1: \documentclass[12pt]{article}
2: \usepackage{graphicx}
3:
4: %% Macros
5: \def\eqref#1{eq.\ (\ref{#1})}
6: \def\figref#1{Fig.~\ref{#1}}
7: \def\dt{{\Delta t}}
8: \def\bz{{\bar{z}}}
9: \def\bzn{{\overline{z_n}}}
10: \def\avg#1{{\langle{#1}\rangle}}
11:
12: % --------------------------------------------------
13: % Title and version
14:
15: \begin{document}
16:
17: \title{Coarse-graining and self-similarity of price fluctuations}
18: \author{Yoshi Fujiwara$^1$ and Hirokazu Fujisaka$^2$\\
19: \\
20: $^1$ Keihanna Research Center,\\
21: Communications Research Laboratory, Kyoto 619-0289, Japan\\
22: $^2$ Department of Applied Analysis and Complex Dynamical Systems,\\
23: Graduate School of Informatics, Kyoto University,\\
24: Kyoto 606-8501, Japan}
25: \date{\today}
26: \maketitle
27: \begin{abstract}
28: We propose a new approach for analyzing price fluctuations in their
29: strongly correlated regime ranging from minutes to months. This is
30: done by employing a self-similarity assumption for the magnitude of
31: coarse-grained price fluctuation or volatility. The existence of a
32: Cram\'er function, the characteristic function for self-similarity,
33: is confirmed by analyzing real price data from a stock market. We
34: also discuss the close interrelation among our approach, the
35: scaling-of-moments method and the multifractal approach for price
36: fluctuations.
37: \end{abstract}
38:
39: \vspace{2em}
40: % --------------------------------------------------
41:
42: Scaling and self-similarity play an important role not only in physics
43: but also in various socio-economic systems including market
44: fluctuations, internet traffic, growth rates of firms and social
45: networks. If phenomenological theory is shown to be applicable to
46: observations, it would guide one to meaningful models and possibly to
47: universal and specific description of such systems without any
48: apparent ``Hamiltonian''.
49:
50: Price fluctuations in speculative markets from physical viewpoint have
51: recently received great attention \cite{SELECTAE,IEP99,TFR00,IEEE99}.
52: Advances have been aided by concepts and techniques derived from
53: statistical physics and also by the new feasibility of analyzing
54: high-frequency financial data. Considerable statistical features have
55: been uncovered that, interestingly, are quite independent of asset
56: type (stock/foreign-exchange) and of markets (New York/Tokyo). In
57: fact, we shall describe below that price fluctuation is spectrally
58: white, but is not statistically independent since its magnitude
59: exhibits a strong correlation in the form of a power law
60: autocorrelation \cite{olsen93,liu99,gopi00} (see also references in
61: \cite{gopi00}). Moreover, some have suggested that price fluctuations
62: might have similar properties in different time-scales, that is
63: partially supported by the multifractal analysis of price fluctuations
64: \cite{mandelbrot97,arnedo98,schmitt99}. In this paper we propose a new
65: approach to characterizing self-similarity in the strongly correlated
66: regime. We actually show the existence of self-similarity and
67: characterize it both qualitatively and quantitatively from a new
68: statistical physical viewpoint. Specifically, we directly evaluate the
69: {\it invariant scaling\/} relation for fluctuations under {\it
70: coarse-graining\/} of asset price returns.
71:
72: Time-scale of our interest is related to the temporal features of the
73: price fluctuations known so far. Let $P(s)$ be the asset price at
74: time-step $s$, then the (logarithmic) return at $s$ for duration $t$
75: is defined by
76: %
77: \begin{equation}
78: r_t(s)=\log(P(s)/P(s-t)) .
79: \label{ret}
80: \end{equation}
81: %
82: Return is equal to the fractional change of price after duration $t$
83: when the change is small. We are interested in the
84: fluctuation $r_t$ for different values of $t$. Each transaction is
85: recorded for every stock in a particular market at the finest
86: time-scale. For the dataset from the New York Stock Exchange (NYSE)
87: used below, the typical time interval between successive transactions
88: was about 10 seconds for stocks with large (trading) volumes.
89:
90: In a regime with time-scales of minutes and longer, considerable
91: statistical features are known to exist \cite{olsen93,liu99,gopi00}.
92: For an appropriate small time-scale $\dt$ (hereafter chosen as one
93: minute), (i) $r_\dt$ has an autocorrelation function (ACF) that almost
94: vanishes beyond several minutes of time lag, whereas (ii) $|r_\dt|$
95: has a long correlation ranging from several minutes to more than
96: several months or even a year. In fact, the ACF for $|r_\dt|$ has a
97: power-law decay that is valid for such a long range of time lags.
98: While (i) is interpretable as the absence of arbitrage using linear
99: prediction, the long memory (ii), known as volatility clustering in
100: economic literature, is not completely understood in its origin.
101:
102: We assume from these observations that self-similarity property which
103: will be formulated shortly holds for time-scales between $t_s$ and
104: $T$, where $t_s$ is the small-, and $T$ the large-scale cut-off. The
105: value of $t_s$ is typically one minute or less, while $T$ may range
106: from several months to one year being regarded as a parameter. Using
107: an appropriate minimal time-scale $\dt$, we focus on the magnitude of
108: the coarse-grained return,
109: %
110: \begin{equation}
111: y_t(s)\equiv
112: \left|\sum_{k=0}^{t/\dt-1} r_\dt(s-k\dt)\right|
113: =|r_t(s)| .
114: \label{defy}
115: \end{equation}
116: %
117: The variable $t$ is the coarse-graining time-scale. \figref{fig:ss}
118: exemplifies in an actual data set the additional property of (iii)
119: self-similarity and intermittency for values of $y_t$ with different
120: time-scales $t$. The main aim of this work is to propose a new approach
121: for analyzing this self-similarity.
122:
123: First, we introduce the coarse-graining step $n$ as
124: %
125: \begin{equation}
126: t=t_n\equiv T\,e^{-n}
127: \label{defn}
128: \end{equation}
129: %
130: for $n=0,1,\ldots,N$ ($N\equiv\log(T/t_s)\gg1$), and $y_n\equiv
131: y_{t_n}$. Next, we assume that the ratio $y_{n+1}/y_n$ is
132: statistically independent of $n$, i.e., statistically steady for $1\ll
133: n\ll N$. This {\it ansatz\/} is the expression of the self-similarity
134: assumption for price fluctuations. That is, by using the equation
135: %
136: \begin{equation}
137: {y_{n+1}\over y_n}=e^{z_n} ,
138: \label{ss}
139: \end{equation}
140: %
141: $z_n$ is statistically steady with respect to $n$ for the similarity
142: region $1\ll n\ll N$. If we introduce the exponent $\bz(t)$ as
143: %
144: \begin{equation}
145: y_t\sim t^{-\bz(t)} ,
146: \label{lexp}
147: \end{equation}
148: %
149: $\bz(t)$ corresponds to the local (H\"older) exponent in multifractal
150: theory. Except for the $t$-dependence of the magnitude $y_t$, the
151: similar temporal evolution of $y_t$ for different values of $t$
152: (\figref{fig:ss}) is the origin of the self-similarity ansatz.
153:
154: It follows immediately from \eqref{ss} that
155: %
156: \begin{equation}
157: y_n=y_T\,\exp\sum_{k=0}^{n-1}z_k=y_T\,\exp(n\bzn) ,
158: \label{yn}
159: \end{equation}
160: %
161: where $\bzn\equiv(1/n)\sum_{k=0}^{n-1}z_k$. We assume that
162: $y_T=y_{n=0}$ has no fluctuation in the sense that one can neglect
163: fluctuation if the coarse-graining scale is beyond the time-scale $T$.
164: Large deviation theory (e.g.\cite{FRISCH}) in probability theory tells
165: us that this self-similarity hypothesis, the statistical independence
166: of exponents $z_k$ in the coarse-graining step $k$, leads us to
167: express the probability density function (PDF) $Q_n(z)$ of the average
168: $\bzn$ as
169: %
170: \begin{equation}
171: Q_n(z)\propto
172: e^{-nS(z)} ,
173: \label{qz}
174: \end{equation}
175: %
176: for $1\ll n\ll N$. Precisely speaking, \eqref{qz} is valid for $n\gg
177: n_c$, where $n_c$ is the correlation step of $z_n$ (e.g. defined as the
178: decay step of the correlation function of $\avg{\delta z_n\delta z_0}$,
179: where $\delta z_n\equiv z_n-\avg{z_n}$). Here the positive function
180: $S(z)$, being independent of $n$ for $n\gg n_c$, is known as a {\it
181: fluctuation spectrum\/} or {\it Cram\'er function}, whose functional
182: form characterizes the self-similar fluctuations and can be immediately
183: determined by observing the price fluctuations. Under the variable
184: transformation given by Eqs.\ (\ref{defn}) and (\ref{yn}),
185: %
186: \begin{equation}
187: z_t(y)=\log(y/y_T)/\log(T/t)
188: \label{zofy}
189: \end{equation}
190: %
191: (denoting $\bzn(y_n)$ as $z_t$ for short),
192: the PDF $Q_n(z)$ is transformed to the PDF of $y_t$, $P_t(y)$, as
193: %
194: \begin{equation}
195: P_t(y)\propto
196: {1\over y}\left({T\over t}\right)^{-S(z_t(y))} .
197: \label{py}
198: \end{equation}
199: %
200:
201: As a dataset for investigating this self-similarity hypothesis, we
202: used the Trade and Quote (TAQ) database of intraday transactions
203: provided by NYSE, Inc. The prices of a particular stock recorded
204: between Jan.~4, 1993 and June 30, 1999 were used. The median of the
205: time differences between quotes was 12~sec. The timing of the quotes
206: was used to construct a time-series of returns $r_\dt$ for
207: $\dt=$1~min, excluding midnights, i.e., the returns corresponding to
208: pairs of closing price and next day's open. The total dataset is
209: comprised of 638,901 records. Next we obtained an ensemble for $y_t$
210: with different time-scales $t$ by using \eqref{defy}. Other stocks
211: with high volumes were analyzed and yielded similar results to those
212: given below.
213:
214: Removing the intraday periodicity from the time-series is a vital but
215: difficult task. We used essentially the same method as that in
216: \cite{liu99} (see also \cite{footnote1}). Note that a daily period for
217: the NYSE from open to close is approximately 390 minutes. The
218: observables adjusted with the intraday periodicity will be referred to
219: as $r_t$ and $y_t$.
220:
221: Let us now calculate how the self-similarity ansatz formulated above
222: actually holds in the range of $t_s\ll t\ll T$. This
223: can be done by directly examining the scaling relation, \eqref{py},
224: rewritten as
225: %
226: \begin{equation}
227: \log[y P_t(y)]/\log(T/t)= -S(z_t(y)) ,
228: \label{scaling}
229: \end{equation}
230: %
231: where we ignored an additional constant term which corresponds to the
232: normalization of the PDF. The result is summarized in
233: \figref{fig:sz}. It shows that the curves for different coarse-graining
234: time-scales $t$ collapse. The value of $T$ is taken to be 40960~min,
235: $t=10,20,40,\ldots$ and 1280~min, and $y_T$ is set to be the mean of the
236: data for the cut-off time-scale $T$. The whole set of curves are shifted
237: vertically so that the minimums are at the origin of the vertical axis
238: (this corresponds to the additional constant ignored in
239: \eqref{scaling}). This invariant, i.e. the $t$-independent scaling
240: relation is the main result of the present work. It supports the
241: self-similarity ansatz and yields important information about the
242: functional form of $S(z)$. Note that the minimum $z_m$ for $S'(z_m)=0$
243: is approximately $-0.5$ demonstrating the diffusive nature of
244: $\avg{y_t}\sim\sqrt{t}$ in \eqref{lexp}.
245:
246: It is easily shown that if the Cram\'er function is a linear function,
247: $S(z)=a(z-z_0)$, in a certain region of $z$, the PDF for $y_t$ takes
248: the power law $P_t(y)\sim y^{-a-1}$ in the corresponding region of
249: $y$. Data analysis shows that the region $dS(z)/dz<0$ can be closely
250: approximated by using $a=-1$. This corresponds to the fact that
251: $P_t(y)$ has a constant value for small fluctuations in $y$. It should
252: be noted that this region, and that for even smaller fluctuations, is
253: subject to the fact that price value is actually discrete (not shown
254: in \figref{fig:sz}), which is of limited interest. On the other hand,
255: the region $dS(z)/dz>0$ describes large fluctuations, which is related
256: to previous work on the scaling of moments reported in
257: Refs.~\cite{mandelbrot97,arnedo98,schmitt99}. That is, the moments
258: $\avg{y_t^q}$ of order $q$ scales with $t$ as $\avg{y_t^q}\propto
259: t^{\phi(q)}$. Such scaling is related to the approach presented here.
260: Actually, the relation between $\phi(q)$ and $S(z)$ can be found if we
261: note that
262: %
263: \begin{equation}
264: \avg{y_t^q}\propto y_T^q\int_{-\infty}^\infty
265: (T/t)^{qz-S(z)}dz ,
266: \label{ytq}
267: \end{equation}
268: %
269: by using \eqref{py}. If we assume that the scaling is of the form,
270: $\avg{y_t^q}\propto t^{\phi(q)}$, and $S''(z)>0$, evaluating
271: \eqref{ytq} by using the steepest-descent method yields
272: %
273: \begin{equation}
274: \phi(q)=\min_z[S(z)-qz] .
275: \label{phi_sz}
276: \end{equation}
277: %
278: Thus $\phi(q)$ and $S(z)$ are related to each other by the Legendre
279: transformation of \eqref{phi_sz}.
280:
281: Our self-similarity formulation originates from Kolmogorov's lognormal
282: theory in the statistical theory of turbulence\cite{K62}, and has been
283: recently applied, from the statistical viewpoint of large deviation
284: theory, to analyze the shell model of turbulence \cite{watanabe00} and
285: the long-time correlation in on-off intermittency
286: \cite{fujisaka00}. Indeed the return $r_t$ for duration $t$
287: corresponds to the velocity difference across a spatial distance
288: arbitrarily chosen in the inertial subrange in turbulence, where the
289: self-similar energy cascade process is observed \cite{K62,FRISCH}. In
290: the study of on-off intermittency which is typically observed right
291: after the breakdown of synchronization of coupled chaos
292: \cite{fujisaka00}, the difference of dynamical variables of chaotic
293: elements also shows self-similar characteristics. In connection with
294: this phenomenon, $r_t$ corresponds to the difference of dynamical
295: variables which exhibits on-off intermittency. In the latter analogy,
296: we note that the corresponding variable possesses similar
297: properties as the above (i)--(iii) for the price fluctuations. The
298: present analysis is useful for these systems which show
299: self-similarity due to strongly correlated fluctuations either in
300: space or in time.
301:
302: \figref{fig:cf} compares $S(z)$ with the result numerically
303: obtained from $\phi(q)$ by the Legendre transformation in the region
304: of $dS(z)/dz>0$. Order $q$ of the moment is increased from 0 to 8,
305: near and beyond which the scaling of moments, $\avg{y_t^q}\propto
306: t^{\phi(q)}$, becomes unclear. This might be related to an ill
307: definition of higher moments caused by the heavy-tail behavior of the
308: PDF for $y_t$ \cite{plerou99}. Near the minimum of $S(z)$ ($z<-0.3$),
309: the curve obtained from $\phi(q)$ can be closely approximated by a
310: quadratic polynomial \cite{mandelbrot97}, and fits reasonably to the
311: curve of $S(z)$. However, it deviates significantly from $S(z)$ for
312: the region of larger fluctuations, where $S''(z)$ approaches zero.
313: This indicates a breakdown in the relation between $\phi(q)$ and
314: $S(z)$, which is based on the concavity of $S(z)$. Actually, if we
315: approximate $S(z)$ in the region $-0.3<z<-0.1$ by using the linear
316: relation, $S(z)=a(z-z_0)$, with $a\sim3$, the result corresponds to
317: the power-law $P_t(y)\sim y^{-4}$. This is compatible with the
318: heavy-tail behavior found in \cite{plerou99}, although more detailed
319: study is needed to clarify the functional form $S(z)$ for this region
320: and those of even larger fluctuations.
321:
322: The Cram\'er function $S(z)$ is related to the multifractal spectrum
323: investigated in \cite{mandelbrot97}. Suppose the time-series with span
324: $T$ is divided into $(T/t)$ intervals each of which has a time span
325: $t$. By evaluating the exponent $\bz(t)$, \eqref{lexp}, in each
326: interval, one obtains the number density $N_t(z)$ for the intervals
327: such that $\bz(t)$ is between $z$ and $z+dz$. The fractal
328: dimension $f(z)$ of the support for those intervals is defined by
329: %
330: \begin{equation}
331: N_t(z)\sim(t/T)^{-f(z)} .
332: \end{equation}
333: %
334: Noting that $Q_t(z)\sim N_t(z)/(T/t)$ results in the relation
335: %
336: \begin{equation}
337: S(z)=1-f(z) .
338: \end{equation}
339: %
340: Thus the multifractal spectrum $f(z)$ is essentially equivalent to the
341: present Cram\'er function $S(z)$. It should be mentioned, however,
342: that the ``dimension'' $f(z)$ can take negative values because, in
343: principle, $S(z)$ can take any positive value. We also stress
344: that the function $S(z)$ can be directly calculated from data without
345: resorting to the scaling of moments and the Legendre transformation as
346: is necessary to evaluate $f(z)$. These facts may suggest that the
347: concept of $S(z)$ is more useful than $f(z)$.
348:
349: In summary, we have found that self-similarity of volatility at
350: different time-scales in the strongly correlated regime can be
351: characterized by direct evaluation of the Cram\'er function. The
352: scaling relation \eqref{scaling} is our main result, which was a
353: consequence of the self-similarity and large deviation theory. We have
354: also shown the relations of the Cram\'er function $S(z)$ to the
355: scaling-of-moments method and to the multifractal spectrum. The
356: functional form of $S(z)$ is compatible with the power-law (``cubic
357: law'') behavior at extremely large deviation where $S(z)$ deviates
358: from quadratic form.
359:
360: Finally we would like to remind the reader that price fluctuates
361: inevitably near the critical point of the apparent equilibrium between
362: demand and supply \cite{sornette99}. The critical point is not
363: observable to any market participants whose expectations and memory
364: give rise to the interesting statistical features of price
365: fluctuations. It is indeed intriguing that the critical fluctuations
366: in market can be well characterized by methods in non-equilibrium
367: statistical physics. While we have not reached an understanding of how
368: market dynamics gives rise to the scaling and self-similarity found
369: here, we believe that our finding and analysis will provide a good
370: basis for phenomenology challenging any models used to understand
371: market dynamics.
372:
373: % acknowledgement
374: \vspace{1em}
375:
376: Y.F. would like to thank H.~Takayasu for various information in this
377: discipline, M.~Aoki for relevant concepts in economics, as well as
378: S.~Maekawa, T.~Watanabe and Y.~Nakayama for discussions. This work
379: is partially supported by Grant-in-Aid for Scientific Research
380: No.~11837009 from the Ministry of Education, Science, Sports and
381: Culture of Japan.
382:
383:
384: % --------------------------------------------------
385:
386: \begin{thebibliography}{00}
387:
388: \bibitem{SELECTAE}
389: B.~B. Mandelbrot, {\em Fractals and Scaling in Finance: Discontinuity,
390: Concentration, Risk} (Springer-Verlag, New York, 1997).
391:
392: % \bibitem{leland94}
393: % W.~E. Leland, M.~S. Taqqu, W. Willinger and D.~V. Wilson, IEEE/ACM Trans. on
394: % Networking {\bf 2}, 1 (1994).
395:
396: % \bibitem{stanley96}
397: % M.~H.~R. Stanley {\it et~al.}, Nature {\bf 379}, 804 (1996).
398:
399: % \bibitem{watts98}
400: % D.~J. Watts and S.~H. Strogatz, Nature {\bf 393}, 440 (1998).
401:
402: \bibitem{IEP99}
403: R.~N. Mantegna and H.~E. Stanley, {\em Introduction to Econophysics:
404: Correlations and Complexity in Finance} (Cambridge Univ. Press, Cambridge,
405: 1999).
406:
407: \bibitem{TFR00}
408: J.~P. Bouchaud and M. Potters, {\em Theory of Financial Risk} (Cambridge Univ.
409: Press, Cambridge, 2000).
410:
411: \bibitem{IEEE99}
412: J.~D. Farmer, Computing in Science \& Engineering {\bf 1} (1999) 26.
413:
414: \bibitem{olsen93}
415: M.~M. Dacorogna, U. A. M\"uller, R. J. Nagler, R. B. Olsen and O. V. Pictet
416: {\it et~al.}, J. Inter. Money and Finance {\bf 12} (1993) 413.
417:
418: \bibitem{liu99}
419: Y. Liu, P. Gopikrishnan, P. Cizeau, M. Meyer, C. -K. Peng and
420: H. E. Stanley, Phys. Rev. E {\bf 60} (1999) 1390.
421:
422: \bibitem{gopi00}
423: P. Gopikrishnan, V. Plerou, Y. Liu, L.A.N. Amaral, X. Gabaix and
424: H.E. Stanley, Physica A {\bf 287} (2000) 362.
425:
426: \bibitem{mandelbrot97}
427: B.~B. Mandelbrot, A. Fisher and L. Calvet, Cowles Foundation Discussion Papers
428: 1164, 1165, 1166, Yale University (1997).
429:
430: % \bibitem{brachet99}
431: % M.~E. Brachet, E. Taflin and J.~M. Tch\'eou, LANL preprint cond-mat/9905169
432: % (1999).
433:
434: \bibitem{arnedo98}
435: A. Arneodo, J.-F. Muzy, D. Sornette,
436: Eur. Phys. J. B {\bf 2} (1998) 277.
437:
438: \bibitem{schmitt99}
439: F. Schmitt, D. Schertzer and S. Lovejoy, Applied Stochastic Models and Data
440: Analysis {\bf 15} (1999) 29.
441:
442: \bibitem{FRISCH}
443: U. Frisch, {\em Turbulence: the legacy of A. N. Kolmogorov} (Cambridge Univ.
444: Press, Cambridge, 1995).
445:
446: \bibitem{footnote1}
447: Briefly, each day in NYSE is divided into bins with an interval of 5
448: minutes, for each of which an average of $y_t(s)$ is calculated for
449: the entire dataset with $s$ being in the bin. Then, $y_t(s)$ is
450: normalized by the average according to the bin of $s$.
451:
452: \bibitem{K62}
453: A.~N. Kolmogorov, J. Fluid Mech. {\bf 13} (1962) 82.
454:
455: \bibitem{watanabe00}
456: T. Watanabe, Y. Nakayama and H. Fujisaka, Phys. Rev. E {\bf 61}
457: (2000) R1024.
458:
459: \bibitem{fujisaka00}
460: H. Fujisaka, H. Suetani and T. Watanabe, Progr. Theor. Phys. Suppl.,
461: No.~139 (2000) 70.
462:
463: \bibitem{plerou99}
464: V. Plerou, P. Gopikrishnan, L. A. N. Amaral, M. Meyer and
465: H. E. Stanley, Phys. Rev. E {\bf 60} (1999) 6519.
466:
467: \bibitem{sornette99}
468: D. Sornette, D. Stauffer and H. Takayasu,
469: %LANL preprint cond-mat/9909439 (1999).
470: in workshop ``Facets of Universality: climate, biodyamics and stock
471: markets'' (Giessen University, 1999).
472:
473: % \bibitem{HNW}
474: % P. Bak, {\em How Nature Works} (Copernicus, New York, 1996).
475:
476: \end{thebibliography}
477:
478: % --------------------------------------------------
479: \clearpage
480: \thispagestyle{empty}
481:
482: \begin{figure}
483: \begin{center}
484: \includegraphics[width=14cm]{fig1.eps}
485: \end{center}
486: \caption{%
487: Time-series of price fluctuations magnitudes $y_t=|r_t(s)|$
488: (normalized by standard deviation of each after removing intraday
489: periodicity) for time-scales $t=10$, 20 and 40~min. Similar temporal
490: evolutions are visible.}
491: \label{fig:ss}
492: \end{figure}
493:
494: \clearpage
495: \thispagestyle{empty}
496:
497: \begin{figure}
498: \begin{center}
499: \includegraphics[width=14cm]{fig2.eps}
500: \end{center}
501: \caption{%
502: Variable $z$ vs. $S(z)$ (Eqs.\ (\protect\ref{zofy}) and
503: (\protect\ref{scaling})) for time-scales $t=10,20,40,\ldots$ and
504: 1280~min, when $T=40960$ min. Plots lie on similar curve, which
505: represents functional form of Cram\'er function $S(z)$. Note that
506: $S(z)$ has minimum at $z\approx -0.5$.}
507: \label{fig:sz}
508: \end{figure}
509:
510: \clearpage
511: \thispagestyle{empty}
512:
513: \begin{figure}
514: \begin{center}
515: \includegraphics[width=14cm]{fig3.eps}
516: \end{center}
517: \caption{%
518: Details of Fig.\ \protect\ref{fig:sz} in region
519: $dS(z)/dz\geq0$ for time-scales up to $t=320$ min. Compare with the
520: result (heavy line) obtained from scaling of moments (of order $0\leq
521: q\leq8$) by Legendre transformation. Broken line is
522: quadratic-polynomial fitting for solid line and an
523: extrapolation (lognormal approximation).}
524: \label{fig:cf}
525: \end{figure}
526:
527: % --------------------------------------------------
528: \end{document}
529: