1: \documentclass[letterpaper,twocolumn,english,floatfix, aps, pre,showpacs]{revtex4}
2: \usepackage[T1]{fontenc}
3: \usepackage[latin1]{inputenc}
4: \usepackage{babel}
5: \usepackage{graphics}
6:
7: \makeatletter
8: \usepackage{graphicx}
9: \usepackage{dcolumn}
10: \usepackage{bm}
11:
12: \makeatother
13: \begin{document}
14:
15: \title{Percolation of frozen order in glassy combinatorial problems}
16:
17:
18: \author{P.M. Duxbury}
19:
20:
21: \email{duxbury@pa.msu.edu}
22:
23:
24: \affiliation{Dept. of Physics \& Astronomy,
25: Michigan State University, East Lansing, MI 48824, USA}
26:
27: \begin{abstract}
28: A local order parameter which is important in the analysis of
29: phase transitions in frustrated combinatorial problems
30: is the probability that a node is frozen
31: in a particular state. There is
32: a percolative transition when an infinite connected cluster
33: of these frozen nodes emerges. In this contribution, we
34: develop theories based on this percolation process
35: and discuss its relation to conventional
36: connectivity percolation.
37: The emergence of frozen order may also be
38: considered to be a form of constraint
39: percolation (CP) which enables us to draw analogies with
40: rigidity percolation and its associated matching problems.
41: We show that very simple
42: CP processes on Bethe lattices
43: lead to the replica symmetric equations
44: for KSAT, coloring and the Viana-Bray model.
45: \end{abstract}
46:
47: \pacs{89.75.-k, 05.50.+q,75.10.Nr}
48:
49: \maketitle
50:
51: \section{Introduction}
52:
53: There is intense
54: interest in the relations
55: between statistical physics
56: and computational complexity, from
57: both the computer science and
58: physics communities\cite{mezard87,dubois01}. This activity
59: has resulted in the application of
60: physics methods to computer science \cite{monasson99}
61: and clever extensions of computer
62: science methods to glassy problems\cite{mezard02a}.
63: One fascinating result which has
64: emerged from these studies
65: is the existence of phase transitions
66: in computational complexity. These
67: phase transitions are continuous
68: in some cases and are discontinuous
69: in others\cite{kirkpatrick94}.
70: More recently the $k-$connectivity
71: and $k-$core\cite{pittel96} problems
72: have attracted interest, for example
73: in designing redundant
74: networks\cite{kirkpatrick02}.
75: $k-$connectivity is the generalisation
76: of the conventional connectivity
77: percolation problem to the requirement
78: of $k-$fold connectivity. That is,
79: a graph is $k$-connected if for each pair of vertices in the graph
80: there exist at least $k$
81: mutally independent paths connecting them.
82: The $k-$core of a graph
83: is the largest subgraph with minimum vertex
84: degree $k$.
85: The Bethe lattice equations for the $k$-core
86: were actually first derived in the
87: context of $k$-bootstrap percolation\cite{chalupa79}.
88: $k$-bootstrap percolation is the percolation
89: process found by recursively deleting all nodes
90: which have connectivity less than $k$.
91: More recently the Bethe lattice $k-$core
92: equations have been used to develop theories for rigidity
93: percolation \cite{moukarzel97,duxbury99,moukarzel03}.
94: In this paper, we give a brief
95: introduction to the connectivity percolation
96: and $g-$ rigidity equations
97: on Bethe lattices and then describe similar percolation
98: processes which are
99: important in the Viana-Bray spin-glass model,
100: the coloring problem and the K-SAT problem.
101: For these problems we develop equations
102: for the probability that a infinite frozen
103: cluster emerges. We then show that in the
104: simplest approximation, this formalism reproduces the
105: replica symmetric equations in
106: a surprisingly straightforward manner.
107:
108: Frozen order is
109: a unifying concept in the analysis of
110: glasses and geometrically frustrated systems
111: in physics\cite{binder86} and also in NP-complete
112: problems in computer science, such as coloring\cite{culberson01}
113: and K-SAT\cite{dubois01}.
114: Frozen long-range order is most
115: easily understood at zero temperature.
116: At zero temperature the paradigm geometry is to
117: fix the variables on a surface of the
118: system and then to test whether
119: these frozen degrees of freedom cause the
120: propagation of frozen order
121: into the bulk of a sample. A spin
122: is frozen only if the spin
123: is fixed or constrained by the spin configurations
124: of its neighbors, as we shall demonstrate
125: explicitly below using the Viana-Bray model.
126: Frozen order may occur even though
127: the variables (e.g. the spins in a spin glass)
128: at each vertex of a graph look random.
129: Furthermore, not all of the variables in the system
130: need to freeze. However for the system
131: to be in the frozen ordered ground state, the frozen component
132: must percolate.
133:
134: The vertex q-coloring problem is equivalent to finding the
135: ground state of the q-state Potts antiferromagnet\cite{wu82}. Each node
136: of a complex graph may have any one of $q$ colors.
137: The objective is to find the color configuration which
138: minimizes the number of edges which have the same
139: color at each end. The propagation of
140: frozen color has many conceptual similarities with
141: the propagation of rigidity in central force networks\cite{moukarzel95,jacobs96}.
142: However there is a key difference which makes the
143: coloring problem NP-complete whereas the rigidity
144: problem is polynomial. The key difference is
145: that the constraints in coloring
146: are distinguishable while the constraints
147: in rigidity percolation are not.
148:
149: Spin glasses and many frustrated antiferromagnets
150: map exactly to problems in the NP-complete
151: class\cite{mezard87}. NP-complete problems are of central
152: interest in computer science (CSE)\cite{garey79} and have
153: motivated many attempts to design quantum
154: algorithms for their efficient solution.
155: The phase transitions which physicists study
156: often correspond to a change in the computational
157: complexity of the corresponding CSE problem.
158: Since these problems are of enormous interest
159: in physics, CSE and also in practical applications
160: it is not surprising that there is a burgeoning
161: of efforts to better understand the phase
162: transition which occurs in NP-complete problems.
163:
164: It is necessary to consider the effects of randomness
165: on physics problems as randomness is present in
166: most magnetic and electronic materials. The CSE interest
167: in random instances is from a different perspective.
168: The motivation is to find ``typical'' problem instances
169: which are then used to test the algorithmic complexity of new algorithms.
170: A result of broad importance is the observation of
171: a phase transition in computational complexity in
172: random satisfiability problems\cite{kirkpatrick94}. The key quantity
173: is the ratio of the number of constraints, $M$, to the
174: number of variables, $N$, and this ratio is $\alpha = M/N$.
175: For $\alpha <\alpha_*$ it is believed that random SAT problems
176: are almost surely in P, while for $\alpha >\alpha_*$
177: random SAT problems are almost surely in NP. In
178: addition there is a phase transition as measured
179: by the number of violated clauses in the
180: optimal solution. For $\alpha < \alpha_c$, the
181: number of violated clauses is of order one, while
182: for $\alpha > \alpha_c$ the number of violated
183: clauses is of order $N$. It is believed that
184: $\alpha_c \ge \alpha_*$.
185:
186: The physics community has applied the replica method
187: to NP combinatorial problems
188: with remarkable success\cite{fu86,kanter87,monasson97,
189: monasson99,hartmann01,mourik02,mulet02}. In addition
190: new algorithms have been developed based on a combination
191: of replica symmetry breaking ideas from physics and
192: belief propagation ideas from the artificial intelligence
193: community\cite{mezard02a,mezard02b,braunstein03}.
194: Though the replica method is an excellent
195: tool, it is quite difficult both technically
196: and intuitively. We show that a simple
197: combinatorial procedure based on percolation ideas
198: can reproduce many of the successes of the replica
199: method. The percolation process occuring at the
200: phase transition can be thought of as either
201: percolation of constraint or percolation of
202: frozen order. In this contribution, we
203: derive the replica symmetric theories
204: for K-SAT, the Viana-Bray model and coloring
205: using percolation concepts.
206:
207: The next section of the paper
208: gives a brief review of the
209: analysis of connectivity percolation on
210: Bethe lattices and random graphs, and
211: also describes its extension to k-connectivity
212: percolation. Section III describes the analysis of the
213: glass transition, at $T=0$, in the Viana-Bray model.
214: Section IV focuses on the coloring problem, while
215: Section V presents an analysis of K-SAT. Section
216: VI contains a brief summary.
217:
218:
219:
220: \section{Connectivity and Rigidity percolation}
221:
222: Percolation on diluted Bethe lattices was analysed by Fisher and Essam\cite{fisher61},
223: who defined the probability that a node is part of the
224: infinite cluster, $T$. They found that
225: the probability that a node
226: is not on the infinite cluster, $Q=1-T$,
227: only requires that all of its connected neighbors also
228: not be part of the infinite cluster, so that,
229: \begin{equation}
230: Q = (1-p(1-Q))^{\alpha}
231: \end{equation}
232: where $p$ is the probability that an edge is present in the
233: Bethe lattice, and $\alpha=z-1$, where $z$ is the
234: co-ordination number of the Bethe lattice.
235: Note that this expression may be written as,
236: \begin{equation}
237: T = \sum_{l=1}^{\alpha} {\alpha \choose l} (pT)^{l}(1-pT)^{\alpha-l}
238: \end{equation}
239: which is more convenient for the generalisation to
240: rigidity percolation.
241: From Eq.~(1), it is easy to show that there is a phase
242: transition at $p_c=1/\alpha$ and that $T\sim (p-p_c)$ near
243: the critical threshold. The phase transition is thus
244: continuous with order parameter exponent one. Somewhat
245: earlier, this transition was also studied in the
246: graph theory community by Erd\"os and R\'enyi\cite{erdos60}. They
247: concentrated on random graphs, which consist of
248: highly diluted complete graphs. A complete
249: graph is a graph where every node is
250: connected to every other node. In fact they
251: defined $p=c/N$, where $c$ is finite and showed
252: that a giant (extensive) connected cluster emerges at $c=1$.
253: They derived an equation for the probability that a node
254: is on the giant cluster, $\gamma$. Their
255: equation is found from Eq.~(2), by taking
256: the limit $p=c/N,\ N = z\rightarrow \infty$, to
257: find $\gamma = 1-e^{-c \gamma}$. Near
258: the critical point $\gamma \sim 2(c-1)/c^2$ so, as
259: expected based on the universality hypothesis,
260: $\gamma$ also has an order parameter exponent of one.
261:
262: Rigidity percolation on Bethe lattices,
263: is described by a simple generalisation of Eq.~(2).
264: In this generalisation, each node has $g$ degrees of freedom.
265: For example if we wish to model rigidity percolation
266: on central force networks, then $g=d$, where $d$ is the
267: lattice dimension. In order to make
268: a giant $g$-rigid cluster, we need to constrain
269: the $g$ degrees of freedom at each node with at least $g$ bonds, so
270: we generalise Eq.~(2) to,
271: \begin{equation}
272: T_g = \sum_{l=g}^{\alpha} {\alpha \choose l} (pT_g)^{l}(1-pT_g)^{\alpha-l}
273: \end{equation}
274: which is the simple generalisation of
275: Eq.~(2) to the requirement of at least $g-$
276: neighbor connections.
277:
278: Eq.~(3) was first invented
279: in the context of a Bethe lattice theory for
280: Bootstrap percolation\cite{chalupa79} and
281: has been used more recently to develop a Bethe lattice
282: theory for rigidity percolation \cite{moukarzel97,duxbury99,moukarzel03}.
283: In the random graph limit, Eq.~(3) reduces to,
284: \begin{equation}
285: \gamma_g = 1 - e^{-c\gamma_g} \sum_{l=0}^{g-1} {(c \gamma_g)^l \over l!}
286: \end{equation}
287: When $g=1$ this gives the Erd\"os-R\'enyi result\cite{erdos60} for the
288: emergence of a giant cluster in random graphs, while
289: for $g>1$, there is a discontinuous onset of a finite
290: solution at a sharp threshold $c_g$\cite{moukarzel97}.
291: Numerical solution of Eq.~(4) indicates that for $g=2$,
292: $c_2=3.3510(1)$. This value has
293: also been found in a recent mathematical analysis\cite{pittel96}
294: of the threshold for the emergence of
295: the giant 3-core on random graphs. The k-core
296: problem is equivalent to the k-bootstrap percolation problem.
297: However the k+1-core is, in general
298: different than the k-rigidity problem, and even
299: on Bethe lattices and random graphs there are some
300: important differences.
301:
302: The most important difference is that
303: for $g$-rigidity, the finite solution $T_g$ is metastable
304: for a range of $c>c_g$\cite{moukarzel97,duxbury99}.
305: The true rigidity transition actually sets
306: in at $c_r >c_g$ and is identified using
307: constraint counting arguments \cite{duxbury99,moukarzel03}.
308: Nevertheless the probability
309: of being on the infinite rigid cluster is
310: correctly found from Eq.~(4), provided $c>c_r$, where
311: $c_r$ is the rigidity threshold\cite{duxbury99,moukarzel03}.
312: As we shall see below
313: the analogous theories for glassy combinatorial problems,
314: in particular the Viana Bray model, K-SAT and q-coloring,
315: provide solutions at the level of the replica symmetric
316: theory. Moreover, as will be described elsewhere,
317: the methodology we introduce here
318: can be used to develop simple and accurate recursive algorithms
319: for these glassy problems on general graphs. In the case
320: of first order transitions, as occurs in
321: q-coloring (with $q\ge 3$) and for K-SAT ($K\ge 3$),
322: the transition point we find below may mark
323: the onset of metastability. In order to
324: find the true threshold we need
325: to find the ground state energy from the
326: order parameter, in a manner similar
327: to that used in rigidity percolation.
328: This is non trivial and
329: will be elucidated elsewhere.
330:
331: \section{Viana-Bray model}
332:
333: We first analyse the onset of frozen order
334: in the Viana-Bray(VB) spin-glass model\cite{viana85},
335: which provides a basic model for disordered and
336: frustrated magnets, such as $Eu_xSr_{1-x}S$\cite{maletta79}.
337: The Hamiltonian for the VB model is,
338: \begin{equation}
339: H = \sum_{ij} J_{ij} S_i S_j
340: \end{equation}
341: where $S_i=\pm 1$.
342: The exchange constants $J_{ij}$ are randomly drawn from the
343: distribution,
344: \begin{equation}
345: D_p(J_{ij}) = p[{1\over 2}\delta(J_{ij}+J) + {1\over 2}\delta(J_{ij}-J)] + (1-p) \delta(J_{ij}),
346: \end{equation}
347: As above we focus on the random graph limit $p = c/N$.
348: We introduce the following probabilities:
349:
350: $P$ = probability a site is frozen in the up state
351:
352: $M$ = probability a site is frozen in the down state
353:
354: $D$ = probability a site is degenerate
355:
356: \noindent In the absence of an applied field and within
357: a symmetric assumption, $P=M$ and $D = 1-2M$. We then need
358: consider only one of these probabilities. However for clarity and
359: for ease of generalisation, we continue to include $M$ and $P$ separately.
360: In terms of these order parameters, the magnetisation is
361: given by, $m = |P-M|$ and the spin glass order parameter is,
362: $q = P+M$. The recurrence formula for $P$, using $p=c/N$ is,
363: $$
364: P = \sum_{k=0}^{\alpha} \sum_{l=k+1}^{\alpha}{\alpha!\over k!l!(\alpha-k-l)!}$$
365: \begin{equation}
366: ({c P\over 2 N} + {c M\over 2 N})^k
367: ({c M\over 2 N} + {c P\over 2 N})^l
368: (1- {c\over N}(M+P))^{\alpha - k - l}
369: \end{equation}
370: This is understood as follows. If a bond connects a site
371: at the lower level to a site at the upper level then
372: the site at the upper level wants to be frozen up:
373: if the connecting bond is ferromagnetic
374: and the lower level spin is frozen up; {\it or }
375: if the connecting bond is antiferromagnetic and the
376: lower level spin is frozen down. This event has
377: probability, $ c P/2N + c M/2N$.
378: Similarly, the probability that a spin at the
379: upper level of the bond wants to be frozen down (negative)
380: is given by, $ cM/2N + c P/2 N$. The newly added spin at the
381: upper level
382: is frozen up if there is a larger number
383: of connections from the upper to the lower level
384: which prefer the frozen up state. The sum in Eq.~(7) is thus
385: restricted to events of this sort.
386: The event $(1-c(P+M)/N)$
387: is the probability that a site at the lower level
388: in the tree is either degenerate or disconnected
389: from the newly added site.
390: In the large $N$ limit, Eq.~(7) reduces to,
391: \begin{equation}
392: q = 2 e^{- c q}\sum_{k=0}^{\infty} \sum_{l=k+1}^{\infty}
393: {({cq \over 2})^{k+l}\over k! l!} = 1 - e^{-cq} I_0(cq)
394: \end{equation}
395: where we have used the fact that we are
396: considering a case where the magnetisation $m=0$.
397: In that case, $M = P = q/2$, where $q$ is the
398: spin glass order parameter. $I_0$ is the
399: spherical Bessel function of zeroth order. The result
400: (8) has been found before within
401: the replica symmetric solution to the
402: Viana-Bray (VB) model(see Eq.~(15) of \cite{kanter87}).
403: Thus symmetric constraint percolation (CP) in the VB model is
404: equivalent to the ground state spin glass transition
405: as found within the replica symmetric approach.
406: The CP approach is attractive because is it
407: is simple, it avoids
408: the mathematical difficulties of the replica method
409: and it is physically transparent. The
410: construction we have used makes it clear that
411: simple connectivity is sufficient
412: to ensure propagation of spin glass order in the VB
413: model. Constraint percolation occurs at $c=1$ and
414: the order parameter approaches
415: zero as $q \sim {4\over 3c^2}(c-1)$, so the
416: CP transition in this case is continuous, with
417: the same exponent as the Erd\"os-R\'enyi transition.
418:
419: \section{coloring}
420: Now we turn to the coloring problem.
421: Our analysis centers on the
422: probability $F_l$ ($l=1,2,..q$), which is the
423: probability that a site is frozen in color $l$.
424: The probability $F_1$ is given by the recursion relation,
425: $$
426: F_1 = \sum_{s=0}^{\alpha} \sum_{k_2=s+1}^{\infty} ..
427: \sum_{k_q=s+1}^{\infty} \sum_{k_{q+1}=0}^{\infty}
428: {\alpha! \over s! k_2! k_3!...k_q!k_{q+1}!}
429: $$
430: \begin{equation}
431: (pF_1)^{s}(pF_2)^{k_2}...(pF_q)^{k_q}(1-p\sum F_l)^{k_{q+1}}\delta (s+\sum_{l=2}^{q+1}k_l-\alpha)
432: \end{equation}
433: This formula is understood as follows.
434: In order for a site to be frozen in the color ``1'', all of the
435: other $q-1$ colors must appear, and be frozen, on one of the
436: connected neighbor sites. In addition the frozen color ``1'' must
437: occur, on these neighbor sites, a strictly smaller number of times than any other
438: frozen color. The probability that a neighbor site
439: is connected and frozen in color ``1'' is $pF_1$. This event
440: occurs $s$ times. We thus have a term $(pF_l)^{s}$ for the color ``1''.
441: A similar term applies for each of the other required $q-1$ frozen neighbor
442: colors, with each of them occuring $k_l$ times.
443: We must also allow for the possibility of events which
444: are not of the type $pF_l$, which leads to the term $(1-p\sum F_l)^{k_{q+1}}$.
445: This probability is summed from $0$ to infinity as it does not
446: have to exist in a configuration in order to ensure that $F_1$ be finite.
447: Note however that $(1-p\sum F_l)$ is by far the most likely
448: event in the random graph limit, where $p\rightarrow c/N$.
449: All of these probabilities are exclusive and independent. We must also allow
450: for all ways of arranging this set of $q+1$ exclusive events amongst the
451: $\alpha$ possible connections between our newly added site
452: and the sites at the lower level in the tree. This leads to the
453: multinomial factor. An equation like (9) occurs for each of the $q$ colors
454: which are allowed. If we assume that all colors have the
455: same probability (which is natural provided there are
456: no symmetry breaking terms),
457: then $F_1 = F_2 = F_l= F/q$. Using this,
458: and taking the random graph limit yields,
459: \begin{equation}
460: F = q e^{-cF}
461: \sum_{s=0}^{\infty} {1\over s!}({cF\over q})^s
462: [\sum_{k=s+1}^{\infty} {1\over k!} ({cF\over q})^k ]^{q-1}
463: \end{equation}
464: This equation is valid for arbitrary $q$ provided $q/N \rightarrow 0$.
465:
466: For $q=2$, we assume that $F$ is continuous near the
467: percolation threshold and expand this expression in
468: powers of $F$ which yields,
469: \begin{equation}
470: F \approx cF -{3\over 4}(cF)^2 + O((cF)^3)
471: \end{equation}
472: This has the solution,
473: \begin{equation}
474: F \approx {4\over 3c^2}(c-1) \ \ c\ge 1
475: \end{equation}
476: This is, other than a prefactor of $4/3c^2$ instead of $2/c^2$,
477: the same as the critical behavior of the giant cluster
478: probability in random graphs\cite{erdos60,fisher61}.
479: For $c$ well away from the transition, we solve
480: Eq.~(10) numerically.
481: The $s$ and $k$ sums are rapidly convergent
482: and for the $c$ range near criticality, only a
483: few terms are required for high accuracy results.
484: From the solution for $F$ we obtain all of the results
485: of interest and they are presented in Fig. 1.
486: The continuous behavior of 2-coloring near
487: threshold is evident from these data.
488:
489: \begin{figure}
490: {\centering \resizebox*{0.8\columnwidth}{!}
491: {\rotatebox{0}{\includegraphics{figures/fig1.eps}}}\par}
492: \caption{The coloring order parameters for $q=2$. The
493: lower two curves are the probability that a site is
494: frozen and colorable, $G$ (the $s=0$ term in Eq.~(10)), and the probability that
495: a site is frozen and frustrated, $H$ (the $s\ge 1$ term
496: in Eq.~(10)). The top curve is the probability that a site has a frozen
497: color $F=G+H$, which is found by solving Eq.~(10) with $q=2$. }
498: \end{figure}
499:
500: \begin{figure}
501: {\centering \resizebox*{0.8\columnwidth}{!}
502: {\rotatebox{0}{\includegraphics{figures/fig2.eps}}}\par}
503: \caption{The coloring order parameters for $q=3$. The
504: lower two curves are the probability that a site is
505: frozen and colorable, $G$ (the $s=0$ term in Eq.~(10)), and the probability that
506: a site is frozen and frustrated, $H$ (the $s\ge 1$ term
507: in Eq.~(10)). The top curve is the probability that a site has a frozen
508: color $F=G+H$, which is found by solving Eq.~(10) with $q=3$.}
509: \end{figure}
510:
511:
512: For $q=3$, an attempt to find a continuous transition
513: by expanding in powers of $F$ fails.
514: Numerical solution of
515: Eq.~(10) is presented in Fig. 2 where it is
516: seen that there is a jump discontinuity in the
517: infinite frozen cluster probability at a sharp threshold.
518: We find that $c_*=5.14(1)$ and that the jump
519: in the order parameter is $\Delta F_c = 0.365(1)$.
520: We thus find that the coloring transition for $q=3$ is first order as has been found in
521: numerical simulations\cite{culberson01} on random graphs.
522: Our coloring threshold is consistent with a recent replica symmetric
523: numerical calculation,
524: which yielded $c_* \approx 5.1$\cite{mourik02}, but is significantly higher
525: than that found in the simulation work of Culberson and Gent\cite{culberson01}
526: where $c_*\approx 4.5-4.7$ or in the numerical work on
527: survey propagation\cite{mulet02}, which yields $c_* \approx 4.42$.
528: Nevertheless the nature of the transition is
529: correctly captured by the simple CP theory.
530: It is also important to note that the
531: solution found here may also
532: be metastable for a range of $c$, as
533: was found in the rigidity case\cite{duxbury99}.
534: The onset of metastability is an important
535: threshold from the point of algorithmic efficiency,
536: as it marks the onset of glassy relaxation dynamics.
537: The coloring theory developed above can be formulated in
538: a very similar way to the formulation of the
539: propagation of the k-core. However there is
540: a critical difference. The constraints in the coloring
541: theory have to be treated as distinguishable, while the
542: constraints in the k-core calculation are indistinguishable.
543:
544:
545: \section{K-SAT}
546:
547: The satisfiability problems we consider ask the question: Given
548: a set of binary variables, $z_i = 0,1$ or equivalently
549: $z_i = True\ or False$, is it possible to
550: satisfy a specified set of constraints on these
551: variables? In the K-SAT case, a typical constraint is
552: of the form,
553: \begin{equation}
554: (z_i \wedge \overline{z}_j \wedge z_k)
555: \end{equation}
556: where $\wedge$ is the logical OR operation and the
557: overline indicates a negated variable. This
558: logical clause is satisfied (SAT) if any one of the
559: variables in the clause is SAT. The variables
560: $z_i$ and $z_k$ are SAT if they are true (T),
561: which we take to be $z_i=z_k=1$, while the
562: variable $\overline{z}_j$ is SAT when $z_j$ is
563: false (F), which corresponds to $z_j=0$. We shall
564: also fix the number of variables in each
565: clause to be $K$, which is the $K$-SAT problem.
566: In these SAT problems we shall
567: randomly choose a set of $M$ clauses like that
568: in Eq.~(13)and try to find the
569: assignment of the binary variables which
570: minimizes the number of violated clauses.
571: Each variable appearing in a clause is negated with
572: probability $1/2$ and
573: the number of variables is $N$. The key
574: ratio is $\alpha = M/N$.
575: We would like to find the threshold for
576: constraint percolation. That is, what
577: is the threshold for the appearance of
578: a giant cluster of clauses where each
579: clause is completely specified or ``frozen''. These
580: completely specified clauses cannot
581: be altered without increasing the
582: total number of violated clauses, so that
583: they are non-degenerate. There are three
584: types of clauses in an optimal assignment
585: of a formula: (i) Clauses that are SAT but are degenerate;
586: (ii) Clauses that are SAT but are frozen;
587: (iii) Clauses that are UNSAT but are degenerate.
588: Only type (ii) clauses propagate constraint.
589: We seek a formula for the probability, $V$, that
590: a variable is frozen and the probability,
591: $F$, that a clause is frozen
592: and SAT.
593:
594: We make a tree construction of the factor
595: graph for the K-SAT problem (see Fig. 3).
596: \begin{figure}
597: {\centering \resizebox*{0.8\columnwidth}{!}
598: {\rotatebox{0}{\includegraphics{figures/r1.ps}}}\par}
599: \caption{The factor graph used to construct the recurrence
600: relations. The circles denote variable
601: nodes, while the square nodes are the clause nodes.
602: $V$ is the probability that a variable node is
603: frozen, while $F$ is the probability that a clause
604: node is frozen (see the text). We assume that
605: a variable at level 1 is frozen and find the
606: probability that a variable at level 3 is frozen.
607: The clause nodes have co-ordination $K$, while the
608: variable nodes have co-ordination $M$}
609: \end{figure}
610: The probability
611: that a {\it variable} is frozen and part
612: of the giant frozen cluster is $V$ and the
613: probability that a {\it clause} is frozen and
614: part of the giant frozen cluster is $F$.
615: The branching of the variable nodes
616: has maximum co-ordination $M$, but
617: the probability that a link actually exists
618: between a node and clause is $p = K/N$. We start
619: by assuming that a variable is
620: frozen at level 1 (see Fig. 3) and then
621: determine the consequences of this assumption
622: at levels 2 and 3.
623:
624: The probability that a clause is frozen, $F$,
625: at level 2, given the probability, $V$, that
626: a variable is frozen at level 1
627: is given by,
628: \begin{equation}
629: F = ({V\over 2})^{K-1}.
630: \end{equation}
631: This equation is understood as follows. In order for
632: a clause at level 2 to be frozen by the
633: variables at level 1,
634: all of the level 1 variables to which it is connected
635: must be frozen and in conflict with the assignment
636: in the clause. This imposes a fixed assignment
637: on the variable 3. This is the only configuration
638: of variables at level 1 which propagates constraint
639: through a clause to level 3.
640: Now we must consider the
641: cummulative effect of all of the clauses
642: which are connected to the variable at level 3.
643: There are $M-1$ such clauses of which a fraction $F$
644: propagate constraint (are frozen) according to the
645: mechanism of the previous paragraph. Some of these
646: frozen clauses propagate the requirement $x$
647: and others propagate the requirement $\overline{x}$.
648: The variable at level 3 then has three possible states,
649: $P$ = positive, $N$ = negative and $D$=degenerate.
650: The state of the level 3 variable is degenerate if
651: the number of constrained connections which
652: favor the positive state ($x$) is the same as the
653: number of connections which favor the negative state ($\overline{x}$).
654: The probability this variable is frozen (ie. either negated
655: or not) is $V=P+N = 1-D$ as we are considering the
656: case where the probability that a variable is negated is
657: $1/2$. It is straightforward to
658: generalise to the case of unequal probabilities.
659: The probability that the node at level 3 is degenerate is
660: then,
661: \begin{equation}
662: D = \sum_{k=0}^{M} {M!\over (k!)^2 (M-2k)!} ({pF\over 2})^{2k} (1-pF)^{M-2k}
663: \end{equation}
664: Where we have used the fact that the probability that a connection occurs between
665: a variable node and a clause node is $p=K/N$. Eq.~(15) is
666: understood as follows. The probability that a clause
667: at level 2 is frozen and connected(ie it propagates constraint),
668: and it requires the variable at level 3 to
669: be $x$ is $pF/2$. The probability that this clause propagates
670: constraint and requires the variable at level 3 to be $\overline{x}$
671: is also $pF/2$. The variable at level 3 is degenerate if these
672: two events occur an equal number of times, hence the term $(pF/2)^{2k}$.
673: The combinatorial factor gives all ways of arranging these events,
674: taking into account that the $x$ and $\overline{x}$ events are distinct.
675: In the thermodynamic limit, using $pM = \alpha K$, we find,
676: \begin{equation}
677: D = e^{-\alpha K F} \sum_{k=0}^{\infty} {1\over (k!)^2} ({\alpha K F\over 2})^{2k}
678: \end{equation}
679: This provides the reccurence formula for $V = 1-D$ which
680: may be written in the form,
681: \begin{equation}
682: V = 1-e^{-\alpha K F} I_0(\alpha K F)
683: \end{equation}
684: where $I_0$ is the spherical Bessel function of zeroth order.
685: Note that $I_0(0) = 1$.
686: For completeness, we note that
687: the probability that the new variable is frozen in the
688: positive (not negated) state is,
689: \begin{equation}
690: P = \sum_{k=0}^{M} \sum_{l=k+1}^M {M!\over (k!)^2 (M-k-l)!} ({pF\over 2})^{k+l}
691: (1-pF)^{M-k-l}
692: \end{equation}
693: The probability that the variable is frozen in the $N$ state
694: is the same as $P$ for the case we are considering, where
695: the variables have equal probability of being negated and
696: not negated.
697:
698: Equations (14) and (18) provide the self
699: consistent theory for the
700: onset of a giant constrained cluster in
701: K-SAT. We now analyse this theory for
702: the two typical cases.\\
703:
704: \noindent {\it The 2-Sat case ($K=2$)}
705:
706: \begin{figure}
707: {\centering \resizebox*{0.7\columnwidth}{!}
708: {\rotatebox{0}{\includegraphics{figures/figks1.eps}}}\par}
709: \caption{The probability that a clause is frozen, $F$,
710: as a function of $a = \alpha$, for 2-SAT.}
711: \end{figure}
712: In this case Eq.~(14) is $F = V/2$. Expanding Eq.~(17)
713: in powers of $F$, we then have,
714: \begin{equation}
715: F = {1\over 2}[1 - (1-2\alpha F + 2 \alpha^2 F^2 +..)(1+\alpha^2F^2))]
716: \end{equation}
717: This has the trivial solution $\alpha =1$. It also has the non-trivial
718: solution
719: \begin{equation}
720: F ={2\over 3 \alpha^2} (\alpha -1)
721: \end{equation}
722: Thus the random 2-Sat giant cluster emerges smoothly
723: at $\alpha = 1$.
724: Numerical calculation of $F$
725: from Eqs. (14) and (17) is presented in Fig. 4.\\
726:
727: \noindent {\it The $K\ge 3$-Sat case}
728:
729: In these cases, Eq.~(14) with (17) do
730: not have a non-trivial solution with a smooth
731: behavior near a critical point. However they do
732: have a non-trivial solution which has a discontinuous
733: onset at a threshold value, $\alpha_c(K)$.
734: This solution is found by iteration of
735: Eq.~(14) and Eq.~(17) and the results are
736: presented in Fig. 5.
737: We find that although the emergence of the giant
738: cluster is discontinous,
739: for any $K>2$
740: the size of the first order jump
741: decreases quite rapidly with increasing
742: $K$. This indicates that the K-SAT transition is
743: weakly first order and that an
744: analytic analysis at large $K$ is possible.
745: The 3-SAT critical value which we find, $\alpha_c(3) \approx 4.6673(3)$,
746: is consistent with the replica symmetric solution\cite{monasson97}
747: for the metastability point, and
748: significantly higher than the numerical values for the K-SAT transition which lie
749: around $4.3$\cite{mezard02a}. The numerical results we have
750: found (using Eq.~(18)) for metastability point and the jump
751: in $F$ at that point are: $\alpha_c(3) = 4.6673(3),\
752: \delta F_c = 0.0680(1)$; $\alpha_c(4) = 11.833(1),\
753: \delta F_c = 0.0341(3)$; $\alpha_c(5) = 29.91(1),\
754: \delta F_c = 0.016(1)$ ; $\alpha_c(6) = 64.1(1),\
755: \delta F_c = 0.0071(1)$.
756:
757: \begin{figure}
758: {\centering \resizebox*{0.7\columnwidth}{!}
759: {\rotatebox{0}{\includegraphics{figures/figks2.eps}}}\par}
760: \caption{The probability that a clause is frozen, $F$,
761: as a function of $a = \alpha$, for the 3-SAT problem}
762: \end{figure}
763:
764: \section{Summary}
765:
766:
767: We have shown that
768: the probability that a site
769: is on the infinite frozen cluster
770: may be calculated using simple combinatorial methods.
771: This provides a general analytic approach
772: to many hard combinatorial problems, and
773: provides a useful complement to the replica method.
774: Although we
775: concentrated on the symmetric theory here,
776: cavity methods\cite{mezard02a} hold promise for
777: generalising this approach to the unsymmetric case,
778: as will be presented elsewhere.
779:
780: The coloring transition is continuous for $q=2$
781: and discontinous for $q\ge 3$, similarly K-SAT
782: is continuous for $K=2$ and discontinuous
783: for $K\ge 3$. In contrast the VB model
784: of glasses has a continuous phase transition.
785: As found in the rigidity
786: percolation problem\cite{moukarzel97}, processes
787: which require more than 2-connectivity
788: in order to propagate constraint
789: have a tendency toward first order
790: transitions. However a
791: counter example is rigidity percolation
792: on triangular lattices, where the
793: rigidity transition is continuous\cite{moukarzel99}.
794: It thus seems a difficult task to determine the
795: conditions which produce continuous as opposed
796: to discontinuous percolation transitions in complex
797: combinatorial problems.
798:
799:
800: \begin{acknowledgments}
801: This work has been supported by the DOE under contract DE-FG02-90ER45418.
802: PMD acknowledges useful discussions with
803: Radu Cojocaru, Remi Monasson and Bart Selman. PMD
804: thanks Chris Farrow for a careful reading of the
805: manuscript.
806: \end{acknowledgments}
807:
808: \bibliographystyle{apsrev}
809: %\bibliography{conp.bib}
810: \begin{thebibliography}{30}
811: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
812: \expandafter\ifx\csname bibnamefont\endcsname\relax
813: \def\bibnamefont#1{#1}\fi
814: \expandafter\ifx\csname bibfnamefont\endcsname\relax
815: \def\bibfnamefont#1{#1}\fi
816: \expandafter\ifx\csname citenamefont\endcsname\relax
817: \def\citenamefont#1{#1}\fi
818: \expandafter\ifx\csname url\endcsname\relax
819: \def\url#1{\texttt{#1}}\fi
820: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
821: \providecommand{\bibinfo}[2]{#2}
822: \providecommand{\eprint}[2][]{\url{#2}}
823:
824: \bibitem[{\citenamefont{M.~M\'ezard and Virasoro}(1987)}]{mezard87}
825: \bibinfo{author}{\bibfnamefont{G.~P.} \bibnamefont{M.~M\'ezard}}
826: \bibnamefont{and} \bibinfo{author}{\bibfnamefont{M.~A.}
827: \bibnamefont{Virasoro}}, \emph{\bibinfo{title}{Spin glass theory and beyond}}
828: (\bibinfo{publisher}{World Scientific}, \bibinfo{year}{1987}).
829:
830: \bibitem[{dub(2001)}]{dubois01}
831: in \emph{\bibinfo{booktitle}{Theoretical Computer Science, volume 265}}, edited
832: by \bibinfo{editor}{\bibfnamefont{O.}~\bibnamefont{Dubois}},
833: \bibinfo{editor}{\bibfnamefont{R.}~\bibnamefont{Monasson}},
834: \bibinfo{editor}{\bibfnamefont{B.}~\bibnamefont{Selman}}, \bibnamefont{and}
835: \bibinfo{editor}{\bibfnamefont{R.}~\bibnamefont{Zecchina}}
836: (\bibinfo{year}{2001}).
837:
838: \bibitem[{\citenamefont{Monasson et~al.}(1999)\citenamefont{Monasson, Zecchina,
839: Kirkpatrick, Selman, and Troyansky}}]{monasson99}
840: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Monasson}},
841: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
842: \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Kirkpatrick}},
843: \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Selman}}, \bibnamefont{and}
844: \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Troyansky}},
845: \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{400}},
846: \bibinfo{pages}{133} (\bibinfo{year}{1999}).
847:
848: \bibitem[{\citenamefont{M\'ezard et~al.}(2002)\citenamefont{M\'ezard, Parisi,
849: and Zecchina}}]{mezard02a}
850: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'ezard}},
851: \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Parisi}}, \bibnamefont{and}
852: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
853: \bibinfo{journal}{Science} \textbf{\bibinfo{volume}{297}},
854: \bibinfo{pages}{812} (\bibinfo{year}{2002}).
855:
856: \bibitem[{\citenamefont{Kirkpatrick and Selman}(1994)}]{kirkpatrick94}
857: \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Kirkpatrick}} \bibnamefont{and}
858: \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Selman}},
859: \bibinfo{journal}{Science} \textbf{\bibinfo{volume}{264}},
860: \bibinfo{pages}{1297} (\bibinfo{year}{1994}).
861:
862: \bibitem[{\citenamefont{Pittel et~al.}(1996)\citenamefont{Pittel, Spencer, and
863: Wormald}}]{pittel96}
864: \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Pittel}},
865: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Spencer}}, \bibnamefont{and}
866: \bibinfo{author}{\bibfnamefont{N.}~\bibnamefont{Wormald}},
867: \bibinfo{journal}{J.\ Comb.\ Theory\ B} \textbf{\bibinfo{volume}{67}},
868: \bibinfo{pages}{111} (\bibinfo{year}{1996}).
869:
870: \bibitem[{\citenamefont{Kirkpatrick et~al.}(2002)\citenamefont{Kirkpatrick,
871: Wilcke, Garner, and Huels}}]{kirkpatrick02}
872: \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Kirkpatrick}},
873: \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Wilcke}},
874: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Garner}}, \bibnamefont{and}
875: \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Huels}},
876: \bibinfo{journal}{Physica} \textbf{\bibinfo{volume}{314}},
877: \bibinfo{pages}{220} (\bibinfo{year}{2002}).
878:
879: \bibitem[{\citenamefont{Chalupa et~al.}(1979)\citenamefont{Chalupa, Leath, and
880: Reich}}]{chalupa79}
881: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Chalupa}},
882: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Leath}}, \bibnamefont{and}
883: \bibinfo{author}{\bibfnamefont{G.~R.} \bibnamefont{Reich}},
884: \bibinfo{journal}{J. Phys. C.} \textbf{\bibinfo{volume}{12}},
885: \bibinfo{pages}{L31} (\bibinfo{year}{1979}).
886:
887: \bibitem[{\citenamefont{Moukarzel et~al.}(1997)\citenamefont{Moukarzel,
888: Duxbury, and Leath}}]{moukarzel97}
889: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Moukarzel}},
890: \bibinfo{author}{\bibfnamefont{P.~M.} \bibnamefont{Duxbury}},
891: \bibnamefont{and} \bibinfo{author}{\bibfnamefont{P.~L.} \bibnamefont{Leath}},
892: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{55}},
893: \bibinfo{pages}{5800} (\bibinfo{year}{1997}).
894:
895: \bibitem[{\citenamefont{Duxbury et~al.}(1999)\citenamefont{Duxbury, Jacobs,
896: Thorpe, and Moukarzel}}]{duxbury99}
897: \bibinfo{author}{\bibfnamefont{P.~M.} \bibnamefont{Duxbury}},
898: \bibinfo{author}{\bibfnamefont{D.~J.} \bibnamefont{Jacobs}},
899: \bibinfo{author}{\bibfnamefont{M.~F.} \bibnamefont{Thorpe}},
900: \bibnamefont{and}
901: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Moukarzel}},
902: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{59}},
903: \bibinfo{pages}{2084} (\bibinfo{year}{1999}).
904:
905: \bibitem[{\citenamefont{Moukarzel}(2003)}]{moukarzel03}
906: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Moukarzel}},
907: \bibinfo{journal}{Preprint} (\bibinfo{year}{2003}).
908:
909: \bibitem[{\citenamefont{Binder and Young}(1986)}]{binder86}
910: \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Binder}} \bibnamefont{and}
911: \bibinfo{author}{\bibfnamefont{A.~P.} \bibnamefont{Young}},
912: \bibinfo{journal}{Rev.\ Mod.\ Phys.} \textbf{\bibinfo{volume}{58}},
913: \bibinfo{pages}{801} (\bibinfo{year}{1986}).
914:
915: \bibitem[{\citenamefont{Culberson and Gent}(2001)}]{culberson01}
916: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Culberson}} \bibnamefont{and}
917: \bibinfo{author}{\bibfnamefont{I.}~\bibnamefont{Gent}},
918: \bibinfo{journal}{Theor.\ Comp.\ Sci.} \textbf{\bibinfo{volume}{265}},
919: \bibinfo{pages}{227} (\bibinfo{year}{2001}).
920:
921: \bibitem[{\citenamefont{Wu}(1982)}]{wu82}
922: \bibinfo{author}{\bibfnamefont{F.~Y.} \bibnamefont{Wu}},
923: \bibinfo{journal}{Rev.\ Mod.\ Phys.} \textbf{\bibinfo{volume}{54}},
924: \bibinfo{pages}{235} (\bibinfo{year}{1982}).
925:
926: \bibitem[{\citenamefont{Moukarzel and Duxbury}(1995)}]{moukarzel95}
927: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Moukarzel}} \bibnamefont{and}
928: \bibinfo{author}{\bibfnamefont{P.~M.} \bibnamefont{Duxbury}},
929: \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{75}},
930: \bibinfo{pages}{4055} (\bibinfo{year}{1995}).
931:
932: \bibitem[{\citenamefont{Jacobs and Thorpe}(1996)}]{jacobs96}
933: \bibinfo{author}{\bibfnamefont{D.~J.} \bibnamefont{Jacobs}} \bibnamefont{and}
934: \bibinfo{author}{\bibfnamefont{M.~F.} \bibnamefont{Thorpe}},
935: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{53}},
936: \bibinfo{pages}{3683} (\bibinfo{year}{1996}).
937:
938: \bibitem[{\citenamefont{Garey and Johnson}(1979)}]{garey79}
939: \bibinfo{author}{\bibfnamefont{M.~R.} \bibnamefont{Garey}} \bibnamefont{and}
940: \bibinfo{author}{\bibfnamefont{D.~S.} \bibnamefont{Johnson}},
941: \emph{\bibinfo{title}{Computers and Intractability}}
942: (\bibinfo{publisher}{W.H. Freeman and company}, \bibinfo{year}{1979}).
943:
944: \bibitem[{\citenamefont{Fu and Anderson}(1986)}]{fu86}
945: \bibinfo{author}{\bibfnamefont{Y.}~\bibnamefont{Fu}} \bibnamefont{and}
946: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Anderson}},
947: \bibinfo{journal}{J.\ Phys.\ A: Math.\ Gen.} \textbf{\bibinfo{volume}{19}},
948: \bibinfo{pages}{1605} (\bibinfo{year}{1986}).
949:
950: \bibitem[{\citenamefont{Kanter and Sompolinsky}(1987)}]{kanter87}
951: \bibinfo{author}{\bibfnamefont{I.}~\bibnamefont{Kanter}} \bibnamefont{and}
952: \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Sompolinsky}},
953: \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{58}},
954: \bibinfo{pages}{164} (\bibinfo{year}{1987}).
955:
956: \bibitem[{\citenamefont{Hartmann and Weigt}(2001)}]{hartmann01}
957: \bibinfo{author}{\bibfnamefont{A.~K.} \bibnamefont{Hartmann}} \bibnamefont{and}
958: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}},
959: \bibinfo{journal}{Theor.\ Comp.\ Sci.} \textbf{\bibinfo{volume}{265}},
960: \bibinfo{pages}{199} (\bibinfo{year}{2001}).
961:
962: \bibitem[{\citenamefont{van Mourik and Saad}(2002)}]{mourik02}
963: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{van Mourik}} \bibnamefont{and}
964: \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Saad}},
965: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{66}},
966: \bibinfo{pages}{56120} (\bibinfo{year}{2002}).
967:
968: \bibitem[{\citenamefont{Mulet et~al.}(2002)\citenamefont{Mulet, Pagnani,
969: Weight, and Zecchina}}]{mulet02}
970: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mulet}},
971: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Pagnani}},
972: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weight}}, \bibnamefont{and}
973: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
974: \bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{89}},
975: \bibinfo{pages}{268701} (\bibinfo{year}{2002}).
976:
977: \bibitem[{\citenamefont{Monasson and Zecchina}(1997)}]{monasson97}
978: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Monasson}} \bibnamefont{and}
979: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
980: \bibinfo{journal}{Phys. Rev. E} \textbf{\bibinfo{volume}{56}},
981: \bibinfo{pages}{1357} (\bibinfo{year}{1997}).
982:
983: \bibitem[{\citenamefont{M\'ezard and Zecchina}(2002)}]{mezard02b}
984: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'ezard}} \bibnamefont{and}
985: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
986: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{66}},
987: \bibinfo{pages}{56126} (\bibinfo{year}{2002}).
988:
989: \bibitem[{\citenamefont{Braunstein et~al.}(2003)\citenamefont{Braunstein,
990: Mulet, Pagnani, Weigt, and Zecchina}}]{braunstein03}
991: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Braunstein}},
992: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mulet}},
993: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Pagnani}},
994: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}}, \bibnamefont{and}
995: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
996: \bibinfo{journal}{Cond-Mat} \textbf{\bibinfo{volume}{0304558}}
997: (\bibinfo{year}{2003}).
998:
999: \bibitem[{\citenamefont{Fisher and Essam}(1961)}]{fisher61}
1000: \bibinfo{author}{\bibfnamefont{M.~E.} \bibnamefont{Fisher}} \bibnamefont{and}
1001: \bibinfo{author}{\bibfnamefont{J.~W.} \bibnamefont{Essam}},
1002: \bibinfo{journal}{J.\ Math.\ Phys.} \textbf{\bibinfo{volume}{2}},
1003: \bibinfo{pages}{609} (\bibinfo{year}{1961}).
1004:
1005: \bibitem[{\citenamefont{Erd\"os and R\'enyi}(1960)}]{erdos60}
1006: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Erd\"os}} \bibnamefont{and}
1007: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{R\'enyi}},
1008: \bibinfo{journal}{Publ.\ Math.\ Inst.\ Hung.\ Acad.\ Sci.}
1009: \textbf{\bibinfo{volume}{5}}, \bibinfo{pages}{17} (\bibinfo{year}{1960}).
1010:
1011: \bibitem[{\citenamefont{Viana and Bray}(1985)}]{viana85}
1012: \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Viana}} \bibnamefont{and}
1013: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Bray}}, \bibinfo{journal}{C}
1014: \textbf{\bibinfo{volume}{18}}, \bibinfo{pages}{3037} (\bibinfo{year}{1985}).
1015:
1016: \bibitem[{\citenamefont{Maletta and Felsch}(1979)}]{maletta79}
1017: \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Maletta}} \bibnamefont{and}
1018: \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Felsch}},
1019: \bibinfo{journal}{Phys. Rev. B} \textbf{\bibinfo{volume}{20}},
1020: \bibinfo{pages}{1245} (\bibinfo{year}{1979}).
1021:
1022: \bibitem[{\citenamefont{Moukarzel and Duxbury}(1999)}]{moukarzel99}
1023: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Moukarzel}} \bibnamefont{and}
1024: \bibinfo{author}{\bibfnamefont{P.~M.} \bibnamefont{Duxbury}},
1025: \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{59}},
1026: \bibinfo{pages}{2614} (\bibinfo{year}{1999}).
1027:
1028: \end{thebibliography}
1029:
1030:
1031:
1032: \end{document}
1033:
1034: