1: %\documentstyle[aps,epsf,rotate,preprint]{revtex}
2: \documentstyle[aps,epsf,rotate,multicol]{revtex}
3: \begin{document}
4: \draft
5:
6: \title{Symmetry Breaking in Stock Demand}
7:
8: \author{Vasiliki Plerou, Parameswaran Gopikrishnan, and
9: H. Eugene Stanley}
10:
11: \address{Center for Polymer Studies and Department of Physics,
12: Boston University, Boston, Massachusetts 02215.\\}
13:
14:
15:
16: \date{\today}
17:
18: \maketitle
19:
20:
21: \begin{abstract}
22:
23: Scale-free distributions and correlation functions found in financial
24: data are reminiscent of the scale invariance of physical observables
25: in the vicinity of a critical point. Here, we present empirical
26: evidence for a transition phenomenon, accompanied by a symmetry
27: breaking, in the investors' demand for stocks. We study the volume
28: imbalance $\Omega$ --- difference between the number of shares traded
29: in buyer-initiated and seller-initiated trades in a time interval
30: $\Delta t$ --- conditioned on $\Sigma$ which is defined as the local
31: first moment of $\Omega$ in $\Delta t$. We find that the conditional
32: distribution $P(\Omega \vert \Sigma)$ undergoes a qualitative change
33: in behavior as $\Sigma$ increases beyond a critical threshold
34: $\Sigma_c$. For $\Sigma <\Sigma_c$, $P(\Omega\vert\Sigma)$ displays a
35: maximum at $\Omega=0$, i.e., trades in $\Delta t$ are equally likely
36: to be buyer initiated or seller initiated. For $\Sigma > \Sigma_c$,
37: $\Omega=0$ becomes a local minimum and two new maxima $\Omega_{+}$ and
38: $\Omega_{-}$ appear at non-zero values of $\Omega$, i.e., trades in
39: $\Delta t$ are either predominantly buyer initiated or predominantly
40: seller initiated. We interpret these results using a Langevin equation
41: with multiplicative noise.
42:
43:
44:
45: \end{abstract}
46: \pacs{PACS numbers: 05.45.Tp, 89.90.+n, 05.40.-a, 05.40.Fb}
47: \begin{multicols}{2}
48:
49: Phase transitions of systems from an `ordered' phase into a
50: `disordered' phase are closely linked to symmetry breaking. For
51: example, in an Ising ferromagnet above its critical temperature, the
52: most probable state (zero net magnetization) possesses the symmetry
53: that leaves it invariant under the flipping of each spin. For
54: temperatures below the critical value, this symmetry is broken, and
55: there is a preferred direction.
56:
57: Here, we present evidence for an analogous transition phenomenon in a
58: financial
59: context~\cite{Takayasu02,Takayasu99,Bouchaud01,Zhang,Mantegna99,Bouchaud00,Farmer99}.
60: Specifically, we study the statistical properties of investors'
61: demand for stocks --- quantified as the imbalance in the number of
62: shares transacted by buyers and sellers over a time interval $\Delta
63: t$. We analyze the probability distribution of demand, conditioned on
64: its local ``noise'' intensity (a variance-like parameter $\Sigma$
65: defined below). We find that for intensities smaller than a critical
66: value $\Sigma_c$, the most probable value of demand is approximately
67: zero --- neither buying nor selling behavior dominates. For
68: intensities larger than that critical value ($\Sigma > \Sigma_c$), two
69: most probable values emerge that are symmetric around zero demand,
70: corresponding to two distinct ``phases'' --- excess demand and excess
71: supply~\cite{Takayasu99}. Under such conditions, the market behavior
72: is either mainly-buying or mainly-selling, spending almost equal
73: amount of time in each state. In other words, exchanging every
74: ``buy'' with a ``sell'' gives the same state below the critical noise
75: intensity, whereas above this threshold, the symmetry of this exchange
76: is broken.
77:
78: In classic critical phenomena, the qualitative change in behavior
79: accompanying a phase transition can be formalized in terms of the
80: extrema of a phenomenological potential, or equivalently in terms of
81: the extrema of the corresponding probability
82: distributions~\cite{Halperin77,Stanley71}. We first follow the latter
83: approach and study the behavior of the probability distribution of
84: demand. Using transactions and quotes data for the 116 most-actively
85: traded stocks~\cite{TAQrec}, we quantify the demand for a stock by
86: calculating the ``volume imbalance'' over a time interval $\Delta t$,
87: defined to be the difference between $Q_{\rm B}$, the number of shares
88: traded in buyer-initiated trades, and $Q_{\rm S}$, the number of shares
89: traded in seller-initiated trades in $\Delta
90: t$~\cite{CLM,LeeReady91,LeeReady91note,Plerou01,Hasbrouck88,Farmer98},
91: \begin{equation}
92: \Omega (t) \equiv Q_{\rm B} - Q_{\rm S} = \sum_{i=1}^N q_i a_i\,.
93: \label{defOmega}
94: \end{equation}
95: Here, the indicator $a_i=1$ for buyer-initiated trades (buy trades)
96: and $a_i=-1$ for seller-initiated trades (sell trades)~\cite{buysell},
97: $q_i$ is the number of shares traded in transaction $i$, and $N \equiv
98: N_{\Delta t}$ denotes the number of trades in $\Delta
99: t$~\cite{TAQrec,blockTrades}.
100:
101: Our analysis of the (unconditional) probability distribution
102: $P(\Omega)$ for each stock shows a single peak around
103: $\Omega=0$. Since previous work shows that the distribution of $q_i$
104: has divergent variance~\cite{Gopi00a,notedistQprime}, we quantify the
105: noise intensity by computing the `local
106: deviation'~\cite{controlparameter}, defined as the centered first
107: moment,
108: \begin{equation}
109: \Sigma (t) \equiv \langle \vert q_i a_i- \langle q_i a_i \rangle \vert \rangle \,,
110: \label{defsigma}
111: \end{equation}
112: where $\langle\dots\rangle$ denotes `local' expectation values
113: computed from all trades in the time interval $\Delta t$. Next, we
114: examine the behavior of the conditional distribution $P (\Omega \vert
115: \Sigma)$ of $\Omega$ for a given value of the local deviation $\Sigma$
116: for $\Delta t=15$ min, [Fig.~1(a)]. For small $\Sigma$, we find that
117: $P(\Omega \vert \Sigma)$ is {\it single peaked} displaying a maximum
118: at $\Omega=0$. When $\Sigma$ exceeds a critical threshold $\Sigma_c$,
119: the behavior of $P(\Omega \vert \Sigma)$ undergoes a qualitative
120: change, and is {\it double peaked} with two new maxima appearing at
121: non-zero values, $\Omega_{+}$ and $\Omega_{-}$, symmetric around
122: zero. Figure~1(a) also shows that the separation between the two
123: maxima increases with $\Sigma$.
124:
125: This qualitative change in the behavior of $P(\Omega)$ implies that
126: for $\Sigma < \Sigma_c$, the most-probable value of demand is
127: approximately zero, and possesses the symmetry that leaves the most
128: probable value invariant under the operation $\Omega \rightarrow
129: -\Omega$, or at the microscopic (trade) level, under the operation
130: ${\rm B}\rightarrow {\rm S}$ of changing every buyer-initiated
131: trade~B, to a seller-initiated trade~S. For $\Sigma > \Sigma_c$, the
132: two most probable values $\Omega_{\pm}$ are non-zero, and the $\Omega
133: \rightarrow -\Omega$ (${\rm B}\rightarrow {\rm S}$) symmetry is
134: broken. In other words, while for $\Sigma < \Sigma_c$ buy and sell
135: trades are equally probable in each time interval (zero demand), for
136: $\Sigma > \Sigma_c$, trades in each time interval are either mostly
137: buy trades (excess demand) or mostly sell trades (excess supply)
138: giving rise to non-zero values of $\Omega_{\pm}$. Identical results
139: can be obtained by conditioning $P(\Omega)$ on the total trade volume
140: in $\Delta t$, $Q(t)\equiv Q_{\rm B} + Q_{\rm S}$.
141:
142: Our finding is analogous to phase transition phenomena in physical
143: systems, where the behavior of the system undergoes a qualitative
144: change at a critical threshold of a control parameter
145: $T$. In such systems, the change in behavior
146: can be quantified by an order parameter $\Psi$ which is identically
147: zero for values of $T$ below (or above as the case may be) a certain
148: critical value $T_{\rm c}$, and becomes nonzero as $T$ crosses $T_{\rm
149: c}$. In our problem, the ``order parameter'' $\Psi$, can be identified
150: by the location of the maxima $\Omega_{\pm}$ of $P
151: (\Omega)$. Figure~1(c) shows that the change in $\Psi$ as a function
152: of $\Sigma$ is described by
153: \begin{equation}
154: \Psi (\Sigma) = \cases{ 0 & $[\Sigma < \Sigma_c]$ \cr
155: \vert \Sigma - \Sigma_c \vert^{\beta} & $[\Sigma > \Sigma_c]$},
156: \end{equation}
157: with $\beta \approx 1$~\cite{notebetahalf}.
158:
159: In the mean-field theory of critical phenomena (Landau-Ginzburg
160: theory), the qualitative change in behavior of the system is
161: attributed to the changes in symmetry of the underlying
162: potential~\cite{Halperin77,Stanley71}. In the following, we pursue an
163: analogous approach to understand our empirical results. Since the
164: transition behavior that we find occurs with change in noise
165: intensity, we follow an approach similar to those used to understand
166: non-equilibrium phase transition
167: phenomena~\cite{Horsthemke,Haken75,Haken78}. We start with expressing
168: the dynamics of $\Omega$ through a {\it deterministic} differential
169: equation,
170: \begin{mathletters}
171: \begin{equation}
172: d\Omega = h_{\lambda}(\Omega)\, dt\,,
173: \label{determinomega}
174: \end{equation}
175: where $\lambda$ is a parameter quantifying the coupling of the system
176: to its environment. Letting $\lambda$ fluctuate randomly with noise
177: intensity $\sigma$, Eq.(\ref{determinomega}) becomes, in general, a
178: {\it stochastic} differential equation with multiplicative
179: noise. Since the form of $h_{\lambda}(\Omega)$ is not known, and since
180: it is not {\it a priori} clear if $\lambda$ is an observable, we
181: describe the dynamics of $\Omega$ through~\cite{Horsthemke}
182: \begin{equation}
183: d\Omega = u (\Omega) dt + \sigma\, v(\Omega) dW_t\,,
184: \label{defdomega}
185: \end{equation}
186: \end{mathletters}
187: where $u\equiv u(\Omega)$ is the drift term, $v\equiv v(\Omega)$
188: reflects the effects of multiplicative noise, $dW_t$ is the standard
189: Wiener differential satisfying $\langle dW_t\,dW_{t^{\prime}} \rangle
190: =\delta (t-t^{\prime}) dt$, and the parameter $\sigma$ quantifies the
191: intensity of the noise term~\cite{lambda,fractional}. The functions
192: $u$ and $v$ may estimated from the data as a ``local'' mean and
193: standard deviation of $d\Omega$~\cite{notediff}. The Fokker-Planck
194: equation corresponding to Eq.~(\ref{defdomega}) which describes the
195: evolution of the transition probability density $\Pi
196: \equiv P(\Omega, t \vert \Omega (t=0))$ is
197: \begin{equation}
198: {\partial \Pi \over \partial t} = -{\partial \over \partial \Omega} [u
199: \Pi] + {\sigma^2 \over 2} {\partial^2 \over \partial \Omega^2} [v^2\Pi]\,.
200: \label{fpe}
201: \end{equation}
202:
203: The stationary probability density $P(\Omega)$ from Eq.~(\ref{fpe}) is
204: \begin{equation}
205: P (\Omega) = {A \over v^2} \exp \left( {2\over \sigma^2}
206: \int^{\Omega} {u \over v^2} dx \right)\,,
207: \label{fpesol}
208: \end{equation}
209: where $A$ is the normalization constant, and $P$ is assumed
210: to be normalizable. Equation~(\ref{fpesol}) can be rewritten as
211: \begin{equation}
212: P (\Omega) = A\, \exp\left(-{V(\Omega)\over\sigma^2}\right)\,,
213: \label{fpesolpot}
214: \end{equation}
215: where the function $V (\Omega)$ takes the meaning of a
216: effective ``stochastic'' potential given by~\cite{Horsthemke}
217: \begin{equation}
218: V (\Omega) \equiv -\left[\int^{\Omega} {u \over v^2} dx
219: - \sigma^2 \ln v
220: \right] \,.
221: \label{defpot}
222: \end{equation}
223: Thus, the extrema of the probability density can be translated into
224: the extrema of the underlying stochastic potential.
225:
226: As the noise intensity of the environment ($\sigma$) changes, the
227: potential could change shape, acquire new minima, and consequently the
228: system could undergo drastic changes in behavior~\cite{noteCONTB}. Can
229: the transition behavior we find empirically for $P(\Omega \vert
230: \Sigma)$ be understood in this framework? To address this question, we
231: must examine the shape of the potential $V (\Omega)$ for
232: different values of $\Sigma$, which monotonically depends on
233: $\sigma$~\cite{noteSigma-sigma}.
234:
235: In order to study $V(\Omega)$ for different $\sigma$ empirically, we
236: first extract the functions $u(\Omega)$ and $v(\Omega)$ and analyze
237: their behavior for different values of $\Sigma$. From
238: Eq.~(\ref{defdomega}), $u$ is the drift term, which for a given
239: $\Sigma$, can be estimated by computing the equal-time expectation
240: value of the change $\Delta
241: \Omega \equiv \Omega (t+\Delta t) - \Omega(t)$ for a given $\Omega$,
242: \begin{equation}
243: u \approx \langle \Delta \Omega \rangle_{(\Omega,\Sigma)}\,.
244: \label{Fest}
245: \end{equation}
246: Similarly, the product $\sigma v$ can be estimated from the ``local''
247: deviation
248: \begin{equation}
249: \sigma\, v \approx \langle \vert \Delta \Omega - \langle \Delta \Omega
250: \rangle_{(\Omega,\Sigma)} \vert \rangle_{(\Omega, \Sigma)}\,.
251: \label{Gest}
252: \end{equation}
253:
254: Figures~\ref{FG}(a) and (b) show $u (\Omega)$ and $v (\Omega)$ for
255: three different values of $\Sigma$. Clearly, the functional form
256: $u(\Omega)$ does not vary with $\Sigma$ and is consistent with a
257: linear behavior for all $\Sigma$. Figure~\ref{FG}(b) shows that, for
258: small $\Sigma$, $\sigma\,v (\Omega)$ is approximately flat, whereas
259: for large $\Sigma$, $\sigma\,v$ acquires a marked `peak' around
260: $\Omega=0$. Except for the smallest $\Sigma$, the functional forms
261: $\sigma v (\Omega)$ for different $\Sigma$ seem to be consistent
262: within a multiplicative factor (related to
263: $\sigma$)~\cite{note-additivenoise}.
264:
265: Next, we shall analyze the extrema of the stochastic potential $V
266: (\Omega)$ for different values of $\Sigma$. From
267: Eqs.~(\ref{fpesolpot})~and~(\ref{defpot}), the extrema of $P(\Omega)$
268: correspond to the roots of the function
269: \begin{equation}
270: F (\Omega) \equiv {dV(\Omega) \over d\Omega} = u -
271: \sigma^2 v {dv \over d \Omega}.
272: \label{defFmaxmin}
273: \end{equation}
274: Figure~\ref{min}(a) shows $ F (\Omega)$ for three different
275: $\Sigma$. For $\Sigma< \Sigma_c$, the function $F (\Omega)$
276: displays only one root at $\Omega=0$. Near $\Sigma = \Sigma_c$, $ F
277: (\Omega)$ displays an inflexion point at $\Omega=0$. For $\Sigma >
278: \Sigma_c$, the one root branches into three: $\Omega=0$, $\Omega_{+}$,
279: and $\Omega_{-}$.
280:
281: Integrating the stochastic ``force'' $ F (\Omega)$, we next find the
282: potential $V (\Omega)$. Figure~\ref{potential} shows the behavior of
283: $V(\Omega)$ with different $\Sigma$. For $\Sigma<
284: \Sigma_c$, the potential has only one minimum at $\Omega=0$, which is
285: consistent with one maximum for $P(\Omega)$ which we find. For $\Sigma
286: \approx \Sigma_c$, the potential appears almost flat: the existing
287: minimum begins to change into a (unstable) local maximum and displays
288: an inflexion point at $\Sigma=0$. For $\Sigma > \Sigma_c$, the
289: potential displays two clear minima at $\Omega_{\pm}$ and a local
290: maximum at $\Omega=0$, consistent with the bimodal nature of the
291: distribution $P(\Omega)$ for $\Sigma > \Sigma_c$.
292:
293: In summary, we investigate the dynamics of the demand $\Omega$ by
294: examining the distribution of volume imbalance $\Omega$ for changing
295: market conditions quantified by the local deviation $\Sigma$. We find
296: that the distribution $P(\Omega)$ is single peaked for $\Sigma$
297: smaller than a critical threshold $\Sigma_c$. For $\Sigma$ larger than
298: $\Sigma_c$, the distribution $P(\Omega)$ changes to a double-peaked
299: distribution. The analog of the order parameter $\Psi$ which describes
300: the above qualitative change in behavior is zero for $\Sigma <
301: \Sigma_c$ and behaves as $\sim \vert \Sigma - \Sigma_c
302: \vert^{\beta}$~for $\Sigma > \Sigma_c$, where $\beta \approx 1$.
303: We have also seen that the dynamics of demand can be understood in
304: terms of a `stochastic' potential which changes its behavior with
305: $\Sigma$. As $\Sigma$ crosses the critical threshold $\Sigma_c$, the
306: system undergoes a transition from a `disordered' state with most
307: probable demand equal to zero to an `ordered' state with two phases,
308: excess demand and excess supply.
309:
310:
311: We thank the National Science Foundation for support.
312:
313: \begin{references}
314:
315: \bibitem{Takayasu02}
316: H.~Takayasu, ed., {\em Empirical Science of Financial Fluctuations}.
317: \newblock Springer, Berlin, 2002.
318:
319: \bibitem{Takayasu99}
320: H.~Takayasu and M.~Takayasu {\em Physica A} {\bf 269}, 24 (1999).
321:
322: \bibitem{Bouchaud01}
323: J.-P. Bouchaud, {\em Quantitative Finance} {\bf 1} 105 (2001).
324:
325: \bibitem{Zhang}
326: Y.~C. Zhang, ``Towards a Theory of Marginally Efficient Markets,'' {\em Physica
327: A} {\bf 269} 30 (1999)
328:
329: \bibitem{Mantegna99}
330: R.~N. Mantegna and H.~E. Stanley, {\em An Introduction to
331: Econophysics: Correlations and Complexity in Finance} ( Cambridge
332: University Press, Cambridge, 2000).
333:
334: \bibitem{Bouchaud00}
335: J.~P. Bouchaud and M.~Potters, {\em Theory of Financial Risk}
336: (Cambridge University Press, Cambridge, 2000).
337:
338: \bibitem{Farmer99}
339: J.~D. Farmer, {\em Computing in Science \& Engineering} {\bf 1}, 26--39 (1999).
340:
341: \bibitem{Halperin77}
342: B.~Halperin and P.~Hohenberg {\em Rev. Mod. Phys} {\bf 49}, 435 (1977).
343:
344: \bibitem{Stanley71}
345: H.~E. Stanley, {\em Introduction to Phase Transitions and Critical Phenomena}
346: (Oxford University Press, Oxford, 1971).
347:
348: \bibitem{TAQrec}
349: For records of trade times, price, and number of shares traded, and the bid-ask
350: quotes, we analyze the data for the 116 most-frequently traded US stocks from
351: the {\it Trades and Quotes database} (New York Stock Exchange, 1994-95) for
352: the 2~yr period 1994-95.
353:
354: \bibitem{CLM}
355: J.~Campbell, A.~W. Lo, and A.~MacKinlay, {\em The Econometrics of Financial
356: Markets} (Princeton University Press, 1997).
357:
358: \bibitem{LeeReady91}
359: C.~M. Lee and M.~J. Ready {\em Journal of Finance} {\bf 46}, 733 (1991).
360:
361: \bibitem{LeeReady91note}
362: Following the procedure of C.~M.~Lee and M.~J.~Ready~\protect\cite{LeeReady91},
363: we use the prevailing quote at least 5~s prior to the trade.
364:
365: \bibitem{Plerou01}
366: V.~Plerou {\it et al}, ``Quantifying Stock
367: Price Response to Demand Fluctuations,'' e-print: cond-mat/0106657.
368:
369: \bibitem{Hasbrouck88}
370: J.~Hasbrouck {\em J. Fin. Econ.} {\bf 22} (1988) 229.
371:
372: \bibitem{Farmer98}
373: J.~D. Farmer, ``Market Force, Ecology and Evolution,'' e-print: adap-org/9812005.
374:
375: \bibitem{buysell}
376: We distinguish {\it buyer-initiated} and {\it seller-initiated} trades defined
377: by which of the two participants in the trade, the buyer or the seller, is
378: more eager to trade. When such a distinction does not exist, we label the
379: trade as {\it indeterminate}. We identify buyer and seller initiated trades
380: using the bid and ask quotes $S_{\rm B} (t)$ and $S_{\rm A} (t)$ at which a
381: market maker is willing to buy or sell respectively~\protect\cite{TAQrec}.
382: Using the mid-value $S_{\rm M} (t)= (S_{\rm A} (t)+ S_{\rm B}(t))/2$ of the
383: prevailing quote~\protect\cite{CLM,Hasbrouck88,LeeReady91}, we label a trade
384: buyer initiated if $S(t)> S_{\rm M}(t)$, and seller initiated if $S(t) <
385: S_{\rm M}(t)$. For trades occurring exactly at $S_{\rm M}(t)$, we use the
386: sign of the change in price from the previous trade to determine whether the
387: trade is buyer or seller initiated, while if the previous trade is at the
388: current trade price, the trade is labelled
389: indeterminate~\protect\cite{LeeReady91}. In our case, an average of $\approx
390: 17\%$ of the trades remain indeterminate.
391:
392: \bibitem{blockTrades}
393: We disregard the first trade of each day as it is most often a block trade.
394:
395: \bibitem{Gopi00a}
396: P.~Gopikrishnan {\it et al}, {\em Phys. Rev. E} {\bf 62,} 4493 (2000).
397:
398: \bibitem{notedistQprime}
399: In Ref.~\protect\cite{Gopi00a} it is shown that $q_i$ have a power-law
400: distribution with exponent $\zeta\approx 3/2$ which corresponds to a
401: divergent second moment.
402:
403: \bibitem{controlparameter}
404: We note that the results presented below are robust to using the `local
405: deviation' of $\Omega$ ($=\Sigma N$) instead of $\Sigma$.
406:
407:
408: \bibitem{notebetahalf}
409: Interestingly, using the variance instead of the local deviation, one obtains
410: the mean-field theory value for the exponent, $\beta=1/2$.
411:
412: \bibitem{Horsthemke}
413: W.~Horsthemke and R.~Lefever, {\em Noise Induced Phase Transitions}.
414: \newblock Springer Verlag, Berlin, 1984.
415:
416: \bibitem{Haken75}
417: H.~Haken, {\em Rev. Mod. Phys.} {\bf 47}, 67 (1975).
418:
419: \bibitem{Haken78}
420: H.~Haken, {\em Synergetics: An Introduction: Nonequilibrium phase transitions
421: and self-organization in physics, chemistry, and biology} (Springer Verlag,
422: Berlin, 1978).
423:
424: \bibitem{lambda}
425: In the literature, the diffusion term in Eq.~(\protect\ref{defdomega}) is
426: attributed to fluctuations of a parameter $\lambda$ coupled to the
427: environment, and $\sigma$ would then correspond to the intensity of
428: environmental fluctuations. It is yet unclear what interpretation to ascribe
429: to $\lambda$ (qualitatively some measure of news coupled with price changes),
430: and the question of whether it is an observable remains open. Therefore, we
431: estimate the drift and diffusion term and examine their behavior as $\Sigma$
432: varies.
433:
434: \bibitem{fractional}
435: Since the number of trades $N$ is distributed as a
436: power-law~\protect\cite{Plerou00}, or equivalently the time interval between
437: trades is distributed as a power-law (with an exponential truncation), the
438: stochastic differential equations should be more precisely
439: fractional~\protect\cite{Halperin77,Mainardi00}.
440:
441: \bibitem{notediff}
442: Our goal is to understand the change in behavior of
443: $P_{\Omega}(\Omega)$ with $\sigma$. Our results hold under the
444: assumption that $\sigma$ changes much slower than the scale at which
445: $\Omega$ changes, so that a stationary $P(\Omega)$ is reached.
446:
447:
448: \bibitem{noteCONTB}
449: For a related discussion in the context of price fluctuations and market
450: crashes see J.-P.~Bouchaud and R.~Cont, {\it Eur. Phys. J.} B {\bf 6}, 543
451: (1998).
452:
453: \bibitem{noteSigma-sigma}
454: In our problem, $\sigma$ is not known; however, $\Sigma$ is known, and it
455: monotonically depends on $\sigma$. Thus, instead of studying $V
456: (\Omega)$ for different $\sigma$, we examine $V(\Omega)$ for different
457: $\Sigma$. We note that although $\Sigma$ depends monotonically on $\sigma$,
458: the converse need not be true. However, it is convenient to assume that
459: changes of $\Sigma$ largely reflect changes in $\sigma$.
460:
461: \bibitem{note-additivenoise}
462: Although, we consider only multiplicative noise in our description, in general
463: there could be additive noise present as well.
464:
465: \bibitem{Plerou00}
466: V.~Plerou {\it et al}, {\em Phys. Rev. E} {\bf 62}, R3023 (2000).
467:
468:
469: \bibitem{Mainardi00}
470: F.~Mainardi, M.~Raberto, R.~Gorenflo, and E.~Scalas, {\em Physica A}
471: {\bf 287}, 468 (2000).
472:
473: \bibitem{Holy97}
474: T.~E. Holy, {\em Phys. Rev. Lett.} {\bf 79}, 3545 (1997).
475:
476: \end{references}
477:
478:
479:
480: \begin{figure}
481: \narrowtext
482: \centerline{
483: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{qpCdevqp15-116-dens.eps}}}
484: }
485: \centerline{
486: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{qpCdevqp15-116-kdens_1.eps}}}
487: }
488: \centerline{
489: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{qpCdevqp15-116-kmax.eps}}}
490: }
491: \caption{(a) Conditional density $P(\Omega \vert \Sigma)$ for
492: varying $\Sigma$. The distribution changes from a single-peaked
493: distribution for small $\Sigma$ (solid line), to a double-peaked
494: distribution for large $\Sigma$ (dashed line). (b) Same as (a) in
495: semi-logarithmic scale. (c) Order parameter, $\Psi$ (position of the
496: maxima of the distribution $P(\Omega \vert \Sigma)$), as a function of
497: $\Sigma$. For small $\Sigma$, $P(\Omega \vert
498: \Sigma)$ displays a single maximum whereas for large $\Sigma$ two
499: maxima appear. Both $\Sigma$ and $\Omega$ have been normalized to zero
500: mean and unit first moment. For a more accurate estimation of the
501: location of the extrema, all densities were computed using the density
502: estimator of~\protect\cite{Holy97}. }
503: \label{qp_dens}
504: \end{figure}
505:
506: \begin{figure}
507: \narrowtext
508: \centerline{
509: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{FG.eps}}}
510: }
511: \caption{(a) The drift part $u(\Omega)$ and (b) the variance part
512: $\sigma\,v(\Omega)$ of Eq~(\protect\ref{defOmega}), estimated from the
513: data as the local mean and local deviation of $\Delta\Omega$. The
514: curves in~(b) have been shifted vertically for clarity.}
515: \label{FG}
516: \end{figure}
517:
518: \begin{figure}
519: \narrowtext
520: \centerline{
521: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{f-ggp-group1-5-11.eps}}}
522: }
523: \caption{Stochastic ``force'' $F(\Omega)$ for three
524: different values of $\Sigma$ calculated from
525: Eq.~(\protect\ref{defFmaxmin}). The derivative $\sigma\,dv/d\Omega$
526: is calculated by first fitting the function $\sigma\,v$ by a third
527: order polynomial. For $\Sigma<
528: \Sigma_c$, the function $ F (\Omega)$ displays only one
529: zero at $\Omega=0$. For $\Sigma \approx \Sigma_c$, the existing root
530: starts to branch into three roots, and for $\Sigma > \Sigma_c$, we
531: find three roots: $\Omega=0$ and $\Omega=\Omega_{\pm}$.}
532: \label{min}
533: \end{figure}
534:
535: \begin{figure}
536: \narrowtext
537: \centerline{
538: \epsfysize=0.7\columnwidth{\rotate[r]{\epsfbox{potential-group1-5-11.eps}}}
539: }
540: \caption{Stochastic potential $V(\Omega)$ for
541: $\Sigma < \Sigma_c$ shows one minimum at $\Omega=0$, whereas for
542: $\Sigma > \Sigma_c$ the potential has a maximum at $\Omega=0$ and two
543: new minima $\Omega_{\pm}$ appear. We compute $V_{\Sigma}(\Omega)$ by
544: integrating $F (\Omega)$ [Fig.~\protect\ref{min}]. The curves have
545: been shifted vertically for clarity.}
546: \label{potential}
547: \end{figure}
548: \end{multicols}
549:
550:
551: \end{document}
552:
553:
554:
555:
556: