1: \documentclass[12pt]{article}
2: \usepackage[a4paper]{geometry}
3: \usepackage{amsmath}
4: \usepackage{amssymb}
5: \usepackage[dvips]{graphicx}
6: \usepackage{psfig}
7:
8: \textwidth 147mm
9: %\textheight 245mm
10: \evensidemargin 10mm
11: \oddsidemargin 10mm
12: \renewcommand{\baselinestretch}{1.0}
13: \setlength{\textfloatsep} {-2mm}
14:
15: \begin{document}
16:
17: \noindent
18: \begin{center}
19: {\large \bf The High Temperature Dynamics of a mean field Potts glass}\\[2mm]
20: Claudio Brangian$^*$, Walter Kob$^\ddagger$\footnote{Author to whom
21: correspondence should be addressed}, and Kurt Binder$^*$\\
22: $^*$ Institute of Physics, Johannes Gutenberg University, Staudinger Weg 7,\\
23: D-55099 Mainz, Germany\\
24: $\ddagger$ Laboratoire des Verres, Universit\'e Montpellier II, F-34095
25: Montpellier, France\\
26:
27: \end{center}
28:
29: \vspace*{7mm}
30: \par
31: \noindent
32: \begin{center}
33: \begin{minipage}[h]{122mm}
34: We use Monte Carlo simulations to investigate the dynamical properties
35: of the infinite range 10 states Potts glass. By analyzing the spin
36: autocorrelation function for system sizes up to $N=2560$, we show that
37: strong finite size effects are present around the predicted dynamical
38: transition temperature. The autocorrelation function shows strong
39: self-averaging at high temperatures, whereas close to the dynamical
40: transition they show the lack of self-averaging.
41:
42: \end{minipage}
43: \end{center}
44:
45: \vspace*{5mm}
46: \par
47: \noindent
48:
49: \section{Introduction}
50: In recent years a new class of disordered spin glass models has been
51: introduced (for a review see Kirkpatrick and Thirumalai 1995) that show
52: strong analogies with the theoretical scenario proposed for the structural
53: glass transition, such as the presence of a dynamical transition at
54: a temperature $T_D$ and a static phase transition (with discontinuous
55: order parameter but without latent heat) at $T_0 < T_D$. Furthermore, the
56: equations of motion for the spin autocorrelation functions are formally
57: analogous to the equations of motion of the density-autocorrelation
58: functions introduced by the mode coupling theory of the glass transition
59: (G\"otze 1989). One example of such spin models is the $p$ states mean
60: field Potts glass with $p>4$ (Kirkpatrick and Wolynes 1987, Kirkpatrick
61: and Thirumalai 1988). The goal of the present paper is to compare the
62: relaxation dynamics of such a system with a {\it finite} size with
63: the dynamics of the system in the thermodynamic limit. The latter has
64: previously been determined at the level of one step replica symmetry
65: breaking (De Santis {\it et al.} 1995). Furthermore we investigate to
66: what extend the time correlation functions are self-averaging or not.
67:
68:
69: \section{Model and Simulations}
70: In the Potts model each spin $\sigma_i$ is a discrete variable that
71: can take one of $p$ different values: $\sigma_i \in \{1,\ldots,p\}$. The
72: Hamiltonian is given by
73: %
74: \begin{equation}
75: H=-\sum_{i,j}J_{ij}(p\delta_{\sigma_i \sigma_j}-1) ,
76: \label{eq1}
77: \end{equation}
78: %
79: i.e. each spin interacts with all the others. The coupling constants $J_{ij}$
80: are taken from a Gaussian distribution
81: %
82: \begin{equation}
83: P(J_{ij})=\frac{1}{\sqrt{2\pi }(\Delta J)} \exp \left[- \frac{
84: (J_{ij}-J_0)^2}{2(\Delta J)^2}\right] .
85: \label{eq2}
86: \end{equation}
87: %
88: We consider the case $p=10$ since for this case the static transition
89: has a strong first order character (jump of the order parameter from zero
90: to $q_0=0.452$) (De Santis {\it et al.} 1995). Note that for $p=2$ we
91: recover the Sherrington-Kirkpatrick model. It has been shown that in
92: order to prevent the system from ordering ferromagnetically, a negative
93: value of $J_0$ has to be chosen (Elderfield and Sherrington 1983,
94: Gross \textit{et al.} 1985, Cwilich and Kirkpatrick 1989) and therefore
95: we have chosen $J_0=(3-p)/(N-1)$. The variance is set to $\Delta J =
96: (N-1)^{-1/2}$ and in the following we will set the Boltzmann's constant
97: $k_b=1$. With these units a numerical solution of the replica equations
98: at the level of one step replica symmetry breaking predicts a dynamical
99: transition at $T_D=1.142$ and a static transition from a paramagnet to
100: a spin glass at $T_0=1.131$ (De Santis \textit{et al}. 1995).
101:
102: We have simulated 5 different system sizes, $N=160,$ 320, 640, 1280
103: and 2560 spins, at various temperatures. In this paper we will focus on
104: temperatures between $T=1.8$ to $T_D=1.142$, i.e. the range above the
105: dynamical transition temperature in the thermodynamic limit. Due to the
106: random nature of the interactions, we have to average all observables
107: not only over the canonical distribution but also over the possible
108: realizations of the disorder given by Eq.~(\ref{eq2}). In the following
109: we will denote this latter average by $\left [ \cdot \right]$. For
110: this we used $500$ different samples for $160$ spins, $100$ for $320,
111: 640 \;\textrm{and}\; 1280$ spins and between $20 \;\textrm{and}\;
112: 50$ for $2560$ spins. The dynamics is generated using the Metropolis
113: algorithm. Starting from a given spin configuration, a spin is picked at
114: random and assigned a new random orientation. If the energy difference
115: between these two states is negative the move is accepted. If it is
116: positive it is accepted only with probability $\exp\left (-\Delta E/T
117: \right)$.
118:
119:
120: \section{Results}
121: We present now our results regarding the analysis of the spin-spin
122: autocorrelation function, defined as
123:
124: \begin{equation}
125: \label{ssaf}
126: C(t)= \frac{p}{p-1} \frac{1}{N}\sum_{i}^{N} \left[ \left\langle
127: \left( {\delta_{\sigma_i(t') \sigma_i(t'+t)}-1/p} \right) \right\rangle .
128: \right]
129: \end{equation}
130:
131: The mean field theory predicts that in the thermodynamic limit the
132: dynamics of the system slows down upon approach of the (dynamical)
133: transition temperature $T_D$. The relaxation time $\tau(T)$ for the time
134: correlation function should show at $T_D$ a divergence of the form $\tau
135: \propto (T-T_D)^{-\Delta}$. At the same time the spin autocorrelation
136: function is predicted to show at intermediate times a plateau with a
137: height $q_{EA}=0.328$ before it decays on the time scale $\tau$ towards
138: zero. At $T=T_D$ the system becomes nonergodic in that the correlation
139: function does not show anymore the final decay, i.e. it stays at
140: the plateau even for infinite times. (For a review of this behavior
141: on a class of spin glass systems see Thirumalai and Kirkpatrick 1995.
142: For the works regarding the Potts glass, see Kirkpatrick and Wolynes,
143: 1987 and Kirkpatrick and Thirumalai, 1988).
144:
145: However, for a {\it finite} system we do no longer expect the sharp
146: ergodic to nonergodic transition at $T_D$, since for $N< \infty$ the
147: relaxation times have to remain finite for all $T>0$. It is therefore
148: of interest to see how the typical relaxation behavior of the system
149: changes if $N$ is increased.
150:
151: In Fig. \ref{fig1} we show the temperature dependence of $C(t)$ for
152: $1280$ spins. From the figure we see that with decreasing temperature
153: the dynamics does indeed slow down. However, even at $T_0<T_D$ we do
154: not see a pronounced sign for the existence of the dynamic transition,
155: since the curves show, instead of the expected plateau at $q_{EA}$ (shown
156: in the figure as solid horizontal line), only a weak shoulder. Thus we
157: conclude that the dynamics of this system is strongly affected by finite
158: size effects even for systems as large as 1280 spins. The reason for this
159: is that the barriers in the free energy that separate one ``basin''
160: of configuration space from a neighboring ``basin'' are apparently
161: not very high, in contrast to the thermodynamics limit in which their
162: height diverges. Note that this $N$ dependence of the dynamics is in
163: marked contrast with the one found in structural glasses since in these
164: systems finite size effects are usually absent if the system has more
165: than a few hundred particles (Kob 1999).
166:
167: To study in more detail how the relaxation dynamics depends on the size
168: of the system we show in Fig.~\ref{fig2} the time correlation function
169: for different system sizes for a high temperature, and at $T=T_D$. At the
170: high temperature the different $C(t)$ fall nicely onto a master curve,
171: i.e. there are no finite size effects. This is in stark contrast to
172: the behavior at $T_D$. At this temperature we find that the relaxation
173: dynamics for the different system sizes depends strongly on $N$ since
174: the large systems relax much slower than the small ones. If one defines a
175: relaxation time $\tau(T)$ via $C(\tau)=0.2$, see horizontal dashed line
176: in the figure, one finds that, at $T_D$, these times show a power-law
177: dependence: $\tau\propto N^{1.5}$ (Brangian \textit{et al.} 2001). Note
178: that the value of 1.5 for the exponent is significantly larger than the
179: estimate 2/3 for the Sherrington-Kirkpatrick model (Bhatt and Young 1992).
180:
181: Using a dynamical finite size scaling Ansatz we have found that {\it
182: in the thermodynamic limit} we expect a divergence of the relaxation
183: time of the form $\tau \propto (T-T_D)^{-2}$ (Brangian \textit{et al.}
184: 2001). Thus the exponent $-2$ is close to the values found for structural
185: glasses (G\"otze 1999).
186:
187: In the remaining of this paper we will concentrate on the self-averaging
188: properties of $C(t)$. Knowing these properties will help to decide whether
189: or not it is necessary to average the results of a simulation over many
190: independent realizations of the disorder even in the case that the size
191: of the system is very large. Suppose that we have determined the thermal
192: average $X_i$ of an observable $X$. (Here $i$ stands for the realization
193: of the disorder.) Following Wiseman and Domany (1998, and references
194: therein) we consider the quantity $R_X$ defined as follows:
195: %
196: \begin{equation}
197: R_X=\frac{[X_i^2]-[X_i]^2}{[X_i]^2}.
198: \label{eq3}
199: \end{equation}
200: %
201: Here $[\cdot ]$ stands again for the average over the disorder. Usually
202: one has the situation of ``strong self averaging'' which means that
203: $R_X \propto 1/N$ for $N\gg 1$. The case $R_X \propto 1/N^{\alpha}$,
204: with $\alpha <1$ is denoted by ``weak self averaging''. Finally the case
205: $R_X = const.$ is called ``non self-averaging''.
206:
207: In Fig.~\ref{fig3} we show the spin autocorrelation function for
208: system size $N=1280$ and for 20 representative samples. From the
209: figure it becomes clear that at high temperatures the sample to sample
210: fluctuations are quite small and that therefore the system is probably
211: self-averaging. For a temperature close to $T_D$ this is, however,
212: not the case in that the fluctuations are now on the same order as the
213: typical relaxation time.
214:
215: To study this effect in a more quantitative way we use the
216: relaxation time $\tau$ as the observable $X$ discussed above. Using
217: thus equation~(\ref{eq3}) to define the quantity $R_{\tau}$ we can
218: investigate the $N$ dependence of $R_{\tau}$. In Fig.~\ref{fig4} we
219: show the temperature dependence of $R_{\tau}N$ for all system sizes
220: investigated (main figure), with error bars that have been determined with
221: the jackknife method (Newman and Barkema 1999). We see that for high
222: temperatures we do indeed find that this quantity goes to a constant of
223: order one, {\it independent of the system size}. Hence we conclude that
224: $R_{\tau}$ is proportional to $1/N$ and that hence the system is strong
225: self-averaging. At low temperatures this is, however, no longer the case
226: since there we see that the product increases with increasing system size
227: and becomes, for the largest systems, as large $O(10^3)$. Thus this is
228: evidence that the system is no longer self-averaging. To investigate this
229: point closer, we plot in the inset $R_{\tau}$ at $T_D$ as a function of
230: $N$. (Note that at this temperature we do not have data for the largest
231: system size since the relaxation time becomes too large.) From this graph
232: we see that the value of $R_{\tau}$ is basically constant within the
233: noise of the data, or shows even a slight trend to increase. Thus this
234: is evidence that at this temperature the system is not self-averaging. We
235: also mention that we expect that for sufficiently large $N$ self-averaging
236: will be recovered for all $T>T_D$, although our data are not conclusive
237: on this issue for $T\le 1.3$, due to the strong finite size effects.
238:
239: To conclude, we have analyzed the dynamics of a $10$ states infinite
240: range Potts glass. Analytical results show that that this is a spin
241: model which resembles in many points structural glasses. We have shown
242: that the mean field scenario can, from a qualitative point of view, also
243: be seen in systems with finite $N$. However, close to the transition
244: temperature dynamical as well as static quantities are strongly
245: affected by finite size effects. In particular we find that the dynamics of the
246: system shows a crossover from a self-averaging behavior to a non self-averaging
247: behavior as the temperature approaches $T_D$.
248:
249: {\bf Acknowledgements:}
250: C.B. was partially supported by the Deutsche Forschungsgemeinschaft,
251: Sonderforschungsbereich 262/D1. W.K. and K.B. are grateful to the
252: German Israeli foundation (GIF) for travel support. We thank the RUS
253: for a generous grant of computing time on the Cray T3E.
254:
255: \newpage
256: \section*{References}
257: \begin{trivlist}
258: %
259: \item[]
260: Bhatt, R. N., and Young, A. P., 1992,
261: \textit{Europhys. Lett.}, \textbf{20}, 59.
262:
263: \item[]
264: Brangian, C., Kob, W., and Binder, K., 2001,
265: \textit{Europhys. Lett.}, \textbf{56}, 756.
266:
267: \item[]
268: Cwilich, G., and Kirkpatrick, T.R.,
269: 1989, \textit{J. Phys. A}, \textbf{22}, 4971
270:
271: \item[]
272: Elderfield, D., and Sherrington, D.,
273: 1983, \textit{J. Phys. C}, \textbf{16}, L497
274:
275: \item[]
276: De Santis, E., Parisi, G., and Ritort, F., 1995
277: {\it J. Phys. A: Math. Gen}. {\bf 28} 3025
278:
279: \item[]
280: G\"otze, W., 1989,
281: \textit{Liquids, freezing and the glass transition},
282: edited by J. P. Hansen, D. Levesque and J. Zinn-Justin,
283: (Amsterdam, North-Holland), pp. 287-503
284:
285: \item[]
286: Gross, D.J., Kanter, I., and Sompolinsky, H.,
287: 1985, \textit{Phys. Rev. Lett.}, \textbf{55}, 304
288:
289: \item[]
290: Kirkpatrick, T. R., and Wolynes, P. G., 1987,
291: \textit{Phys. Rev. B}, \textbf{36}, 5388.
292:
293: \item[]
294: Kirkpatrick, T. R., and Thirumalai, D., 1988,
295: \textit{Phys. Rev. B}, \textbf{37}, 5342.
296:
297: \item[]
298: Kirkpatrick, T. R., and Thirumalai, D., 1995,
299: \textit{Transp. Theory Stat. Phys.}, \textbf{24}, 927.
300:
301: \item[]
302: Kob, W, 1999,
303: \textit{J. Phys.: Condens. Matter}, \textbf{11}, R85.
304:
305: \item[]
306: Newman, M. E. J., and Barkema, G. T., 1999,
307: \textit{Monte Carlo Methods in Statistical Physics},
308: (Oxford: Oxford University Press).
309:
310: \item[]
311: Wiseman, S., and Domany, E., 1998,
312: \textit{Phys. Rev. E}, \textbf{58}, 2938.
313:
314: \end{trivlist}
315:
316:
317: %\newpage
318: \newpage
319: \begin{figure}[f]
320: \psfig{file=fig1.eps,width=13cm,height=9.5cm}
321: \caption{Spin autocorrelation function $C(t)$ versus $t$ (measured
322: in units of Monte Carlo steps per spin) for $1280$ spins at
323: various temperatures. The horizontal solid line shows the position of the
324: Edward Anderson order parameter in the thermodynamic limit.}
325: \label{fig1}
326: \end{figure}
327:
328: \begin{figure}[f]
329: \psfig{file=fig2.eps,width=13cm,height=9.5cm}
330: \caption{Spin autocorrelation functions for different system sizes at two
331: different temperatures, $T=1.8$ and $T=1.142=T_D$. The horizontal solid line shows
332: the position of the Edward Anderson order parameter in the thermodynamic
333: limit. The dashed line is used to define the relaxation time $\tau(T)$. }
334: \label{fig2}
335: \end{figure}
336:
337: \begin{figure}[f]
338: \psfig{file=fig3.eps,width=13cm,height=9.5cm}
339: \caption{Correlation functions for different realization of disorder.
340: System with 1280 spins, temperature $T=1.5$ and $T=1.142=T_D$}
341: \label{fig3}
342: \end{figure}
343:
344: \begin{figure}[f]
345: \psfig{file=fig4.eps,width=13cm,height=9.5cm}
346: \caption{Plot for the scaled quantity $R_X \cdot N$ as a function of
347: temperature; the inset shows $R_X$ as a function of the
348: system size at $T=1.142=T_D$}
349: \label{fig4}
350: \end{figure}
351:
352: \end{document}
353: