. (14) for physically-realistic pa- coupled CGLE model of Eqs. (17a,b) is the
rameter values [24; see also 45,46]. However, influence of the convective concentration field on
the correspondence between theory and experi- the growth and propagation of TW. It was first
ment in this regime is as yet only qualitative. observed in numerical integrations of the full
Before a quantitative comparison can be made, Navier-Stokes equations that TW convection
three physical effects must be incorporated into causes lateral solute fluxes which lead to varia-
the CGLE model. The first is the stabilizing tions in the vertical concentration gradient in
nonlinear interaction between oppositely-prop- regions where the TW amplitude fields vary
agating TW. In most of the experiments reported strongly in space [50]. These convectively-
herein, the left- and right-TW components of the pumped concentration gradients in turn affect
signal exhibit similar signal levels and dynamical the buoyancy forces in the fluid layer and hence
behavior, so that their interaction cannot be the growth of the TW. This effect leads to slow
ignored as in Eq. (14). In experiments in rectan- propagation of stable TW pulses for 0 ~ - 0 . 0 7
gular containers of moderate length [47], this [14], and a reduction in drift velocity was also
interaction causes first one TW component and noticed in the propagation of erratic bursts in
then the other to dominate, leading to persistent, dispersive chaos [13]. Riecke [26] has derived an
quasiperiodic dynamics. This effect is modeled equation which governs the evolution of the slow
by writing an analog of Eq. (14) for the left-TW code C ( x , t) corresponding to the convectively-
component A L and including cubic cross terms in pumped concentration field1:
both equations [48]:
OtC = d e O~xc + acC + h 2 0x(IARI 2 - I A L I b
%(0, + s 8~)A R = e(1 + iCo)_A~
+ (ha + h3C)(I-4RI + 1.4LI2) • (18)
+ ~2(1 + icl) 02AR
a c < 0, so that C(x, t) is damped and thus slaved
+ g(1 + iC2)IARI2AR
to the convective amplitude fields. The reduction
+ h(1 + ic3)l.4Ll2A~. (17a) of the buoyant forces by the growth of the

%(0, - s Ox)A L = e(1 + iCo)_,tL

+ ~ ( 1 + ic,) 021z~L
+ g(1 + iC2)[/~LI~3L 1
The concentrationequations written in Ref. [26] contain a
• • •

term h4(A R OxA R - A L OxAL) which has not been retained

+ h(1 + ic3)]_ l 3L. (lYb) here. Since C is real, h4 must vanish.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 175

c o n c e n t r a t i o n field is m o d e l e d by rewriting the 6. Overview of dynamical states

stress p a r a m e t e r e in Eqs. (17a,b) as e + f C +
f2 C2 + f 3 axC. T h e e x p e r i m e n t a l l y - o b s e r v e d re- B e f o r e discussing detailed e x p e r i m e n t a l re-
d u c t i o n o f T W drift velocity by the c o n c e n t r a t i o n sults, it will be useful to briefly s u m m a r i z e all of
field is m o d e l e d by changing the g r o u p velocity the dynamical states o b s e r v e d in these experi-
f r o m the linear value s to s + s2C. Qualitatively, ments and to explicitly define symbols and ab-
this velocity r e d u c t i o n is not o b s e r v e d in the very breviations used in their description. This sum-
first dispersive-chaos states e n c o u n t e r e d a b o v e m a r y is p r e s e n t e d with reference to Fig. 1, T a b l e
o n s e t in these experiments, but r a t h e r b e c o m e s 1, and the subsections of Sections 7 and 8 below.
a p p a r e n t only at intermediate values of e - see
Section 7.3 below. T h u s , we expect that numeri-
cal integrations o f the perturbative m o d e l of LINEAR (a)
TW REGIME
Eqs. (17a,b) and (18) should be able to re-
p r o d u c e this effect in a w a y which can be COUNTERPROPAGATING
REGIME
DISPERSIVE-CHAOS COEXISTENCE
c o m p a r e d with e x p e r i m e n t by varying p a r a m e - REGIME REGIME
I I I
ters. £a ~:b
~11 BO QOO I O O ~ O oo O O OA~O~ A A A A A A
T h e final effect missing f r o m the coupled- 0 5 10 15 ~' 20
I I I I I
C G L E m o d e l that we have d e v e l o p e d so far o 0.005 [Link] o.o15 s 0.020
consists of h i g h e r - o r d e r n o n l i n e a r terms. It is
well k n o w n that c o n v e c t i o n is triggered in this (b)

system by a subcritical bifurcation [12], and thus

we m u s t h a v e g > 0 in Eq. (14). Ordinarily, the
existence of T W states of b o u n d e d amplitude
w o u l d require the addition of a stabilizing high-
e r - o r d e r t e r m to this equation. H o w e v e r , it is
k n o w n that the strong n o n l i n e a r dispersion in DISPERSIVE REGIMES
-20 -110 { I 10 I 20 £'
this system causes the solutions of Eq. (14) to be i i I~ II
-0.02 -0.01 0 0,01 0.02 S
b o u n d e d even w i t h o u t a stabilizing nonlinearity Es S* ~a £b £c
[24,46]. In addition, K a p l a n et al. [51,52] have Fig. I. Dynamical regimes. (a) Symbols representing data
f o u n d in numerical simulations of Eq. (14) that sets acquired in the four main dynamical regimes studied in
the c h a r a c t e r of the basic burst p h e n o m e n o n in this work are extended along axes of e and e'~1000e.
dispersive chaos is n o t altered by the addition of Time-independent linear TW, represented by an open
square, are seen exactly at e = 0. The "counterpropagating
a saturating quintic term. Thus, h i g h e r - o r d e r regime", represented by closed squares, is observed in the
t e r m s m a y not be necessary for a p h e n o m e n o - range 0 < e' < e" = 1.9. Dispersive chaos is seen in the range
logical description of the experimental results. e" < e ' < e~ = 11.1 and is represented by closed circles. The
"coexistence regime" (open triangles) extends over the range
Eqs. (17a,b) and (18), with the altered g r o u p e~< e ' < e~.= 19.2. Some dynamical states exhibit behavior
velocity and stress p a r a m e t e r discussed above, intermediate between these last two regimes and are repre-
r e p r e s e n t a c o u p l e d - C G L E m o d e l which in- sented by the closed triangles and half-filled circles in the
range 9 . 7 < e ' < 12.2. For clarity, some points have been
c o r p o r a t e s all of the qualitative physical effects omitted from this diagram. (b) Schematic bifurcation dia-
that can be discerned in o u r data. O u r suggestion gram showing the transition at e = 0 from the quiescent state
is that this full m o d e l be studied numerically and to the dynamical regimes shown in part (a) and the sub-
sequent transition, at e~ = 19.2, to states in which the entire
quantitatively c o m p a r e d with o u r data. W e will
cell is filled with spatially-uniform rolls. These uniform
e x p l o r e this suggestion f u r t h e r in Section 9.2 extended states consist of steady rolls for e ' > s * ' = - 0 . 6 3
below. and of traveling waves for e*' > e' > e'~~ -16.
176 P. Kolodner et al. / Physica D 85 (1995) 165-224

Table 1
Dynamical states
Section State e' range Symbol

7.1 linear TW 0 []
7.1 nonlinear onset transients 0.4-1.7
7.2 counterpropagating wave packets 0-1.9 [], •
7.3 dispersive chaos 1.9-11.1 •
7.3, 7.4 part dispersive, part "coexistence ''a 10.3-11.0
7.4 regular coexistence regime a 11.1-19.2 b
7.4 erratic "coexistence" 9.7-12.2 •
7.5 extended states 19.2 -~c
a In the coexistence regime, TW and steady rolls coexist in different parts of the cell.
b States in the coexistence regime were studied up to e' = 16.5; the upper limit e~ = 19.2 is based on an extrapolation.
c The transition to cell-filling extended states at e'c = 19.2 is hysteretic. Once created, these states persist below onset.

In addition to using a consistent set of symbols these and the boundaries ea, b which separate
t h r o u g h o u t this p a p e r (with some exceptions regimes are shown in Fig. l a and in Table 1. Fig.
n o t e d in the text and figure captions), we have l b puts this information in the context of a
a t t e m p t e d , as much as possible, to illustrate all schematic bifurcation diagram. As e is increased
of our observations with a c o m m o n subset of f r o m below the saddle node e s ~ - 0 . 0 1 6 past
data runs, and to cross-reference all quantitative e = 0, the system undergoes a transition from the
m e a s u r e m e n t s on a given run with a figure that quiescent state to the dynamical regimes in Fig.
shows the qualitative behavior of that run. For la. Increasing e through eC causes the system to
example, nine dispersive-chaos data sets are m a k e a transition to a state in which the entire
illustrated in Fig. 14. These same data sets are cell is filled with spatially-uniform rolls, shown as
analyzed in several subsequent figures, and their the u p p e r branch in this diagram. This state
captions refer explicitly to Fig. 14. persists u p o n further changes in e, making a
Before starting this summary, we m a k e one transition to a spatially-uniform T W state below
r e m a r k about the values of the control parame- e*' = - 0 . 6 3 and then back to the quiescent state
ter e quoted in this paper. Since e ~ 1 for all of below e s.
our data sets, we have found it m o r e easily Section 7.1 deals with linear a n d n o n l i n e a r
readable to cite values for e ' ~ 1000e in the text onset transients. Small-amplitude, unidirectional
rather than the measured control p a r a m e t e r e. T W can be produced in the forms of spatially-
This representation of a factor 1000 by a prime is uniform T W or wave packets. Studying the
reserved exclusively for the control p a r a m e t e r e evolution of these states exactly at onset allows
and is also used for values which are assigned the linear dispersive properties of small-am-
subscripts. Thus, for example, the transition to plitude T W to be measured. Following the
steady, spatially-uniform roll is found in these weakly-nonlinear evolution of unidirectional T W
experiments to occur at ec=0.01922(22); this just above onset allows the cubic coefficients in
m e a s u r e m e n t is also quoted as e~ = 19.22(22). the C G L E in Eq. (14) to be measured. Thus,
Fig. l a is a schematic diagram i n which the onset transients can be used to extract all of the
symbols representing individual data sets are coefficients in that model. M e a s u r e m e n t s m a d e
extended along axes of e and e'. In these in these experiments are represented by open
experiments, e was varied f r o m below e = 0 to square symbols and are summarized in Table 2.
a b o v e e = ec, and, in so doing, we encountered The first persistent, nonlinear state observed
four main dynamical regimes which are n a m e d in in this system, after onset transients have de-
Fig. la. Each regime is assigned its own symbol; cayed, consists of pairs of broad, slightly
 P. Kolodner et al. / Physica D 85 (1995) 165-224 177

Table 2 poral duration of the T W bursts seen in the

Coefficients in the complex Ginzburg-Landau equationa
dispersive-chaos regime leads to a transition, at
Coefficient Experiment Theory t

e b = l l . 1 , to a regime in which we observe

r~ 1 9.81(19) 9.161(8) persistent, steady convective rolls occupying a
s 0.7445(4) 0.843(6)
spatial region which is centered on a fixed
$fflc~ 4.5(2.4) -0.678(36)
r~l~ 1.51(5) 1.3547(11) location in the cell. The spatial extent of this
7~i~c 1 0.012(5) 0.028(5) region expands linearly with e as e is further
Tolg +60(26) - increased, filling the cell with steady rolls at
~0 gc2 -488(25) -

c2 -7.5(3.2) -9.84 e~ = 19.2. Outside the region of steady rolls,

rco 1.18044(6) 1.0297(7) strong bursts of oppositely-propagating T W are
k¢ 3.0505(5) 3.11825 seen. The domain of e over which such states of
wo 3.143828(25) 2.743(21)
coexistence of steady rolls and T W is observed is
aThese coefficients are listed in the combinations in which referred to as the "coexistence regime". Surpris-
they are measured. The theoretical values have been com-
puted for a laterally-infinitesystem, c2 is obtained by dividing ingly, the appearance of the T W bursts in this
the measured values of r ol gc2 and r olg. The last three linear regime is accurately periodic in time in most
parameters do not enter directly into the CGLE. cases. Such regular coexistence states are repre-
sented in graphs by open triangular symbols.
asymmetric wave packets of small-amplitude T W Erratic states in this regime, in which the nomi-
which propagate around the cell in opposite nally steady rolls actually exhibit slow, irregular
directions and at high velocity. These c o u n t e r - dynamics, are represented by closed triangles.
p r o p a g a t i n g q u a s i l i n e a r w a v e p a c k e t s persist up Additionally, some data sets acquired just below
to e" = 1.9 and are represented in graphs by open the transition point e'b = 11.1 exhibit dispersive-
and closed square symbols. We use the shortened chaos dynamics some of the time and "coexist-
term "counterpropagating regime" to refer to ence" dynamics at other times. These are repre-
this dynamical regime. It is described in Section sented by half-filled circles. The states in the
7.2. A partial account of these observations has coexistence regime were described in part in
been given in Ref. [18]. Ref. [19], where they were referred to as
Section 7.3 describes d i s p e r s i v e chaos. This traveling-wave/steady-roll (TW/SR) states. They
state consists of the erratic, repeated appearance are fully described in Section 7.4.
and subsequent abrupt decay of spatially-local- The transition from the coexistence regime to
ized bursts of TW. As e is increased through the the spatially-extended state of uniform, steady
range 1.9 = e'a ~< e' ~< e'b = 11.1, these bursts ap- rolls at e" = 19.2 is hysteretic. As e is sub-
pear more frequently in time and more densely sequently decreased, the steady rolls persist. A t
in space. Their spatial extent shrinks, and their e' = e*' = - 0 . 6 3 , the rolls begin to move, creat-
temporal duration increases with increasing e. ing a uniform pattern of TW. This behavior is
Additionally, as e is increased, the T W bursts qualitatively consistent with that exhibited by
exhibit lower and lower temporal frequencies extended states at more negative separation ratio
and begin to appear in the same spatial loca- [15]. See Section 7.5 for a brief recapitulation of
tions. States in the dispersive-chaos regime are these observations.
represented by full circles in the graphs below. Each of the subsections of Section 7 men-
Ref. [13] describes previous observations of this tioned so far deals with an individual dynamical
regime in a less well-developed experimental state seen in its own range of e. In each subsec-
apparatus. tion, we present only enough quantitative diag-
The trend of decreasing temporal frequency, nostic information to characterize the state being
correlated spatial location, and increasing tem- described. A general discussion of the e depen-
 178 P. Kolodner et al. / Physica D 85 (1995) 165-224

dence of the behavior of all of the observed where A % = o ) - w 0 , with w0 the Hopf fre-
dynamical states is postponed until Section 8. quency. Neglecting nonlinear terms, the disper-
sion of the nth mode is

7. Characterization of dynamical states Aoo = TolCo e + s A k , - %- , 2~ 0 C l Ak ] (20a)

(20b)
7.1. L i n e a r a n d n o n l i n e a r o n s e t transients
Thus, fitting the measured values of % and % to
Observations of small-amplitude, unidirection- --1 2
parabolas in Ak yields the coefficients s, TO ~0Cl,
al, linear TW exactly at onset have been used to and %-t ~0-slnce
2 • e ~ 0 , the terms "rolCo e in Eq.
measure the linear dispersive properties of TW (20a) and "role in Eq. (20b) turn out to be
in this system and to track slow drifts in the negligible. Fig. 2 shows the results of a series of
apparatus. Allowing such states to evolve slowly such experiments and the fits to Eqs. (20a,b). To
just above onset produces weakly-nonlinear compensate for small frequency drifts over the
transients which, for sufficiently small amplitude,
are quantitatively described by the C G L E of Eq.
(14). The further evolution of such nonlinear t i i r i i
transients causes the TW amplitude to grow too 0.2 (a)
$
large for Eq. (14) to be quantitatively valid;
however, it is still illuminating to compare model z 0,1
l.U
and observation qualitatively. [3z

Experimentally, we begin as in Refs. [14,53] u.

o 0
by launching a narrow, unidirectional, small-am- >-
o
z
plitude wave packet at e = 0 and acquiring data w
-0.1
as it evolves dispersively into a spatially-uniform
state of unidirectional TW. The time series at
-0.2
individual spatial points in the system exhibit i i i i i i

oscillations near the Hopf frequency o)0 which i i i i t i

0 (
are amplitude modulated by the recurrent drift
of the wave packet past the point of observation.
-0.02
A temporal Fourier transform of one of these tJA

time series exhibits as many as seven sharp <

rv- -0.04
'l-
sidebands at frequencies denoted %. Narrow-
band temporal demodulation of the nth sideband et -0.06 i
yields a decaying amplitude whose dependence
on time is exponential. -0.08
The evolution of the small-amplitude, uni-
J I
directional state created in this experiment is 2.8 2'.9 3; 3., 312 3'.3
quantitatively described by the linear part of WAVE NUMBER

Eq. (14). To analyze these data, we substitute Fig. 2. Linear dispersion obtained from the evolution of
small-amplitude wave packets exactly at onset. (a) T h e
into Eq. (14) a solution corresponding to the difference A~o between linear oscillation frequency and the
initial wave packet: H o p f frequency % is plotted as a function of wave n u m b e r .
The curve is a fitted parabola. (b) T h e dimensionless growth
AR(x, t) = ~ A , exp[%t + i ( A % t - Aknx)] , rate 3' is plotted vs, wave n u m b e r . T h e wave n u m b e r at the
n
peak of the fitted parabola corresponds almost exactly to the
(19) quantized m o d e in which n = 40 wavelengths fill t h e cell.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 179

weeks which separated the runs in these experi- As noted in Section 3, a variation of the
ments, Fig. 2a shows the difference frequencies complex demodulation program described in
A% rather than the directly measured frequen- Ref. [27] allows a spatially-averaged frequency
cies wn. The linear frequency is nearly disper- and growth rate to be extracted from full space-
sionless: ro1£2c1=0.012(5), Locating the ex- time measurements made during this transient.
tremum of the parabola fit to yn(k) yields the This method is more precise than the measure-
critical wavenumber: k c = 3.0505(5), almost ex- ments reported in Ref. [13], in which temporal
actly equal to the wavenumber k40 = 3.0475 of demodulation of the signals measured at single
the n = 40 quantized mode; Interpolating the fit points in space was used. As in that work, we fit
of the frequency to k c defines the Hopf fre- Aw a n d y to linear functions of A 2, obtaining the
quency w0 = 3.143828(25). The curvature of the complex cubic coefficient in the C G L E from the
marginal curve is r0-1 ~02 = 1.51(5). Table 2 lists slopes of these fits: r o l g C 2 = - 4 4 8 ( 2 5 ) and
these values, along with those of other coeffi- -1
r0 g = +60(26). As can be seen in Eqs. (21a,b),
cients in Eq. (14). We postpone a comparison the intercepts of these linear fits depend on the
with theoretical predictions and a discussion of linear parameters, the wavenumber, and e. By
uncertainties until all the measurements on making measurements at several values of e and
transients have been described. fitting the intercepts to linear functions of e, we
After the dispersive evolution of the linear extract the coefficients r o l c o and %1. The first
wave packet is complete, the servo control of these values i s quite imprecise; the second
described in Ref. [27] is used to measure the matches the direct measurement quoted above.
critical Rayleigh number: rco=1.18044(6). In The transient experiments just described have
addition, by raising and lowering e and observing allowed measurements of all the coefficients in
the resulting growth and decay of the TW the C G L E in Eq. (14). These measurements
amplitude as in Ref. [12], we directly measure have been carried out in the limit o f small
the coefficient %1 in Eq. (20b): TO 1 = 9.81(19). amplitude, on states in which the nonlinear
The next transient experiment consists of interaction with stray, oppositely-propagating
simply increasing e to a small, positive value TW is negligible. Thus, we expect these results
( e ' = 0.4 to 1.7) and observing the nonlinear to be a useful basis for quantitative modeling of
evolution of the unidirectional, spatially-uniform more complex nonlinear states.
TW state created in the linear transient just It is useful to pause at this point to discuss the
analyzed. As mentioned in Section 3 and in Ref. precision of these measurements and their rela-
[13], the growth of such transients is so slow that tionship to theoretical predictions. The theoret-
there is a long period during which the non- ical values for the linear parameters listed in
linearity affects the growth rate and oscillation Table 2 have been computed using the p r o g r a m
frequency of the TW without affecting their of Cross and Kim [22] and have been assigned
spatial structure. Such a state can be represented uncertainties corresponding to the imprecision in
by a single term in the expansion in Eq. (19) (let the properties of the experimental fluid. The
it be represented without subscripts). Substitut- errors quoted for the measured values in Table 2
ing this into Eq. (14) and retaining the cubic represent statistical uncertainties only. As de-
nonlinear term, we obtain tailed in Ref. [15], the absolute uncertainties of
measurements of length, velocity, frequency,
A w = rolCo e + s A k - TO1~2C1Ak 2 + ~'olgc2 A2 , and Rayleigh number in this system are about
(21a) -+1%, -+1%, -+2%, and -+2%, respectively. To
within these absolute uncertainties, the differ-
'Y = z o l ( E __ ~2 Ak 2) + r o l g a z " (21b) ences between the measured and computed
180 P. Kolodner et al. / Physica D 85 (1995) 165-224

values of rco and w0 can be ascribed to the effects } I L I I

of the narrow width of the cell, as measured at
= - 0 . 2 3 9 in a rectangular cell of variable width
[54]. To the extent that such width effects are
understood [55], these measurements may be
said to be consistent with theory. In contrast, the
effect of cell width on the critical wavenumber k c
depends strongly on separation ratio [54,56]; this
effect is not quantitatively understood. The
group velocity s is affected by the dependence of
kc on both width and ~0.
We have no way of assessing the effect of cell
width on the other linear parameters listed in
Table 2. Of these, the values for T0-1 and To 1C02
can be said to agree reasonably well with theory.
The linear dispersion is quite weak in this sys-
tem; thus the discrepancy between the measured
and c o m p u t e d values of TO -1 ~0cl
2 is probably
unimportant. This is definitely true in the case of
the parameter Co, which is scaled away in theo-
retical analysis anyway [43].
The values of the nonlinear parameters 7olgC2 I I I } I
0° 90 ° 180 ° 270 ° 360 °
and To lg depend on the sensitivity of the optical
POSITION IN CELL
system and are thus not directly comparable with
Fig. 3. Hidden-line plots of the demodulated right-TW
theory. However, their ratio, c 2, is independent amplitude AR(x, t) show the nonlinear evolution of small-
of the optics and is in agreement both with that amplitude, unidirectional TW at e' =0.38(5). Each frame
q u o t e d in Ref. [13] and with the trend of recent covers a time interval 24000 sec=468%. Bottom: initial
nonlinear burst. The initial condition consisted of small-
measurements of a different nature made over a amplitude, right-going TW of spatially-uniform amplitude.
range of separation ratios in a rectangular con- Top: subsequent burst later in the same run. The initial
vection cell of width 2.0 [52]. Comparison with condition for this burst was the spatially-localized, small-
amplitude wave packet produced by the first burst. The
the value c 2 = - 9 . 4 computed for a laterally-
double-humped structure is a generic feature caused by the
infinite system [57] suggests that c 2 decreases as development of a trailing-edge phase defect- see discussion
a function of decreasing cell width. relating to Fig. 7 below. Note the small-amplitude wave
The quantitative results presented so far were packet propagating through the system after the collapse of
the double-humped burst.
obtained from observations of small-amplitude
T W before nonlinearity causes a detectable
change in spatial structure. Further evolution of region of depressed amplitude expands in space,
such a state leads to gross qualitative changes in causing the T W region to shrink to a narrow,
structure which result in the production of weak wave packet. The details of this spatial
spatially-localized bursts of TW. This is illus- collapse are hidden behind the burst in the
trated in the hidden-line plots in Fig. 3. As bottom frame of Fig. 3 but are discussed below
shown in the lower frame of this figure, an with reference to Fig. 6.
initially spatially-uniform state develops an am- The subsequent evolution of the small-am-
plitude null (here centered at location 10 °) once plitude wave packet caused by the first burst in
its amplitude grows sufficiently strong. T h i s this run differs strongly from that of the
 P. Kolodner et al. / Physica D 85 (1995) 165-224 181

spatially-uniform state in which the run began. this process took so little time that the nearly-
As shown in the top frame of Fig. 3, the narrow spatially-uniform first burst seen in this run was
wave packet evolves into a characteristic double- never encountered - only double-humped bursts.
h u m p e d burst. This burst also collapses, produc- Further improvements in the uniformity of this
ing another weak, narrow wave packet - this can cell would probably only change the time and
now be seen behind the double-humped burst. amplitude scales of this first burst, not its quali-
The subsequent evolution of the system consists tative behavior,
of several more sequences of burst and collapse The data in Fig. 4 show that the initial,
before evolving into the counterpropagating re- dominantly-right-TW state in this run actually
gime discussed below. contained a small left-TW component. The left-
Fig. 4 shows the t i m e evolution of the T W amplitude profile remained spatially uniform
spatially-averaged left- and right-TW compo- and weak until it was suppressed to near zero by
nents of the transient in Fig. 3. The initial, the first burst of the right-TW component. In our
nearly-spatially-uniform T W grow to very high estimation, the amplitude of the left-going T W
spatial amplitude before collapsing, as compared was too small to affect the nonlinear evolution of
to that of the subsequent three double-humped the first right-TW burst. After the first right-TW
bursts. The amplitude to which a spatially-uni- burst, the left-TW component slowly grew in
form state can grow before something triggers its amplitude, without however erupting into bursts.
spatial collapse appears to be determined by the This slow growth may indicate that the bursting
nonuniformities in the system: in a very uniform right-going T W are being weakly reflected by
system, inhomogeneities take a long time to inhomogeneities in the cell, but an intrinsic
grow up and destabilize the state, allowing its process unrelated to nonuniformities may actual-
amplitude to grow to high values before its ly be the cause. In any event, at the end of the
spatial structure is distorted. In previous experi- run in Fig. 4, the left and right T W have equal
ments in a cell of much worse uniformity [13], spatially-averaged amplitude and are starting to
burst together. The system ultimately evolves
into the counterpropagating regime. This
TIME ('Cv)
scenario can also be described in terms of sym-
500 1000 1500 2000 2500 3000
0.10 I I I I L I I metry breaking: the initial state exhibited trans-
lational invariance which was broken by the
2 0.08 formation of spatially-localized bursts. This
0-
< scenario has been observed in several runs.
~" 0.06
Careful observations of the spatial structure of
the T W in the first burst in the runs of Figs. 3
<
rr and 4 reveals that the destabilization of the
t.u 0 . 0 4
>
< spatially-uniform initial state can be qualitatively
._J explained on the basis of the C G L E in Eq. (14).
-- 0.02
F-
< The T W amplitude and wave-number profiles
shown in Fig. 5 illustrate the first deviations of
0
0 40 80 120 160 this state from spatial uniformity. As time
TIME (1000 sec) proceeds, the amplitude profiles in the top frame
Fig. 4. The spatially-averaged right-TW (solid curve) and of Fig. 5 grow up and develop spatial gradients.
left-TW (dashed curve) amplitudes are plotted vs. time for In the bottom frame, we see that the
the run shown in Fig. 3. Note the difference between the
shapes and peak heights of the first burst and the subsequent wavenumber profile is growing more distorted as
bursts. time passes; further, note that wavenumber
182 P. Kolodner et al. / Physica D 85 (1995) 165-224

(0 t + s Üx) A k = Ox[a -2 0x(a 2 Ak)] - c z O~a 2 ,

0.035
(23a)
0.030 (O, + s Ox)a = (e - ~ k 2 ) a + 02 a + a 3 . (23b)
b-
-n 0.025 With these results it is easy to see the mechanism
0_
of destabilization operating in Fig. 5: the
0.020 - c 2a~a 2 term in Eq. (23a) causes the
t
wavenumber deviation Ak to grow in regions
where there is an amplitude gradient. This in
3.100 turn affects the ( s - Ak a) a term in Eq. (23b),
3.075
reducing the linear component of the amplitude
n'-
uJ growth and thus reinforcing the amplitude gra-
3.050
dient. Because of the convective derivatives,
z
LU 3 . 0 2 5
these effects are advected downstream at the
>
< group velocity. This is what is seen in Fig. 5.
3.000
Specifically, since c 2 < 0 in this system, Ak grows
2.975 t I I I I negative downstream of regions where 0xa2< 0;
0o 90 ° 180 ° 270 ° 360 °
POSITION IN C E L L
thus, the growing amplitude gradient in the
Fig. 5. The demodulated right-TW amplitude (top) and region between locations 310 ° and 15 ° in Fig. 5
wave-number (bottom) profiles are shown for 10 time steps causes the wave number to decrease in the
(3.90%) in the run of Figs. 3 and 4, starting at time 20800sec. downstream region between locations 320 ° and
As time progresses, the amplitude profiles in the top frame 50 ° .
grow in magnitude, and the wave-number profiles become
more distorted. The wave-number profile is driven Fig. 6 shows that these effects are also active
nonuniform in regions of amplitude gradients, in accord with in the collapse of the high-amplitude part of the
the CGLE model. first burst in this run. These amplitude and wave-
number profiles were recorded just after the first
gradients are growing in regions where the amplitude peak in Fig. 4. During this time
amplitude profile exhibits gradients. This un- interval, the edges of the amplitude profile move
stable mutual reinforcement of the growth of inwards, causing the T W region to collapse in
amplitude and wavenumber gradients is exactly space. At the same time, the center of the
the mechanism of destabilization caused by convecting region is dropping in amplitude.
strong nonlinear dispersion in the C G L E . This Because these amplitude profiles are shrinking,
can be shown by following the analysis of Kaplan they were hidden behind the large burst in the
et al. [51,52]. We start with Eq. (14), rescaling so lower burst in the lower frame of Fig. 3. Note
that TO= ~o = g = 1. Since c o can be scaled out of that the wave-number profile is again becoming
the equation, and since Icl1 < 1 for these experi- strongly distorted in the regions where the am-
ments, we set Co,1 to zero. This gives plitude gradients are strong; this leads to the
2 N
formation of a p h a s e d e f e c t , consisting of a
(0, + s 0~)AR = sAR + OxAR singularity in the wavenumber profile and a cusp-
+ (1 + iC2)IARIaAR. (22) shaped null in the amplitude profile, at the
trailing (left) edge of the convecting region. In
Writing AR = a ( x , t) exp(i~b(x, t)), with real the center of the T W region, where the am-
a ( x , t) and qS(x, t), and defining Ak = - 0 x & for plitude profile is fairly flat, there is only weak
right-going TW, we obtain distortion of the wave-number profile.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 183

0.12 I I I I I
0.04
0.10
w 0.03
w 0.08
t-., ,,7
I.-
}- 0,06
-7 "7 0,02
13_
<~0.04 ,<
0,01
0.02
I

I I I I I

rr 4.0
rr 4.0 11 w
3.5 133
:E)
z 3.5
Z 3.0
I..LI
>
,< <I:
2.5 3.0

2.0 ~ [ I r I I
0o 90 ° 180 ° 270 ° 360 ° 0° 90 ° 180 ° 270 ° 360 °
POSITION IN CELL POSITION IN CELL
Fig. 6. The demodulated right-TW amplitude (top) and Fig. 7. The demodulated right-TW amplitude and wave-
wave-number (bottom) profiles are shown for 11 time steps number profiles are shown for 10 time steps in the run of
(4.29%) in the run of Figs. 3 and 4, starting at time 22240 sec. Figs. 3 and 4, starting at time 74640 sec. As time proceeds, a
As time proceeds, the amplitude profiles shrink spatially, and phase defect forms in the trailing (left) edge of the localized
the amplitude decreases in the bulk of the shrinking burst. burst, as indicated by the development of a cusp-shaped null
This process is accompanied by further distortion of the in the amplitude profile and a singularity in the wave-number
wave-number profile, culminating in a phase defect at loca- profile. By this process, the burst splits in two, and the
tion 60°. For clarity, the first three time steps have been trailing burst eventually dominates. For clarity, the last time
deleted from the top frame, and the last time step has been step has been deleted from the bottom frame.
deleted from the bottom frame.

(which was h i d d e n b e t w e e n the two bursts in the

T h e trailing-edge phase defect seen in Fig. 6 is t o p of Fig. 3), the initial burst splits into two;
a characteristic f e a t u r e of w e a k l y - n o n l i n e a r T W s u b s e q u e n t l y , the leading burst p r o p a g a t e s for-
c o n v e c t i o n [27,56] and is widely seen in disper- wards and decays, leaving only the trailing burst.
sive chaos. I n particular, this defect f o r m a t i o n is This process of a trailing-edge phase defect
responsible for the d o u b l e - h u m p e d shape of the causing d o u b l e bursting events is seen in all of
b u r s t seen in the t o p of Fig. 3. This is illustrated the dynamical states discussed in this paper.
in Fig. 7, which shows the amplitude and wave-
n u m b e r profiles for a short time interval b e t w e e n 7.2. Counterpropagating quasilinear w a v e
the d o u b l e bursts in the t o p of Fig. 3. A t the packets
b e g i n n i n g of this time interval, the T W am-
plitude profile h a d a s m o o t h pulse shape with a G i v e n sufficient time, transients like that in
m a x i m u m amplitude at location 248 ° . A s time Figs. 3 - 7 always d e c a y into pairs of c o u n t e r -
p r o c e e d s , the trailing edge of this profile de- p r o p a g a t i n g quasilinear wave packets for e ' ~<
velops a c u s p - s h a p e d amplitude null, accom- 1.9. Fig. 8 shows a state of this type for e ' =
p a n i e d b y a singularity in the w a v e - n u m b e r 0.65(7). T h e p o w e r s p e c t r u m c o m p u t e d f r o m the
profile. B y the f o r m a t i o n of the amplitude null time series signal m e a s u r e d at any single spatial
184 P. Kolodner et al. / Physica D 85 (1995) 165-224

I ] I I I

21
°°[Link]~"!f ~a ' '"/ i ' 't
18 o oo5t: ", ,/ ' , , ,,

0.015 l ' ~ ~ ' ' ~ '

o 12
//%

V--

°.7f \
'c't
LLI
l I I I I

I I I I I o ,,"A
0° 90 °
POSITION
180 °
IN CELL
270 ° 360 °
o o,o
Fig. 8. The spatiotemporal temporal evolution of the coun-
terpropagating-wave-packet state is shown for e ' = 0.65(7).
This hidden-line plot was produced by simple demodulation
W
>
.< °7f
of the shadowgraph signal at the measured mean
wavenumber, with no separation of left- and right-TW. The
o.o15~ , . .... .,, j
time axis extends over an interval 24000 sec = 410%. This
state is periodic in both space and time.
o
(d)
0.005 I
o- r" , "-, 7
point in this state exhibits discrete, narrow peaks
which are due to both amplitude and frequency 0.015
modulation (see below). At the locations x 1 and
0.010
x 2 of the peak amplitudes (81 ° and 261 ° in the
particular run of Fig. 8), these spectral peaks 0.005
occur at frequencies which can be fit with the 0
two-parameter form to. = too + n Ato, with n an
3.2 1 , ¢-~1. ~ , , , • ,-~ ~ 42
integer, to within the experimental uncertainty.
, \ / / (f) 41 =
Full complex demodulation of the shadow- rr 3.1
graph signals yields the amplitude and wave-
number profiles in Fig. 9, showing explicitly that
WW
>,-n
.<
. ~
Z
3.0 /\ ,' ~ 40 ~ w

this state consists of a pair of wave packets which 2.9

I I I"""
38 Z
180 ° 0° 180 ° 360 °
circulate around the cell at approximately con-
P O S I T I O N IN C E L L
stant velocity and no long-term change in shape.
Fig. 9. Frames (a)-(e) show the spatial profiles of the right-
The wave-packet amplitude profiles are slightly (fuU curves) and left-TW (dashed) amplitude profiles at times
asymmetric (see below), and the wave-number separated by 25.0% for the counterpropagating-wave-packet
profiles exhibit a strong gradient. The temporal state in Fig. 8, with time increasing from (a) to (e). For
clarity, 540 ° of angle are shown. This state consists of two
f r e q u e n c y m e a s u r e d at a s i n g l e s p a t i a l p o i n t is
slightly-asymmetric wave packets which circulate around the
o b s e r v e d t o e x h i b i t a c o r r e s p o n d i n g c h i r p , so cell in opposite directions. Frame (f) shows the strong
t h a t t h e p h a s e v e l o c i t y in e a c h w a v e p a c k e t is gradients in the wavenumber profiles corresponding to the
approximately constant. amplitude profiles in (e). The right-hand vertical axis in (e)
shows the wavenumbers corresponding to spatially-uniform
The splitting Ato in the discrete temporal states with n wavelengths filling the cell. Linear TW exactly
spectrum measured at o n e o f t h e t w o l o c a t i o n s at onset relax to the n = 40 state, as shown by the arrow.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 185

xl, 2 where the wave packets cross corresponds to 0.8 i i i i

a t i m e - a v e r a g e d group velocity og=Aoo/Ak, (a)

w h e r e Ak=2~r/F~, with F~ = 8 2 . 4 7 the m e a n >-
F-
0.7
circumference of the cell. (The group velocity o
H~q
FE34
can also be m e a s u r e d directly by following the >

m o t i o n of the individual wave packets, but the i_[B4 HB4

0 0.6
tr
spectral m e a s u r e m e n t is easier and m o r e pre- (9

cise.) As shown in Fig. 10a, vg decreases smooth- F£}4 1434

ly f r o m the linear value measured at onset. The 0.~ I I

largest value of Vg m e a s u r e d in these experi-

1.04 -- I
m e n t s is only 7 % below the linear group (b)
velocity 2. By following the paths of points of
constant T W phase in s p a c e - t i m e , a phase ve- 1.02 I-U-I

locity vpu can also be measured. As shown in S

LU
Fig. 10b, Vph also decreases smoothly f r o m the >
tU 1.00
0'3
linear value at onset and only drops a few !
13_
percent over the entire range of e studied. Vph is
0.98
slightly lower when m e a s u r e d during times when
_r I I F
left- and right-going wave packets overlap (full 0 0.0005 0.0010 0.0015

squares in Fig. 10b) than when the two wave

packets are far from one another (open squares). Fig. 10. The group velocity (a) and phase velocity (b) are
shown as functions of the control parameter e for counter-
Fig. 11 shows the evolution with e of the propagating-wave-packet states. Both velocities decrease
t i m e - a v e r a g e d amplitude and w a v e - n u m b e r pro- smoothly from the linear values measured exactly at onset.
files of the right-going wave packet. These were The open squares in (b) represent measurements made when
the left and right wave packets were diametrically opposite in
calculated by shifting the profiles c o m p u t e d by the cell. The full squares were measured when the wave
complex demodulation at time t to the left by an packets overlapped. The nonlinear interaction between the
a m o u n t Vgrt--thus transforming to a comoving opposite TW components causes a slight decrease in the
f r a m e - and then time averaging. A t the lowest phase velocity in this case.
e, T W fill the cell with an amplitude profile that
exhibits only a weak, sinusoidal spatial modula- value of 0.12 near onset, which is also compar-
tion. T h e w a v e - n u m b e r profile also exhibits a able to the p e a k amplitude seen in the weakest
gentle modulation. As e is increased, the am- persistent dispersive-chaos state at e ' = 1.95(7).
plitude profiles grow stronger, narrower and By comparison, Fig. 11 shows that the strongest
m o r e asymmetric, and the w a v e - n u m b e r gra- wave packet exhibits a p e a k amplitude of only
dients b e c o m e stronger. Importantly, the wave 0.020.
packets seen in these experiments are substan- The small T W amplitudes and small group and
tially w e a k e r than any other n o n l i n e a r state in phase velocities observed in these experiments
this system. The amplitude profiles of slow, are indications that nonlinearities are weak in
spatially-uniform, nonlinear T W have an average the counterpropagating r e g i m e - i n Ref. [18],
these states were referred to as quasilinear.
W h e n the two wave packets are far f r o m one
a The linear group velocity used in Ref. [18] has been another, they p r o p a g a t e with nearly the linear
discovered to be in error by about 2%. Thus, the difference group and phase velocities and grow up at the
between the group velocities at e = 0 and e = 0.00065 has
been corrected from the 9% value in Ref. [18] to 7% here, linear growth rate ro*e. W h e n they pass through
and Fig. 10 has been slightly altered. each other, the stabilizing cross terms in the
186 P. Kolodner et al. / Physica D 85 (1995) 165-224

x/d states exhibit the same amplitudes and group and

-40-30-20-10
i i i i
0
i
10
~
20
i
30
i
40
i
phase velocities, a quantitative, first-principles
0.020
description of this state could be attained. The
w 0.015 fitted parameter h ( l + i c 3 ) could be compared
0.010
with the theoretical value computed by Sch6pf
<
and Zimmerman [23,24].
0.005 Counterpropagating wave-packet states are
observed for 0.18(7)-- e'-< 1.66(7). At the
upper end of this range, these states can be
3.3 I I I ] I somewhat irregular. An example is shown in Fig.
,-,- 3.2 12, where the left- and right-TW amplitude
3.1
profiles at e' = 1.23(7) are shown in hidden-line
z format. In the very middle of the bottom frame
w
> 3.0
<
2.9
I I
2.8 i i i [ i
0° 90 ° 180 ° 270 ° 360 °
20
P O S I T I O N IN C E L L

Fig. 11. T h e time-averaged fight-TW amplitude (top) and

16
w a v e - n u m b e r (bottom) profiles are shown for counter-
propagating-wave-packet states at several values of e. From
w e a k e s t to strongest amplitude profiles: e' = 0 . 1 8 ( 7 ) , e ' = 12

0.43(7), e' = 0 . 6 5 ( 7 ) , e' = 0 . 9 5 ( 7 ) , and e ' = 1.66(7). With

increasing e, the amplitude profiles become stronger, nar- 8

rower, and m o r e asymmetric. T h e corresponding wave-num-

ber profiles develop stronger gradients with increasing e. 4

go
§ I i I I I
coupled-CGLE of Eqs. (17a,b) cause their am-
E i I i
plitudes to be slightly damped, compensating the
V--
linear growth and leading to a regular state of 20
small amplitude and no long-term change in
16
wave-packet shape. Other signatures of these
weak nonlinear effects are quite difficult to 12
detect. In particular, we have not been able to
measure any modulations in the height or the 8
width of the wave packets that are phase-coher-
4
ent with their circulation.
Because of the weakness of the nonlinearity in 0 I I p I
the counterpropagating regime, we propose that 0o 90 ° 180 ° 270 ° 360 ° 90 °
P O S I T I O N IN CELL
these observations can be quantitatively de-
scribed by the coupled CGLE model of Eqs. Fig. 12. Hidden-line plots showing the evolution of the left-
T W (top) and right-TW (bottom) amplitude profiles in a
(17a,b), with no other nonlinear terms. The only counterpropagating-wave-packet state at e ' = 1.23(7). T h e
parameter in this model that has not been vertical axis extends over a time 20000sec=390"r v. For
measured is the complex coupling coefficient clarity, 540 ° of a,,gle are shown. In the center of the b o t t o m
frame, the right-going wave packet develops a trailing-edge
h ( l + i % ) . We suggest that, by varying this
phase defect and splits into two wave packets, by the same
parameter in numerical integrations of this m e c h a n i s m seen the top of Fig. 3 and in Fig. 7 above. Over
model until the experimental and numerical the same time period, the left-TW remains fairly regular.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 187

of Fig. 12, the right-going wave packet splits into i i i I I

two, and the leading wave packet decays. This
scenario is the same as that illustrated in Fig. 7. 25

In addition, the peak amplitude of the right wave

packet is not steady: observe the abrupt decrease 20
G"
in amplitude in the upper left-hand corner of the
bottom frame of Fig. 12. By contrast, the left- § 15
going wave packet (upper frame of Fig. 12) uJ
displays regular behavior during this run. While --~
I- 10
the hidden-line presentation of the TW am-
plitude profiles in Fig. 12 reveals this instance of
gross dynamical irregularity quite well, we have
found in general that the irregularities in the
1 I I I
counterpropagating regime are quite weak and 0° 90 ° 180 ° 270 ° 360 °

difficult to detect using spectral or demodulation P O S I T I O N IN C E L L

techniques. Dynamical analysis based on the Fig. 13. Hidden-line plot showing the evolution of the total
BOD reveals these weak effects more clearly demodulated T W amplitude for a state of dispersive chaos at
e' = 4.8 over a period of 25000 sec = 488%. Transient bursts
[41]. Part of this analysis is presented in Section
of both left- and right-going TW are seen.
8.5 below.
The transitions to and from the counter-
propagating regime are nonhysteretic. Increasing Fig. 14. Fig. 14a shows the first well-equilibrated
e' from 1.66 to 1.95 causes a transition to dispersive-chaos state studied in this system, at
persistent dispersive chaos. The reverse transi- s' = 2.00(6). This run exhibits infrequent, broad,
tion is seen on dropping e' from 1.97 to 1.78. double-humped bursts in both TW components.
Thus, we quote the upper limit of the counter- Note that the TW phase velocity changes very
propagating regime as e'a = 1.87(12). As detailed little in these bursts, except in the right-TW
in the Discussion section below, we expect that, burst at time 26000 sec. Also note that there are
in a system of different length F~, this e limit substantial regions in which both left and right
would scale as l/F+. We presume that the lower TW are seen. As e is increased (Figs. 14b,c), the
limit of the counterpropagating regime is exactly bursts become more localized in space and
•=-0. appear more frequently in time and more den-
sely in space. Regions of lower phase velocity
7.3. Dispersive chaos become more apparent, and regions of overlap
between left and right TW grow smaller. These
Fig. 13 shows a hidden-line plot of the full TW trends continue as e is further increased and are
amplitude field in a state of dispersive chaos at particularly apparent in Fig. 14e (e ' = 6.34(11)).
e ' = 4.8. This representation shows very clearly At higher e (Figs. 14f,g; e'=7.44(5) and
that this state consists of repeated bursts in both 8.66(5)), a new trend is observed: the TW bursts
the left- and right-TW components. However, begin to last longer in time, to appear in
the duration of the bursts, the phase velocity of spatially-correlated locations, and to exhibit ex-
the underlying TW, and the general qualitative tremely low phase velocity for long periods of
evolution of this state with e are better revealed time. The evolution of these features culminates
in space-time diagrams. A series of nine dis- in the run of Fig. 14i at s' = 10.98(8); this is the
persive-chaos data sets, separated in e' by ap- highest value of e at which dispersive-chaos
proximately 1.1, is illustrated in this format in dynamics are exhibited, although we classify this
188 P. Kolodner et al. / Physica D 85 (1995) 165-224

I J
I I I I I
32
32

261
28

24
24
Y

20
%-
2O
SS
o
O 16 o 16
oo
LLI
LU I'--

12 12

/ S
S

e 0 b
i I I I I I I I I
0° 90 ° 180 ° 270 ° 360 ° 0 90 ° 180 ° 270 ° 360 °
P O S I T I O N IN C E L L P O S I T I O N IN C E L L

Fig. 14. Space-time diagrams of dispersive chaos. The vertical axis extends over 32000 sec = 625%. Blank regions represent
domains where the local TW amplitude was lower than a threshold which varied from 0.030 to 0.035 as e was increased through
the dispersive-chaos regime. (a) e ' = 2.00(6); (h) e ' = 3.23(5).

run as exhibiting coexistence-regime dynamics disappearance of regions in which both left and
part of the time as well (There are three such right T W are seen as e is increased is likely due
part "dispersive, part coexistence" data sets). to the increasing importance of the nonlinear
Thus, the nearly-steady convective rolls between cross terms in Eqs. (17a,b) as the T W amplitudes
locations 120 ° and 190 ° in the bottom half of the increase. The decrease of the T W phase velocity
frame coexist with very long-lived bursts of much with e is undoubtedly caused by the stronger
faster TW which appear in nearby locations; in pumping of the convective concentration field in
the top half of the frame, the regular TW bursts Eq. (18) at higher T W a m p l i t u d e - a n effect
characteristic of the coexistence regime are seen. which has been explored in detail for the case of
Many of the features seen in Fig. 14 can be spatially-uniform T W [15,50]. Localized T W
interpreted qualitatively in terms of physical bursts tend to "dig a hole" in the concentration
effects discussed in Section 5 above. The gradual f i e l d - i.e. to locally destroy the diffusive con-
 P. Kolodner et al. / Physica D 85 (1995) 165-224 189

I I I I
I I I ] L
32 32

28 28

24 24

20 2O

o
o
oo 16 o 16
o
LU
I.U
I--

I- 12 12

d
0 c
I [ i I I
I I I I I 0o 90 ° 180 ° 270 ° 360 °
0o 90 ° 180 ° 270 ° 360 °
POSITION IN CELL
POSITION IN C E L L
Fig. 14c. e ' = 4.07(6). Fig. 14d. e ' = 5 . 1 6 ( 5 ) .

centration gradient caused by the Soret effect in Section 5, will be a successful strategy in re-
the previously quiescent fluid. After the T W producing the observed dynamics. In particular,
burst collapses, this "concentration hole" dif- it seems reasonable to expect at least a
fuses away only slowly, due to the small Lewis semiquantitative description of the dynamical
n u m b e r in this system. For sufficiently high states seen at the lowest values of e in the
Rayleigh number, the increased buoyancy in this dispersive-chaos regime.
localized region causes a new burst to grow up Fig. 15 shows the spatial structure of a right-
there faster than diffusion can erase the local T W burst taken from the run of Fig. 14i. The
concentration hole. Thus, subsequent bursts leading-edge amplitude shoulder and the strong
grow preferentially in the locations of previous wave-number gradient are strongly reminiscent
bursts. The fact that these effects turn on gradu- of similar features exhibited by stable pulses at
ally as e is increased through the dispersive m o r e negative qJ [14,56]. In the unsteady dy-
regime suggests that the program of modeling namical states exhibited at this separation ratio,
them by adding terms to the C G L E , as in these features build up in time at the beginning
190 P. Kolodner et aL / Physica D 85 (1995) 165-224

m
I I I I I I I I I I
32 32

28 28

24 24

20 20

O 16 16
O v
O

LLI
I-
t-- 12 12

Y////

f
I I I I I I I I I i
0° 90 ° 180 ° 270 ° 360 ° 0° 90 ° 180 ° 270 ° 360 °

POSITION IN CELL POSITION IN C E L L

Fig. 14el ~'= 6.34(11). Fig. 14f. e ' = 7.44(5).

of the burst event, starting from a very broad rolls; in the rest of the cell, strikingly regular T W
amplitude profile and a much flatter wave-num- bursts are observed. We have performed most of
ber profile. This evolution was also noted in Ref. the analysis of these states by removing the
[52]. When the burst decays, trailing-edge phase steady-roll component with a temporal hipass
defects are seen. filter. The structure of the T W component iso-
lated in this way is shown in hidden-line format
7.4. Coexisting traveling waves and steady rolls in Fig. 17. Although it is not directly apparent in
this figure, the T W bursts in this state exhibit the
The evolution in the dispersive-chaos regime same trailing-edge phase defect documented in
towards longer-lived, correlated regions of ever- Figs. 3 and 7, and this results in the same burst-
slower T W documented in Fig. 14 ultimately splitting scenario. As before, this process is
leads to the coexistence regime. Fig. 16 shows a hidden between bursts in this figure.
s p a c e - t i m e diagram of a state in this regime for The time series measured at a single spatial
e' = 12.12(5). The region between locations 190 ° point in the demodulated amplitude profile of
and 280 ° is occupied by almost perfectly steady Fig. 17 exhibits a monoperiodic oscillation re-
 P. Kolodner et al. Physica D 85 (1995) 165-224 191

I I ! I L
32 I I I I I
32

28
28

24 ////;
24

20 2O

O 16
o
O o 16
O
[iJ
UA Y
I--
12 12

h~ ~.
I I I L~ I I I I I
0° 90 ° 180 ° 270 ° 360 ° 0° 90 ° 180 ° 270 ° 360 °
POSITION IN C E L L
POSITION IN C E L L

Fig. 14g. e'= 8.66(5). Fig. 14h. e ' : 9.34(7).

flecting the timing of the T W bursts. The Fourier allows the nature of the T W bursts to be iden-
transform of this time series exhibits a sharp tified. This is illustrated in Fig. 19, which shows
fundamental spectral line and a few harmonics. the profiles of the amplitudes of the T W and
Strikingly, the modulation frequencies measured steady-roll components (top frame) and of the
on the left and right halves of Fig. 17 are quite modulation frequency o)R of the right-TW com-
different. Fig. 18 shows the sum of the left and ponent (bottom frame) in this state. The steady-
right modulation spectra for this run, with peaks roll component was isolated from the raw
labeled L and R. The L peaks are absent from shadowgraph signal using lopass temporal fil-
the R spectrum, and vice versa. The mean tering at each spatial point. Its amplitude profile
modulation frequency in this run is Wmod 0.232, = was computed at each time step using simple
corresponding to the 1400-sec burst frequency spatial demodulation, and the time average of
seen in Fig. 16. these profiles is shown as the dotted curve. The
The difference between the L and R modula- left- and right-TW amplitude profiles shown as
tion frequencies in the state of Fig. 17, combined dashed and full curves in Fig. 19 were computed
with the sensitivity of demodulation techniques, using a "double demodulation" technique based
192 P. Kolodner et al. / Physica D 85 (1995) 165-224

I I I I I
32 i i i i
0.08

LU 0 . 0 6
tm

28
_a
t2.
0.04

0.02
24

4.0 I I I I

2O
~muJm 3.5

o z 3.0
o
o 16 LLI
>

ttl g 2.5

t-
2.0 p r I
12 -10 -5 0 5 10
POSITION, x/d

Fig. 15. TW amplitude and wave-number profiles for the

right-TW burst at location 48 ° and time 6000 sec in the run of
Fig. 14i at e ' = 1 0 . 9 8 ( 8 ) . The characteristic asymmetric
amplitude profile and strong wave-number gradient are seen.

I I I I I

16
i
i I I I 4
®
0° 90 ° 180 ° 270 ° 360 °
POSITION IN CELL
12
Fig. 14i. e ' = 10.98(8).
"S

on complex demodulation of the raw signal. We o

@
o
first computed the left- and right-TW amplitude o

profiles whose sum is displayed in Fig. 17. Then, LU

each of these signals was temporally demodu- N

lated at ~O,~oa at each spatial point, yielding the
time-averaged amplitude and modulation-fre-
quency profiles. The amplitude profiles obtained
in this manner reflect convectively-amplified TW
which propagate from the middle of the TW
region towards the edges of the steady-roll
I I I I I
region, where their amplitudes decay. These 0° 90 ° 180 ° 270 ° 360 °
profiles also reveal weak left-going (right-going) POSITION IN CELL
TW on the right (left) side of the TW region. Fig. 16. Coexisting state of TW bursts and steady rolls at
e ' = 12.12(5). In this run, persistent steady rolls fill the
Simply time-averaging the TW profiles in Fig. 17
region between angular locations 190 ° and 280 ° . Outside the
yields similar results. The more complicated region of steady rolls, TW bursts appear at regular intervals.
technique, however, allows the time-averaged Time domain: 16000 sec = 312%.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 193

I L I I 0.03 I t I

" t i
16 0.02

<

o -~ 0.01
,<
o

0.25
I I
>-
O
0.24
I
270 ° 360 ° 90 ° 180 °
rr
P O S I T I O N IN C E L L u. 0.23

Fig. 17. This hidden-line plot shows the evolution of the sum
of the left- and right-TW amplitude profiles computed from 0.22
the temporal hipass component of the signal in the state of
0
Fig. 16. With this processing, the steady rolls are removed, I I I I
0.21
and regular bursts of left- and right-going TW are seen. In 270 ° 360 ° 90 ° 180 °
this and the related Figs. 19, 24, and 26, the horizontal axis is P O S I T I O N IN C E L L
shifted so that the TW region is centered in the frame of the
figure. Fig. 19. Top frame: time-averaged amplitude profiles for the
state of Figs. 16-18. For clarity, the horizontal axis has been
shifted with respect to Fig. 16. Solid curve: right-TW am-
spatial profile of the modulation frequency of plitude profile. Dashed curve: left-TW amplitude profile.
each TW component to be separately and accu- Dotted curve: profile of the amplitude of the steady rolls,
reduced in vertical scale by a factor 5.5 and cut off at low
rately computed. The right-TW modulation-fre- amplitude for clarity. Bottom frame: spatial profile of the
quency profile in the bottom of Fig. 19 shows modulation frequency of the right-TW component. The weak
that the weak right-going TW on the left of the right-going TW on the left of the TW region exhibit a
different modulation frequency (~oe = 0.2245) than the strong
right-going TW on the right (o9R = 0.2402). They are caused
10-5 by reflections of left-going TW from the edge of the region of
steady rolls.
10 - 6

10 - 7 figure e x h i b i t t h e s a m e modulation frequency

LU ~oL = 0 . 2 2 4 5 as t h e strong l e f t - g o i n g T W o n that
g~
.J 10-8
<
t',r side o f t h e cell. O n t h e right side o f t h e cell, b o t h
10 - 9
T W c o m p o n e n t s are m o d u l a t e d at O)g = 0 . 2 4 0 2 .
co
± T h u s , w e c a n c o n c l u d e that t h e w e a k T W are
.~ 10 - t 0 reflections o f t h e strong, i n c i d e n t T W f r o m t h e
o
e d g e s o f t h e r e g i o n o f s t e a d y rolls. T h e r e f l e c t i o n
10-11
c o e f f i c i e n t is r o u g h l y r ~ 0 . 1 . T h e r e f l e c t e d T W ,
0 0.2 0.4 0.6 0.8 1.0
DIMENSIONLESS FREQUENCY
h o w e v e r , are c o m p l e t e l y a t t e n u a t e d b y t h e inci-
dent TW.
Fig. 18. This figure shows the sum of the spatially-averaged
temporal Fourier spectra of the demodulated left- and right- T h e T W profiles s e e n in this state are r e m i n -
TW amplitudes of Fig. 17. Demodulation removes the carrier i s c e n t o f t h e profiles s e e n in t h e r e g u l a r " b l i n k -
signal due to the underlying TW, revealing sharp peaks in the ing states" s e e n in short, r e c t a n g u l a r cells at t h e
spectra that are caused by the regular appearance of TW
bursts. The left- and right-TW modulation frequencies differ s a m e s e p a r a t i o n ratio [47], e x h i b i t i n g t h e ap-
by about 7% in this run. p r o x i m a t e l y e x p o n e n t i a l g r o w t h in t h e T W direc-
194 P. Kolodner et al. / Physica D 85 (1995) 165-224

tion that is characteristic of convective instabili- 8

i i i i i

ty. The T W amplitudes decay near the edges of

the steady-roll r e g i o n - s u c h "healing regions" 8
are also seen in blinking states [47,48]. However,
unlike the weakly-nonlinear blinking states, the LU
2
incident T W in the coexistence regime are F-

strongly nonlinear and completely absorb the I

reflected waves. The source of the T W bursts i

8
must therefore lie in the center of the T W
region. This situation is similar to that encoun-
tered by S c h r p f and Rehberg [58], who studied
T W convection in a slot-shaped cell with absorb-
ing endwalls. Curiously, however, they observed
erratic "blinking" attributable to the convective
o
iii

V-
4

0
°N
I I I I

amplification of small fluctuations, rather than

i i i
8
the regular T W bursts seen here. This difference
is puzzling. IN
Fig. 20 shows the evolution of the dynamics in o

o 4
the coexistence regime with e. The most obvious .C.
uJ
feature is the decrease with e of the length FTw :s 2
I.--
of the T W region. With further increase in e, this
_ L I
trend causes the cell to fill with steady rolls. The 00 90 ° 180 ° 270 °

bursting appears to be regular in all three of P O S I T I O N IN C E L L

these runs, although a more detailed analysis of Fig. 20. E v o l u t i o n of the dynamics in the coexistence regime
the run in Fig. 20c at e ' = 15.35(5) shows these with Rayleigh n u m b e r . E a c h frame covers a time d o m a i n
8000 sec = 156rv. (a) e ' = 11.31(5); (b) e ' = 12.93(5); (c) e ' =
events to be slightly e r r a t i c - see below. 15.35(5). With increasing e, the T W region shrinks.
The regions of steady rolls in the three runs
shown in Fig. 20 are centered at the same spatial
location. As mentioned above, this is caused by by time-averaging the results of spatial demodu-
convective "self-trapping": once a region of lation of the lopass component of the signal at
steady rolls is formed, it "digs a hole" in the each time step. In this figure, they have been
convective concentration field which pins it spa- shifted horizontally, so that the center of the
tially. Subsequent small changes in e, as in the region of steady rolls always lies at location 180 ° .
series of experiments in Fig. 20, change only the At the highest values of e, these rolls exhibit a
size of the region, but not its location. However, bulk wavenumber of 3.22-3.23 which increases
in experiments in which the pattern was sharply at the edges of the steady-roll region.
quenched by reducing e far below onset and then Fig. 22 shows quantitative measures of the
re-made by increasing e again, the region of dynamical behavior in the coexistence regime as
steady rolls reappeared at apparently random functions of e. The decrease in the length FTW of
positions. These locations do not seem to be the T W region is approximately linear in e,
related to fixed inhomogeneities in the convec- intersecting the FTW = 0 axis at e~ = 19.22(22)
tion cell. with slope/3 = - 8 . 3 2 ( 3 3 ) x 103. We take ec to be
Fig. 21 shows t h e spatial structure of the the threshold for onset of the state of spatially-
steady rolls seen in the coexistence regime. As uniform, steady rolls. The dynamics of the T W
mentioned above, these profiles were computed bursting exhibits a complex dependence on e and
 P. Kolodner et al. / Physica D 85 (1995) 165-224 195

LU
Q

<
0.15

0.10

0.05
i i I i i

°°I
70t
50
40
3O
z~
£x
A
(a)

20
i
0 0.24( i

4.00 (b)
i i i 1
0.23~
rr
3.75 .~ ,,x
LU
m 3.50 E 0.230
S
3.25 0.225 A
LU
>,< 3.00
0.220 i
2.75
i
i i i i 0.02 I-
2.50 (c)
0o 90 ° 180 ° 270 ° 360 ° 0.01 ~-
POSITION IN CELL
Fig. 21. Amplitude (top) and wave-number (bottom) profiles
T °°°5I
, 0.0021-
of the steady-roll component in the coexistence regime. The [Link] t
center, middle, and outer curves were recorded at e ' = 0.0005 ~-
1 1 . 3 1 ( 5 ) , 12.93(5), a n d 15.35(5), respectively. For clarity,
0.0002~- t I
these profiles were shifted horizontally to be centered in the 0.011 0,013 0.015 0.017
frame of the figure. E

Fig. 22. Quantitative measures of the evolution of states in

experimental history. As seen in Fig. 22b, the the coexistence regime with increasing Rayleigh number, (a)
modulation frequency O9=oa = (wL + fOR)/2 de- Length FTw of the TW region. A linear fit to these data
intersects the FTw = 0 axis at e'C = 19.22(22). ( b ) Average
creases as e is increased. Fig. 22c shows the burst frequency O~mod = (WL + ~0R)/2. (C) D i f f e r e n c e [oJL -- o~R[
difference IWL- tOR] between the left- and right- between the burst frequencies of the left- and right-TW
TW burst frequencies, computed from temporal components. Error bars correspond to the uncertainties in
the identification of the centers of spectral lines.
Fourier analysis of the spatially-averaged left-
and right-TW amplitude profiles. Below e ' = 15,
these time series exhibit sharp spectral lines at dynamics in the coexistence regime quite clearly.
frequencies which differ by fractional amounts We compute them from two single-channel sig-
ranging from 0.1% to 7%. At e' = 15.35(5), the nals AL,l~(t ) obtained by averaging the left- and
left and right spectral lines are broad and sub- right-TW amplitudes over the spatial regions
stantially overlap, giving ItoL - WRI = 0.001 +- near the left and fight ends of the TW region. We
0.012. This value is represented by the large compute the phase of the left-TW signal by
error bar without a data point in Fig. 22c. At demodulating AL(t ) at ~Omod, thus determining
e ' = 16.50(5), the bursts appear so erratically times t , , n = 1, 2 , . . . , at which this phase passes
that it is not possible to identify a modulation through a chosen value ~b. We then plot AR(t . +
frequency for either TW component - see Fig. 24 6) vs. AR(tn), where the delay time 6 is chosen
below. These data are presumably not reproduc- to be about 1/4 of a modulation cycle. The
ible in detail. In particular, o~L - ~ o R does not qualitative nature of the results does not depend
have a consistent sign. This suggests that the on the exact values chosen for q5 or 6. Fig. 23a
left/right modulation-frequency difference is not shows the Poincar6 section computed for a
due to a spatially-fixed experimental inhomo- regular "coexistence" state at e ' = 13.70(5). In
geneity. this run, the left-right modulation-frequency
Poincar6 sections reveal the nature of the difference was very small: [tOL - - WRI =
196 P. Kolodner et al. / Physica D 85 (1995) 165-224

0.04 exhibit no phase correlation at all, and the

(a) Poincar6 section consists of an ellipse. A similar
0.03 section is obtained for the state of Figs. 16-19,
for which iO~L-- (ORI = 0.01571(15). In the state at
+ 0.02 e' = 15.35(5) in Fig. 23c, the left- and right-TW
rr
< modulation phases are not only uncorrelated but
0.01 ~ O separately erratic. This behavior stands out
clearly in the Poincar6 section but is otherwise
I I I
0.01 0.02 0.03 0.04 hard to detect: Fig. 23c corresponds to the state
AR (tn)
shown in Fig. 20c, which appears visually regu-
0.04
(b)
lar.
Erratic behavior in the coexistence regime can
0.03
take several forms. In Fig. 23c, while the T W
• e•
+ bursts are somewhat irregular, the steady rolls
0,02
4, ,,4 are essentially motionless, as they are in states
which exhibit regular T W b u r s t s - c o m p a r e Fig.
0.01
20c with Figs. 20a,b. A more extreme case of
I L I this type is shown in Fig. 24, in which very
0.01 0.02 0.03 0.04
AR (tn)
steady rolls coexist with T W bursts that are so
0.04 irregular that no value for t0mod can be extracted,
(c) and no Poincar6 section can be computed.
0.03
• • o•
"Coexistence" states which exhibit truely or
~, • o nearly steady rolls are grouped together and
+
represented with open triangles in this paper,
= 0.02

•
':., ,:.., .
OO
0.01
I I I I

0.01 0. 2 0.03 0.04

20 -
AR (tn)

Fig. 23. Poincar6 sections of TW dynamics in the coexistence

15-
regime. (a) e ' = 13.70(5): monoperiodic modulation with
weak phase drift. (b) e ' = 14.46(5): quasiperiodic modula- o
o
tion. (c) e' = 15.35(5): erratic modulation (data of Fig. 20c). 10-
For clarity, the open and closed triangles used elsewhere to
F-
represent these states have been replaced with small, closed
circles.

0.00034(10). The points in the Poincar6 section I

270 ° 360 ° 90 ° 180 °
are clustered in a region whose smaller dimen- POSITION IN CELL
sion represents the noise level. The slight elonga-
Fig. 24. Hidden-line plot of the hipass component of the
tion of this patch of points represents a very signal in the erratic coexistence state at e' = 16.50(5). Here,
weak drift in the relative phase of the left and the appearance of TW bursts is so irregular that an average
right components. In Fig. 23b, computed for a modulation frequency cannot be determined. However, the
steady rolls in this state are truly motionless (they are
state at e ' = 1 4 . 4 6 ( 5 ) for which IWL--WRI---- removed by the hipass filtering). Time domain: 20000 sec =
0.0015(5), the left- and right-TW modulations 390r v.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 197

even if their T W bursts do not exhibit regular I I I [

dynamics. In contrast, it was observed in some
runs that the nominally steady rolls themselves 20
exhibit irregular dynamics. An example of such
behavior is given in Fig. 25. H e r e , the "steady"
rolls move very slowly and erratically, exhibiting o8
occasional defects, and the region of these
" s t e a d y " rolls as a whole sometimes moves and/ V-

or changes in spatial extent. We reserve the

n a m e erratic coexistence states and closed tri-
angular symbol for states in which the "steady"
rolls are irregular in this way. Fig. 26 shows the I
360 ° 90 ° 180 ° 270 °
demodulated T W amplitude profiles for the POSITION IN CELL

hipass component of another erratic coexistence Fig. 26. H i d d e n - l i n e p l o t of t h e h i p a s s c o m p o n e n t of t h e

state. In a third class of dynamical states, the signal in a n e r r a t i c s t a t e in t h e c o e x i s t e n c e r e g i m e at e ' =
system exhibits dispersive-chaos dynamics part of 12.15(10). Time domain: 20000sec = 390%.
the time and coexistence dynamics at other
times. The state at e ' = 10.98(8) in Fig. 14i is such a state; two other runs of this type were
recorded, just below the transition to the coexist-
ence regime. These three runs are represented
I I I I I
by half-open circles in this paper. The possibility
of exhibiting these two new types of erratic
16
dynamical behavior near the transition between
dispersive chaos and the coexistence regime
gives this transition a hysteretic character. This is
12
discussed further in Section 9.1.

~0
o 7.5. E x t e n d e d , spatially-uniform n o n l i n e a r states
o
v
LH Very brief observations were made in this

il
regime. Setting e above e' ~ 20 caused the cell to
fill with n = 40 pairs of steady convective rolls.
This steady state persisted as e was subsequently
decreased. Decreasing e sufficiently caused a
transition to a state of uniform TW; their oscilla-
tion frequency was found to be a linear function
of e over the narrow range - 6 . 8 < e ' < - 2 . 3 ,
I I I I [
0o 90 ° 180 ° 270 ° 360 ° with slope do)/de = - 7 3 ( 3 ) and intercept e * ' =
POSITION IN CELL - 0 . 6 3 ( 1 4 ) . Comparison with detailed measure-
Fig. 25. S p a c e - t i m e d i a g r a m of a n e r r a t i c s t a t e in t h e ments at 41 = - 0 . 1 2 7 gives an approximate value
c o e x i s t e n c e r e g i m e at e ' = 11.97(8). T h e v e r t i c a l axis c o v e r s
of e~ = - 1 6 for the saddle node, below which the
a n i n t e r v a l 16000 sec = 312%. In this run, T W b u r s t s a p p e a r
i r r e g u l a r l y in t i m e , a n d t h e r e g i o n of s t e a d y rolls e x p a n d s a n d system makes a transition back to the quiescent
c o n t r a c t s in s p a c e , e x h i b i t i n g s p a t i o t e m p o r a l defects. state.
198 P. Kolodner et al. / Physica D 85 (1995) 165-224

8. Statistical analyses of dynamical states Exploiting the lack of an amplitude cusp at

Ax = 0 and the reasonably flat shape of the wave-
8.1. Correlation functions number profile there, we fit the main body of the
correlation function using the function Cs(Ax ) =
The oscillatory curve in Fig. 27a shows the A exp(-2Ax2/o -2) cos(k~c). This fit worked well
time-averaged spatial correlation function C~(Ax) for all TW data in this work 3. Fig. 28 shows the
computed from the dispersive-chaos data at s' -- fitted values of the spatial width cr and the
3.23(5) displayed in Fig. 14b. Basically, this central wave number k as functions of e. The
measurement reveals the time-averaged spatial width initially shows a sharp decrease with e in
burst shape in dispersive chaos. To quantify this the dispersive-chaos regime (full circles), quan-
shape, we computed the amplitude and wave- tifying a trend noted in the discussion of Fig. 14
number profiles for this correlation function above. Thereafter, o- increases again, decreasing
using spatial demodulation; these are the smooth
curves in Figs. 27a and 27b, respectively. Moving
away from &x = 0, the amplitude profile decays I I I
approximately as a Gaussian until the amplitude
10
drops by about a factor of 3 from its value at the
peak; further away from the peak, sharp minima $ A
9
are caused by phase defects. These appear as D

singularities in the wave-number profile. 8

@.
• ~ ~
t i I i J [ I I
1
1.0
[ I I
< 3.5
A
0.5 A
O 0.1 A
z n- 3.4 A ~ AA (b)
W
I.-- 0 CO
5w 3.3
rn-r --0.5 0.01 Z OOQ'
O
~ 3.2
0 <
- 1 .( I I r I I I I I
a.1
3.6 I I I t I I
(b) 3.0 I I I
3.4 0 0.005 0.010 0.015
rr C
m 3.2
Fig. '28. The width ~r (a) and mean wave number k (b)
7 3.0
~ extracted from fits to the time-averaged spatial correlation
W function Cs(Ax ) are plotted as functions of stress parameter
~ 2.8
e. The different symbols represent different dynamical states,
2.6 as listed in Table 1. The uncertainties were estimated by
2.4 I
performing the fits over spatial domains of varying sizes. For
-50 -40 -30 -20 -10 0 10 20 30 40 50 clarity, widths for states of counterpropagating wave packets,
S P A T I A L S E P A R A T I O N AX which range from 15 to 35, are excluded from part (a). Note
Fig. 27. Time-averaged spatial correlation function Cs(&r ) the systematic increase of wave number with e.
computed from the dispersive-chaos data in Fig. 14b at
e' = 3.23(5). (a) The oscillatory curve is C,(Ax). The smooth
curve is the demodulated amplitude profile of the correlation 3 Spatial correlation functions computed in the coexistence
function, plotted with a logarithmic vertical scale. The main regime exhibited a Gaussian envelope only when temporal
body of this profile is approximately Gaussian. (b) hipass filtering was applied. If the steady rolls w e r e not
Wavenumber profile of Cs(A~c). The singularities are phase removed in this manner, their contribution dominated the
defects. correlation function.
 P. Kolodner et al, / Physica D 85 (1995) 165-224 199

once more in the coexistence regime. The wave 0.6 i i i i L i i i i

number exhibits a consistent increasing trend (a)
with e, or, more properly, with TW amplitude. o_ 0.5
F-
This effect was noted previously in connection 5[.u
with "repeated transients" in experiments on the rr 0.4
n-
O
same fluid in a short, rectangular cell [59] and is O
09
09 0.3
a general feature of weakly-nonlinear TW con- O
n"
vection. This dependence of wave number on O 0.2
TW amplitude explains the presence of phase I F I I I I I I I
-40 -30 -20 -10 0 10 20 30 40
defects in the spatial correlation function ob- SPATLALsEPARATION Ax/d
served in Fig. 27 and many other dispersive-
i i i I
chaos data sets: the presence of high-amplitude
0.08 ( b ) ~
TW at Ax = 0 is strongly correlated with the
presence of small-amplitude TW far away. These ~3 0.06
have a lower wave number; hence, there must be
phase defects at intermediate distances. ~ 0.04
Fig. 29 shows time-averaged spatial cross-cor-
0.02
relation functions computed from the demodu- I
lated left- and right-TW amplitude profiles of 0° 90 ° 160° 270 ° 360 °
several runs in the dispersive-chaos regime. With POSITION iN CELL
increasing e, the cross-correlation function grows Fig. 30. Physical origin o f the s h a p e o f the cross-correlation
weaker and exhibits an increasing asymmetry. function, c o m p u t e d f r o m dispersive-chaos data at e ' =
Fig. 30 elucidates the meaning of this asymmetric 8.66(5). (a) T i m e - a v e r a g e d spatial cross-correlation function.
(b) D a s h e d curve: left-TW a m p l i t u d e profile o f an isolated
correlation shape. We started in Fig. 30a with the T W burst f o u n d just after the time p e r i o d displayed in Fig.
cross-correlation function computed for the dis- 14g. Full curve: right-TW a m p l i t u d e profile w h o s e cross-
persive-chaos data, set at e ' = 8.66(5) displayed correlation with the dashed c u r v e d yields the curve in part
(a). A s indicated by the n o t a t i o n " x 10", this profile corre-
in Fig. 14g. We then searched the original left-
sponds to w e a k , convectively-amplified right-going T W w h i c h
TW amplitude profile for a typical, isolated TW are strongly attenuated as they pass t h r o u g h the left-TW
burst.

0.8 i i i i i i I i i
burst-this is shown as the dashed curve in Fig.
30b. Deconvoluting these two functions using a
0.6
fast-Fourier-transform algorithm yielded the
solid curve in Fig. 30b-this is the right-TW
5
amplitude profile whose cross-correlation with
0.4 the dashed profile gives the curve in Fig. 30a.
o9
09 Because of the normalization of our cross-corre-
O
n-
O lation functions, the amplitude of this profile is
0.2 arbitrary. In fact, the amplitude of the right-TW
component in the raw data is quite weak at the
I I I I I 110 ~ I [
-40 -30 -20 -10 0 20 30 40 time instant of Fig. 30b. We have labeled this
SPATIAL SEPARATION Ax/d curve "x 10" to emphasize this fact. The expla-
Fig. 29. T i m e - a v e r a g e d spatial cross-correlation functions
nation for the shape of this amplitude profile is
c o m p u t e d f r o m dispersive-chaos data. (a) e ' = 2.00(6) (Fig.
14a); (b) e ' = 3.23(5) (Fig. 14b); (c) e ' = 5.16(5) (Fig. 14d); reasonably straightforward: weak, right-going
(d) e ' = 7.44(5) (Fig. 14f); (e) e ' = 9.34(7) (Fig. 14h). TW propagate near the linear group velocity in
200 P. Kolodner et al. / Physica D 85 (1995) 165-224

this system and are convectively amplified- grow narrower (open triangles in Fig. 28a) and
hence the increasing right-TW amplitude as one move closer together (Fig. 20). Thus, the cross-
moves to the fight from location 0°. However, correlation functions exhibit well-defined peaks
they suffer a strong decrease in amplitude as which move closer to Ax = 0 and grow narrower
they propagate through a strong left-TW burst, as e increases.
due to the competitive cross terms in Eqs. Fig. 32 shows the space-averaged temporal
(17a,b). This is the cause of the weak, asymmet- correlation function Ct(At) computed from the
ric right-TW amplitude profile. As e is de- raw signal in three dispersive-chaos data sets.
creased, the spatial growth rate of the weak TW The envelope of C t ( A t ) decays rapidly with At;
component decreases, so that its spatial am- this decay becomes faster as e increases. The
plitude profile gets flatter, and its cross-correla- very short phase correlation time is caused by
tion with a sharp peak in the opposite TW the nonlinear dependence of TW frequency on
component also grows flatter. Thus, the amplitude in this system: as a typical burst forms
asymmetry in the cross-correlation function in in space and grows in amplitude, its frequency
Fig. 29 tells us that the typical state of the system drops. Thus later TW are not in phase with
in this regime consists of a sharp, strong peak in earlier TW. This interference becomes more
one TW component which controls the convec- pronounced as e is increased. Extracting an
tive amplification of weak TW in the opposite
component.
Fig. 31 shows time-averaged spatial cross-cor- I L I
1.0 (a)
relation functions computed from the demodu-
lated left- and right-TW amplitude profiles of 0.5

several runs in the coexistence regime. In these 0

states, the opposite TW components always
-0.5
occur in spatially-separated bursts; as e is in-
creased and the TW region shrinks, these bursts I I I I I I

i i i
1.o
v (b)
[ i I [ I~1 I I I I
a g d
z 0.5
0.8 o_
i-
z
5 0
o_ kU
rr
0.6 ~- -0.5
5
kU
o
g£ t I I
rr
O0 0.4 I i i
03 1.0 (c)
O)
0
n-
O 0.5
0.2

I
0 10 20 30 ~40 -30 -20 -10 0
-0.5
SPATIAL SEPARATION Ax/d
I I I

Fig. 31. Time-averaged spatial cross-correlation functions -40 -30 -20 -10 ll0 210 3~0 40
computed from the hipass component of data taken in the TIME DELAY At
coexistence regime. (a) e ' = 11.31(5) (Fig. 20a); (b) e ' = Fig. 32. Space-averaged temporal correlation function Ct(At)
12.12(5) (Figs. 16-19); (c) e' = 12.93(5) (Fig. 20b); (d) s' = computed from the raw data in the regime of dispersive
13.70(5); (e) e ' = 14.46(5); (f) e ' = 15.35(5) (Fig. 20c); (g) chaos. (a) e ' = 2.00(6) (Fig. 14a); (b) e ' = 4.07(6) (Fig. 14c);
e ' = 16.50(5) (Fig. 24). For clarity, the horizontal axis has (c) e ' = 6.34(11) (Fig. 14e). The phase-correlation time is
been shifted 180 °. short and decreases rapidly with e.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 201

accurate value for the phase correlation time ~pc In order to eliminate the interference which
from these correlation functions is quite difficult limits the coherence of temporal correlation
because the correlation decays so quickly. We functions computed from the raw signals, we
used a definition of ~-p~based on the first moment have also performed these calculations starting
of the rectified correlation function: from the demodulated T W amplitudes. Fig. 34
n! shows the sum of the space-averaged temporal
correlation functions computed f r o m the left-
~[Link],(AtDI
and right-TW amplitude profiles in three dis-
Ypc = "=' nl (24)
persive-chaos data sets. In all cases, Ct(At )
E Ic,(at.)l decays to a level of about 0.3 at large delays At.
n=l
This behavior reflects that essentially trivial fact
H e r e , Atn = n 6t, with 6t the time step. The sums that the presence of a burst at zkx = 0 is strongly
in this definition exclude the n = 0 term in order correlated with the existence of bursts at other
to eliminate the contribution of rectified noise to times; since demodulation has stripped off the
that channel, and they are cut off at a channel carrier frequency from the signal, distant bursts
n u m b e r denoted n~, to minimize the dominating cannot interfere to product an asymptotic value
effect of noise at large delays, n t is determined of zero. However, these three data sets differ in
iteratively as a multiple arp¢, and the multiplier the manner in which C t approaches this
a has been varied as a test. Fig. 33 shows the asymptote. The two curves for e ' = 2.00(6) and
values of 7p¢ obtained using s e v e r a l different 6.34(11) exhibit a sharp initial decay whose
values of a, plotted vs. e. ~-p~ depends sys- characteristic time A t e 3 0 represents the dura-
tematically on the value of a, so that these tion of a typical burst. For larger delays, these
measurements cannot be considered quantita- two curves differ. The run at e ' = 2.00(6) con-
tive. However, %c clearly decreases systematical- tained so few bursts that they can be seen
! !
ly with e. An exponential fit ~-p¢o~exp(-e /epc ) individually in the dotted curve, and their
would yield a value of about 3.3 for e'pc, and this characteristic double-humped shape can be dis-
n u m b e r is not too sensitive to a.
I I ~ f J v F~
1.0 1.0~- r r-

10 0 0.6 ~-~~-~
9
| Z \ 0.4
Io
7 F-O [ ~,~ 0 50 100 150 200
r- 6 50.6 ~ A t
LU ~s
55
©
~ 4 | • o
rr | I 0.4 _ ..........y ~ . ,.. "-.,..
O ,.. -.~._~ .- .. .~.......~ .
O 3 • |
0.2 ~ I r I I
D- 250 500 750 1000 1250
2 • | TIME DELAY At
Fig. 34. Space-averaged temporal correlation functions
L L L ~ L
0.002 0.003 0.005
0-004 0.006 0.007 0.008 Ct(At ) for dispersive-chaos states, computed from demodu-
E lated TW amplitude profiles, Each curve represents the
Fig. 33. Dependence of the phase-correlation time ~-vcon e in average of the left- and right-TW components. The insert
the regime of dispersive chaos, computed using different shows the behavior at short delay time. Dotted curves:
values of the multiplier c~: circles: c~ = 4.4; squares: c~ = 4.8; e ' = 2.00(6) (Fig. 14a). Full curves: e ' = 6.34(11) (Fig. 14e).
triangles: c~ = 5.2; inverted triangles: a = 5.6. Dashed curves: e' = 10.98(8) (Fig. 14i).
202 P. Kolodner et el. / Physica D 85 (1995) 165-224

cerned. There are many more bursts in the data tions in these curves represent the periodic
set at e' = 6.34(11) represented by the full curve; appearance of TW bursts. The slowly-decaying
their shapes average out more completely, envelope in Fig. 35a represents the regular,
producing smaller features in the correlation nearly-phase-coherent left- and right-TW bursts
function. The dashed curve, computed for the seen in the data set of Fig. 23a, whose Poincard
data set of Fig. 14i at e' = 10.98(8), decays much section is quasi-point-like. The correlation func-
more slowly, reflecting the presence in the data tion in Fig. 35b exhibits slow modulations due to
set of long-lived bursts of very slow TW as well fact that the left- and right-TW bursts exhibit
as short bursts of fast TW. different modulation frequencies-see Fig. 23b.
Fig. 35 presents temporal correlation functions Fig. 35c corresponds to a data set in which the
computed from the demodulated TW amplitude left- and right-TW bursts are not modulation-
profiles for regular states in the coexistence phase-coherent-see Fig. 23c. The difference
regime. These data sets correspond to the Poin- between the dynamics of these last two states
card sections shown in Fig. 23. The fast oscilla- shows up much more clearly in the Poincar6
sections in Figs. 23b,c than in the correlation
functions in Figs. 35b,c.
I I i I I Fig. 36 shows the corresponding results for
1.0
coexistence states exhibiting strongly erratic TW
0.9 dynamics. In Fig. 36a, the correlation drops to

0.8

I I i I , I t i I
0,7
1.0 (a) ~o . 9 1 °
0.6 I I [
0.9 ~ o.a
~. 1.o t, (b) 0.7
0.8 o.6
z) 0.9

~< 0.8

+
oO 0.7
0.6

1.0
t[
I

/
I
I
I
I
I
I
I
i
I
I
(C)
I
!
O

~1
1.o-
I
(b)
I

I
I

I
,o
]

I
. . . .
I

I J

O 0.9
0.9 o 0.9 c~ 0.8
0.7
0.8 0.8
06
0j
0 20 40 60 80
At
0.7
0.7
0.6
0 250 500 750 1000 1250
0.6-
TIME DELAY At I I I I
Fig. 35. Space-averaged temporal correlation functions 0 250 500 750 1000 1250
Ct(At ) computed from demodulated TW amplitude profiles of TIME DELAY At
the hipass component of raw data in the coexistence regime. Fig. 36. Space-averaged temporal correlation functions
Each curve represents the average of the left- and right-TW Ct(At) for erratic states in the coexistence regime, computed
components. The data here correspond to the Poincar6 from demodulated TW amplitude profiles of the hipass
sections in Fig. 23: (a) e' = 13.70(5): monoperiodic modula- component of the raw data. Each curve represents the
tion with weak phase drift. (b) e' = 14.46(5): quasiperiodic average of the left- and right-TW components. (a) e ' =
modulation. (c) e' = 15.35(5): erratic modulation. 16.50(5) (Fig. 24); (b) e ' = 12.15(10) (Fig. 26).
 P. Kolodner et al. t Physica D 85 (1995) 165-224 203

an asymptotic value of ~0.65 within/st = 20 and gions of weak, linear TW in all data sets. As e is
exhibits no coherent oscillatory features-com- increased, longer-lived bursts of slower and
pare this with the original TW amplitude profiles slower TW appear, and this causes the temporal
shown in Fig. 24. The correlation function in Fig. power spectra to extend to lower and lower
36b, computed from the amplitude profiles in frequencies. In the coexistence regime, the
Fig. 26, exhibits residual fast oscillations, due to power spectrum of the raw data extends all the
the weakly-coherent, nearly-periodic emission of way down to w = 0; however, hipass filtering
TW bursts in the original data. The slower remoVes the lowest frequencies from the signal.
modulations seen in this curve are caused by Fig. 38 shows the first few moments of the
erratic dynamics exhibited on very long time
scales by the nominally steady rolls.
, L '(a)
8.2. Temporal power spectra >-
O
30
z
I IJ
:~ 25
O
Fig. 37 shows the space-averaged temporal I.W
e:, 2.0 ~
power spectra computed from the raw signal in Z
iil
A
/s

three dispersive-chaos data sets. All of these w

1.5
signals exhibit spectral power extending up to I I r
1.0
the Hopf frequency % = 3.1438; there are re-
n- el (b)
b--
a • @ •
0.8
. . . . I . . . . i (a)
• A A
o• •e
n- AA A
I- 0.6 I° A A
0
LIJ
0..
;.*
eD
0o
1 0.4
I I i
== I i I
1.0 ,, ~' ~, (c)
AA A
~3 . . . . i . . . . I (b) m 0.5
E
>- z
I-
2 h£
,,=,
-0.
,,g .
E -1 (
40
I P i

rr
(d)
ua 3 . . . . I . . . . i . . . . i (c)~ 3.5
03
• AAA
0') 3.0
O 00 A A zx
2 I-
rr 2.5 A

• " .:~.......~'.
1 2.0

1.5 P I I
0 0.005 0010 0 015
0 1 2 3
FREQUENCY Fig. 38. Moments of space-averaged temporal power spectra.
Fig. 37. Space-averaged temporal power spectra computed (a) mean frequency; (b) width; (c) skewness; (d) kurtosis.
from the raw signal in the regime of dispersive chaos. (a) Higher moments were not calculated for the counter-
e ' = 2 . 0 0 ( 6 ) (Fig. 14a); (b) e ' = 5 . 1 6 ( 5 ) (Fig. 14d); (c) e ' = propagating regime, where line spectra are observed. In the
9.34(7) (Fig. 14h). As e is increased, the spectrum extends coexistence regime (triangles), these measures are distorted
from the Hopf frequency w0 =3.1438 down to lower and by the hipass filter used to separate the TW and steady-roll
lower frequencies. components.
204 P. Kolodner et al. / Physica D 85 (1995) 165-224

temporal power spectra, plotted vs. e. We re- from Fig. 22b. fb exhibits a weakly decreasing
strict our comments here to the region below trend over much of this regime as well.
e ' = 12; in the coexistence regime above this
value, the spectra are distorted by the hipass 8.3. Probability distributions
filtering. The spectra exhibit consistent trends
with increasing e: the mean frequency (Fig. 38a) Fig. 40a shows the probability distribution
drops, the spectral width (Fig. 38b) increases, functions (PDF's) computed from the data in
and the skewness (Fig. 38c) increases. The trend two runs in the dispersive-chaos regime (squares:
of the kurtosis (Fig. 38d) is not as clear; compare
with the values 3.0 and 1.8 exhibited by a
Gaussian and a flat distribution, respectively. I I I t i

10-1
Another useful measure of the temporal spec- 3-
~,, 10 -2
trum is the frequency fb of TW bursts. This is
10 "3
plotted as a function of s in Fig. 39. In the "1
dispersive-chaos regime, we measured fb simply 10_4

by counting the number of bursts in each data o lO-5

O..
set. In this regime, fb increases with e. In the 10-6

counterpropagating regime, the dynamics cannot b o;5 [Link] o. s o. o

SIGNAL AMPLITUDE l ul
be said to consist of repeated bursts, but we 0.10
identified a burst frequency anyway: fb = 2Vg~/F~, (b) A
0.08 AA
where Vg~ is the group velocity plotted in Fig. A
A

~0.06
10a, and F+ = 82.47 is the circumference of the A •

cell. In this regime, fb decreases with e. In the >

0.04
oS•
coexistence regime, we set fb =~%od/Tr, where 0.02
the angular modulation frequency Wmod is taken
i i I q
I I I I
14
- (c) "
12

co 10
I I I I 03 °0o °
0.08 o 8
~AAA I-
& cc
23 6
w' A•

o
4 22 . . . . .
0.06
Z o
I.U 0.005 0,010 0.015
O 000
LU O0
rr Fig. 40. Part (a) shows probability distribution functions
" 0.04
b- ( P D F ' s ) computed from the absolute values lul of the raw
or
signals u(x, t), normalized by setting the integral of the P D F
00
m to unity. Squares: first dispersive-chaos state at e ' = 200(6)
0.02 O00 (Fig. 14a). Circles: dispersive-chaos state at e' = 8.66(5) (Fig.
14g). Triangles: erratic state in the coexistence regime at
I I I I e ' = 15.35(5) (Fig. 20c). N o t e that open symbols different
0 0,005 0,010 0.015 from those listed in Table 1 have been used. A s s is
E increased, the distribution extends to higher and higher
Fig. 39. The dimensionless burst frequency fb is plotted as a amplitudes. Parts (b) and (c) show the second (b) and fourth
function of e. In the dispersive chaos regime, fb is computed (c) m o m e n t s of these P D F ' s , plotted as functions of e.
by counting the number of bursts in each data set. In the Moments were not computed for the counterpropagating
counterpropagating and coexistence regimes, fb is computed regime- they were periodic rather than erratic. A s e in-
from the temporal modulation spectrum by dividing the creases, the kurtosis drops from a very high value to one
angular modulation frequency by ~r. closer to the value 1.8 exhibited by a flat distribution.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 205

e' = 2.00(6); circles: e' -- 8.66(5)) and one in the (25) applies to individual bursts, then the P D F
coexistence regime (triangles: e ' = 15.35(5)). In computed from an entire dispersive-chaos data
processing this last data set,we retained the set should also exhibit this form. To test this
contribution of the dc signal due to the steady scaling, we observe that, if Eq. (25) holds, then
rolls. As e increases from the lowest dispersive- we should observe
chaos data set to the highest data set in the
1/AP(A) ~ e + 2gA 2 . (26)
coexistence regime, the PDF evolves from a
nearly exponential to an almost flat shape. These That is, plots of 1/AP(A) v s . A 2 for various
trends are reflected in the variance and kurtosis values of e should appear as straight lines with
of the PDF, which are plotted vs. e in Figs. 40b slopes that are independent of e. We find partial
and c. The variance increases with increasing e, confirmation of the scaling of Eqs. (25) and (26)
reflecting an essentially monotonic increase in in the dispersive-chaos regime. Fig. 41a shows
spectral energy with e - see Figs. 51-53 below. PDF's for four dispersive-chaos data sets plotted
The kurtosis decreases from very high value at in this format. At sufficiently low values of e, the
low e to a one close to the value 1.8 exhibited by curves do indeed exhibit linear behavior at
a flat distribution at the upper end of the intermediate amplitudes. However, the slopes of
coexistence regime. the linear regions decrease with increasing e.
In Ref. [52], Kaplan et al. present a PDF Another potentially important scaling property
computed from the amplitude A(x, t) extracted can be ruled out by analysis of PDF's computed
by demodulation of data corresponding to the from dispersive-chaos data. Plotting P(lu]) vs. lul
dispersive evolution of a single, unidirectional on log-log paper, as shown in Fig. 41b, we
TW burst in an otherwise erratic state. Their observe that the PDF's obtained at different
experiment was carried out in a long, rectangular values of e cannot be made to overlap by simply
cell, using a fluid similar to ours, at e ' - - 8 - shifting the axes of the graphs. This means that
midway between our runs of Figs. 14f and g. we cannot find a universal function f and scaling
They suggest that, on the basis of an analytical functions a(e) and b(e) for which
solution of the C G L E in the form of a single
P(lu]) = a(e) f(]u I/b(e)). (27)
burst, the PDF for the TW amplitude should
have the form The PDF's computed from the demodulated TW
amplitudes, as in Fig. 41a, also fail to obey the
1
P(A) ~ A(e + 2gA 2) ' (25) scaling of Eq. (27).

where g is the real part of the cubic coefficient in 8.4. Comoving time averages
the C G L E . They point out that this scaling
should be valid only for sufficiently small A, Comoving time-averaged patterns are com-
since dissipative effects not modeled by the puted by shifting the instantaneous pattern to a
C G L E become important at high signal am- flame of reference moving at velocity v m and
plitudes. We note that this scaling should also be then averaging in time. Thus, the comoving
invalid near zero amplitude: in selecting the part time-averaged pattern of space-time signal u(xj,
of a data set which covers a single burst, the ti), j= I , . . . , M , i = I , . . . , N is:
fraction of space-time at near-zero amplitude N
which is excluded is in fact arbitrary. Uv~(Xj) = Z 12(Xi -- Umti, t i ) / N . (28)
Since dispersive chaos consists of the evolution i-1
of TW bursts whose overlap in space-time is We have computed the comoving time-averaged
small, we would expect that, if the scaling of Eq. pattern as a function of comoving velocity v m for
206 P. Kolodner et al. / Physica D 85 (1995) 165-224

100000 With this normalization, a time-invariant pattern

(a) o
propagating at a steady velocity v m would exhibit
80000
° 0 v e
Urms(Vm) = 1, while a completely random signal
e o
~'-60000 s~ 4o°° would give Urms(Vm)=N-U2=0.0123. Fig. 42
E" m ~
o%
o
~ shows the rms time-averaged pattern amplitude
40000
Urms(Vm) as a function of comoving velocity v~
20000 ~ oO2°
for the three dispersive-chaos data sets whose
temporal power spectra are shown above in Fig.
0g 0.005
I
0.010
t
0.015
I _ _
0.020
I
37. At the lowest value of e (Fig. 42a), where
A2 there is a reasonably well-defined average T W
I I L phase velocity @h = 1.03 (see Fig. 14a), the time-
averaged amplitude is enhanced well above the
10"2 ~ % expected value N -I/2 for Vm~-+--6ph. At these
_~.
_= 10-3
values of Vm, the lines of constant T W phase in
~" 10_4

0.05 I I

10.6~- (b) .= o
I I I D~o 0.04
10-3 10-2 10-1
lul 0,03
Fig. 41. Tests of scaling of PDF's computed from data in the 0.02
dispersive-chaos regime. For clarity, open symbols different
from those listed in Table 1 have been used. These symbols 0.01 L
are consistent with those used in Fig. 40a. (a) To test for the 0
scaling of Eqs. (25) and (26), we plot 1lAP(A) vs. A 2, where _g o.030
P(A) represents the sum of the PDF's computed from the
m 0.025
demodulated left- and right-TW amplitudes: Squares: e ' = (3
2.00(6) (Fig. 14a); lozenges: e' = 4.07(6) (Fig. 14c); inverted <
EE
0.020
LU
triangles: e ' = 6.34(11) (Fig. 14e); circles: e ' = 8.66(5) (Fig. >
<
o.o15
14g). To show more clearly the linear regime in the PDF's at
0.010
intermediate amplitudes, these data have been smoothed
slightly. (b) To test for the scaling of Eq. (27), we present a ,,~ 0.005
log-log plot of the PDF's computed from the absolute values ~E 0
lul of the raw signals u(x, t), for three of the data sets in part o
z 0.03
(a). T h e evolution of the shape of the PDF as e increases
rules out of the scaling of Eq. (27).
0.02

several dispersive-chaos data sets. To measure 0.01

their amplitudes quantitatively, we compute an
0
rms space average Urns(Ore), normalized to the -1 0 1
variance of the data values graphed in Fig. 40b: COMOVING VELOCITY

Urms(Vm) = _ (Uvm(Xj))2/m
>,,2/ Ur, (29)
Fig. 42. The rms spatially-averaged amplitude Urrns(Vm) of
the comoving time-averaged pattern is plotted as a function
of comoving velocity v m for the three dispersive-chaos states
shown in Fig. 37, (a) e' = 2.00(6) (Fig. 14a); (b) e' = 5.16(5)
where (Fig. 14d); (c) e ' = 9.34(7) (Fig. 14h). As e is increased, the
region of appreciable pattern amplitude extends from the
linear phase velocity Ivphl = 1.0306 down to lower and lower
ur = (u(xj, ti))2/NM (30) velocities, mirroring the behavior of the temporal spectra in
i, -1 Fig. 37.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 207

individual left-TW or right-TW bursts line up in I I I I I

space, leading to constructive interference and o
w
0.001

thus to an enhanced time average. As e is

n'<
increased, the resulting peaks broaden and move wz 0
>(.9
towards lower Vm, mirroring the behavior seen in ILl
the space-averaged temporal power spectra. -0.001

The top frame of Fig. 43 shows the time- I J I I

averaged pattern computed at the peak value 0.0015 I r I I--
v m = +0.998 for e' = 2 . 0 0 ( 6 ) . The bottom two
w
frames show the amplitude and wavenumber o 0.0010
},.-
profiles of this time-averaged pattern; these were J
13..
computed using spatial demodulation. The time- 0.0005
,<
averaged pattern consists of three separate am-
plitude bumps which are separated by phase I I
defects. This time-averaged pattern is not coher- I I I
ent. The pattern computed for v m = - 0 . 6 4 6 at
w 4.0-
e ' = 5.16(5) in Fig. 44 is even less coherent. At m

this e, we have seen that TW bursts appear more z 3.0

LLI
>
,<
2.0

[ I I I
w 0.001 0° 90 ° 180 ° 270 ° 360 °
(..9
POSITION IN CELL
rr<
uJz 0 Fig. 44. T i m e - a v e r a g e d pattern in the dispersive-chaos state
>(.9
<,g at e ' = 5 . 1 6 ( 5 ) ( F i g . 1 4 d ) , as c o m p u t e d in a f r a m e m o v i n g at
kU
v = - 0 . 6 4 6 . This pattern consists o f m a n y m u t u a l l y - i n c o h e r -
--~ - 0 . 0 0 1
I- ent bursts separated by phase defects.
I I I

0.0015
frequently than at e' = 2.00(6), and that they are
uJ

I-
.._J
13..
0.0010

:~ 0 . 0 0 0 5
,<
1 narrower in space. Because of this, there are
many more phase defects in this time-averaged
pattern.
The incoherent nature of the time-averaged
I
patterns in Figs. 43 and 44 is partly due to the
rapid loss of coherence in individual T W bursts
4.5 - I I I I ~--
n-"
that was seen in the raw temporal correlation
u.I
m 4.0 functions in Fig. 32. Another important contri-
bution is due to the fact that different T W bursts
z 3.5
w are not mutually coherent. When shifted into the
>
,< 3.0
comoving frame, these bursts end up in random
2,5
-i t r i ~_ locations with uncorrelated phases. If they all
0° 90 ° 180 ° 270 ° 360 ° were to end up at the same location with un-
POSITION IN CELL
correlated phases, then they would interfere
F i g . 4 3 . T i m e - a v e r a g e d pattern in the first dispersive-chaos
more completely, and the rms pattern amplitude
state at e ' = 2 . 0 0 ( 6 ) ( F i g . 1 4 a ) , as c o m p u t e d in a f r a m e
m o v i n g at v = + 0 . 9 9 8 . T w o p h a s e defects separate regions o f would be closer t o N -1/2. The residual excess
c o h e r e n t TW. rms amplitude and the noisy spatial phase seen
208 P. Kolodner et al. / Physica D 85 (1995) 165-224

in the actual time-averaged patterns are the

(a) f
s y m p t o m s of the random, burst-like nature of 0.5

the spatiotemporal signals in dispersive chaos.

0.4 _ ~ d

8.5. Biorthogonal decomposition

b

As discussed in Section 4 above, the B O D is a ?~

I
0.3 a

>-
decomposition of the signal u(x, t) into a sum of n
i i i i i
20
i i L I
40
I I I I I F
60
I r i r
80
0
rr
products of spatial and temporal eigenfunctions TIME DOMAIN "T ( 1 0 0 0 sec)

i i i , L L I ~ L '
called topos and chronos, respectively. F r o m the 0.7

spectrum of this decomposition, we can extract

global quantities which characterize the spectral 0.6 ~ 9
complexity of the signal. The reader should be
aware that we have applied the B O D separately
0.5 k
to each data set, thus obtaining a new set of I
rn
topos and chronos for each value of e studied. n
0,4 r , i t i i _ r i i i L
T h e evolution of the topos and chronos with e 0 20 40 60 80 100 120
I i
140

has not yet b e e n explored in this work. TIME DOMAIN "T ( 1 0 0 0 sec)

Before presenting detailed descriptions of the Fig. 45. The saturation behavior of the global entropy H is
dynamical behavior measured using B O D tech- illustrated by plotting H(T) as a function of the length T of
the time domain used for the calculation, s increases mono-
nique, we discuss an important systematic issue. tonically on moving through the curves from a to n. (a)
In applying this technique, we must ensure that Counterpropagating regime. Curve a: e ' = 0.18(7); curve b:
the duration T of each data set is long enough e ' = 0.43(7); curve c: e ' = 0.65(7) (data of Fig. 8); curve d:
e' = 0.95(7); curve e: e' = 1.43(7); curve f: s' = 1.66(7). The
that further data acquisition would not affect oscillations in H(T), most easily seen in curves b and c
results of the decomposition. We assess this by between times 6000 and 20000 sec, correspond to the circula-
computing the spectral energy E and entropy H tion of the wave packets around the experimental cell. (b)
Upper curves: dispersive-chaos data. Curve g: e'=2.00(6)
for subsets t E (0, T) of the time domain r ( T <
(Fig. 14a); curve h: e' =4.07(6) (Fig. i4c); curve i: e'=
T ) , and studying their saturation with increasing 6.34(11) (Fig. 14e); curve j: e ' = 10.00(5). Lower curves:
as a function of e. Typical results are shown in coexistence regime. Curve k: e' = 12.12(5) (Figs. 16 and 17);
Figs. 45 and 46. Fig 45a shows the saturation of curve 1: e' = 13.70(5) (Fig. 23a); curve m: e' = 14.46(5) (Fig.
23b); curve n: e' = 15.35(5) (Figs. 20c and 23c).
the spectral entropy for several data sets in the
counterpropagating regime. After an initial
sharp increase f r o m 0 at small T, H ( T ) exhibits tral entropy for several data sets in the dis-
t e m p o r a l oscillations corresponding to the circu- persive-chaos regime (curves g - j ) and in the
lation of the wave packets around the cell. These coexistence regime (curves k - n ) . In the disper-
oscillations die out as time progresses; H ( T ) sive-chaos regime, the entropy is seen to increase
exhibits further slow evolution for some data sharply from 0 at very small T and then to
sets. To average out the oscillations and to continue increasing very slowly before saturat-
ensure saturation of subsequent slow drifts, all ing. On average, the time required for saturation
data sets had length >55000 sec, and we quote increases with e in this regime. All of our
global results c o m p u t e d from subsets of length T dispersive-chaos data sets were sufficiently long
equal to the time taken for the wave packets to (length 132000 sec) for the entropy to be com-
m a k e exactly nine round trips. pletely saturated. In addition to these long-time
Fig. 45b illustrates the saturation of the spec- trends, H ( T ) exhibits jumps with increasing T,
 P. Kolodner et al. / Physica D 85 (1995) 165-224 209

800
103 (a)

600

z 400
w
1°1t
t -Q
Lu
10-3 / I I I I I I r
-3
(5
200
105 ~ ~ ' ' ' '
~ (b)

103 ~ %
o 0 20 40 60 80 1oo 120 140
TIME (1000 sec) <= 101

Fig. 46. T h e global energy E ( T ) is plotted as a function of 10-1

t h e length T of the time d o m a i n used for the calculation.
C u r v e a: e ' = 0 . 9 5 ( 7 ) ; curve b: e ' = 2 . 0 0 ( 6 ) (data of Fig. 10-3
14a); curve c: e ' = 3.23(5) (Fig. 14b); curve d: e ' = 4.07(6) I I I I I I I
0 30 60 90 120 150 180
(Fig. 14c); curve e: e ' = 6.34(11) (Fig. 14e); curve f: e ' =
rl
7.44(5) (Fig. 14f); curve g: e ' = 10.00(5); curve h: e ' =
F i g . 47. E i g e n v a l u e s p e c t r a p r o d u c e d b y B O D analysis. (a)
10.98(8) (Fig. 14i).
Dispersive-chaos data. From bottom to top: s' =2.00(6)
( F i g . 14a); e' = 3 . 2 3 ( 5 ) ( F i g . 14b); e' = 4 . 0 7 ( 6 ) ( F i g . 14c);
s ' = 5.16(6) (Fig. 14d); e' = 6.34(11) (Fig. 14e); e' = 7.44(5)
(Fig. 14f); e' =9.34(7) (Fig. 14h). Successive curves are
corresponding to the inclusion of individual T W
displayed upwards by a factor of 2 for clarity. (b) Coexist-
bursts in the time domain. Because new bursts ence regime. From top to bottom: s' = 12.12(5) (Figs. 16 and
are not coherent with previous bursts, they add 17); e ' = 12.93(5) (Fig. 20b); e ' = 13.70(5) (Fig. 23a); e ' =
to the complexity of the data set and increase the 14.46(5) (Fig. 23b); e ' = 15.35(5) (Figs. 20c and 23c); s ' =
16.50(5) (Fig. 24). Successive curves are displaced down-
spectral entropy. For the regular data sets in the wards by a factor of 4 for clarity.
coexistence regime shown in curves k - n in Fig.
45b, the entropy saturates rapidly. Irregular
states in this regime, however, required an
e n o r m o u s time to saturate; in these cases, we curve. A t the lowest value of e, the spectrum
used data sets of length 140000 sec instead of exhibits an inflection point separating a rapid,
132000 sec. nearly exponential decay at low n f r o m a flatter
Fig. 46 shows the evolution of the spectral region a t high n. As e increases the, spectra
energy E with time domain T for several values b e c o m e smoother, flatter, and stronger; the
of e in the dispersive-chaos regime. E ( T ) in- spectral entropy increases with e in this state,
creases in jumps caused by the inclusion of new and t h e s p e c t r a l energy increases. In Fig. 47b,
T W bursts. Averaging over these jumps, an we show spectra c o m p u t e d in six r e g u l a r states in
approximately linear trend with T is observed in the coexistence regime, with e now increasing
almost all cases. T h e saturation of the " p o w e r " from the top curve to the b o t t o m curve. In this
or global slope of these curves is much faster dynamical state the spectral entropy decreases
than that of the entropy. with e. In some of these curves, quasi-two-fold
Fig. 47a shows seven fully-saturated spectra degeneracies can be seen, as represented by
c o m p u t e d in the dispersive-chaos regime, with e pairs of nearly equal adjacent eigenvalues. Such
increasing f r o m the b o t t o m curve to the top near-degeneracies are reminiscent of the pres-
210 P. Kolodner et al. / Physica D 85 (1995) 165-224

ence of TW in the system4; they are present in the square of the norm of the vector ~t(x) in the
all the BOD spectra computed in this work but space x(X).
are visible in only a few instances in Figs. 47a,b Fig. 48 shows the profiles of the spatial and
because of the small format of these graphs. temporal energy and entropy for a state in the
In Figs. 48-50, we illustrate the dynamics counterpropagating regime at e' =0.95(7). In
observed in the counterpropagating, dispersive- this state, the spatial energy and entropy Es(t )
chaos, and coexistence regimes by plotting the and Hs(t ) (Figs. 48a,b) exhibit fast temporal
spatial and temporal energies and entropies oscillations at twice the frequency of the underly-
extracted from the BOD. Recall from Section 4 ing TW; these are modulated on a slower time
the spatial energy and entropy, Es(t ) and Hs(t),
are functions of instantaneous time t that reflect
the complexity of the spatial pattern recorded at
that time. The introduction of an ordered spatial
_~o.o8~
structure into a previously featureless pattern at o-w<rrl-(9' -<0"02
)~ I
a given time step will cause a sudden increase in ~ 0.01F
the spatial energy and a corresponding decrease 0/ I I I I I
in the spatial entropy, as compared to the values
computed at the previous time step. These .~ 3:: 0.80
signatures, which we see in all of our data, are < ~ 0.75~
essentially the opposite of the jumps seen in the
~i 0.70F / ' I I I I
growth of both the global energy E ( T ) and
0 8 16 24 32 40
entropy H ( T ) shown in Figs. 45 and 46. There, TIME (1000 sec)
new bursts are added in with all previous bursts i ]
as the computation is carried out over larger ,~d °'15I
time domains T. The spectral energy and en-
tropy associated with this growing collection of
mutually incoherent dynamical structures both
rr>_ 0,I0
oo
o,o5
t
O/f I I f I
tend to increase in jumps with 3?. Analogous 0.85 t i
comments apply in the spatial domain to the `5=~o.6o~
o~
temporal energy and entropy, ET(x) and HT(x). n 0 0,75 ~-
~rr ] I I I]l
t
Lul- l
At each spatial point, these functions measure 1-~ 0,70 ~-
the complexity of the time series signal measured 0.65 r I
0 9Po
° i°
180 ,°
270 360°
at that point, independently of the signal at other POSITION IN CELL
spatial points. Recall also that ET(x) is the Fig. 48. The top two frames show the instantaneous spatial
square of the norm of the vector 7/x(t) which energy Es(t ) (a) and spatial entropy Hs(t ) (b) as functions of
represents the dynamics of the system in the time for a state in the counterpropagating regime at e ' =
0.95(7). For clarity, these plots and those in Figs. 49 and 50
temporal configuration space x(T), and Es(t ) is cover only half of the temporal duration of the data set. The
slow modulations (period 3533 sec) correspond to the propa-
gation of the wave packets halfway around the cell. The fast
carrier oscillations (frequency 0.0191 Hz) have twice the
4 The relationship between quasi-twofold degeneracies and frequency of the underlying T W and are due to beating in
the presence of TW can be appreciated most easily in the standing-wave regions of space-time. The bottom two curves
case when the B O D reduces to a Fourier transform. In this show the temporal energy ET(X) and the temporal entropy
case, the topos and chronos are just sines and cosines, and to HT(X) as functions of spatial position. Both profiles show
construct a T W from products of such functions requires not localized oscillations at twice the average T W wavenumber,
just one term but two: c o s ( o ) t - k x ) = c o s ~ o t c o s k x + caused by standing waves produced in regions where the
sin tot sin kx. wave packets overlap.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 211

scale with a modulation period equal to half the o.5 ( ~

time taken for the wave packets to circulate ...at.u~
<>-
0.4
around the cell. Careful inspection shows that N~ 0.3
D-UI
¢o~ 0.2
the fast carrier oscillations in Hs(t ) are out o f
0.1
phase with those in Es(t ). These observations are 0
0.851 ~ i i I d l I
easily explained. Twice per round trip, the wave :z~ (b)
packets overlap and interfere to produce local-
ized standing waves. The nature of standing _~.>-
~
°8°
0.75
I
waves is to oscillate in time at the carrier
frequency between a featureless pattern (which 0.70 [-
0
I i
10 20
i
30
I
40
r
50
i
60
has low spectral energy and high entropy) and a TIME (1000 sec)

coherent pattern of spatial oscillations (which has &5|i i i i =

high spectral energy and low entropy). Thus, at

~ 2.5
the times when the wave packets overlap, Hs(t ) rr>..
0(.9 2.0
Q.n-
and Es(t ) exhibit fast, out-of-phase temporal :~w
uJz
1.5
t--tu 1.0
oscillations. Since the BOD is not sensitive to 0.5 I I I I i
the sign of the signal, the spatial pattern is in
effect rectified, and the temporal oscillations in ~.>.~o.~or
Ea-(t) and HT(t ) occur at twice the TW fre- 0 ° - 0.78 F
~rr°-O

quency. At other times during the wave-packet w }- 0,76 F /

circulation, the pattern consisting of spatially- ~"=' 0.74 I- I I ] I

0 90 ° 180 ° 270 ° 360 °
separated left and right TW which do not inter- POSITION IN CELL
fere; thus at those times in the modulation cycle,
Fig. 49. The spatial energy (a), spatial entropy (b), temporal
Es(t ) and Hs(t ) attain intermediate values and energy (c), and temporal entropy (d) are plotted for the
exhibit only weak oscillations on the fast time dispersive-chaos data set of Fig. 14f at e ' = 7.44(5). The
scale. spatial energy (a) and entropy (b) exhibit bursts and dips,
respectively, caused by the appearance of TW bursts. The
These features are also reflected in the spatial temporal energy and entropy exhibit large-scale spatial
profiles of the temporal energy and entropy structure due to the finite number of bursts in the data set.
HT(x) and ET(X) plotted in Figs. 48c,d. These
profiles once again exhibit out-of-phase oscilla-
tions, localized in regions centered at the points the dispersive-chaos state shown in Fig. 14f (e' =
x 1 and x 2 where the wave packets overlap and 7.44(5)). In this regime, the spatial energy Es(t )
produce standing waves. Since the spatial phase exhibits erratic bursts corresponding to the ap-
of the standing waves is not exactly the same pearance of new TW bursts (Fig. 49a). Some of
each time the wave packets pass through each these exhibit the characteristic double-humped
other, the spatial oscillations in ET(X) and HT(x) shape noted in connection with Figs. 3, 4i and 7
are washed out by averaging over many circula- above. Because there is little spatial overlap of
tion cycles. Thus, the contrast in the time-aver- left- and right-TW bursts in this regime, the fast
aged oscillations is weak, and the profiles appear oscillations in Es(t ) caused by standing waves in
noisy. In the configuration-space representation the counterpropagating regime are essentially
of these dynamics, r/x(t) makes large excursions absent in dispersive chaos. Each TW burst adds
from the origin at the locations xl, a and ap- a coherent spatial structure to the pattern; thus,
proaches the origins as one moves away from the bursts in Es(t ) coincide with dips in Hs(t )
X1,2" (Fig. 49b). These characteristics of the disper-
Fig. 49 shows the corresponding profiles for sive-chaos regime have been noted previously
212 P. Kolodner et al. t Physica D 85 (1995) 165-224

[60]. The profile of the temporal energy ET(X ) of a regular state at e' = 13.70(5) in the coexist-
shown in Fig. 49c exhibits large-scale spatial ence regime. The spatial energy and entropy
structure. The temporal entropy HT(X) (Fig. shown in Figs. 50a,b were computed using tem-
49d) shows corresponding nonuniformities; once poral-hipass-filtered versions of the data, to
again, regions of high temporal energy exhibit eliminate the influence of the localized steady
low temporal entropy. The reason for the rolls on the dynamics. These profiles exhibit
nonuniformity of the dynamics is that, even out-of-phase oscillations at a modulation fre-
though this data set is long enough to completely quency corresponding to the regular emission of
saturate the BOD in the sense defined in Figs. 45 TW bursts. Since the left- and right-TW bursts
and 46, it still contains only a finite number are spatially separated in this state, there are no
(approximately 120) of TW bursts. At the end of standing waves and thus no oscillations on faster
this section, we will have more to say about the time scales in Hs(t ) or Es(t ). Including the lopass
nonuniformities revealed by these time-averaged temporal component of the data shifted these
spectral measures. profiles vertically without altering their oscillat-
Fig. 50 illustrates the spatiotemporal dynamics ory behavior. The temporal dynamics of this
state are illustrated in Figs. 50c,d. Here we show
0,4/ I I I t t I t HT(x ) and Ea-(x) computed from both the tempo-
ral-hipass-filtered version without steady rolls
<>. 6.3 t
(labeled " H F " in these graphs) and the spatial-
<rr
~
a- w 6.2 t (a hipass-filtered version which retains the steady-
roll component. The strong spatial oscillations in
0.1 I I I
, bI iI iI iI tL 1I T - - the profiles computed from the spatially-filtered
CO 0 . 7 0
_11 data correspond to the steady rolls. When this
)0 -g~ 0.66 component is removed from the signal, the
temporal energy drops to zero and the entropy
0.62
I I I i F increases nearly to 1 in this spatial region, values
o lo 20 30 ~o ~o 60
TIME (1000 sec)
consistent with the weak, noisy output of the
temporal hipass filter used to suppress the
~ f f 80f ~(e) I I I steady-roll component. In the TW region, the
rr~ 60 shape of the temporal energy curve is the same
O~
o-n- 40 for both versions and reflects the convective
wz
I-- w 20 growth of TW bursts as they propagate out from
o
1.0 | ~ ) j I I the location 50°. Once again, the profile of the
(d) , HF ~ ' ~ x ~ - ~ ~
temporal entropy is out of phase with that of the
0 o- 0.6
a.O
temporal energy. It may seem strange that the
w~7 0.4 two temporal-entropy curves differ in shape in
F-LU
0.2
I I I I I
the TW region even though the dynamics are the
0 90 ° 180 ° 270 ° 360 = same in that region for both versions of the data.
POSITION IN CELL
This is an artifact of the definition of HT(x) in
Fig. 50. The spatial energy (a), spatial entropy (b), temporal
energy (c), and temporal entropy (d) are plotted for a data
Eqs. (12) and (13), which depends on the
set in the coexistence r e g i m e a t e' = 13.70(5) (Fig. 23a). In behavior of the topos qSn(x) derived from the
(c) and (d), the curves labeled " H F " were computed from dynamics integrated over the entire system.
data that had been passed through a temporal hipass filter to Fig. 51 shows the evolution of the global
suppress the steady rolls seen between locations 160 ° and
310% The other two curves were computed using data in properties of the BOD with stress parameter e.
which the steady-roll component was retained. All the results in these graphs were computed
 P. Kolodner et al. / Physica D 85 (1995) 165-224 213

I J I I in Fig. 51a). A fit of the data in the range

(a) A
uJ 1000
>-
&
2.0<e'<11.0 to a functional form E ( e ' ) =
qDgAA A A
(.9
750 A
E 0 e ~'/~ gave an inverse log slope ~ = 4.23(15)
z 500 • ~ og,. A and a prefactor E 0 = 62.1(3.5). Below, it will be
useful to recall that the fitted spectral energy
25o
increases by a factor F = 6.63 on increasing e'
I F I from 2.0 to 10.0 in the dispersive regime. The
decrease of the global energy in the coexistence
0.7
I(b) ~q, oennO ; o qVO o ° ° ° Q °1~o~, regime above s ' = 1 3 is associated with the
0.6 • Ak
shrinking of the T W region documented in Fig.
F'-
z 0.5
DD • A A 20.
LIJ
A The behavior of the global entropy H(e),
m 0.4
£ shown in Fig. 51b, is markedly different for the
_1~ I I [
three dynamical regimes. In the counter-
propagating regime, H increases linearly with e.
i i
3o (e) .. • The scatter in these points is noteworthy and is
omo~ D° ~ discussed below. In the dispersive-chaos regime,
2o ~eq • • H ( e ' ) increases linearly with e' up to e' ~ 8 and
g
then flattens out slightly. A fit of the data in the
lO A&A range 2 . 0 < e ' < 8 . 0 to a linear dependence
A
A
¢ ¢ t
H ( e ' ) = H o + ( e / e ~ ) gave an inverse slope e H =
o
0.005 0.010 0.015 85(5) and an intercept H 0 = 0.5800(37). Below,
E
we will use some specific values taken from this
Fig. 51. Global measures of dynamical behavior, derived
from BOD analysis, are plotted vs. e. (a) Global energy E;
fit: H = 0.604 at e ' = 2.0, and H = 0.698 at e ' =
(b) Global entropy H; (c) Number of terms ng0 in the BOD 10.0. In the coexistence regime, H(e) decreases
expansion required to recover 90% of the spectral energy of again. As the T W region shrinks in this regime,
the original data. In the coexistence regime, the calculations
the T W bursts become more ordered, in a
were performed using temporal-hipass-filtered data to isolate
the TW dynamics and suppress contributions from steady spectral sense, even though their timing becomes
rolls. The counterpropagating, dispersive-chaos, and coexist- erratic at the highest values of e s t u d i e d - s e e
ence regimes are readily distinguished by the e dependence Fig. 24 above.
of these global quantities.
In reconstructing a dynamical state using the
BOD, one would often like to truncate the
using temporal-hipass-filtered data, so that, in expansion in Eq. (5) at the lowest possible
the case of states which exhibited regions of order. The global entropy is an indirect measure
steady rolls (triangles and half-filled circles in of the number of terms required to reconstruct
Fig. 51), the contribution of the steady-roll the data to within a desired precision. A more
c o m p o n e n t to the dynamics has been suppressed. direct measure is simply to find the limit nf < N
The effect of retaining this component in these for which the energy sum in Eq. (6) equals a
data sets is explored below in Fig. 53. The given fraction f of the global energy E, as in Ref.
spectral energy E, shown in Fig. 51a, exhibits an [9]. Fig. 51c shows the e dependence of ng0, the
increasing trend up to e' ~ 12 in the coexistence number of terms required to recover 90% of the
regime. E(e) is a linear function in the counter- spectra energy in this sum. rtg0 depends on the
propagating regime (open squares) but increases stress parameter in a way that is similar to the
approximately exponentially in the dispersive- dependence of the global entropy. Indeed, in a
chaos regime (circles and first triangular symbols plot of ln ng0 vs. H, the data points in the
214 P. Kolodner et al. / Physica D 85 (1995) 165-224

different r e g i m e s are found to lie on slightly independently (this is not too surprising, given
offset, nearly parallel straight lines. For the the extremely weak nonlinear coupling between
w e a k e s t dispersive-chaos data, n 9 0 ~ 1 8 . The them and the tendency for isolated TW to burst
d y n a m i c s of this state e v o l v e in a configuration and collapse as the amplitude increases). The
s p a c e w h o s e d i m e n s i o n is of this order. BOD technique has thus revealed the presence
In the discussion of Fig. 12 in S e c t i o n 7.2 of a symmetry-breaking bifurcation which es-
a b o v e , it w a s n o t e d that s o m e of the data sets caped detection by the statistical measures dis-
acquired in the counterpropagating r e g i m e ex- cussed in Sections 8.1-8.4 above. The greatest
hibit w e a k l y erratic dynamics. T h e B O D has L - R energy difference is exhibited by the state
turned o u t to b e a v e r y sensitive w a y to reveal of Fig. 12 at e' --- 1.23(7). Fig. 51b shows that the
such b e h a v i o r . In Fig. 52, w e plot for t h e s e data raw data in this run exhibited an anomalously
the d e p e n d e n c e on e of the global energy and high global entropy as well.
e n t r o p y calculated n o t f r o m the raw d a t a but In Fig. 53, we compare the global energy and
s e p a r a t e l y f r o m the d e m o d u l a t e d left- and right- entropy computed from data sets at high e in
T W amplitudes. For e ' < 0.95, b o t h c o m p o n e n t s which the low-temporal-frequency component
exhibit identical global e n t r o p y and energy.
H o w e v e r , the e n t r o p i e s of the t w o c o m p o n e n t s
differ f r o m all data a b o v e this value, and the
5OOO (a) A
e n e r g i e s begin to separate for e' ~ 1.23, indicat-
w A
ing that the t w o T W c o m p o n e n t s are e v o l v i n g 4000 a
A
a
3000
w

2000 a
20 I I £
(a) • 1000
A
[] ~ []
15 A
I
::,.-
C.5 113 1
Cc
LLI 10
Z (b)
ILl IE] -1- 0.6
wE]
O Lx A A
rr
I-- • A a A
0.4
0 i I
0.30 I I m A A
a
O 0.2
(b) i a
0.25 A
>- 113 0.0 I I
0.010 0.012 0.014 0.016
rr 0.20
rl
z
LU Fig. 53. Effect of retaining or suppressing the contributions
0.15 of steady rolls on the global energy (a) and entropy (b) in the
ir] coexistence regime. In both of these graphs, the small
0.10 I I symbols are taken from Fig. 51 and represent calculations
0.0005 0.0010 0.0015 performed on temporal-hipass-filtered data in which the
~3
steady-roll component is suppressed. The large symbols
Fig. 52. Evolution with e of the global energy (a) and represent the results of calculations performed on data in
entropy (b) computed from demodulated left-TW (closed which this component was retained. Thus, the difference
symbols) and right-TW (open symbols) amplitude profiles in between the trends of the small and large symbols may be
the counterpropagating regime. For e' --<1.05, the data points associated with the steady-roll component. As e is increased,
have been displaced horizontally to avoid overlap. Above the growing region of large-amplitude, steady-rolls makes a
s' = 1.1, the two TW components begin to evolve indepen- large, positive contribution to the spectral energy and causes
dently. the entropy to decrease.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 215

J
due to slow TW and steady rolls is retained (a)
0.8
(large symbols) with those computed using tem- 3

poral hipass filtering to remove that component >-

(5
0.6

of the signal (small symbols, copied from Figs.

~ 0.4
51a,b). The presence of high-amplitude steady
rolls increases the spectral energy and decreases
the entropy in all cases. As e increases in the
n<
co
0.2 H
coexistence regime, the region of steady rolls
grows in space, and its contribution to the energy 0.75
grows correspondingly larger. Since the steady (b)
:z~ 0 . 7 0
rolls exhibit very weak dynamical behavior in
regular coexistence states (open triangular sym- ¢r 0.65

bols), the spectral entropy decreases rapidly at z 0.60

..-I
the highest values of e. <

Because of the marked differences in spectral 03

0.55

entropy and energy associated with erratic TW

bursts and regions of steady rolls, switching
0.50
0 2 '0 40 0 8'0 '
1O0 120
TIME (1000 sec)
between dispersive-chaos and "coexistence" dy-
Fig. 54. The spatial energy (a) and the spatial entropy (b)
namics (in data sets represented by half-filled reveal switching between dynamics characteristic of the
circles in the graphs in this paper) causes corre- dispersive-chaos and coexistence regimes at e' = 10.97(8). At
sponding changes in the time evolution of the time 8000 sec, low-entropy, "coexistence" dynamics gives
way to a high-entropy dispersive-chaos state; the dynamics
spatial energy and entropy Ex(t ) and Hx(t ). Fig. switches back at t = 101000 sec. These dynamical changes are
54 shows an example, at e ' = 10.97(8). At the signaled by opposing jumps in the spatial entropy and
beginning of this run (t < 8000 sec), a region o f energy. In (a), the curve marked " H F " was computed using
temporal-hipass-filtered data to suppress the contribution of
nearly steady convective rolls was in the process steady rolls.
of disappearing from the system. This is reflected
in a small increase in the spatial entropy in Fig.
54b. The spatial energies computed retaining the dynamics. We have seen that, in individual data
steady-roll component and suppressing it with sets, the spatial profiles of the temporal energy
temporal hipas s filtering (labeled " H F " in Fig. Ew(x) and entropy HT(X) may exhibit strong
54a) were noticeably different until this transi- spatial variations-see Figs. 49c,d above. How-
tion. In the resulting pure dispersive-chaos state, ever, if we add together the profiles computed
the dynamics looked qualitatively similar to that from many data sets, then these spatial varia-
exhibited in Figs. 49a,b. Then, at time t = 101000 tions might be expected to average to zero,
sec, a large TW burst evolved into another leaving only variations associated with
region of steady rolls, producing a sharp drop in nonuniformities in the experimental cell, the
the spectral entropy and a jump in the spatial optical system, or in the data processing. We
energy computed from the unfiltered data. After have used 24 data sets in the dispersive-chaos
this transition, regular TW bursts dominate the regime to evaluate this idea, and Fig. 55 shows
evolution of the temporal-hipass-filtered spatial the results. The exponential dependence on e of
energy. the global energy E observed in this regime
A final piece of information which can be suggests that the proper way to compute an
extracted from the B O D analysis concerns the averaged, normalized temporal energy ~ET(X) is
spatial uniformity of the experimental system, as to average the logarithms of the individual ET(X)
reflected in the average spatial structure of the profiles. We scale this mean profile by ln F,
216 P. Kolodner et al. / Physica D 85 (1995) 165-224

uJ'-~ 0.2, ~ ~ ~ ,
ized average to obtain the profile shown in Fig.
55b.
The averaged energy and entropy profiles in
Figs. 55a,b exhibit evidence of weak, residual
dynamical activity averaged over all available
. . . . dispersive-chaos data sets. T o the extent that the
features in the profiles BET(X) and BHT(X) are
mirror images, they reflect the presence of co-
LUI~ 0.2, I ~ I I I
herent spatial structures with high spectral
energy and low spectral entropy. The peak (dip)
in BET(X) (AHT(X)) at location 200 ° in Figs. 55a,b
<~-01
represent a structure whose spectral energy is
n-O / "1 equivalent to about a tenth of the variations seen
~-02L , i , , ,
<111
I--
in E(e') or H(e') over the range 2,0 < e' < 10.0.
How can we decide whether this represents
50L i L ~ ~ t persistent, dynamical behavior, experimental in-
• / (e)
homogeneities, or just statistical fluctuations?
c,O
First, let us note in comparison that the features
~o in the profiles of Figs. 49c,d, when scaled in the
same way as those in Figs. 55a,b, represent a
0° 90 ° 180 ° 270 ° 360 °
structure of about 5 times stronger spatial vari-
POSITION IN C E L L ation. Since we have arranged 24 ~ 52 data sets
Fig. 55. The normalized temporal energy ~ET(X) (a) and to produce Figs. 55a,b, it seems likely that the
entropy BHx(x) (b), computed by averaging the results of 24 features therein simply represent normal statisti-
dispersive-chaos data sets, are plotted vs. spatial position. A
cal fluctuations of erratic T W bursting in disper-
region of weak, residual dynamical activity centered at
location 190° survives this averaging. The sum of these two sive chaos and not experimental inhomogen-
profiles, normalized by the slope 1/ell, yields the profile ~edy. cities. Second, for a direct calibration of the
shown in (c). strength of these features against the stress
parameter e, we can multiply them by Ae =
where F = 6.63 is the factor by which the fit to 0.008, the width of the range 2.0 < e ' < 10.0 to
E(e') varies over the range 2 . 0 < e ' < 10.0. Fig. which they have been normalized. Thus, the
55a shows this averaged, normalized profile, with features of height 0.1 in Figs. 55a,b would
its spatial mean subtracted. Fig. 55b shows an correspond to local inhomogeneities in e of size
analogous result for the averaged, normalized 0.1 Ae = 8 × 10 -4. This is 2 - 3 times larger than
temporal entropy BHT(X). In this case, since the the measured nonuniformity of the experimental
global entropy increases approximately linearly cell: Berms ~< 4 x 10 -4. T o reduce the contribu-
with e in the dispersive-chaos regime, we calcu- tions of true dynamics to below this experimental
late BHT(X) by computing the arithmetic mean of resolution, we would have to average together
the temporal entropy profiles. Once again, we several hundred dispersive-chaos data sets. This
divide by a factor D in order to scale this profile is not feasible. Finally, let us note that, if we add
to the total entropy variation measured over this the profiles in Figs. 55a,b, we cancel the contri-
e range: using the numbers given in the fit to Fig. butions of coherent dynamics to the spectral
52b, we obtain D = 0.651/(0.698-0.604) = 6.91. energy and entropy. Fig. 55c shows this sum,
As in the case of the mean energy profile, we scaled to equivalent e v a r i a t i o n : BEdyn(X)=
have subtracted the spatial mean of this normal- Ae(BET(X) + BHT(X)). This profile represents the
 P. Kolodner et al. / Physica D 85 (1995) 165-224 217

averaged contribution to the temporal energy havior. Above a threshold ea' = 1.9, the counter-
and entropy of all effects- such as noise- which propagating regime gives way to dispersive
do not exhibit the high-energy, low-entropy chaos. In this regime, both TW components
signature of coherent spatiotemporal structures. exhibit irregular sequences of growth, bursting,
~edyn(X) is spatially uniform to well within our and collapse. With increasing e, the global be-
experimental resolution. havior of this state becomes more erratic, and
the spatiotemporal character of the TW bursts
themselves evolves in ways which have been
9. Discussion documented in detail, principally by computing
correlation functions. At the highest values of e
In this paper, we have reported detailed and for which this bursting behavior is seen, t h e
precise observations of one-dimensional TW system exhibits extended periods in which local-
patterns in a well-characterized, stable, and ized regions of extremely slow or stationary
uniform experimental system. We have explored convective rolls are seen. Above a threshold
!
and documented all of the dynamical states % = 11.1, this behavior culminates in the coexist-
possible in an annular geometry of a particular ence regime, in which part of the cell is filled
size at this separation ratio. We have argued that with perfectly steady rolls, while the rest of the
a coupled-Ginzburg-Landau-equation model cell exhibits TW bursts whose timing is strikingly
should accurately describe at least the first ob- regular. As e is further increased in this regime,
served nonlinear states. Finally, we have ex- the spatial extent of the TW region decreases,
plored various techniques for the analysis of vanishing at e~ = 19.2. Above this threshold, the
these data. We have divided our discussion of cell is filled with steady convective rolls of
this work into three parts, each of which is spatially-uniform amplitude. This transition is
assigned a separate subsection. In Section 9.1, hysteretic; e must be reduced to below e~ ~- -16
we briefly summarize the experimental observa- in order to induce a subsequent transition back
tions. In Section 9.2, we discuss the prospects for to the quiescent state.
quantitatively modeling this system with the In all of these dynamical states, TW tend to
CGLE. Finally, in Section 9.3, we present a take the form of localized bursts or wave packets
discussion of the statistical characterization of which have an asymmetric structure similar to
the observed dynamics. the one documented in Fig. 15. There is a
universal tendency for the wave-number gradient
9.1. S u m m a r y o f experimental observations seen in this structure to develop into a trailing-
edge phase defect. By this process, the burst
As e is increased above onset, this system splits into two, and the leading burst decays.
evolves through three dynamical regimes before This scenario, documented in Figs. 3, 6, 7, and
making a final transition to a spatially-uniform 12, is seen in all the dynamical states described
state of steady convective rolls. In the first in this paper and has been observed in other
regime, the so-called counterpropagating regime, weakly-nonlinear [27,59] and strongly-non-linear
the dynamical states consist of pairs of broad, [56] TW states as well.
quasilinear wave packets which circulate around The cause of the trailing-wave phase defect is
the cell at high velocity and with no long-term different at low and high convective amplitudes.
change in spatial structure. With increasing e, In weakly-nonlinear TW convection, as in the
these wave packets become narrower, stronger, case of the bursts seen at the lowest values of e
and more asymmetric, and they begin to exhibit in dispersive chaos, the negative sign of the
weakly erratic and independent dynamical be- nonlinear dispersion coefficient c 2 in the CGLE
218 P. Kolodner et al. / Physica D 85 (1995) 165-224

causes the low-amplitude T W in the trailing edge be obeyed at least approximately. We note that
of a burst to have a higher phase velocity than the existence of an amplitude threshold for
the higher-amplitude T W downstream. This nonlinear bursting has been suggested in Refs.
leads to phase slippage and the formation of a [51,521.
defect. A t high convective amplitudes, it is the The transition from the dispersive-chaos re-
effect of the convection on the background gime to the coexistence regime is complex and
concentration field that causes high-amplitude appears to be hysteretic, although we do not
T W downstream to travel more slowly than the have enough data to make a definitive statement.
low-amplitude T W upstream. Judging from Fig. 51b, there appears to be range
Some of the transitions between dynamical of control parameters, 1 0 ~< e' <~ 13, over which
states in this system are observed to be hy- both kinds of dynamics are possible. This may
steretic, and some are reasonably expected to represent hysteresis, but it is also possible that
depend on the system size F~. e = 0 corresponds the system takes even longer to switch its dy-
to the threshold of convective instability in this namical state than we have been willing to wait,
system and is independent of F~ in an annular so that some of these observations are actually of
geometry [11,12,47]. The transition between the very long-lived transients. The dependence of
counterpropagating regime and dispersive chaos the limits of existence of the coexistence state on
has been carefully checked for hysteresis, which the size of the system follows from the observa-
is absent to almost within our experimental tion of Fig. 20a that the extent of the T W region
resolution. By an argument similar to that which is this state decreases linearly with e: FTW =
accounts for the size dependence of the thres- f i ( e - - e c ) , where e~ =19.22(22) and /3 =
hold for small-amplitude T W convection in a - 8 . 3 2 ( 3 3 ) × 10 3. Increasing e causes the T W
rectangular cell exhibiting lossy reflections at the region to shrink to zero width at ec, independent
endwalls [12], we expect this threshold to vary as of F~. Thus, the threshold for the transition to
1/F~. This argument is as follows: The spatially-uniform steady convection is indepen-
quasilinear wave packets propagate at the group dent of the size of the system.
velocity Vgr in Fig. 10a and grow exponentially at The transition back to dispersive chaos upon
approximately the linear rate %1e. Thus, be- decreasing e takes place when the T W region
tween collisions, the wave packets grow by a expands to fill the system, at 6b = e c d- F~//3. This
factor exp(eF~/2Vgr% ). If this growth factor is value and that of the threshold ea computed in
greater than some threshold 3'b, then the wave the previous paragraph coincide at F~ = 8.5(6)
packets will grow too strong in amplitude be- and at 152(6). For F~ < 8.5, we expect that the
tween collisions for the weak nonlinearity of the counterpropagating regime will evolve directly
collision to suppress their further nonlinear evo- into the coexistence regime upon increasing e
lution into dispersive bursting. Thus, the thres- above ea, so that no dispersive chaos will be
hold for the transition from counterpropagating seen. For F~ > 152, we expect that, once the
to dispersive-chaos dynamics is e a = a/F~, with transition from dispersive chaos to the coexist-
°z=2VgrT01n Yb- Our measurements of e'a = ence regime has been triggered by increasing e
1.87(12) at F~ = 82.47 implies that a = eaF~ = above ea, this state will persist upon subsequent
0.154(10) and thus that ln yb = 6.05(24) if we reduction of e; the next transition will be back to
extrapolate the data in Fig. 10a to obtain Vgr = the counterpropagating regime, bypassing the
0.050(2) at ea. The application of this linear dispersive-chaos regime. However, it may also
argument to a nonlinear state is not justified in turn out that the tendency to form regions of
detail, but because the nonlinearity in this state very slow T W is so strong in this system that
is quite weak, we would expect this prediction to dispersive chaos will be suppressed entirely for
 P. Kolodner et al. / Physica D 85 (1995) 165-224 219

F~ > 152, even when slowly increasing e from tions of the c o u p l e d - C G L E model of Eqs.
below. (17a,b) could be fit to the experimental data in
It would be quite interesting to extend these the counterpropagating regime by varying the
experiments to different parameter regimes. At complex coupling parameter h(1 + ic3) so as to
smaller 10[, where the nonlinear dispersion is match the velocities measured in Figs. 10a,b. (In
stronger [23,24,51,52], we would expect to see this connection, we note that Malomed [62] has
stronger dispersive bursting. Moving towards the shown that the simplest c o u p l e d - C G L E model
domain of stable confined states at larger Iq~l does indeed possess solutions in the form of
[14,56], we would expect to see more persistent oppositely-propagating wave packets of time-in-
T W bursts. Some evidence supporting these dependent shape.) We fully expect that all the
expectations was seen in the experiments re- coefficients in this model, up to cubic order, can
ported in Ref. [13]. It is worth noting that, be accurately known. The question is whether an
despite the long absolute time scales and large extension of this model can be used to quantita-
convection cell used in these experiments, our tively describe the more complicated dynamical
system is still small. The largest number of T W behavior observed at higher e.
bursts encountered in our longest data set was For this purpose, we have written down in Eq.
only about 200, and we have never observed (18) and the associated definitions of the group
more than about 4 bursts in the cell at one time. velocity and stress parameter a c o u p l e d - C G L E
Thus, it would also be useful to extend the model which incorporates the effects of the
s p a c e - t i m e domain of these observations. This convective concentration field C(x, t) on the
could be accomplished by using a thinner cell of evolution of TW. Modeling this effect has re-
the same lateral dimensions, and by acquiring quired no less than nine new coefficients and is
even longer data sets. The remarks of the previ- based on the behavior of a field which cannot be
ous paragraph, however, may mean that disper- measured 5 [63]. Thus, it would appear at first
sive chaos simply cannot be seen in a system that blush to be unrealistic in the extreme to hope to
is too much longer than our present cell. We also fit such a model to our data in any sensible way.
note that evidence of dispersive bursting has However, we have noted in Sections 5 and 7.3
b e e n observed in experiments in rectangular cells above that the qualitative effects of the convec-
of intermediate length [51,52,61]. tive concentration field are weak at the lowest
values of e in the dispersive-chaos regime, and
9.2. M o d e l i n g the system with the C G L E that they appear to turn on gradually as we move
up through this regime. So it would seem that
The complex G i n z b u r g - L a n d a u equation in this effect represents a very small perturbation to
Eq. (14) has been shown to accurately describe the weakest dispersive chaos states and could
the weakly-nonlinear transient behavior of small- thus be modeled by setting all but the most
amplitude, unidirectional T W in this experiment. essential new coefficients to zero. Modeling
All the coefficients in this equation have been experimental data in this regime by adjusting the
extracted by fitting formulae derived from the remaining parameters until numerical integra-
C G L E to experimental measurements. The re- tions match the data might then have a realistic
sulting coefficients match the values computed
for an ideal (laterally-infinite) geometry with
generally good accuracy; in the case of some of 5The convective concentration field in nonlinear TW convec-
these coefficients, we can ascribe the differences tion has been measured [63]. However, based on the strength
of the signals seen in those experiments, .the concentration
to the effect of the width of the experimental field in the dispersive-chaos regime is too weak to be
cell. We have suggested that numerical integra- measured.
220 P. Kolodner et al. / Physica D 85 (1995) 165-224

prospect of success. If so, then more terms in revealed systematic trends in the structure of the
Eq. (18) could be turned on as needed to match T W bursts that constitute dispersive chaos. The
the data at higher e. This is the incremental spatial width and central wave number in Fig. 28
approach that we propose for modeling these are good examples of the kind of precision that
data with the C G L E . can be obtained using these diagnostic tools.
It may be that a program of adjusting parame- However, these techniques have their limits. For
ters for a best fit of a numerical integration of example, the asymmetric structure typically seen
the C G L E to the measured signal u(x, t) is in T W bursts (Fig. 15) easily evades detection by
hopeless. The Lyapunov spectrum of the system Cs(Ax ). In addition, we have not found any way
may be such that the numerical solution will to estimate the average temporal duration of the
diverge from the measured signal faster than the T W bursts in dispersive chaos that is more
m e a n time between dynamically significant reliable than simply inspecting the s p a c e - t i m e
events, even for a correct parameter set. In this diagrams of Fig. 14. This is but one aspect of a
case, another basis for a detailed comparison more general failure: we do not have a result for
between numerical integrations and experimen- the general, average spatiotemporal shape of a
tal data will have to be found. The spatio-tempo- T W burst in dispersive chaos at any value of e.
ral statistical measures presented herein may be The biorthogonal decomposition has been a
a reasonable starting point for this comparison. powerful tool for studying some aspects of the
T o summarize, we have described a compli- dynamics of dispersive chaos. H e r e , we have
cated model and have proposed that it be fit to concentrated on global properties of the B O D
complex data by varying parameters. We suggest spectra, deferring our studies of the s p a c e - t i m e
that this program be carried out in stages, adding dynamics in the characteristic spaces x ( X ) and
more coefficients to fit increasingly-complicated x ( T ) and of bifurcations until a later publication
data as e is varied. We expect that at least the [41]. Some of the features in the B O D spectra
first few steps of this program will be successful. (Fig. 47) are related to general symmetry prop-
The essential question is: why do this? An erties of the system. For example, it was pointed
important component of the answer is the size of out in the discussion of Fig. 47 that quasi-two-
the system, as discussed in the paragraphs above. fold degeneracies in the B O D spectra are related
E v e n doubling the size of our experimental to the presence of T W in the dynamics u(x, t).
a p p a r a t u s - a difficult p r o p o s i t i o n - will not bring Exact degeneracy is not observed because the
us into the regime in which many T W bursts can dynamics breaks the space-time translation sym-
be seen in the cell at a single time. The only metry. It was also pointed out that the spectrum
realistic way to explore dispersive chaos in a computed in the dispersive-chaos regime ex-
truly large system is to do it computationally. hibited a nearly exponential decay law. Ex-
ponential decays in the B O D spectra have been
9.3. Characterization of the data shown to be manifestation of space-time quasi-
symmetries [30,31], of which the group of dila-
We have attempted to extract as much quan- tion symmetries is an example. Does dispersive
titative and qualitative information as possible chaos exhibit such a quasi-symmetry? In answer-
from our data using the tools of classical statisti- ing this question, it is useful to recall the case of
cal analysis. Demodulation techniques, used in fully developed, three-dimensional turbulence.
various guises, have been extremely useful, In this case, the quasi-symmetry is precisely the
because they extract exactly the complex T W space-time scale (dilation) invariance of the
amplitudes that are described by the C G L E Navier-Stokes equations, satisfied by the flow
model. Correlation functions and spectra have itself in its "self-similar" regime. This scale
 P. Kolodner et al. / Physica D 85 (1995) 165-224 221

invariance manifests itself in the exponential operator V in Eq. (4), reveals that such a lack of
decay of the BOD spectrum, as well as in the smoothness can only be due to the appearance of
existence of a (stretching) map relating the a new degeneracy in the BOD spectrum as a
different topos and another (stretching) map parameter is varied. Since the appearance of a
relating the different chronos. This quasi-symme- new degeneracy causes a decrease in the spectral
try and the resulting exponential BOD spectrum entropy H, a bifurcation necessarily occurs at a
are exhibited independently of the homogeneity local maximum of H, This is indeed the case at,
of the flow. If in addition the turbulence is for example, the boundary between the disper-
homogeneous, then topos coincide with Fourier sive-chaos and coexistence regimes in these ex-
modes, and the BOD spectrum corresponds to periments. While a local entropy maximum is a
Kolmogorov's k -5/3 spectrum law [31,64,65]. necessary condition and therefore is a good
What does this let us conclude about dispersive indicator of the possibility of a bifurcation, it is
chaos? On the one hand, it is clear that TW not a sufficient condition, and a more detailed
convection always exhibits a rather narrow range analysis of the transition is needed. This question
of wave numbers and thus that dispersive chaos will be addressed in Ref. [41].
is far from the Kolmogorov spectrum of fully- The spatial and temporal energies and en-
developed, homogeneous turbulence. On the tropies Es(t), Hs(t), ET(X), and HT(X) defined in
other hand, dispersive chaos is clearly not Eqs. (8)-(13) are less "global" properties of the
homogeneous. Therefore, the nearly-exponential dynamics than the spectral energy E and entropy
decay seen in the BOD spectrum may still reflect H. These functions reveal features which Can be
an underlying quasi-symmetry. Whether this is directly related to the properties of TW bursts
the c a s e - a n d whether dispersive chaos obeys ao and other details of the dynamical behavior. For
scale invariance in a space-time sense-still example, the growth of individual TW bursts is
remains to be revealed by a detailed study of reflected in sudden increases in the spatial
topos and chronos. energy correlated with dips in the spatial en-
The global complexity of the spatiotemporal tropy. This is the characteristic signature that a
dynamics is reflected in the spectral energy E coherent spatiotemporal structure has appeared
and entropy H computed from the BOD spectra. in the system.
As shown in Figs. 51a,b, transitions between the Beyond these measures of dynamical complex-
three main dynamical states seen in these experi- ity, there remain two important open problems
ments are clearly reflected in the e dependence in our characterization of this system. First, as
of these global quantities. The global entropy is mentioned above, while we can easily appreciate
clearly more sensitive to these transitions than the qualitative spatiotemporal shape of the TW
the global energy. Such transitions have been bursts in dispersive chaos from their shapes in
observed in the BOD spectra of coupled-map the space-time diagrams in Fig. 14, we have not
lattices [20,66] and indeed were predicted in our found a way to quantitatively extract the average
early work on the present system [60]. An open spatiotemporal shapes of these bursts; In par-
question is whether they correspond to bifurca- ticular, there is not a one-to-one correspondence
tions in the mathematical sense. A spatiotempor- between a burst and a BOD mode, a situation
al bifurcation [39,40,67], in analogy with Poin- partially due to the traveling character of the
car6's terminology for bifurcations in temporal bursts. This is analogous to the problem of
dynamical systems, corresponds to a lack of identifying coherent structures in turbulence.
smoothness in the dependence of the flow on the The second open issue relates to the application
stress parameter. The perturbation theory of of the tools of the theory of dynamical systems to
linear operators, applied to the biorthogonal these d a t a - a point mentioned in the very first
222 P. Kolodner et al. / Physica D 85 (1995) 165-224

paragraph of this paper. The BOD is a method and D.J. Thomson for providing invaluable com-
for representing complex spatiotemporal dynam- puter software. R.L. expresses his sincere
ics as the orbit described by a vector in a spatial gratitude to his colleagues from the Levich
characteristic space as time evolves and as the Institute for their hospitality during his visit
orbit described by a corresponding vector in a when a part of this work was performed. He and
temporal characteristic space as the spatial loca- N.A. gratefully acknowledge the support of
tion moves. These two spaces are the spatiotem- NATO grant No. 900265, N.A. also thanks
poral counterparts of the notion of phase space ONR and NSF for their support through NSF/
in temporal dynamical systems theory. Although PYI award MSS89-57462 and ONR grant No.
these two spaces are the smallest linear spaces N00014-90-J-1554.
which contain the dynamics of u(x, t) (viewed as
a set of spatial or temporal vectors), a difficult
problem presented by most of our data is one of References
high dimensionality: the weakest dispersive-
chaos data set exhibits a dimension which is [1] For a comprehensive summary of progress in the study
larger than n90 = 18 f r o m Fig. 51c. The essential of chaotic systems, see E. Ott, Chaos in Dynamical
Systems (Cambridge Univ. Press, Cambridge, 1993).
challenge in understanding dispersive chaos lies [2] For a unified discussion of many pattern-forming sys-
in finding ways to characterize high-dimensional tems, see M.C. Cross and P.C. Hohenberg, Rev. Mod.
dynamical systems. Although it is clear that we Phys. 65 (1993) 851.
[3] P.C. Hohenberg and B.I. Shraiman, Physica D 37
have already significantly decreased the dimen-
(1989) 109.
sionality of the dynamics by the application of [4] N.B. Tufillaro, R. Ramshankar and J.P. Gollub, Phys.
the BOD itself, as compared to other bases such Rev. Lett. 62 (1989) 422;
as the Fourier basis [40,41,68,69], it is also J.P. Gollub and R. Ramshankar, in: New Perspectives
in Turbulence, L. Sirovich, ed. (Springer, New York,
obvious that a further reduction is needed if 1991) p. 165.
progress is to be made. As a start on this [5] S.W. Morris, E. Bodenschatz, D.S. Cannell and G.
problem, we have begun a dynamical-systems Ahlers, Phys. Rev. Lett. 71 (1993) 2026;
M. Assenheimer and V. Steinberg, Nature 367 (1994)
analysis of data in the counterpropagating re- 345.
gime, where 3 ~<n90 ~ 12. The dimension of the [6] M. Rabaud, Y. Couder and S. Michalland, Eur. J.
dynamics is reduced still further by analyzing the Mech. B 10 (1991) 253.
[7] J.M. Flesselles, A.J. Simon and A. Libchaber, Phys.
demodulated TW amplitude fields rather than
Rev. A 40 (1991) 1.
the raw d a t a - s e e Fig. 52. [8] J. Bechhoefer, A.J. Simon and A. Libchaber, Phys.
Furthermore, the BOD localizes the subspace Rev. A 40 (1989) 2042.
in which a spatiotemporal bifurcation may take [9] S. Ciliberto and M. Caponeri, Phys. Rev. Lett. 64
(1990) 2775;
place to that spanned by degenerate eigenfunc- M. Caponeri and S. Ciliberto, Physica D 58 (1992) 365.
tions, therefore providing a technique for still [10] F. Daviaud, J. Lega, P. Berg6, P. Coullet and M.
further dimension reduction similar to that of the Dubois, Physica D 55 (1992) 287.
[11] P. Kolodner, D. Bensimon and C.M. Surko, Phys. Rev.
center manifold theorem of temporal dynamical Len. 60 (1988) 1723;
systems theory. A report on our initial progress D. Bensimon, P. Kolodner, C.M. Surko, H. Williams
on this problem is in preparation [41]. and V. Croquette, J. Fluid Mech. 217 (1990) 441.
[12] P. Kolodner, A. Passner, C.M. Surko and R.W. Walden,
Phys. Rev. Lett. 56 (1986) 2621.
[13] P. Kolodner, J.A. Glazier and H. Williams, Phys. Rev.
Acknowledgements Lett. 65 (1990) 1579;
J.A. Glazier, P. Kolodner and H. Williams, J. Stat.
Phys. 64 (1991) 945.
P.K. thanks P.C. Hohenberg and H. Green- [14] P. Kolodner, Phys. Rev. A 44 (1991) 6448, 6466.
side for illuminating discussions and M.C. Cross [15] P. Kolodner, Phys. Rev. A 46 (1992) 6431.
 P. Kolodner et al. / Physica D 85 (1995) 165-224 223

[16] M.S. Heutmaker and J.P. Gollub, Phys. Rev. A 35 M.N. Glauser, S.J. Leib and W.K. George, in: Turbul-
(1987) 242. ent Shear Flow, Vol. 5 (Springer, Berlin, 1987);
[17] P. Kolodner, H. Williams and C. Moe, J. Chem. Phys. A. Glezer, Z. Kadiogliu and J. Pearlstein, Phys. Fluids
88 (1988) 6512. A 1 (1989) 1363;
[18] P. Kolodner, Phys. Rev. Lett. 69 (1992) 2519. A. Deane, G.E. Karniadakis, I.G. Kevrekidis and S.A.
[19] P. Kolodner, Phys. Rev. E 48 (1993) 665. Orszag, Phys. Fluids A 3 (1991) 2337;
[20] N. Aubry, R. Guyonnet and R. Lima, J. Stat. Phys. 63 P. M0in and R.D. Moser, J. Fluid Mech. 200 (1989)
(1991) 683. 471.
[21] A.C. Newell, in: Nonlinear Wave Motion, Lectures in [38] D.S. Broomhead and G.P. King, Physica D 20 (1986)
Applied Mathematics,Vol. 15, A.C. Newell, ed. (AMS, 217.
Providence, 1974) p. 157. [39] N. Aubry and R. Lima, Spatio-temporal Symmetries,
[22] S.J. Linz and M. L~cke, Phys. Rev. A 35 (1987) 3997; Levich Institute Report NO. 9306026 (1993), J. Stat.
B.J.A. Zielinska and H.R. Brand, Phys. Rev. A 35 Phys., submitted.
(1987) 4349; [40] N. Aubry, M.P. Chauve and R. Guyonnet, Phys. Fluids
E. Knobloch and D.R. Moore, Phys. Rev. A 37 (1988) 6 (1994) 2800.
860; [41] S. Slimani, N. Aubry, R. Lima and P. Kolodner, in
M.C. Cross and K. Kim, Phys. Rev. A 37 (1988) 3909. preparation.
[23] W. Sch6pf and W. Zimmerman, Europhys. Lett. 8 [42] A.C. Newell, in: Proc. Workshop on Propagation in
(1989) 41; Phys. Rev. E 47 (1993) 1739. Systems Far From Equilibrium, J.E. Wesfreid et al.,
[24] W. Sch6pf, PhD thesis, Universitfit Bayreuth, 1992 eds. (Springer, Berlin, 1988) p. 122.
(unpublished). [43] W. van Saarloos and P.C. Hohenberg, Phys. Rev. Lett.
[25] P. Collet and J.-P. Eckmann, Commun. Math. Phys. 132 64 (1990) 749; Physica D 56 (1992) 303; Physica D 69
(1990) 139. (1993) 209 (E).
[26] H. Riecke, Phys. Rev. Lett. 68 (1992) 301; [44] G. Ahlers, L.I. Berge and D.S. Cannell, Phys. Rev.
H. Rieeke, Physica D 61 (1992) 253. Lett. 70 (1993) 2399.
[27] P. Kolodner and H. Williams, in: Proc. NATO Ad- [45] C.S. Bretherton and E.A. Spiegel, Phys. Lett. A 96
vanced Research Workshop on Nonlinear Evolution of (1983) 152.
Spatio-temporal Structures in Dissipative Continuous [46] W. Sch6pf and L. Kramer, Phys. Rev. Lett. 66 (1991)
Systems, F.H. Busse and L. Kramer, eds., NATO 2316.
Advanced Study Institute, Series B2, Vol. 225 (Plenum, [47] P. Kolodner and C.M. Surko, Phys. Rev. Lett. 61
New York, 1990) p. 73. (1988) 842;
[28] B.J. Gluckman, P. Marcq, J. Bridger and J.P. Gollub, J. Fineberg, E, Moses and V. Steinberg, Phys. Rev.
Phys. Rev. Lett. 71 (1993) 2034; Lett. 61 (1988) 838;
L. Ning, Y. Hu, R.E. Ecke and G. Ahlers, Phys. Rev. P. Kolodner, C.M. Surko and H. Williams, Physica D 37
Lett. 71 (1993) 2216. (1989) 319;
[29] N. Aubry, Theor. Comp. Fluid Dyn. 2 (1991) 339. V. Steinberg, J. Fineberg, E. Moses and I. Rehberg,
[30] N. Aubry, R. Guyonnet and R. Lima, J. Nonlin. Sci. 2 Physica D 37 (1989) 359.
(1992) 183. [48] M.C. Cross, Phys. Rev. A 38 (1988) 3593; Physica D 37
[31] N. Aubry, R. Guyonnet and R. Lima, J. Stat. Phys. 67 (1989) 315.
(1992) 203. [49] W. Sch6pf and W. Zimmermann, Phys. Rev. A 41
[32] M. Lo~ve, Probability Theory (Van Nostrand, New (1990) 1145.
York, 1965). [50] W. Barten, M. Lficke and M. Kamps, Phys. Rev. Lett.
[33] J.L. Lumley, in: Atmospheric Turbulence and Radio 66 (1991) 2621;
Wave Propagation, A.M. Yaglom and V.I. Tatarski, eds. M. Lficke, W. Barten and M. Kamps, Physica D 61
(Nauka, Moscow, 1967) p. 166; (1992) 183.
J.L. Lumley, Stochastic Tools in Turbulence (Academic [51] E. Kaplan, E. Kuznetsov and V. Steinberg, Europhys.
Press, New York, 1972). Lett. 28 (1994) 237.
[34] G. Berkooz, P. Holmes and J i . Lumley, Ann. Rev. [52] E. Kaplan, E. Kuznetsov and V. Steinberg, Phys. Rev.
Fluid Mech. 25 (1993) 539. E 50 (1994) 3712.
[35] N. Aubry, P. Holmes, J.L. Lumley and E. Stone, J. [53] P. Kolodner and J.A. Glazier, Phys. Rev. A 42 (1990)
Fluid Mech 192 (1988) 115; 7504;
S. Sanghi and N. Aubry, J. Fluid Mech. 247 (1993) 455. J.A. Glazier and E Kolodner, Phys. Rev. A 43 (1991)
[36] L. Sirovich, Q. Appl. Math 5 (1987) 561, and references 4269.
therein. [54] P. Kolodner, unpublished.
[37] G. Berkooz, P. Holmes and J.L. Lumley, J. Fluid Mech. [55] J.M. Luijkx and J.K. Platten, J. Non-Equilib. Ther-
230 (1991) 75; modyn. 6 (1981) 141.
D.H. Chambers, R.J. Adrian, E Moin, D.S. Stewart [56] E Kolodner, Phys. Rev. E 50 (1994) 2731.
and H.J. Sung, Phys. Fluids 31 (1988) 2573; [57] W. Sch6pf, private communication.
224 P. Kolodner et al. I Physica D 85 (1995) 165-224

[58] W. Sch6pf and I. Rehberg, Europhys. Lett. 17 (1992) K.D. Eaton, D.R. Ohlsen, S.Y. Yamamoto, C.M.
321. Surko, W. Barten, M. Liicke, M. Kamps and P.
[59] P. Kolodner, Phys. Rev. E 47 (1993) 1038. Kolodner, Phys. Rev. A 43 (1991) 7105.
[60] S. Slimani, N. Aubry, P. Kolodner and R. Lima, in: [64] W. Lian and N. Aubry, in FED-Vol. 151 (ASME)
Bifurcation Phenomena and Chaos in Thermal Convec- (1993) p. 129.
tion, eds. H.H. Bau, L.A. Bertram and S.A. Korpela, [65] A.N. Kolmogorov, Dokl. Acad. Nauk SSSR 30 (1941)
HTD-Vol. 214, Book No. G00699 (ASME, New York, 301.
1992) p. 39. [66] R. Lima, Chaos 32 (1992) 315.
[61] V. Steinberg and E. Kaplan, in: Proc. NATO Advanced [67] N. Aubry, F. Carbone, R. Lima and S. Slimani, J. Stat.
Research Workshop on Spontaneous Formation of Phys. 76 (1994) 1005.
Space-Time Structures and Criticality, eds. T. Riste and [68] N. Aubry and W. Lian, Lectures in Applied Mathe-
D. Sherrington (Kluwer, 1991) p. 207. matics, Vol. 29 (1993) p. 71.
[62] B.A. Malomed, Phys. Rev. E 46 (1993) 3841. [69] M.P. Chauve and P. Le Gal, Physica D 58 (1992) 407.
[63] B.L. Winkler and P. Kolodner, J. Fluid Mech. 240
(1992) 31;