1: \documentclass[12pt,oneside]{article}
2: \usepackage{graphicx}
3: \linespread{1.6}
4: %
5: % ABBREVIAZIONI
6: %
7: \def\no{\noindent}
8: \def\be{\begin{equation}}
9: \def\ee{\end{equation}}
10: \def\ba{\begin{array}}
11: \def\ea{\end{array}}
12: \def\bea{\begin{eqnarray}}
13: \def\eea{\end{eqnarray}}
14: \def\<{\langle}
15: \def\>{\rangle}
16: \def\~{\tilde}
17: \def\s{\sigma}
18: \def\l{\lambda}
19: \def\a{\alpha}
20: \def\b{\beta}
21: \def\g{\gamma}
22: \def\o{\omega}
23: \def\t{\tau}
24: \def\n{\eta}
25: \def\bs{{\bar{s}}}
26: \def\hs{{\hat{s}}}
27: \newtheorem{theorem}{Theorem}
28: \newtheorem{proposition}[theorem]{Proposition}
29: \newtheorem{lemma}[theorem]{Lemma}
30: \newtheorem{corollary}[theorem]{Corollary}
31: \newtheorem{definition}[theorem]{Definition}
32: \newtheorem{remark}[theorem]{Remark}
33: %
34: %%%%% definizioni Mirko
35: \newcommand{\brac}[1]{\< #1\>}
36: \newcommand{\av}[1]{\mbox{Av}\left(#1\right)}
37: \newcommand{\zn}{Z_N}
38: \newcommand{\nbs}{\sqrt{N}\beta E_\sigma}
39: \newcommand{\nbt}{\sqrt{N}\beta E_\tau}
40: \newcommand{\SN}{\Sigma_N}
41: \newcommand{\mod}{\,\,\mbox{mod }}
42: \newcommand{\legendre}[2]{\left(\begin{array}{c}#1\\#2\end{array}\right)}
43: %%%% fine definizioni Mirko
44: %
45: %
46: \begin{document}
47: %
48: %
49: \title{Energy Landscape Statistics of the \\
50: Random Orthogonal Model} \maketitle
51: %
52: \begin{center}
53: %
54: \author{M. Degli Esposti, C. Giardin\`a, S. Graffi}\\
55: %{\large M. Degli Esposti, C. Giardin\`a, S. Graffi}\\
56: {\small\it
57: Dipartimento di Matematica \\ Universit\`a di Bologna,\\
58: Piazza di Porta S. Donato 5, 40127 Bologna, Italy \\
59: }
60: %
61: \end{center}
62: \date{}
63: %
64: %\begin{abstract}
65: %\end{abstract}
66: %
67: \abstract{The Random Orthogonal Model (ROM) of
68: Marinari-Parisi-Ritort \cite{MPR1,MPR2} is a model of statistical
69: mechanics where the couplings among the spins are defined by a
70: matrix chosen randomly within the orthogonal ensemble. It
71: reproduces the most relevant properties of the Parisi solution of
72: the Sherrington-Kirckpatrick model. Here we compute the energy
73: distribution, and work out an extimate for the two-point
74: correlation function. Moreover, we show exponential increase of
75: the number of metastable states also for non zero magnetic field.
76: }
77:
78:
79:
80:
81: \section{Introduction: review of the model and outlook}
82:
83: Random (symmetric) matrices out of a given ensemble can be taken
84: as interaction matrices for Ising spin models. The most famous
85: example is the Sherrington-Kirckpatrick (SK) model of spin
86: glasses, where the elements are i.i.d. Gaussian variables with
87: properly normalized variance. Aim of this paper is
88: to discuss a very specific example of these spin glass models,
89: which also share some interesting connections with number theory,
90: and show how random matrix theory could be useful to investigate
91: its
92: properties.
93:
94: For the sake of simplicity, let us start with a very
95: concrete
96: question: let $N\geq 1$ be a positive integer and denote
97: $\SN$ the
98: space of all possible configurations of $N$ spin
99: variables
100:
101: $$
102: \SN=\{\s=(\s_1,\ldots,\s_N)\,\,,\,\,\s_j=\pm
103: 1\},\quad\vert\SN\vert=2^N.
104: $$
105: Given $k=1,\ldots,N-1$, denote $C_k$ the correlation function:
106: $$
107: C_k(\s)\,=\,\sum_{j=1}^{N}\s_j\s_{j+k},\quad\mbox{where }
108: j+k:=(j+k-1\mod N) +1,
109: $$
110: and define the Hamiltonian function
111: $$
112: H(\s)\,=\,\frac{1}{N-1}\sum_{k=1}^{N-1} C_k^2
113: $$
114: For each N the ground state of the Hamiltonian $H$ can be looked at
115: as the binary
116: sequence with lowest autocorrelation and finding it
117: will have
118: some relevant practical applications in efficient
119: communication
120: (see \cite{Be} and references in \cite{MPR1}).
121:
122: It is remarkable that no concrete procedure for reproducing the
123: ground state for generic $N$ is known, but {\it ad hoc}
124: constructions based on number theory exist for very specific
125: values of $N$: if $N$ is prime number with $N=3\mod 4$, then the
126: sequence of the Legendre
127: symbols\footnote{$\left(\frac{j}{N}\right)=1$, if $j=x^2\,\mod N$
128: and $-1$ otherwise.} ($\s_N=1$)
129: $$
130: \s_j:=\left(\frac{j}{N}\right)=j^{\frac{1}{2}(N-1)}\mod N\,,\quad
131: j=1,\ldots,N-1
132: $$
133: gives the ground state of the system
134: \cite{DGGI,MPR1}.
135:
136: Through the use of the discrete Fourier transform, it is not
137: difficult to see \cite{MPR1,PP} that the previous problem is in
138: fact equivalent to finding the ground state for the so called {\it
139: Sine model}, which represents our starting point:
140: $$
141: H(\s)\,=\,-\frac{1}{2}\sum_{i,j=1}^N J_{ij}\s_i\s_j
142: .
143: $$
144: Here $J$ is the following $N\times N$ real symmetric orthogonal
145: matrix with almost full connectivity:
146:
147: $$
148: J_{ij}\,=\,\frac{2}{\sqrt{1+2N}}\,\sin\left(\frac{2\pi\,ij}{2N
149: +1}\right),\quad i,j=1,\ldots,N
150: $$
151: \vskip.3cm\noindent
152:
153: Here again, if $2N+1$ is prime and $N$ odd,
154: the Legendre symbols $\s_j=j^N\mod 2N+1$ give the ground state of
155: the system for these very specific values of $N$.
156:
157:
158: A natural approach is to extend the study of the ground state to
159: the more general thermo-dynamical behavior of the model in terms
160: of the inverse temperature $\beta=\frac{1}{T}$. As usual, the two
161: basic objects are: $$ \mbox{\it the partition function}\quad
162: Z_J(\beta)\,:=\,\sum_{\sigma\in\SN} e^{-\beta H(\sigma)}, $$ and
163: {\it the free energy density (at the thermodynamical
164: limit)} $$ f_J(\beta)\,=\,\lim_{N\to\infty}-\frac{1}{\beta N} \log
165: Z_J(\beta) $$\vskip0.3cm\noindent
166:
167: It is important to remark now that even if {\it there is no
168: randomness in the system}, the ground state of the model {\it
169: looks like} an output of a random number generator and the numeric
170: of its thermo-dynamical properties resemble the one of disordered
171: systems. This observation was in fact the starting point of an
172: approach developed in \cite{MPR1,MPR2,PP} where this model is
173: seen as a particular realization of a disordered model where the
174: coupling matrix is chosen at random out of a suitable set of
175: matrices:
176: \begin{definition} The {\it Random Orthogonal Model} (ROM) with magnetic field $h\geq0$ is the
177: disordered system with energy \be\label{hamilton}
178: H_J(\s)\,=\,-\frac{1}{2}\sum_{ij}J_{ij}\s_j\s_i\,+
179: \, h\sum_{j}\s_j, \ee
180: where
181: the coupling matrix $J$ is chosen randomly in the set of
182: {\it orthogonal symmetric matrices}:\footnote{In the ROM model
183: generic matrices have non zero diagonal elements. Often these
184: terms will be set to zero and orthogonality will be reconstructed
185: in the large $N$ limit.}
186: $$
187: J\,=\,ODO^{-1},
188: $$
189: Here $O$ is a generic orthogonal matrix and $D$ is diagonal with
190: entries $\pm 1$. The numbers $\pm 1$ are the eigenvalues of $J$.
191:
192:
193: The natural probability measure $\mu$ on this set is
194: the product of
195: the canonical Haar measure on the orthogonal group
196: by the
197: discrete measure on the diagonal terms.
198: \end{definition}
199: \vskip .5cm We will use the notation $\brac{\cdot}$ to denote the
200: average with respect the measure $\mu$. In particular, we are
201: interested in the {\it quenched} (i.e., the
202: average is performed after taking the logarithm) free energy density:
203:
204: \be\label{quenchedf}
205: \brac{f_J(\beta)}\,=\,-\lim_{N\to\infty}\frac{1}{\beta N}
206: \brac{\log Z_J(\beta)} \ee
207:
208:
209:
210:
211:
212:
213:
214: Average over the ROM disorder is performed by the following
215: fundamental formula, which has been obtained by adapting the
216: results in \cite{IZ} (see also \cite{BG}) valid for the unitary
217: case to the orthogonal one \cite{MPR2}.
218: For any $N\times N$
219: symmetric matrix $A$:
220: \begin{eqnarray} \brac{\exp \left\{ tr
221: \left(\frac{JA}{2}\right)\right\}} &=& \exp \left\{ N tr \left(
222: G\left(\frac{A}{N}\right)\right)\right\} +R_N(A), \nonumber\\
223: &\cong&
224: \exp \left\{ N \sum_{j=1}^N
225: G\left(\lambda_j\right)\right\}\label{main} \end{eqnarray} where
226: $R_N\to 0$ in the thermodynamical limit $N\to\infty$, the
227: $\lambda_j$'s are the (real) eigenvalues of $\frac{1}{N}\cdot A$
228: and $G(x)$ is given by $$ G(x) = \frac{1}{4}\left[ \sqrt{1+4x^2}
229: -\ln \left(\frac{1+\sqrt{1+4x^2}}{2}\right)-1\right] $$
230:
231:
232:
233: The same formula is exact for the SK model, i.e Gaussian
234: independent symmetric couplings, with $$ G_{SK}(x)=\frac{x^2}{4}
235: $$ Note that $G(x) = G_{SK}(x) + o(x)$. For example, up to the
236: 10-th order
237: $$
238: G(x)\,=\,\frac{x^2}{4} - \frac{x^4}{8} + \frac{x^6}{6} -
239: \frac{5\,x^8}{16} +
240: \frac{7\,x^{10}}{10} + O(x^{11})
241: $$
242:
243: The ROM model has been chosen in a such a way that, at least for
244: not too small temperature, the {\it deterministic} Sine model and
245: the one with quenched disorder share a common behavior. More
246: precisely, the couplings are always of order $N^{-\frac{1}{2}}$;
247: the diagrams contributing to the thermodynamical limit of the high
248: temperature expansion for the free energy density have
249: the same topology and they can all be
250: expressed in terms of positive powers of the trace of the
251: couplings. By construction, the high temperature expansion of the
252: free energy density $f_J(\beta)$ in powers of $\beta$ is then
253: independent of the particular choice of the symmetric orthogonal
254: matrix $J$ and it does coincide with the annealed average w.r.t.
255: $\mu$. In particular \cite{PP}:
256: $$
257: -\beta \brac{f_J(\beta)}\,=\, \log 2 + G(\beta).
258: $$
259:
260:
261:
262:
263: Besides SK and in general the large class of $p$-spin models,
264: where couplings have a gaussian distribution, the ROM model
265: provide another interesting class of disordered mean-field spin
266: glass. This model has received considerable interest in recent
267: years, especially in the contest of structural glass transition.
268: Indeed it can been seen as the random version of a wide class of
269: models (for example the fully frustrated Ising model on a
270: hypercube or the above mentioned sine model) which despite having
271: a non-random Hamiltonian display a strong glassy behavior
272: \cite{BM,MPR1,DGGI}. This model has been studied in the framework
273: of replica theory \cite{MPR2}, where it was shown that replica
274: symmetry is broken and there are many equilibrium states available
275: to the system. Mean field (TAP) equations have been derived for
276: this model by resumming the high temperature expansion and the
277: average number of solutions of these equations has been studied in
278: ref. \cite{PP}.
279:
280: It is a well established fact that the observed properties of
281: mean-field spin glass models are due to the large number of {\it
282: metastable states} the system possesses. Despite being not fully
283: justified from a mathematical point of view, the Parisi scheme of
284: breaking replica symmetry furnishes a clear picture of equilibrium
285: statistical properties: states with similar macroscopic behavior
286: have vastly different spin configurations, and have large
287: relaxation times for transition between them. As a consequence,
288: the ground state is accessible only on very long time scales. It
289: is worth mentioning
290: that rigorous results validating the Parisi solution
291: have been
292: accumulating in recent times.
293:
294: For example, Guerra and Toninelli\cite{G1} have proved the existence
295: of thermodynamical limit, i.e. the existence of
296: the limit for quenched
297: average of the free energy (eq.
298: \ref{quenchedf}). See also
299: \cite{CDGG} where the result has been
300: extended to general
301: correlated gaussian random energy models.
302: Finally, more recently
303: \cite{G2}, Guerra showed that the
304: Parisi Ansatz represents
305: at least a lower bound for the quenched
306: average of the free
307: energy.
308:
309: However there is not yet an unambiguous way to identify
310: those metastable
311: states which are relevant for thermodynamics in the
312: infinite volume
313: limit. At zero temperature the metastable states can
314: be defined as
315: the states {\it locally stable} to single spin
316: flips (definition recalled in Section 3 below) and
317: the calculations
318: are relatively straightforward. Complete analysis of
319: the typical
320: energy of metastable states and of the effects of the
321: external
322: field have been undertaken both for the SK model
323: \cite{TE,Ro,D} and
324: for general $p$-spin model \cite{OF}. The zero
325: temperature dynamics
326: for the deterministic {\it Sine model} has
327: been instead studied in
328: \cite{DGGI}.
329:
330: At non-zero temperature the identification is less obvious and
331: most studies \cite{BrM,Ri} rely on the counting of the number of
332: solutions to the celebrated TAP equations \cite{TAP}. According to
333: the general belief, one can associate to each metastable state a
334: solution of the TAP equation, but the inverse is not true: a TAP
335: solution corresponds to a metastable state only if it is separated
336: from other solution by a barrier whose height diverges with the
337: volume.
338:
339: It appears, however, that even the calculations at zero
340: temperature in a presence of external field have not been carried
341: out. One expects, in analogy with SK model, the existence of an AT
342: line \cite{AT} indicating the onset of replica-symmetry breaking.
343: In this paper we study at length the effects of the magnetic field
344: on the structure of local optima of the energy landscape. We are
345: able to use these results to shed further light on the nature of
346: the AT instability at zero temperature.
347:
348:
349: In the next Section \ref{statistics} we study the statistics of energy levels over the
350: whole configuration space. We compute energy distribution of a
351: generic spin configuration and the pair correlations for a given
352: couple of spin configurations with a fixed overlap. In Section
353: \ref{metastable} we analyze metastable states at zero temperature,
354: also in a presence of an external field.
355:
356:
357:
358:
359:
360:
361:
362: \section{Statistics of energy levels}
363: \label{statistics}
364:
365: We start by analyzing the statistical features of the landscape
366: generated by the energy function (\ref{hamilton}). In this section
367: we will always consider zero magnetic field $h=0$. Let us begin
368: with the energy distribution for a single fixed configuration.
369:
370: \subsection{Distribution of energy}
371:
372: Let $\s = (\s_1,\s_2,\ldots,\s_N)$ denote a given configuration
373: with energy $H_J(\s)$. The probability $P_\s(E)$ is then given
374: by: $$ P_\s(E) \,:=\, \< \delta(E - H(J,\s))\> $$
375:
376: \noindent Due to gauge invariance, the probability $P_\s(E)$ does
377: not depend on the spin configuration $\s$ and it will be denoted
378: just by $P(E)$, in fact: $H(J,\s) = H(J',\s')$ and $P(J) = P(J')$
379: where $J_{ij}' = J_{ij}\s_i\s_i'\s_j\s_j'$.
380:
381: \noindent Introducing the integral representation for $\delta$
382: function
383:
384: $$ \delta(x - x_0) = \frac{1}{2\pi i}\int_{-i \infty}^{+i \infty}
385: dk \; e^{k \; (x-x_0)} $$
386: we get
387:
388: $$ P(E) = \frac{1}{2\pi i}\int_{-i \infty}^{+i \infty} dk \; e^{kE}
389: \; \<e^{\frac{1}{2} \sum_{i,j =1}^{N}k \; J_{ij}\s_i\s_j}\> $$
390:
391: \noindent and we can apply formula (\ref{main}) to average over
392: disorder considering the matrix $A_{ij} = k\s_i\s_j$.
393:
394: It is easy to prove that $A$ admits only one non-zero, simple
395: eigenvalue $\lambda=kN$, so that
396: $$P(E) = \frac{1}{2\pi i}\int_{-i
397: \infty}^{+i \infty} dk \; \exp \left[N \left(\frac{kE}{N} +
398: G(k)\right)\right] $$
399:
400: \noindent In the large-$N$ limit the integral can be evaluated
401: using the saddle-point method. Clearly, the equation
402: $$
403: \frac{E}{N} + G'(k)\,=\,\frac{E}{N} + \frac{k}{1 + {\sqrt{1 +
404: 4\,k^2}}}=0
405: $$
406: admits the solution $\bar{k}= \frac{2\,E\,N}{4\,E^2 - N^2}$.
407:
408: This gives:
409:
410: \begin{eqnarray}
411: P_{ROM}(E) &\sim& \exp\left[N\left(\frac{\bar{k}E}{N} +
412: G(\bar{k})\right)\right]\label{rom1}\\
413: &=& \left(1-\left(\frac{2E}{N}\right)^2\right) ^{N/4}\nonumber\\
414: &\sim&\
415: \exp\left[-\frac{E^2}{N}-2\frac{E^4}{N^3}-\frac{16}{3}\frac{E^6}{N^5}+\cdots\right]\nonumber\\
416: \end{eqnarray}
417:
418:
419: \noindent apart for an unimportant constant, not predicted by the
420: saddle-point. As a comparison, in the case of SK model one finds
421: exactly the gaussian distribution: $$ P_{SK}(E) \sim \exp
422: \left(-\frac{E^2}{N}\right) $$
423:
424:
425:
426: To check the validity of formula (\ref{main}) which has been used
427: to average over disorder, we computed $P_{ROM}(E)$ for a relative
428: small ROM (N=100) numerically. For a given spin configuration,
429: random disorder realizations $J=ODO^{-1}$ were generated by using
430: an orthogonal
431: matrix $O$ obtained from a
432: gaussian matrix by applying Gram-Schmidt orthogonalization
433: algorithm and coin tossing for the diagonal $D$. The resulting
434: distribution of energies was binned and is shown as the data
435: points in Fig. (\ref{fig1}).
436:
437: \begin{figure}[ht]
438: \includegraphics[width=4in,angle=270]{fig1.eps}
439: \caption{\baselineskip=13pt {\small { Probability distribution
440: function $P_{ROM}(E)$ for ROM model (full curve). Simulation for a
441: $N=100$ ROM (data points). For a fixed spin configuration, $10^6$
442: realizations of disorder were generated.}}} \protect\label{fig1}
443: \end{figure}
444:
445: \noindent
446: %The agreement between analytical calculation and this
447: %simulation is very good, so we consider this a fair test of
448: %formula (\ref{main}).
449:
450: As it should be, the support of $P_{ROM}(E)$ is almost all
451: in the interval $[-N/2,N/2]$. Indeed, the orthogonality of
452: $J$ imposes simple bound on the energy of any spin configuration:
453: the lower bound $-N/2$ (resp. upper bound $N/2$) is reached if and
454: only if $\s$ is an eigenvector of $J$ relative to eigenvalue $+1$
455: (resp. $-1$).
456:
457:
458:
459:
460: \subsection{Two-point Energy Correlation}
461:
462: We consider now the probability $P_{\s,\tau}(E_1,E_2)$ that two
463: configurations $\s,\tau\in\SN$ have energies $E_1$ and $E_2$
464: respectively. Gauge invariance implies that this probability can
465: only depend on the overlap between the two configurations:
466:
467: $$ q(\s,\tau) = \frac{1}{N}\sum_{i=1}^N \s_i\tau_i $$
468:
469: \noindent Proceeding as before, we get: \bea P_{\s,\tau}(E_1,E_2)
470: & = & \< \delta(E_1 - H(J,\s)) \;
471: \delta(E_2 - H(J,\tau))\>
472: \nonumber\\
473: & = & \frac{1}{(2\pi i)^2}\int_{-i \infty}^{+i \infty} dk_1 \;
474: \int_{-i \infty}^{+i \infty} dk_2 \;
475: \exp (k_1 E_1+k_2 E_2) \nonumber\\
476: & & \<\exp \left(\frac{1}{2} \sum_{i,j =1}^{N} J_{ij}
477: (k_1 \s_i\s_j + k_2 \tau_i\tau_j)\right)\>
478: \eea
479:
480: \vspace{0.5cm} \noindent Consider now the matrix $A_{ij} = k_1
481: \s_i\s_j + k_2 \tau_i\tau_j$ which has two non-zero simple
482: eigenvalues
483:
484: $$ \lambda_{\pm} = \frac{N}{2}\left[(k_1+k_2) \pm
485: \sqrt{(k_1-k_2)^2 + 4k_1k_2q^2} \right]
486: $$
487:
488: \noindent Applying formula (\ref{main}) we obtain: \bea
489: P_{\s,\tau}(E_1,E_2) & = & \frac{1}{(2\pi i)^2}\int_{-i
490: \infty}^{+i \infty} dk_1 \;
491: \int_{-i \infty}^{+i \infty} dk_2 \nonumber \\
492: & & \exp \left[ N \left( \frac{k_1 E_1}{N} + \frac{k_2 E_2}{N}
493: + G(\lambda_{+}/N) + G(\lambda_{-}/N) \right) \right
494: ]\nonumber
495: \eea
496:
497: The saddle-point method yields the equations:
498: $$
499: \frac{E_j}{N} +\frac{1}{N}\,
500: G'(\frac{\lambda_+}{N})\,\frac{\partial\lambda_+}{\partial k_j}
501: \,+\,\frac{1}{N}\,
502: G'(\frac{\lambda_-}{N})\,\frac{\partial\lambda_-}{\partial
503: k_j}\,=\,0,\,\,\qquad j=1,2
504: $$
505:
506:
507:
508: For the SK model, one immediately find:
509: $$
510: \frac{E_1}{N}+\frac{1}{2}(k_1+k_2
511: q^2)=0,\quad\frac{E_2}{N}+\frac{1}{2}(k_2+k_1 q^2)=0
512: $$
513: with solutions:
514: $$
515: k_1=\frac{2(E_1-E_2 q^2)}{N(-1+q^4)},\quad k_2=\frac{2(E_2-E_1
516: q^2)}{N(-1+q^4)}.
517: $$
518: This yields the well known \cite{De} ($ \s,\t\in\SN$ fixed, with
519: overlap $q$):
520: \begin{eqnarray} P_{SK}(E_1,E_2) &=& \left(\frac{\sqrt{1-q^4}}{N\pi}\right)\exp \left
521: [-\frac{(E_1+E_2)^2}{2N(1+q^2)}\right]
522: \exp \left
523: [-\frac{(E_1-E_2)^2}{2N(1-q^2)}\right]\nonumber\\
524: &=&
525: P_{SK}\left(\frac{E_1+E_2}{\sqrt{2(1+q^2)}}\right)\cdot
526: P_{SK}\left(\frac{E_1-E_2}{\sqrt{2(1-q^2)}}\right).
527: \end{eqnarray}
528: For asymptotically uncorrelated configurations $q=0$, one clearly
529: get a product measure, whereas one recover complete degeneracy
530: when $q=1$: \be \label{sk1}P_{SK}(E_1,E_2)\,=\, P_{SK}(E_1)\cdot
531: P_{SK}(E_2),\quad q=0,
532: \ee
533: and \be \label{sk2}P_{SK}(E_1,E_2)\,=\,
534: P_{SK}(E_1)\cdot\delta\left(E_2-E_1\right),\quad q=1.
535: \ee
536:
537: In generale, one has
538: $$
539: \int_{-\infty}^{+\infty}\,\int_{-\infty}^{+\infty} E_1E_2\,
540: dP_{SK}(E_1,E_2)\,=\,\frac{Nq^2}{2}.
541: $$
542:
543:
544:
545:
546: For the ROM model, it is immediate to see that the analog of
547: (\ref{sk1}) and (\ref{sk2}) hold true with the single energy
548: distribution $P_{ROM}(E)$ given by (\ref{rom1}). For generic value
549: of $0<q<1$, a first crude estimate is achieved by using the
550: stationary points of the gaussian approximation and
551: $G(x)=\frac{x^2}{4}-\frac{x^4}{8}$ to evaluate the exponent. This
552: yields:
553: $$
554: P_{ROM}(E_1,E_2)\sim P_{SK}(E_1,E_2)\cdot Exp[-\Phi_q(E_1,E_2)],
555: $$
556: where
557: \begin{eqnarray}
558: \Phi_q(E_1,E_2)\,&:=&\,-2\frac{\, -8\,E_1^3\,E_2\,q^4 -
559: 8\,E_1\,E_2^3\,q^4 +
560: E_1^4\,\left( 1 + 2\,q^2 - q^4 \right) }{N^3\,
561: {\left( -1 + q^2 \right) }^2\,{\left( 1 + q^2 \right)
562: }^4}\nonumber\\
563: &&+\frac{
564: {E_2}^4\,\left( 1 + 2\,q^2 - q^4 \right) +
565: 2\,{E_1}^2\,{E_2}^2\,q^2\,
566: \left( 2 - q^2 + 4\,q^4 + q^6 \right)}{N^3\,
567: {\left( -1 + q^2 \right) }^2\,{\left( 1 + q^2 \right)
568: }^4}\nonumber
569: \end{eqnarray}
570:
571: Further corrections can be now calculated, but we do not known
572: a systematic way of doing it at all orders.
573:
574:
575:
576: \section{Zero temperature metastable states}
577: \label{metastable}
578:
579: Metastable states at zero temperature are defined as the
580: configurations whose energy can not be decreased by reversing any
581: of the spins \cite{DGGI}. Since the energy change $\Delta E_i$
582: involved in flipping the spin at site $i$ is given by $$ \Delta
583: E_i = 2 \left(\sum_{j}J_{ij}\s_i \s_j + h\s_i\right) $$ the
584: constraint a configuration $\s$ must satisfy in order to be
585: metastable is $$\sum_{j}J_{ij}\s_i \s_j + h\s_i > 0 \quad\quad
586: \forall i=1,\ldots, N$$
587:
588: \noindent The average number of metastable configurations $\<{\cal
589: N}(e,h)\>$ with a given energy density $e=E/N$ is then
590:
591: \bea \<{\cal N}(e,h)\>
592: & = & \< \sum_{\{\s\}}
593: \prod_{i=1}^N
594: \left[ \int_{0}^{\infty} d\lambda_i \;
595: \delta \left(\lambda_i - \sum_{j}J_{ij}\s_i\s_j - h\s_i\right)
596: \right] \nonumber \\
597: & & \delta\left( Ne + \frac{1}{2} \sum_{i,j}J_{ij}\s_i\s_j
598: + h \sum_{i} \s_i \right) \>
599: \eea
600:
601: \noindent One should really calculate the average value of the
602: logarithm of the number of metastable states, this being the
603: extensive quantity, and hence introduce replicas; indeed, as
604: pointed out in \cite{BrM}, the introduction of a uniform magnetic
605: field should introduce strong correlations between the metastable
606: states. However, we shall proceed with the direct calculation of
607: $\<{\cal N}(e,h)\>$ as it suffices to bring out the most relevant
608: features of the problem.
609:
610: \noindent Introducing integral representations for $\delta$
611: functions we have \bea \<{\cal N}(e,h)\>
612: & = & \sum_{\{\s\}}
613: \int_{-i\infty}^{+i\infty} \frac{dz}{2\pi i} \;
614: e^{z Ne} e^{z h \sum_{i} \s_i} \nonumber \\
615: & & \prod_{i=1}^N
616: \left[ \int_{0}^{\infty} d\lambda_i \;
617: \int_{-i \infty}^{+i \infty} \frac{dk_i}{2\pi i} \;
618: \right]
619: e^{\sum_{i} k_i(h\s_i-\lambda_i)} \;
620: \< e^{ \sum_{i,j} J_{ij}
621: (\frac{z}{2}\s_i\s_j + k_i\s_i\s_j}) \> \nonumber
622: \eea
623:
624: \noindent
625:
626: \noindent To apply the formula (\ref{main}) for averaging over
627: disorder we define the matrix $A_{ij} = \left( \frac{z}{2} + k_i
628: \right) \s_i\s_j + \left( \frac{z}{2} + k_j \right)\s_j\s_i$. The
629: non-zero eigenvalues of $A_{ij}$ are easily calculated and read
630:
631: $$\mu_{\pm} = \sum_i \left( \frac{z}{2} + k_i \right) \pm \sqrt{N
632: \sum_i \left( \frac{z}{2} + k_i \right)^2} $$
633:
634: \noindent so that we obtain:
635:
636: \bea \<{\cal N}(e,h)\> & = & \sum_{\{\s\}}
637: \int_{-i\infty}^{+i\infty} \frac{dz}{2\pi i}\;
638: e^{z Ne} e^{z h \sum_{i} \s_i}\prod_{i=1}^N
639: \left[ \int_{0}^{\infty} d\lambda_i \;
640: \int_{-i \infty}^{+i \infty}\frac{dk_i}{2\pi i} \;
641: \right] \;
642: e^{\sum_{i} k_i(h\s_i-\lambda_i)} \nonumber\\
643: & & \exp\left\{N\left[
644: G\left(\frac{1}{N}
645: \sum_i \left( \frac{z}{2} + k_i \right) + \sqrt{\frac{1}{N}
646: \sum_i \left( \frac{z}{2} + k_i \right)^2} \right)
647: \right.\right. + \nonumber\\
648: & & \left.\left.
649: G\left(\frac{1}{N}
650: \sum_i \left( \frac{z}{2} + k_i \right) - \sqrt{\frac{1}{N}
651: \sum_i \left( \frac{z}{2} + k_i \right)^2} \right)
652: \right]\right\}
653: \eea
654:
655: \noindent Performing now the trace over spin configuration,
656: defining
657:
658: $$v = \frac{1}{N}\sum_i \left( \frac{z}{2} + k_i \right)
659: \quad\quad\quad w = \frac{1}{N}\sum_i \left( \frac{z}{2} + k_i
660: \right)^2 $$
661:
662: \noindent and imposing the constraints via two Lagrange
663: multipliers, we have
664:
665: \bea \<{\cal N}(e,h)\> & = & \frac{1}{(2\pi i)^3}
666: \int_{-i \infty}^{+i \infty}dz \;
667: \int_{-i \infty}^{+i \infty}dv \;
668: \int_{-i \infty}^{+i \infty}dw \;
669: \int_{-i \infty}^{+i \infty}dx \;
670: \int_{-i \infty}^{+i \infty}dy \;
671: \nonumber \\
672: & & \exp\left\{
673: N [ ze + \frac{zx}{2} + \frac{yz^2}{4}]
674: \right\} \nonumber \\
675: & & \exp\left\{
676: N [ -xv -yw + G(v + \sqrt{w})+G(v - \sqrt{w})]
677: \right\} \nonumber \\
678: & & \prod_{i=1}^N
679: \left[ \int_{0}^{\infty} d\lambda_i \;
680: \int_{-i \infty}^{+i \infty}\frac{dk_i}{\pi i} \;
681: e^{y k_i^2 + k_i(x-\lambda_i + yz)}
682: \cosh{(h(z + k_i))}\right]
683: \eea
684:
685: \noindent The integrals over the $k_i$ are now gaussian \bea
686: \<{\cal N}(e,h)\> & = & \frac{1}{(2\pi i)^3}
687: \int_{-i \infty}^{+i \infty}dz \;
688: \int_{-i \infty}^{+i \infty}dv \;
689: \int_{-i \infty}^{+i \infty}dw \;
690: \int_{-i \infty}^{+i \infty}dx \;
691: \int_{-i \infty}^{+i \infty}dy \;
692: \nonumber \\
693: & & \exp\left\{
694: N [ ze + \frac{zx}{2} + \frac{yz^2}{4}]
695: \right\} \nonumber \\
696: & & \exp\left\{
697: N [ -xv -yw + G(v + \sqrt{w})+G(v - \sqrt{w})]
698: \right\} \\
699: & & \prod_{i=1}^N
700: \left[ \int_{0}^{\infty} d\lambda_i \;
701: \frac{1}{2\sqrt{\pi y}} \left(
702: e^{hz} e^{-\frac{(x+yz-\lambda_i+h)^2}{4y}} +
703: e^{-hz}e^{-\frac{(x+yz-\lambda_i-h)^2}{4y}}
704: \right)\right]\nonumber
705: \eea
706:
707: \noindent and the integrals over the $\lambda_i$ can be performed
708: in terms of the complementary error function
709:
710: $$ \mbox{erfc(x)} = \frac{2}{\sqrt{\pi}}
711: \int_{x}^{\infty} e^{-t^2}\,dt
712: $$
713: \noindent so that we find: \bea \label{saddle} \<{\cal N}(e,h)\>
714: & = & \frac{1}{(2\pi i)^3}
715: \int_{-i \infty}^{+i \infty}dz \;
716: \int_{-i \infty}^{+i \infty}dv \;
717: \int_{-i \infty}^{+i \infty}dw \;
718: \int_{-i \infty}^{+i \infty}dx \;
719: \int_{-i \infty}^{+i \infty}dy \;
720: \nonumber \\
721: & & \exp\left\{
722: N [ ze + \frac{zx}{2} + \frac{yz^2}{4}]
723: \right\} \nonumber \\
724: & & \exp \left\{N \left[ -xv -yw + G(v + \sqrt{w})+G(v - \sqrt{w})
725: \right.\right.\\
726: & & + \ln \left(\frac{1}{2}
727: \left(e^{hz}\mbox{erfc}\left(-\frac{x+yz+h}{2\sqrt{y}}\right) +
728: e^{-hz}\mbox{erfc}\left(-\frac{x+yz-h}{2\sqrt{y}}\right)
729: \right) \right) \left.\left.\right]\right\} \nonumber
730: \eea
731:
732: \noindent As usual the calculation is concluded by carrying out a
733: saddle-point integration. The r.h.s. of Eq. (\ref{saddle}) is to
734: be extremized with respect to the five variables $z,v,w,x,y$.
735:
736:
737: \subsection{Total number of metastable states}
738:
739:
740: Here we study the total number of metastable states ${\< \cal
741: N}(h) \>$ (irrespectively of the energy) as a function of the
742: field. Writing
743: \be\label{Ah}
744: \log \< {\cal N}(h)\> = A(h)N + B(h), \ee $A(h)$, in the
745: thermodynamical limit $(N\rightarrow \infty)$, can be calculated
746: by setting $z=0$ in Eq. (\ref{saddle}), which becomes:
747:
748: \bea \<{\cal N}(h)\> & = & \frac{1}{(2\pi i)^2}
749: \int_{-i \infty}^{+i \infty}dv \;
750: \int_{-i \infty}^{+i \infty}dw \;
751: \int_{-i \infty}^{+i \infty}dx \;
752: \int_{-i \infty}^{+i \infty}dy \;
753: \nonumber \\
754: & & \exp \left\{N \left[ -xv -yw + G(v + \sqrt{w})+G(v - \sqrt{w})
755: \right.\right.\nonumber \\
756: & & + \ln \left(\frac{1}{2}
757: \left(\mbox{erfc}(-\frac{x+h}{2\sqrt{y}}) +
758: \mbox{erfc}(-\frac{x-h}{2\sqrt{y}})\right)
759: \right)
760: \left.\left.\right]\right\}\label{saddle1}
761: \eea In the case of the SK model one recover the well now
762: one-variable saddle-point equation \cite{D}:
763: $$
764: x\,=\,\frac{\exp[-x^2/2] \cosh(h x)}{\int_{-x}^{\infty}
765: \exp[-t^2/2]\cosh(h t)\,dt}
766: $$
767: If $x_c$ is the solution to the previous equation:
768: $$
769: A_{SK}(h)\,=\,\log(2)\,-\,\frac{1}{2}(x_c^2+h^2)\,+\,\log\left(\frac{1}{(2\pi)^{1/2}}\int_{-x_c}^{\infty}
770: \exp[-t^2/2]\cosh(h t)\,dt\right),
771: $$
772: in particular $A_{SK}(0)\sim 0.199$, whereas for large $h$ one has
773: $x\sim\left(\frac{2}{\pi}\right)^{1/2}\, e^{-h^2/2}$ and
774: consequently $A_{SK}\sim\frac{1}{\pi}e^{-h^2}$ (see
775: Fig.(\ref{metask})).
776: \begin{figure}[ht]
777: \includegraphics[width=3in]{metask1.eps}
778: \caption{\baselineskip=13pt {\small { $A_{SK}(h)$ }}}
779: \protect\label{metask}
780: \end{figure}
781:
782:
783:
784: We now turn to the ROM model. We first perform a numerical
785: investigation by doing an exhaustive enumeration of spin
786: configurations and keeping track of metastable states. The
787: system-size dependence of $\log \< {\cal N}(h)\>$ is plotted for
788: different values of $h$ in Fig. (\ref{fig2}, left). The data are
789: fitted to formula (\ref{Ah}), ignoring possible finite size
790: corrections. The resulting $A_{ROM}(h)$ are showed in Fig.
791: (\ref{fig2}, right) as data points.
792:
793: \noindent Moreover, the saddle point equations corresponding to
794: (\ref{saddle1}) were solved numerically, and the result is shown
795: by the solid curve in (Fig. (\ref{fig2}, right)). The agreement
796: between theory and simulations is very good in spite of the fact
797: that we used admittedly small systems ($N<30$).
798:
799: \begin{figure}[!h]
800: \includegraphics[width=2.1in,angle=270]{fig2a.eps}
801: \includegraphics[width=2.1in,angle=270]{fig2b.eps}
802: \caption{\baselineskip=13pt {\small { Numbers of metastable
803: states. (Left): Their dependence on system size $N$ at different
804: magnetic field, see legend. (Right): Data points shows the field
805: dependence of $A_{ROM}(h)$ obtained from the fits, while the full
806: curve indicates the analytical results in the thermodynamical
807: limit. }}} \protect\label{fig2}
808: \end{figure}
809:
810:
811: \noindent As one would expect, metastable states disappear as the
812: magnetic field is increased, since it introduces a tendency
813: towards ferromagnetic behavior. Most of the processes are the
814: confluence of a metastable state to another with a larger drop of
815: free-energy.
816:
817: Note that we have $A_{ROM}(0)\sim 0.285$, while the asymptotic
818: behavior for large magnetic field $h$ does coincide with the
819: gaussian case (see Fig.(\ref{romfig1},Fig.(\ref{romfig2} )).
820: \begin{figure}[!h]
821: \includegraphics[width=3in]{metaskrom2.eps}
822: \caption{\baselineskip=13pt {\small { $A_{SK}(h)$ (bottom
823: blue),$A_{ROM}(h)$ (middle red ), $\frac{1}{\pi}e^{-h^2}$ (top)
824: .}}} \protect\label{romfig1}
825: \end{figure}
826: \begin{figure}[!h]
827: \includegraphics[width=3in]{romhlarge.eps}
828: \caption{\baselineskip=13pt {\small { Plot of $e^{h^2}\cdot
829: A_{ROM}(h)$ for values of the magnetic field $h$ between $1$ and
830: $5$.}}} \protect\label{romfig2}
831: \end{figure}
832:
833: This indicates that $A_{ROM}(h)$ still remain non-zero for
834: arbitrarily large $h$ and hence for any finite value of the
835: external magnetic field the number of metastable states grow
836: exponentially with the system size $N$. As pointed out by \cite{D}
837: for the SK model, this result is in agreement with the observation
838: that the AT instability occurs for all finite $h$ at zero
839: temperature.
840:
841: \thanks{{\bf Acknowledgements:} This work has been partially
842: supported by the European Commission under the Research Training
843: Network MAQC,
844: n° HPRN-CT-2000-00103 of the IHP Programme.}
845:
846:
847:
848:
849:
850:
851: \newpage
852:
853:
854: \begin{thebibliography}{}
855:
856:
857: \bibitem{AT} J.R.L. de Almeida and D.J. Thouless,
858: ``Stability of the Sherrington-Kirkpatrick solution of a spin
859: glass model'', {\em J. Phys. A: Math. Gen.} {\bf 11} 983 (1978)
860:
861: \bibitem{BG} E. Brezin and D.J. Gross,
862: ``The External Field Problem in the Large $N$ Limit of QCD'', {\em
863: Phys. Lett.} {\bf 97B} 120 (1980)
864:
865:
866: \bibitem{BM} J.P. Bouchaud and M.Mezard,
867: ``Self induced quenched disorder: a model for the glass
868: transition'', {\em J. Physique I} {\bf 4} 1109 (1994)
869:
870: \bibitem{Be} J. Bernasconi,``Low Autocorrelation Binary Sequences: Statistical Mechanics
871: and Configuration Space Analysis'', {\em J. Physique } {\bf 48}
872: 559 (1987)
873: \bibitem{BrM}
874: A.J. Bray and M.A. Moore, ``Metastable states in spin glasses'',
875: {\em J. Phys C: Solid State Phys.} {\bf 13} L469 (1980)
876:
877: \bibitem{CDGG}
878: P. Contucci, M. Degli Esposti, C. Giardin\`a, S. Graffi,
879: ``Thermodynamical limit for correlated gaussian random energy
880: models'', {\em cond-mat/0206007}
881:
882:
883: \bibitem{D} D.S. Dean,
884: ``On the metastable states of the zero-temperature SK mode'', {\em
885: J. Phys. A: Math. Gen.} {\bf 27} L889 (1994)
886:
887: \bibitem{De} B. Derrida,
888: ``Random-energy model: An exactly solvable model of disorder
889: systems'', {\em Phys. Rev. B} {\bf 24} 2613 (1981)
890:
891:
892:
893: \bibitem{DGGI}
894: M. Degli Esposti, C. Giardina', S. Graffi and S. Isola,
895: ``Statistics of energy levels and zero temperature dynamics for
896: deterministic spin models with glassy behaviour'', {\em J. Stat.
897: Phys.} {\bf 102} 1285 (2001)
898:
899:
900: \bibitem{G1}
901: F. Guerra, F.L. Toninelli, ``The thermodynamical limit in mean
902: field spin glass model'', {\em cond-mat/0204280}, to appear on22
903: {\em Comm. Math. Phys.}
904:
905:
906: \bibitem{G2}
907: F. Guerra, ``Broken replica symmetry bounds in the mean field spin
908: glass model'', {\em cond-mat/0205123}
909:
910: \bibitem {IZ} C. Itzykson and J.B. Zuber,
911: ``The Planar Approximation (II)'', {\em J. Math. Phys.} {\bf 21}
912: 411 (1980).
913:
914:
915: \bibitem {MPR1} E.Marinari, G.Parisi and F.Ritort,
916: ``Replica field theory for deterministic models: binary sequences
917: with low autocorrelation'', {\em J. Phys. A: Math. Gen.} {\bf 27}
918: 7615 (1994).
919:
920:
921: \bibitem{MPR2} E.Marinari, G.Parisi and F.Ritort,
922: ``Replica field theory for deterministic models II: A non-random
923: spin glass with glassy behaviour'', {\em J. Phys. A: Math. Gen.}
924: {\bf 27} 7647 (1994)
925:
926:
927: \bibitem{OF} V.M. de Oliveira and J.F. Fontanari,
928: ``Landscape statistics of the $p$-spin Ising model'', {\em J.
929: Phys. A.: Math. Gen.} {\bf 30} 8445 (1997)
930:
931:
932: \bibitem{PP} G. Parisi and M. Potters,
933: ``Mean-field equations for spin models with orthogonal interaction
934: matrices'', {\em J. Phys. A.: Math. Gen.} {\bf 28} 5267 (1995)
935:
936:
937: \bibitem{Ri} H. Rieger,
938: ``The number of solutions of the Thouless-Anderson-Palmer
939: equations for $p$-spin-interaction spin glasses'', {\em Phys. Rev.
940: B} {\bf 46} 14655 (1992)
941:
942:
943: \bibitem{Ro} S.A. Roberts,
944: ``Metastable states and innocent replica theory in an Ising spin
945: glass'', {\em J. Phys. C: Solid State Phys.} {\bf 14} 3015 (1981)
946:
947:
948: \bibitem{TE} F. Tanaka and S.F. Edwards,
949: ``Analytic theory of the ground state properties of a spin glass:
950: I. Ising spin glass'', {\em J. Phys. F.: Metal Phys.} {\bf 10}
951: 2769 (1980)
952:
953: \bibitem{TAP} D.J. Thouless, P.W. Anderson, R.G. Palmer
954: ``Solution of a `solvable model of a pin glass' '', {\em Phil.
955: Mag.} {\bf 35} 593 (1977)
956:
957: \bibitem{Vi} N.J. Vilenkin, {\em Special functions and the theory
958: of group representation, Translations of Mathematical Monographs},
959: {\bf 22}, American Math. Soc., Providence, Rhode Island, U.S.A.
960:
961:
962:
963:
964: \end{thebibliography}
965:
966:
967: \end{document}
968: