1: \documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
2: \usepackage{graphicx}% Include figure files
3: \usepackage{epsfig}% Include figure files
4: \usepackage{dcolumn}% Align table columns on decimal point
5: \usepackage{bm}% bold math
6:
7: \begin{document}
8: \title{Statistics and Characteristics of Spatio-Temporally \\
9: Rare Intense Events in Complex Ginzburg-Landau Models}
10: \author{Jong-Won Kim$^1$ and Edward Ott$^{1,2}$}
11: \affiliation{$^1$Department of Physics, and Institute for Research in
12: Electronics and Applied Physics,
13: University of Maryland, College Park, Maryland 20742 \\
14: $^2$Department of Electrical and Computer Engineering,
15: University of Maryland, College Park, Maryland 20742}
16: \date{\today}
17:
18: \begin{abstract}
19: We study the statistics and characteristics of rare intense events in two
20: types of two
21: dimensional Complex Ginzburg-Landau (CGL) equation based models. Our numerical
22: simulations show finite amplitude collapse-like solutions which approach the
23: infinite amplitude solutions of the nonlinear Schr\"{o}dinger (NLS) equation
24: in an appropriate parameter regime. We also determine the probability
25: distribution function (PDF) of the amplitude of the CGL solutions, which is
26: found to be approximately described by a stretched exponential distribution,
27: $P(|A|) \approx e^{-|A|^\eta}$, where $\eta < 1$. This non-Gaussian PDF
28: is explained by the nonlinear characteristics of individual bursts combined
29: with the statistics of bursts.
30: Our results suggest a general picture in which an incoherent background of
31: weakly interacting waves, occasionally, `by chance', initiates intense,
32: coherent, self-reinforcing, highly nonlinear events.
33: \end{abstract}
34: \pacs{02.30.Jr, 03.65.Ge, 05.45.-a, 52.35.Mw}
35: \maketitle
36:
37: \section{Introduction}
38: Many spatio-temporal dynamical systems show rare intense events.
39: One example is that of large height rogue ocean waves \cite{Osborne1}.
40: Another example occurs in recent experiments on parametrically forced surface
41: waves on water in which high spikes (bursts) on the free surface form
42: intermittently in space and time \cite{Lathrop}.
43: Other diverse physical examples also exist (\it e.g.\rm, tornados, large
44: earthquakes, etc.). The characteristic feature of rare intense events
45: is an enhanced tail in the event size probability distribution function. Here,
46: by enhanced we mean that the event size probability distribution function
47: approached zero with increasing event size much more slowly than is the case
48: for a Gaussian distribution. Thus these events, although rare, can be much
49: more common than an expectation based on Gaussian statistics would indicate.
50: The central limit theorem implies Gaussian behavior for a quantity that results
51: from the linear superposition of many random independent contributions.
52: Non-Gaussian tail behavior can result from strong nonlinearity of the events,
53: and enhancement of the event size tail might be expected if
54: large amplitudes are nonlinearly self-reinforcing.
55: Such nonlinear self-reinforcements is present in the nonlinear
56: Schr\"{o}dinger (NLS) equation.
57: In particular, the two dimensional NLS equation,
58: \begin{equation}
59: \frac{\partial A}{\partial t} = - i \alpha|A|^{2} A -i \beta
60: \bigtriangledown ^2 A.
61: \label{eq:nls}
62: \end{equation}
63: exhibits ``collapse" when the coefficients $\alpha$ and $\beta$ have the same
64: sign \cite{Bartuccelli}.
65: In a collapsing NLS solution the complex field approaches
66: infinity at some point in space, and this singularity occurs at finite time.
67: The NLS is conservative in that it can be derived from a Hamiltonian,
68: $\partial A / \partial t = - i \delta H / \delta A^* $, where $ H[A, A^*] =
69: \frac{1}{2} \int [\alpha |A|^4 + \beta |\bigtriangledown A|^2] d {\bf x} $.
70: In the case of nonconservative dynamics, inclusion of lowest order dissipation
71: and instability terms leads to the complex Ginzburg-Landau (CGL) equation
72: \cite{Levermore}. The CGL equation has been studied as a model for
73: such diverse situation as chemical reaction \cite{Chemical},
74: Poiseuille flow \cite{Poiseuille},
75: Rayleigh-B\'{e}rnard convection \cite{Rayleigh}, and
76: Taylor-Couette flow \cite{Taylor}.
77: In the limit of zero dissipation/instability the CGL equation approaches the
78: NLS equation. For small nonzero dissipation/instability, the CGL equation
79: displays a solution similar to the NLS collapse solution, but with a large
80: finite (rather than infinite) amplitude at the collapse
81: time \cite{Bartuccelli}.
82: Furthermore, over a sufficiently large spatial domain, these events occur
83: intermittently in space and time. Thus, in this parameter regime, the CGL
84: equation may be considered as a model for the occurrence of rare intense events.
85:
86: In this paper we study the statistics and characteristics of rare intense
87: events in a two-dimensional CGL-based model.
88: The probability distribution function (PDF) of the amplitude of
89: the solutions is observed to be non-Gaussian in our
90: numerical experiments. This non-Gaussian PDF is explained by the nonlinear
91: characteristics of individual bursts combined with the statistics of bursts.
92: The model equation we investigate is
93: \begin{eqnarray}
94: \frac{\partial A}{\partial t} &=& \pm A - (1+i \alpha)|A|^{2} A +
95: (1-i \beta) \bigtriangledown ^2 A \nonumber \\
96: & & +(\delta_r + i \delta_i) A^*,
97: \label{eq:cgl}
98: \end{eqnarray}
99: where $(\delta_r + i \delta_i) A^*$ is a parametric forcing term \cite{Park}.
100: We will consider two cases: one without parametric forcing ($\delta_r =
101: \delta_i = 0$) in which case the plus sign is chosen in front of the first
102: term on the right-hand side of (\ref{eq:cgl})
103: (Eq. (\ref{eq:cgl}) is then the usual CGL equation),
104: and one with parametric forcing, in which case the minus sign is chosen.
105: As previously discussed, we choose our parameters so that our model,
106: Eq. (\ref{eq:cgl}), is formally close to the NLS equation (\ref{eq:nls}).
107: That is, we take $\alpha, \beta \gg 1, \delta_r, \delta_i$, and for our
108: numerical solutions we will restrict attention to the case $\alpha = \beta$.
109: Note that the coefficient $\pm 1$ for the first term, as well as the ones in
110: $(1+i \alpha)$ and $(1 + i \beta)$ represent no loss of generality, since
111: these can be obtained by suitable scaling of the time($t$), the dependent
112: variable($A$), and the spatial variable(${\bf x}$).
113: In Section II, we discuss the amplitude statistics of our two-dimensional
114: CGL models with and without the parametric forcing term. We find that the
115: PDFs are approximately described by a stretched exponential distribution,
116: $P(|A|) \approx \exp(-|A|^{\eta})$, where $\eta$ is less than 1.
117: In Section III, we investigate the characteristics of individual bursts. We
118: compare our numerical CGL results with known collapse solutions of the NLS
119: equation. The maximum amplitude obtained by many bursts (or the ``event size"
120: statistics) is discussed in Section IV. Section V presents further discussion
121: and conclusions.
122:
123: Our results lead us to the following picture for the occurrence of rare
124: intense events in our system. Linear instability and nonlinear wave saturation
125: lead to an incoherent background of small amplitude waves. This background
126: is responsible for the observed \it small \rm $|A|$ Gaussian behavior of our
127: probability distribution functions. When, by chance, the weakly interacting
128: waves locally superpose to create conditions enabling nonlinear, coherent
129: self-reinforce, a localized, collapse-like event is initiated. Collapse takes
130: over, promoting large, rapid growth and spiking of $A$. This is followed by
131: a burn-out phase in which the energy in rapidly dissipated due to the
132: generation of small scale structure by the spike. We believe that elements of
133: the above general picture may be relevant to a variety of physical situations
134: where rare intense events occur (\it e.g.\rm, the parametrically driven water
135: wave experiments in Ref. \cite{Lathrop}).
136: \begin{figure}[t]
137: \epsfig{file=nopara1.eps,width= 8.5 cm}
138: \caption{Solutions of the CGL model.
139: (a) Snapshot of the amplitude $|A|$ for $L = 20 \pi$, $\alpha = \beta = 30$,
140: $\Delta t = 10^{-5}, \delta_r = \delta_i = 0$ and a $256 \times 256$ grid.
141: (b) Amplitude profile versus time. Solid line indicates $|A|_{max}$ of
142: the whole system, while thick solid lines indicate maximum amplitude of the
143: localized events("bursts"). The dashed line indicates the average amplitude
144: of $|A|$ over the whole system $|A|_{avg} \sim 0.3$.}
145: \label{fig:pdf_11}
146: \end{figure}
147:
148: \section{Amplitude Statistics}
149:
150: \subsection{2D model without a parametric force ($\delta_r = \delta_i = 0$)}
151:
152: We first consider Eq.(\ref{eq:cgl}) with $\delta_r = \delta_i = 0$ and
153: with the plus sign in the first term on the right hand side of the
154: equation. Figure \ref{fig:pdf_11}(a) shows a representation of $|A({\bf x},t)|$
155: [from numerical solution of Eq. (\ref{eq:cgl})] at a fixed instant $t$, where
156: large values are indicated by darker grey shades. As a function of time,
157: the localized dark shades occur in an seemingly random manner, become darker
158: (\it i.e., \rm increase their amplitude) and then go away (become light).
159: Furthermore, the maximum amplitudes also display apparent randomness.
160: [see Fig. \ref{fig:pdf_11}(b)].
161: As shown in the next section (Sec. III), although the occurrence of these
162: intense events is apparently erratic in time and
163: space, individually these events are highly
164: coherent. In this section, we will study the statistical properties of
165: $A({\bf x},t)$.
166:
167: Our numerical solutions of (\ref{eq:cgl}) employ periodic boundary
168: conditions on a $256 \times 256$ grid.
169: We choose the parameters, $\alpha$ and $\beta$, large enough
170: ($\alpha = \beta = 30$) so that the solutions of our model are close to
171: solutions of the NLS equation. We choose the time step
172: small enough to satisfy the condition for unconditional stability
173: of our second-order accurate time integration($\Delta t = 10^{-5}$).
174: We use random initial condition (at $t=0$, we generate random values for
175: amplitudes and phases at each grid point).
176: Localized structures, "bursts", develop very rapidly and occur throughout the
177: spatial domain. The typical life time of a burst ($\delta t$) is approximately
178: 0.2 time units. The maximum amplitudes of bursts are different for different
179: burst events.
180:
181: Imagining that we choose a space-time point $({\bf x},t)$ at random, we now
182: consider the probability distribution functions for $|A|$ (the magnitude of
183: $A$), $A_r = Re[A]$ (the real part of $A$), and $A_i = Im[A]$ (the imaginary
184: part of $A$). We denote these distribution functions $P(|A|), P_r(A_r),
185: P_i(A_i)$, and we compute them via histogram approximations using the values
186: of $|A|, A_r,$ and $A_i$ from each of the $256 \times 256$ grid points at many
187: time frames \cite{footnote}. We find that these distributions are independent
188: of the periodicity length $L$ used in the computation as long as it is
189: sufficiently large compared to the spatial size of a burst, but is not so large
190: that spatial resolution on our $256 \times 256$ grid becomes problematic.
191: In our computations of $P, P_r$ and $P_i$, we choose $L = 20 \pi$.
192:
193: \begin{figure}[t]
194: \epsfig{file=nopara2.eps,width= 8.5 cm}
195: \caption{Probability distribution functions obtained from numerical solution
196: of Eq.(2) using the same parameters as in Fig. 1. The circles are data for
197: $P_r$ and $P_i$, while the pluses are the probability distributions $P_r'$
198: and $P_i'$ obtained from the phase randomized amplitude.}
199: \label{fig:pdf_21}
200: \end{figure}
201:
202: Figures 2 show the numerically computed probability distributions $P_r(A_r)$
203: [Fig. 2(a)] and $P_i(A_i)$ [Fig. 2(b)] plotted as open circles. Since
204: Eq. (\ref{eq:cgl}) with $\delta_r = \delta_i = 0$ is invariant to the
205: transformation
206: $ A \rightarrow A \exp(i\phi)$ (where $\phi$ is an arbitrary constant), we
207: expect the distribution $P_r$ and $P_i$ to be the same to within the
208: statistical accuracy of their determinations. This is born out by Figs. 2.
209: In order to highlight the essential role that nonlinearity plays in
210: determining these distribution functions, we have also computed them after
211: randomizing the phases of each Fourier component. That is, representing
212: $A({\bf x},t)$ as
213: \begin{equation}
214: A({\bf x},t) = \sum_{{\bf k}} a_{{\bf k}}(t) \exp (i {\bf k}\cdot{\bf r}),
215: \end{equation}
216: where ${\bf k} = (2m\pi/L, 2n\pi/L)$, we form a new amplitude,
217: $A'({\bf x},t)$ as
218: \begin{equation}
219: A'({\bf x},t) = \sum_{{\bf k}} a_{{\bf k}}(t)
220: \exp (i {\bf k}\cdot{\bf r} + i \theta_{{\bf k}}),
221: \end{equation}
222: where for each ${\bf k}$, the angle $\theta_{{\bf k}}$ is chosen randomly
223: with uniform probability density in $[0, 2\pi]$. The probability distribution
224: functions for the real and imaginary parts of the randomized amplitudes
225: $A'$ are shown as pluses in Figs. \ref{fig:pdf_21}.
226: Note that by construction, $A$ and $A'$ have
227: identical wavenumber power spectra. Due to the random phases, $A'$ at any
228: given point $\bf{x}$ can be viewed as a sum of many independent random numbers
229: (the Fourier components). Hence the $P_r$ and $P_i$ distributions are
230: expected to be Gaussian, $\log P_{r,i} \sim [(const.) - (const.) A_{r,i}^2]$.
231: This is confirmed by the semi-log plots of Figs. 2, where the data plotted as
232: pluses can be well-fit by parabolae.
233:
234: The above comparisons with the phase randomized variable
235: $A'$ are motivated by imagining the hypothetical situation where the amplitude
236: is formed by the superposition of many noninteracting linear plane waves.
237: In that case we would have an amplitude field of form
238: \begin{equation}
239: \sum_{{\bf k}} b_{{\bf k}}(t) \exp (i {\bf k}\cdot{\bf r}+i \omega_{{\bf k}}t).
240: \end{equation}
241: Because $\omega_{{\bf k}}$ is different for different ${\bf k}$, the phases
242: become uncorrelated for sufficiently large time $t$.
243:
244: \begin{figure}[t]
245: \epsfig{file=nopara3.eps,width= 4.5 cm}
246: \caption{Probability distribution functions before randomizing the phases of
247: the solutions (o) and after randomizing the phases (+). Note the horizontal
248: axis is $|A|^\eta$, where the exponent $\eta=0.8$ is chosen to yield
249: approximately linear dependence of $\log[P(|A|)]$ on $|A|^\eta$ for large
250: values of $|A|$.}
251: \label{fig:pdf_31}
252: \end{figure}
253:
254: Comparing the data from $A$ and $A'$ in Figs. 2, for small values of $A_r$
255: and $A_i$, the PDFs are nearly Gaussian. This can be attributed to near linear
256: behavior of small amplitude waves. On the other hand, for the tails of the
257: distributions, we note substantial enhancement relative to the Gaussian
258: distributions. These must be due to coherent phase correlations resulting from
259: nonlinear interaction of different wavenumber components of $A$. Such phase
260: coherence is implied by the observed coherent localized burst structures.
261:
262: Figure 3 shows the numerically obtained distribution $P(|A|)$ plotted as
263: circles and the probability distribution for the phase randomized amplitude
264: $|A'|$ plotted as pluses.
265: Again, the enhancement of the large amplitude tail
266: is evident. Note that the vertical axis in Fig. 3 is logarithmic, while the
267: horizontal axis is $|A|^{\eta}$. Here we choose the power $\eta = 0.8$ so that
268: the large $|A|$ data in this plot are most nearly fit by a straight line. That
269: is, we attempt to fit $P(|A|)$ using a stretched exponential. The slope of the
270: dashed straight line in the figure is chosen to match the large $|A|$ data.
271: Thus, over the range of $|A|$ accessible to over numerical experiment, we
272: find that the enhanced large $|A|$ tail probability density is roughly fit by
273: a stretched exponential,
274: \begin{equation}
275: P(|A|) \sim \exp(-\zeta |A|^{0.8}).
276: \end{equation}
277:
278: \subsection{2D model with parametric forcing ($\delta_r,\delta_i \ne 0$)}
279:
280: We now report similar results for the case of parametric forcing, Eq.
281: (\ref{eq:cgl}) with $\delta_r, \delta_i \ne 0$ and the minus sign chosen
282: in the first term on the right side of (\ref{eq:cgl}). In this case, instability
283: of small amplitude waves is caused by the parametric forcing (nonzero
284: $\delta_{r,i}$) and the $-A$ term represents a linear wavenumber independent
285: damping. This model for parametric forcing was introduced \cite{Corellet}
286: and has been used to model various situations. One such situation is that of
287: periodically forced chemical reactions \cite{Hagberg}. Our motivation for
288: considering this model is the Faraday experiments on strong parametric forcing
289: of surface water waves in Ref. \cite{Lathrop}. In that work intermittent
290: formation of large localized surface perturbations results in splash and
291: droplet formation.
292:
293: Parameters for our numerical simulations are the same as in Sec. IIA except
294: that now $\delta_r = \delta_i = 5$. Again coherent localized structures,
295: "bursts", develop rapidly and occur intermittently throughout the spatial
296: domain, Fig. \ref{fig:pdf_12}(a). As in Sec. IIA, the typical life
297: time of a burst is less than 0.2 time units, and the maximum amplitudes of
298: bursts are different for different bursts.
299:
300: \begin{figure}[t]
301: \epsfig{file=para1.eps,width= 8.5 cm}
302: \caption{Solutions of the model with parametric forcing.
303: (a) snapshot of $|A|$. Dark regions have high amplitudes.
304: (b) $P(|A|)$ versus $|A|^\eta$, where $\eta = 0.8$. (See captions in Fig.
305: \ref{fig:pdf_31}.)}
306: \label{fig:pdf_12}
307: \end{figure}
308:
309: The PDF, $P(|A|)$ again shows a stretched exponential tail with exponent
310: $\eta = 0.8$, Fig. \ref{fig:pdf_12}(b).
311: The circles in Figs. \ref{fig:pdf_22} show the PDFs of the real and imaginary
312: parts of $A$, while the pluses are data for the PDFs after randomizing the
313: phase. The shape of the PDFs around the central part are nearly Gaussian.
314: In contrast, at large amplitude the PDFs are significantly non-Gaussian.
315: A major difference with the case $\delta_r = \delta_i = 0$ is that,
316: with parametric forcing, the model is not invariant to
317: $A \rightarrow A e^{i\phi}$, and thus $P_r$ and $P_i$ may be expected to
318: evidence differences not present for $\delta_r = \delta_i = 0$.
319: This is seen to be the case in Figs. \ref{fig:pdf_22}.
320:
321: \begin{figure}[b]
322: \epsfig{file=para2.eps,width= 8.5 cm}
323: \caption{$P_r(A_r)$ and $P_i(A_i)$ with parametric forcing. The numerical
324: parameters are the same as those for Fig. \ref{fig:pdf_12}.
325: The circles are PDFs for $A_r$ and $A_i$ before randomizing the phases of the
326: solutions, while pluses are PDFs after randomizing the phases.}
327: \label{fig:pdf_22}
328: \end{figure}
329:
330: \section{Characteristics of Individual Burst Events}
331:
332: Solutions of the CGL equation with large $\alpha$ and $\beta$ may be expected
333: to have features in common with solutions of the NLS equation.
334: It is known that the NLS equation yields localized events which develop finite
335: time singularities where the amplitude becomes infinite ~\cite{Levermore}.
336: While it is difficult to understand the dynamics of the solutions of CGL
337: equation from direct rigorous analysis, the solutions of the
338: NLS equation are relatively well understood. Thus, we analyze the dynamics of
339: individual CGL bursts guided by the known localized self-similar collapsing
340: solution of the NLS equation.
341:
342: The NLS equation has a special solution \cite{Fibich} of the form
343: \begin{equation}
344: A = e^{i\theta t}R(r), ~~~~r=\sqrt{x^2+y^2},
345: \end{equation}
346: where the radial function $R(r)$ satisfies
347: \begin{eqnarray}
348: \label{eq:radial}
349: \left( \frac{\partial ^2}{\partial r^2} + \frac{1}{r}
350: \frac{\partial}{\partial r} \right)R - \xi R + R^3 = 0, \\ \nonumber
351: ~~ \left| \frac{\partial R}{\partial r} \right|_{r=0} = 0, ~~R(\infty) = 0,
352: \end{eqnarray}
353: where $ \alpha = \beta$, $\xi = \theta / \beta$ .
354: Since (\ref{eq:nls}) is invariant under the scaling transformation
355: \cite{Lemesurier},
356: \begin{equation}
357: A({\bf x},t) \longrightarrow L(t)^{-1} A({\bf \kappa}, \tau)
358: {\rm e}^{iL \dot{L}|{\bf \kappa}|^2/4},
359: \label{eq:rescalea}
360: \end{equation}
361: where $L(t)$ tends to zero as $t^* \longrightarrow t$, $t<t^*$ and
362: \begin{equation}
363: {\bf \kappa} = \frac{{\bf x}}{L(t)}, ~~~~
364: \tau = \int_0^t\frac{1}{L^2(s)}ds,
365: \label{eq:rescalex}
366: \end{equation}
367: a family of collapsing solutions of the NLS is given by the rescaled solution
368: of Eq. (\ref{eq:radial}).
369:
370: \begin{figure}[t]
371: \epsfig{file=burst1.eps,width= 8.5 cm}
372: \caption{Self-similar bursts. (a) Enlarged plots of a burst at
373: $t_1 (\bigcirc) = 10.448, $ at grid point $(x,y) = (15.2,22.1),$ and
374: $t_2 (\triangle) = 10.530,$ at grid point $(51.1,25.5),$ and
375: $t_3 (\square)= 10.644 $ at grid point $(57.4,38.8)$.
376: (b) Scaled profiles, where $|\tilde{A}|= |A|/L$, $\tilde{x} = x/L$,
377: and $L = |A|_{max}$ at $t_1$, $t_2$, and $t_3$. The solid line represents
378: the radial solution of Eq. (\ref{eq:radial}).}
379: \label{fig:bursts}
380: \end{figure}
381:
382: With these considerations, we test the expected approximate self-similarity of
383: individual bursts observed in our numerical solutions of Eq. (\ref{eq:cgl}).
384: We consider three typical bursts that occur at different times and spatial
385: positions. In particular, we choose these three as the three dark regions
386: in Fig. \ref{fig:pdf_11}(a) whose spatial maxima have the time dependences
387: shown as thick solid lines in Fig. \ref{fig:pdf_11}(b).
388:
389: In Fig. \ref{fig:bursts}(a) we plot the $x$-dependence of $|A|$ at constant
390: $y$ for each of these bursts at the time that they reach their maximum
391: amplitude (the positions in $x$ of the maxima have been shifted to $x=0$
392: and the constant $y$ value for each is at the location of $|A|_{max}$).
393: Note that, when they reach their maxima, the three bursts have different
394: amplitudes and width. We rescale the amplitude and spatial coordinate as
395: suggested by (\ref{eq:rescalea}) and (\ref{eq:rescalex}),
396: $|\tilde{A}|=|A|/L$ and $\tilde{x}=x/L$, and we take $L = |A|_{max}$ (which
397: normalizes $|\tilde{A}|_{max}$ to one).
398: The resulting data are plotted in Fig. \ref{fig:bursts}(b) along with the
399: solution of Eq. (\ref{eq:radial}).
400: [We again rescale $R(r)$ using (\ref{eq:rescalea}) and (\ref{eq:rescalex}),
401: and we note that, after this rescaling, the result is independent of the value
402: of $\xi=0.1$ in Eq. (\ref{eq:radial}).] The three burst profiles show evidence
403: of collapsing onto the theoretical radial solution.
404:
405: Now, we consider the time dependence of a single burst. We select the burst
406: at the grid point $(x,y) = (51.1,25.5)$ (see caption to Fig. \ref{fig:bursts})
407: and investigate the evolution of its shape and height.
408: We display profiles of the burst at 5 different times in
409: Fig. \ref{fig:single1}(a). Rescaling each profile using Eq. (\ref{eq:rescalea})
410: and Eq. (\ref{eq:rescalex}), and defining $L$ in the same way as before,
411: the four profiles at the first four times approximately collapse onto the
412: radial solution of (\ref{eq:radial}) as shown in Fig. \ref{fig:single1}(b).
413: When the burst reaches its maximum amplitude, the amplitude at some distance
414: away from the center becomes zero (see the amplitude profile at $t=t_3$).
415: After that, the center decays very rapidly
416: (see the amplitude profile at $t=t_5$).
417: (Note that in this section and the next section, we present numerical results
418: for (\ref{eq:cgl}) with $\delta_r =\delta_i = 0$. However, we have also
419: verified that the CGL model with parametric forcing also has similar self
420: similarity properties.)
421:
422: \begin{figure}[t]
423: \epsfig{file=burst2.eps,width= 8.5 cm}
424: \caption{Self-similarity of single burst. (a) Enlarged plots of a burst at
425: $t_1 (\bigcirc) = 10.50$, $t_2 (\square) = 10.52$, $t_3 (\triangle)= 10.53$,
426: $t_4 (\diamond) = 10.54$, and $t_5 (\times) = 10.55$.
427: (b) Scaled profiles, where $|\tilde{A}|= |A|/L$, $\tilde{x} = x/L$,
428: and $L = |A(r)|_{max}$ at $t_1, t_2, t_3 $, and $t_4$. The solid line
429: represents the solution of Eq. (\ref{eq:radial}).}
430: \label{fig:single1}
431: \end{figure}
432:
433: \section{Statistics of Bursts}
434:
435: The self similar properties of bursts implies that the solutions of the
436: CGL model consist of self-similar bursts of various maximum heights.
437: Thus, we expect that the enhanced tail
438: (the deviation from the Gaussian distribution)
439: of the amplitude probability distribution $P(|A|)$ can be understood by
440: the statistics of bursts. In particular, we consider $g(h)$, the frequency of
441: bursts which have maximum height $h$, and a distribution $P_j(|A|)$ defined
442: for an individual burst (burst $j$), as follows.
443:
444: We define the time interval for each burst as the time between when its peak
445: value exceeds $2|A|_{avg}$ and when its peak value drops below $2|A|_{avg}$
446: [typically the time duration of a burst is less than $0.2$, see
447: Fig. \ref{fig:pdf_11}(b)]. Here $|A|_{avg}$ is the space average of $|A|$ over
448: the entire spatial grid of the simulation at each time $t$ [$|A|_{avg}$ is
449: approximately constant at about $0.3$ over all time steps in the simulation,
450: see the dashed line in Fig. \ref{fig:pdf_11}(b)]. Consistent with the
451: observation that a typical burst has radial symmetry, we define the domain of
452: the burst to be a circular region of radius $r_{eff}$ centered at the burst
453: maximum, where $r_{eff}$ is the maximum radius of a circle such that the
454: average of $|A|$ over the perimeter of the circle is greater than $2|A|_{avg}$
455: (typically, $1.23 \le r_{eff} \le 4.91$).
456: In Fig. \ref{fig:stat}(a) we show the distribution $P_j(|A|)$ for the three
457: bursts in Fig. \ref{fig:pdf_11}(b) (thick solid lines). The first burst
458: ($j=1$) has $h = 4.52$ and is plotted as the open circles in
459: Fig. \ref{fig:stat}(a); the second burst ($j=2$) has $h = 6.31$ and is plotted
460: as the open triangles; and the third burst ($j=3$) has $h = 3.85$ and is
461: plotted as the open squares.
462: These distributions are obtained from histograms of the
463: values of $|A|$ at grid points in the domains and time steps in the duration
464: of each of these bursts.
465: We obtain $g(h)$ by counting the number of bursts which have maximum heights
466: between $h$ and $h + \Delta h$, where $\Delta h = 0.2$
467: [see Fig. \ref{fig:stat}(b)]. (In Fig. \ref{fig:stat}(a) $P_j(|A|)$ is not
468: plotted for $|A| < 2|A|_{avg}$, since, by our procedure this range lacks
469: meaning, and since we are interested in the behavior at large values of $|A|$.)
470: We note that the $P_j(|A|)$ in Fig. \ref{fig:stat} all approximately coincide
471: for $|A|<h$. Thus the only characteristic of the bursts on which $P_j(|A|)$
472: depends is the maximum burst amplitude $h$ at which $P_j(|A|)$ goes to zero.
473: To incorporate this fact, we replace $P_j(|A|)$ by the notation $P_h(|A|)$.
474:
475: \begin{figure}[t]
476: \epsfig{file=stat.eps,width= 8 cm}
477: \caption{Statistics of localized events("bursts").
478: (a) $P_{h_j}(|A|)$ at three different times:
479: $t_1 (\bigcirc) = 10.448$ and $h=4.52$, $t_2 (\triangle) = 89.0$ and
480: $h=6.31$, and $t_3 (\square) = 94.0$ and $h=3.85$.
481: (b) The frequency of bursts which have maximum height $h$, $g(h)$. The inset
482: indicates $C(h)$ versus $h$ defined in Eq. (\ref{eq:normal}).}
483: \label{fig:stat}
484: \end{figure}
485:
486: The above suggests that $P(|A|)$ can be obtained from the
487: following approximation \cite{Iwasaki}
488: \begin{equation}
489: P(|A|) \sim \int_0^{\infty} g(h)P_h(|A|)dh,
490: \label{eq:approx}
491: \end{equation}
492: where $P_h(|A|)$ is a probability distribution of a single burst whose temporal
493: maximum amplitude is $h$. Since $P_h(|A|)$ vanishes for $ |A| > h$ and
494: because the $P_h(|A|)$ approximately coincide for $|A|<h$
495: [see Fig. \ref{fig:stat}(a)], we approximate $P_h(|A|)$ as
496: \begin{eqnarray}
497: P_h(|A|) &\sim& C^{-1}(h) \theta (h-|A|) P_*(|A|), \\
498: \label{eq:ph}
499: C(h) &=& \int_0^h P_*(|A|)d|A|,
500: \label{eq:normal}
501: \end{eqnarray}
502: where $C(h)$ is a normalization factor [$C(h) \sim 1$, when $h>1$; see
503: the inset on Fig. \ref{fig:stat}(b)],
504: $\theta(h-|A|)$ is a step function, and $P_*$ is the distribution that we
505: numerically compute at the largest value of $h$ that we considered
506: ($h_{max}=6.31$). Using (\ref{eq:approx}) and (\ref{eq:ph}), we can further
507: approximate $P(|A|)$ as
508: \begin{eqnarray}
509: P(|A|) &\sim& \int_0^{\infty} C^{-1}(h)
510: \theta(h-|A|)g(h)P_*(|A|)dh \nonumber \\
511: &\sim& P_*(|A|) \int_{|A|}^{\infty} C^{-1}(h) g(h) dh.
512: \label{eq:gh}
513: \end{eqnarray}
514: (The integral in (\ref{eq:gh}) is the cumulative frequency of bursts which
515: have maximum height greater than $|A|$.)
516:
517: Figures \ref{fig:stat} show the numerically obtained $g(h)$ and
518: $P_*(|A|)$. Inserting $P_*(|A|)$ into Eq.(\ref{eq:normal}) and
519: Eq.(\ref{eq:gh}), we obtain the prediction for $P(|A|)$ plotted as pluses in
520: Fig. \ref{fig:tot} for $|A|>2|A|_{avg}$. This appears to agree well with the
521: $P(|A|)$ obtained from our numerical solutions of (\ref{eq:cgl})
522: (open circles). (Note that we shift the predicted $P(|A|)$ (pluses) to the
523: $P(|A|)$ (open circles) obtained from (\ref{eq:cgl}) after removing data points
524: for $|A|<2|A|_{avg}$.)
525:
526: \begin{figure}[t]
527: \epsfig{file=pdfs.eps,width= 4.5 cm}
528: \caption{$P(|A|)$ versus $|A|$. Circles represent $P(|A|)$ obtained directly
529: from our numerical solutions of (\ref{eq:cgl}), while pluses represent
530: $P(|A|)$ obtained using $P_*(|A|)$ and $g(h)$ from Fig. \ref{fig:stat},
531: and Eqs. (\ref{eq:approx})-(\ref{eq:gh}).}
532: \label{fig:tot}
533: \end{figure}
534:
535: \section{Conclusion}
536:
537: We find that the large $A$ behavior of the PDF obtained from our CGL
538: solutions is approximately described by a stretched exponential form,
539: $P(|A|) \approx e^{-|A|^\eta}$, where $\eta < 1$.
540: In addition, for small $A$, $P_r(A_r)$ and $P_i(A_i)$ are approximately
541: Gaussian, as is the case for a random linear superposition of waves.
542: We also observe the self similar properties of individual bursts, which allow
543: us to consider the large amplitude behavior of our CGL solutions as composites
544: of coherent self similar bursts. Based on this we explain the observed
545: non-Gaussian $P(|A|)$ using the nonlinear characteristics of individual
546: bursts $P_h(|A|)$ combined with the statistics of burst occurrences $g(h)$.
547:
548: These results lead us to conjecture the following picture of rare intense
549: events in our model. Linear instability leads to a background of relatively
550: low amplitude waves that are weakly interacting and result in a
551: random-like, incoherent background and low $|A|$ Gaussian behavior of
552: $P_r(A_r)$ and $P_i(A_i)$. When, by chance, this incoherent behavior results
553: in local conditions conductive to burst formation, nonlinear, coherent,
554: self-reinforcing collapse takes over and promotes a large growth and spiking
555: of $A$. We believe that this general mechanism may be operative in a variety
556: of physical situations in which rare intense events occur (\it e.g.\rm,
557: the water wave experiments of Ref. \cite{Lathrop}).
558:
559: We thank D. P. Lathrop for initial discussion and for attracting our
560: attention to the subject of rare intense events. We thank P. N. Guzdar for his
561: advice on numerics. This work was supported by the Office of Naval Research
562: (Physics) and by the National Science Foundation (PHY0098632).
563:
564: \begin{references}
565: \bibitem{Osborne1} A. R. Osborne, M. Onorato, and M. Serio, Phys. Lett. A
566: \bf 275 \rm, 386(2000).
567: \bibitem{Lathrop} J. E. Hogrefe, \it et al\rm., Physica D
568: \bf 123 \rm 183(1998); B. W. Zeff, \it et al\rm., Nature \bf 403 \rm 401(2000).
569: \bibitem{Bartuccelli} M. Bartuccelli, \it et al.\rm, Physica D \bf 44 \rm,
570: 421(1990). They draw a phase diagram figures for NLS equation (Fig. 1) and
571: CGL equation (Fig. 2).
572: \bibitem{Levermore} C. D. Levermore and M. Oliver, Lect. Appl. Math. \bf 31 \rm
573: 141(1996).
574: \bibitem{Chemical} Y. Kuramoto, \it Chemical Oscillations, Waves and
575: Turbulence \rm, Series in Synergetics \bf 19 \rm, Springer, New York, 1984.
576: \bibitem{Poiseuille} K. Stewartson and J.T. Stuart, J. Fluid Mech. \bf 48
577: \rm, 529(1971).
578: \bibitem{Rayleigh} A.C. Newell and J.A. Whitehead, J. Fluid Mech.
579: \bf 38 \rm, 279(1969).
580: \bibitem{Taylor} G. Ahlers and D.S. Cannell, Phys. Rev. Lett. \bf 50 \rm,
581: 1583(1983).
582: \bibitem{Park} H.-K. Park, Phys. Rev. Lett. \bf 86 \rm 1130(2000); also see
583: the references therein.
584: \bibitem{footnote} $P(|A|)$ is to be contrasted with the probability
585: distribution of a single burst amplitude $P_h(|A|)$ to be discussed
586: in Sec. IV.
587: \bibitem{Corellet} P. Corellet, Phys. Rev. Lett \bf 56 \rm, 724(1986).
588: \bibitem{Hagberg} C. Elphick, A. Hagberg, and E. Meron, Phys. Rev. Lett
589: \bf 80 \rm, 5007(1998).
590: \bibitem{Fibich} G. Fibich and D. Levy, Phys. Lett. A \bf 249 \rm, 286(1998).
591: \bibitem{Lemesurier} B. LeMesurier, \it et al.\rm, Physica D
592: \bf 32 \rm, 210(1988).
593: \bibitem{Iwasaki} H. Iwasaki and S. Toh, Prog. Theo. Phys. \bf 87\rm, 1127
594: (1992).
595: \end{references}
596:
597: \end{document}
598: