1: \documentstyle[prl,aps,floats,twocolumn,graphics]{revtex}
2:
3: \author{F. Krzakala and O.C. Martin}
4: \title{Spin and link overlaps in 3-dimensional spin glasses}
5: \address{Laboratoire de Physique Th\'eorique et Mod\`eles Statistiques,
6: b\^at. 100, Universit\'e Paris-Sud, F--91405 Orsay, France.}
7: \date{\today}
8: \draft
9:
10: \begin{document}
11:
12: \maketitle
13: \begin{abstract}
14: Excitations of three-dimensional spin glasses are computed numerically.
15: We find that one can flip a finite fraction of an $L~\times~L~\times~L$
16: lattice with an $O(1)$ energy cost, confirming the mean field picture
17: of a non-trivial spin overlap distribution $P(q)$. These low energy
18: excitations are not domain-wall-like, rather they are topologically
19: non-trivial and they reach out to the boundaries of the lattice.
20: Their surface to volume ratios decrease as $L$ increases and may
21: asymptotically go to zero. If so, link and window overlaps between
22: the ground state and these excited states become ``trivial''.
23: \end{abstract}
24:
25: \pacs{75.10.Nr, 75.40.Mg, 02.60.Pn}
26:
27: %\paragraph*{Introduction ---}
28: Spin glasses~\cite{Young98} are currently at the center of a hot debate.
29: One outstanding question is whether there exists
30: macroscopically different valleys whose
31: contributions simultaneously dominate the partition function.
32: At zero temperature, given the ground state configuration,
33: this leads one to ask whether
34: it is possible to flip a finite fraction of the spins and
35: reach a state with excess energy $O(1)$. From a mean-field
36: perspective~\cite{MezardParisi87b},
37: one expects this to be true since it happens in the
38: Sherrington-Kirkpatrick (SK) model. However it
39: is very unnatural in the context of the droplet~\cite{FisherHuse88}
40: or scaling~\cite{BrayMoore86} approaches where the characteristic
41: energy of an excitation grows with its size.
42: Recently it has been suggested that
43: the energy of an excitation may grow with its size $\ell$ as in
44: the droplet scaling law, $E(\ell)\approx \ell^{\theta_l}$, but only for
45: $\ell \ll L$, and that for $\ell=O(L)$ the energies cross
46: over to a different law,
47: $E \approx L^{\theta_g}$, where $L$ is the size of the
48: system~\cite{HoudayerMartin00b}.
49: The first exponent, $\theta_l$ ($l$ for local), may be given by
50: domain wall estimates, $\theta_l \approx 0.2$, while
51: the second exponent, $\theta_g$ ($g$ for global), could be given by the
52: mean field prediction, $\theta_g = 0$.
53: In this ``mixed'' scenario,
54: one has coexistence of the droplet model at finite length scales
55: and a mean-field behavior (if $\theta_g = 0$) for
56: system-size excitations ($\ell \approx L$ for which a finite fraction
57: of all the spins are flipped).
58:
59: The purpose of this article is to provide numerical evidence that such a
60: mixed scenario is at work in the three-dimensional
61: Edwards-Anderson
62: %Edwards-Anderson~\cite{EdwardsAnderson75}
63: spin glass. We have determined ground states
64: and excited states for different lattice sizes and have
65: analyzed their geometrical properties. The qualitative picture
66: we reach is that indeed $\theta_g \approx 0$.
67: System-size constant energy excitations are not
68: artefacts of trapped domain walls caused by periodic
69: boundary conditions, they are intrinsic to
70: this kind of frustrated system. The energy landscape of the
71: Edwards-Anderson model then consists
72: of many valleys, probably separated by large energy barriers.
73: Extrapolating to finite temperature, this picture leads to
74: a non-trivial equilibrium spin overlap
75: distribution function $P(q)$.
76:
77: Given the geometric
78: properties of our excitations, we suggest a new scenario
79: for finite dimensional spin glasses: if
80: the surface to volume ratios of these large
81: scale excitations go to zero in the large $L$ limit,
82: then the replica symmetry
83: breaking will be associated with a {\it trivial} link
84: overlap distribution function $P(q_l)$.
85: % (tending towards a delta function in the large volume limit).
86: We have coined this scenario TNT for
87: trivial link overlaps yet non-trivial spin
88: overlaps. Such a departure from the standard mean field picture
89: might hold in any dimension $d \ge 3$.
90:
91: \paragraph*{The spin glass model ---}
92: We consider an Edwards-Anderson Hamiltonian
93: on a three-dimensional $L \times L \times L$ cubic lattice:
94: \begin{equation}
95: \label{eq_H_EA}
96: H_J(\{S_i\}) = - \sum_{<ij>} J_{ij} S_i S_j .
97: \end{equation}
98: The sum is over all nearest neighbor spins of the lattice. The
99: quenched couplings $J_{ij}$ are independent random variables,
100: taken from a Gaussian distribution of zero mean
101: and unit variance. For the boundaries,
102: we have imposed either periodic or free boundary conditions.
103: Although in simulations of
104: most systems it is best to use periodic boundary conditions so
105: as to minimize finite size corrections,
106: the interpretation of our data is simpler
107: for free boundary conditions. It may also be useful to note
108: that if boundary conditions matter in the infinite
109: volume limit, free boundary conditions are the experimentally
110: appropriate ones to use.
111:
112: \paragraph*{Extracting excited states ---}
113: The problem of finding the ground state of a spin glass
114: is a difficult one. In this study we use a previously
115: tested~\cite{HoudayerMartin99b} algorithmic procedure
116: which, given enough computational ressources,
117: gives the ground state
118: with a very high probability for lattice sizes up to
119: $12 \times 12 \times 12$. (Since our $J_{ij}$s are continuous,
120: the ground state is unique up to a global spin flip.)
121: Our study here is limited to sizes $L \le 11$;
122: then the rare errors in obtaining the ground states
123: are far less important than our statistical errors
124: or than the uncertainties in extrapolating our results
125: to the $L \to \infty$ limit.
126:
127: Our purpose is to extract low-lying excited states to see whether
128: there are valleys as in the mean field picture or whether the characteristic
129: energies of the lowest-lying large scale excitations grow with $L$ as
130: expected in the droplet/scaling picture. Ideally,
131: one would like to have a list of all the states whose excess
132: energy is below a given cut-off. However, because there is a
133: non-zero density of states associated with droplets (localized
134: excitations), this is an impossible task for the sizes of
135: interest to us. Thus instead we extract our
136: excitations as follows. Given the
137: ground state (hereafter called $C_0$), we
138: choose two spins $S_i$ and $S_j$ at random and force their
139: relative orientation to be opposite from what it
140: is in the ground state. This constraint can be implemented by replacing
141: the two spins by one new spin giving the orientation of the first spin,
142: the other one being its ``slave''.
143: We then solve for the ground state $C$ of this modified spin glass.
144: The new state $C$ is necessarily distinct from $C_0$ as at least
145: one spin ($S_i$ or $S_j$) is flipped. That flipped spin may drag along
146: with it some of its surrounding spins, forming a droplet of
147: characteristic energy $O(1)$. In the droplet picture, this
148: is all that happens in the infinite volume limit. However if there
149: exist large scale excitations with $O(1)$ energies,
150: then $C$ may be such
151: an excitation if its energy is below that of {\it all} the droplets
152: containing either $S_i$ or $S_j$.
153:
154: \paragraph*{Statistics of cluster sizes ---}
155: Let $V$ be the number of sites of the cluster defining
156: the spins that are flipped when going from $C_0$ to $C$
157: (by symmetry, $V$ is taken in $\lbrack 1, L^3/2 \rbrack$).
158: If $P(V)$ is the probability to have an event of size $V$,
159: the droplet and mean field
160: pictures lead us to the following parametrization:
161: \begin{equation}
162: \label{eq_P_V}
163: P(V) = (1-\alpha) P_l(V) + \alpha P_g(V/L^3) .
164: \end{equation}
165: Here, $P_l$ and $P_g$ are
166: normalized probability distributions associated
167: with the droplet events ($V$ fixed, $L \to \infty$)
168: and the global events ($V = O(L^3)$).
169: If large scale excitations
170: have energies $O(L^{\theta_g})$, the ratio
171: $\alpha / (1 - \alpha )$ of the two contributions
172: should go as $L^{-\theta_g}$.
173: In the droplet/scaling picture, the global part
174: decreases as $L^{-\theta_l}$; that
175: is slow since $\theta_l \approx 0.2$.
176: In contrast, in the mean field
177: scenario, both the $V$ finite and the $V$ growing
178: linearly with $L^3$ contributions converge with non-zero weights,
179: $0 < \alpha < 1$, albeit with $O(L^{-\theta_l})$ finite
180: size corrections.
181:
182: Given that the usable range in $L$ is no more than a
183: factor of two so that $L^{-\theta_l}$ does not vary much,
184: measurements of $P(V)$ on their
185: own are unlikely to provide stringent tests.
186: Nevertheless, consider the probability
187: $Q(v,v')$ that $V/L^3$ is in the interval $\lbrack v,v' \rbrack$.
188: Up to finite size corrections,
189: $Q(v,v') = \alpha \int_{v}^{v'} P_g(x) dx$.
190: In our computations, we
191: have used $5 \le L \le 11$, averaging for each $L$ over
192: 2000 to 10000 randomly generated samples of the
193: $J_{ij}$. For each sample, we determined the ground state, and then
194: obtained $3$ excitations by choosing successively at random
195: $3$ pairs of spins ($S_i$,$S_j$).
196: We find that $Q(v,v')$ decreases slowly with $L$ for both
197: periodic and free boundary conditions, as expected in
198: the droplet and mean field pictures. Because
199: $\theta_l$ is small, when we perform fits of the form
200: $Q(v,v') = A + B L^{-\mu}$, we are not able
201: to exclude $A = 0$ nor $A \ne 0$ with any significant
202: confidence, so a more refined method of analysis is
203: necessary: we will thus consider the geometrical properties
204: of the events.
205:
206: Before doing so, note that the statistical error on
207: $Q(v,v')$ depends on the number of large scale events found
208: in the $\lbrack v, v' \rbrack$ interval.
209: If the spin $S_i$ or $S_j$ has a small local field,
210: there is a good chance that the corresponding event will
211: have $V=1$, thereby reducing the statistics of the
212: ``interesting'' events. To amplify our signal of
213: large $V$ events, we did not consider such spins and focused
214: our attention on spins in the top $25$ percentile when
215: ranked according to their local field. All of our data
216: was obtained with
217: that way of selecting $S_i$ and $S_j$. (Naturally, $P(V)$ and
218: $Q(v,v')$ depend on this choice, but the general
219: behavior should be the same for any choice.)
220:
221: \paragraph*{Topological features of the clusters ---}
222: Our claim that $\theta_g < \theta_l$ can be credible
223: only if our large scale excitations
224: are different from domain-walls
225: (whose energies are believed to
226: grow as $L^{0.2}$). It is thus useful to
227: consider geometrical characterizations of the excitations
228: generated by our procedure. Figure~\ref{fig_cuboid}
229: shows a typical cluster found for a $12^3$ lattice. It contains 622 spins
230: and its (excitation) energy is 0.98 which is $O(1)$.
231: The example displayed is for free boundary conditions which
232: permits a better visualization than periodic boundary
233: conditions.
234:
235: The cluster shown touches
236: many of the 6 faces of the cube, and the same is true
237: for the complement of that cluster.
238: \begin{figure}
239: \begin{center}
240: \resizebox{0.7\linewidth}{!}{\includegraphics{fig_cuboid.eps}}
241: \end{center}
242: \caption{Example of excitation found for a $12^3$ lattice with
243: free boundary conditions.}
244: \label{fig_cuboid}
245: \end{figure}
246: Such a cluster has a very non-trivial topology and is thus very far
247: from being domain-wall like. This motivates the following
248: three-fold classification of the events
249: we obtain when considering free boundary
250: conditions. In the first class, a cluster
251: and its complement touch all $6$ faces of the cube. In the
252: second class, a cluster touches at most $3$ faces
253: of the cube. The third class consists of all other events.
254: Finite size droplets should asymptotically always fall into
255: the second class, albeit with finite size corrections
256: of order $L^{-\theta_l}$.
257:
258: Does the first class constitute a non-zero
259: fraction of all events?
260: At finite $L$, we find the following fractions: $23.3\%$ $(L=5)$,
261: $23.9\%$ $(L=6)$, $25.1\%$ $(L=7)$, $24.4\%$
262: $(L=8)$, $25.0\%$ $(L=9)$, $25.7\%$ $(L=10)$,
263: and $26.0\%$ $(L=11)$.
264: The trend of these numbers suggests that the first class does
265: indeed encompass a finite fraction
266: of all the events when $L \to \infty$. We also considered
267: the scaling of cluster sizes with $L$.
268: Fig.~\ref{fig_touch} shows $Q(v,1/2)$
269: as a function of $v=V/L^3$,
270: restricted to events belonging to the first class.
271: (The $v=0$ values are the fractions
272: we just gave above.) The curves for different $L$ show
273: a small drift, $Q(v,1/2)$ {\it growing} with $L$.
274: We consider this drift to be a finite size effect and
275: that the correct interpretation of our data is
276: $\theta_g \approx 0$, in agreement with the mean field picture.
277: Our conclusion is then that
278: as $L \to \infty$, there is a finite probability of having an
279: $O(1)$ energy excitation that is non-domain-wall like,
280: the cluster and its complement touching {\it all} faces of
281: the cube.
282:
283: \begin{figure}
284: \begin{center}
285: \resizebox{0.7\linewidth}{!}{\includegraphics{fig_Q1class.pstex}}
286: \end{center}
287: \caption{Integrated probability $Q(v,1/2)$ of events in the first
288: class. (From bottom to top, $L=5, 7, 9, 11$.}
289: \label{fig_touch}
290: \end{figure}
291:
292: \paragraph*{Surface to volume ratios ---}
293: Obviously the mean field picture obtained by extrapolating results
294: from the SK or Viana-Bray spin glasses
295: cannot teach us anything about the topology
296: of excitations for three-dimensional
297: lattices. But mean field may serve
298: as a guide for other properties such as the link overlap
299: $q_l$ between ground states and excited states. In the SK
300: model, the {\it spin} overlap
301: $q \equiv \sum S_i S_i' / N$ and the {\it link} overlap
302: $q_l\equiv \sum (S_i S_j)(S_i' S_j') / (N (N-1)/2)$ satisfy
303: $q_l = q^2$, and both $q$ and $q_l$ have non-trivial distributions.
304: Extrapolating this to our three dimensional system, the mean field
305: picture predicts that the clusters associated with large scale
306: excitations both span the whole system (as we saw with
307: free boundary conditions) and are {\it space filling}.
308: Quantitatively, this implies that their surface grows as the
309: total volume of the system, {\it i.e.}, as $L^3$.
310:
311: To investigate this question, we have measured the surface
312: of our excitations, defined as the number $S$ of links connecting
313: the corresponding cluster to its complement.
314: (Then $q_l = 1- 2 S /3 L^3$.)
315: In figure~\ref{fig_surface} we show the
316: mean value of $S / L^3$ as a function of $L$ for
317: $v=V/L^3$ belonging to the three intervals
318: $\rbrack 0.20, 0.25 \rbrack$,
319: $\rbrack 0.30, 0.35 \rbrack$, and
320: $\rbrack 0.40, 0.45 \rbrack$.
321: The data shown are for free boundary
322: conditions, but the results are very similar
323: for periodic boundary conditions. The most striking feature is that the
324: curves decrease very clearly with
325: $L$. For each interval, we have fitted the data to
326: $\langle S \rangle / L^3 = A + B/L^{\mu}$ and to a
327: polynomial in $1/L$. Of major interest is the value of the constant
328: because it gives the large $L$ limit of the curves.
329:
330: Table~\ref{tab_chi2s} summarizes the quality of the
331: fits as given by their $\chi^2_r$ (chi squared per degree of
332: freedom). In all cases the fits are reasonably good; this
333: is not so surprizing because our range of $L$ values is small.
334: The most reliable fits are obtained using a quadratic polynomial in $1/L$,
335: this functional form leading to a smooth and
336: monotone behavior of the parameters
337: and to small uncertainties in the parameters.
338: For the large $L$ limits, these fits give
339: $A=0.22$, $A=0.27$ and $A=0.30$ for the three intervals.
340: (We do not give results for the
341: linear fits which on the contrary are very poor.) The constant
342: plus power fits also have good $\chi^2_r$ but the $A$s obtained were
343: small and {\it decreased} with $v$; also they
344: had large uncertainties and seemed to be compatible with $A=0$. Because
345: of this, we also performed fits of the form
346: $\langle S \rangle / L^3 = B/L^{\mu}$. These are displayed in
347: Fig.~\ref{fig_surface} and lead to $\mu \approx 0.30$ (the
348: exponent varies little from curve to curve),
349: again with reasonable $\chi^2_r$s. Because of this, we feel we
350: cannot conclude from the data that the surface to volume ratios
351: tend towards a non-zero asymptote. What can be said is that
352: this asymptote seems to be small, and that it will be difficult to
353: be sure that it is non-zero without going to larger
354: values of $L$.
355:
356: \begin{figure}
357: \begin{center}
358: \resizebox{0.7\linewidth}{!}{\includegraphics{fig_surf.pstex}}
359: \end{center}
360: \caption{Mean value of surface to $L^3$ ratios for
361: $v=V/L^3$ in intervals around $0.225, 0.325, 0.425$ (bottom to top) using
362: free boundary conditions. Curves are pure power fits.}
363: \label{fig_surface}
364: \end{figure}
365:
366: \begin{table}
367: \begin{center}
368: \begin{tabular}{cccc}
369: Interval & $A + B/L^{\mu}$ & $A + B/L + C/L^2$ & $B/L^{\mu}$\\
370: \hline
371: $\rbrack 0.20,0.25 \rbrack$ & 0.6 & 0.6 & 2.0\\
372: $\rbrack 0.30,0.35 \rbrack$ & 1.1 & 1.1 & 1.5 \\
373: $\rbrack 0.40,0.45 \rbrack$ & 0.7 & 0.9 & 0.6 \\
374: \end{tabular}
375: \end{center}
376: \caption{Chi squared per degree of freedom for the fits to the data
377: of Fig.~\ref{fig_surface}.}
378: \label{tab_chi2s}
379: \end{table}
380:
381:
382: \paragraph*{Discussion ---}
383: For the three dimensional Edwards-Anderson spin glass model,
384: we have presented numerical evidence
385: that it is possible to flip a finite
386: fraction of the whole lattice at an energy cost of $O(1)$, corresponding
387: to $\theta_g \approx 0$ as predicted by mean field. This
388: property transpired most clearly through the use of
389: free boundary conditions, allowing one to conclude that
390: $\theta_g \approx 0$ is not an artefact of trapped domain walls caused by
391: periodic boundary conditions. Extrapolating to finite
392: temperature, we expect the equilibrium $P(q)$ to be non trivial as in the
393: mean field picture.
394:
395: The other messages of our work concern the nature of
396: these large scale excitations whose energies are $O(1)$. First,
397: using free boundary conditions, we found them to be
398: topologically highly non-trivial:
399: with a finite probability they reach the boundaries
400: on all 6 faces of the cube. Thus they are not domain-wall-like,
401: rather they are sponge-like. Second, our data (both for
402: periodic and free boundary conditions) indicate very clearly that
403: their surface to volume ratios decrease as $L$ increases. The
404: most important issue here is whether or not these ratios decrease to zero
405: in the large $L$ limit. Although our data are
406: compatible with a non-zero limiting value as predicted
407: by mean field, the fits were not conclusive
408: so further work is necessary.
409:
410: If the surface to volume ratios turned out to go to zero, we would
411: be lead to a new scenario that we have coined ``TNT''. In the standard
412: mean field picture, the surface to volume ratios cannot go
413: to zero; indeed
414: in the SK and Viana-Bray spin glass models
415: there are {\it no} spin clusters with
416: surface to volume ratios going to zero.
417: However, in finite dimensions, one can have
418: surface to volume ratios going to zero, in which case
419: $q_l \to 1$. This property would then lead to a non-trivial
420: $P(q)$ but to a trivial $P(q_l)$.
421: This trivial-non-trivial (TNT)
422: scenario does not seem to have been proposed previously.
423:
424: Perhaps the most dramatic consequence of this new scenario is for window
425: overlaps in spin glasses: because in TNT one is asymptotically always
426: in the bulk of an excitation,
427: correlation functions at any finite distance will show
428: no effects of replica symmetry breaking. That this may arise in
429: fact is supported by work by Palassini and
430: Young\cite{PalassiniYoung99a}
431: who showed that certain window overlaps seemed to become
432: trivial as $L \to \infty$.
433: (See also~\cite{Middleton99} for a similar discussion in
434: two-dimensions.) These authors
435: referred to this property as evidence for
436: a ``trivial ground state structure''.
437: But in our picture the {\it global} (infinite distance)
438: structure is non-trivial, as indicated by
439: $\theta_g = 0$, in sharp contrast to the droplet/scaling
440: picture. Also, in very recent work~\cite{YoungPrivate},
441: Palassini and Young have extended their previous investigations
442: and have extracted excited states by a quite different
443: method from ours, and
444: they find that their data is compatible with
445: the TNT scenario.
446: Naturally, there is also evidence in {\it favor}
447: of the non-triviality
448: of window overlaps~\cite{MarinariParisi99b}.
449: Nevertheless, we believe that our mixed scenario is
450: a worthy candidate to describe the physics of short range
451: spin glasses. Furthermore, its plausibility should not
452: restricted to $3$ dimensions, it could hold in
453: all dimensions greater than $2$. (Note that in $d=2$, excitations
454: are necessarily topologically trivial.) An important indication
455: of this was obtained by Palassini and Young whose
456: computations~\cite{YoungPrivate} favor the TNT
457: scenario over the droplet picture
458: in the 4-dimensional Edwards-Anderson model.
459:
460: %Note: systematic errors due to not obtaining the true ground state
461: %are far smaller than statistical errors or uncertainty in the
462: %extrapolation to large sizes.
463:
464: \paragraph*{Acknowledgements ---}
465: We thank J.-P. Bouchaud and M. M\'ezard for very stimulating
466: discussions and for their continuous encouragement, and
467: M. Palassini and A.P. Young for letting us know about their
468: work before publication. Finally, we thank
469: J\'er\^ome Houdayer; without his superb work on the genetic
470: renormalization approach~\cite{HoudayerMartin99b}, this numerical
471: study would not have been possible.
472: F.K. acknowledges support from the
473: MENRT, and O.C.M. acknowledges support from the Institut Universitaire de
474: France. The LPTMS is an Unit\'e de Recherche de
475: l'Universit\'e Paris~XI associ\'ee au CNRS.
476:
477: \bibliographystyle{prsty}
478: \bibliography{../../../Bib/references}
479:
480: \end{document}
481:
482: