cs0503082/paper.tex
1: \documentclass[numreferences]{article}
2: 
3: %\usepackage{latexsym,amsmath,amsfonts}
4: 
5: \usepackage{times}
6: %\usepackage{palatino}
7: 
8: \usepackage{amssymb}
9: 
10: \usepackage{epsfig}
11: %\usepackage{psfig}
12: 
13:   \renewcommand*{\Pr}{\mathop{\mathrm{Prob}}}
14: 
15: 
16: 
17:   \def \qedbox{\hfill\vbox{\hrule\hbox{\vrule
18: height1.3ex\hskip0.8ex\vrule}\hrule}}
19: 
20: 
21: 
22:   \newcommand{\goesto}{\rightarrow}
23:   \newtheorem{theorem}{Theorem}
24:   \newtheorem{definition}{Definition}
25:   \newtheorem{proposition}{Proposition}
26:   \newtheorem{lemma}{Lemma}
27:   \newtheorem{cor}{Corollary}
28:   \newcommand*{\proof}{\noindent {\bf Proof}.\,\,}
29: 
30: 
31: 
32: \def\NP{\rm NP}
33: \def\AND{\wedge}
34: \def\OR{\vee}
35: \def\oper{\circ}
36: \def\goesto{\rightarrow}
37: \def\implies{\Rightarrow}
38: \def\zeroone{\{0,1\}}
39:  \def\sstar{\zeroone^{*}}
40: \def\L{\langle}
41: \def\R{\rangle}
42: \def\HYP{\hbox{-}}
43: \def\IFF{\leftrightarrow}
44: \def\Ldef{\buildrel \rm def \over \leftrightarrow}
45: \def\Edef{\buildrel \rm def \over =}
46: \def\almostall{\hbox{\rlap{$_{\thinspace\forall}$}{$^{^\infty}$}}}
47: \def\infoften{\hbox{\rlap{$_{\thinspace\exists}$}{$^{^\infty}$}}}
48: \def\N{{\bf N}}
49: \def\setdelta{\bigtriangleup}
50: \def\cminus{\dot{-}}
51: \def\plusminus{\pm}
52: \def\PR{{\rm Pr}}
53: \def\HSAT{{\rm HORN}\hbox{-}{\rm SAT}}
54: \def\PUR{{\rm PUR}}
55: \def\beginproof{\noindent{\bf Proof.}\quad}
56: \def\endproof{}
57: \def\qed{\hfill$\Box$\newline\vspace{5mm}}
58: 
59: \newcommand{\eqref}[1]{Eq.~(\ref{#1})}
60: \newcommand{\CSP}{\mathrm{CSP}}
61: \newcommand{\SAT}{\mathrm{SAT}}
62: \newcommand{\cont}{\mathit{cont}}
63: \newcommand{\SATneg}{\mathrm{SAT}^\mathrm{(neg)}}
64: \newcommand{\Var}{\mathit{Var}}
65: \newcommand{\opt}{\mathit{opt}}
66: \newcommand{\Cl}{\mathit{Cl}}
67: 
68: % \newtheorem{lemma}{Lemma}[section]
69: % \newtheorem{theo}[lemma]{Theorem}
70: %\newtheorem{prop}[lemma]{Proposition}
71: %\newtheorem{fact}[lemma]{Fact}
72: %\newtheorem{example}{Example}
73: %\newtheorem{obs}{Observation}
74: \newtheorem{claim}{Claim}
75: %\newtheorem{defi}{Definition}
76: %\newtheorem{obs}[lemma]{Observation}
77: 
78: \newenvironment{PROOF}{\noindent{\bf Proof:}}{{\qed}}
79: \newtheorem{conj}{Conjecture}
80: 
81: \bibliographystyle{unsrt}
82: 
83: %\setpapersize{USletter}
84: %\setmarginsrb{1in}{1in}{1in}{1in}{0pt}{0mm}{0pt}{0mm}
85: 
86: \begin{document}
87: 
88: %\begin{article}
89: 
90: %\begin{opening} 
91: 
92: \title{Spines of Random Constraint Satisfaction Problems: Definition and Connection with Computational Complexity}
93: \author{Gabriel Istrate\footnote{istrate@lanl.gov. CCS-DSS, Los Alamos National Laboratory, Mail Stop M 997, Los Alamos, NM 87545.}, 
94: Stefan Boettcher\footnote{stb@physics.emory.edu. Physics Department, Emory University, Atlanta, GA 30322}, Allon G.~Percus\footnote{percus@ipam.ucla.edu. CCS-3, Los Alamos National Laboratory, Los Alamos, NM 87545
95: and UCLA Institute for Pure and Applied Mathematics, Los Angeles, CA
96: 90049}} 
97: 
98: \maketitle
99: %\runningtitle{Spines of Random CSP}
100: %\runningauthor{Istrate, Boettcher, Percus}
101: %\date{}
102: \begin{abstract} 
103: We study the connection between the order of {\em phase transitions in
104: combinatorial problems} and the complexity of decision algorithms for
105: such problems.  We rigorously show that, for a class of random
106: constraint satisfaction problems, a limited connection between the two
107: phenomena indeed exists.  Specifically, we extend the definition of the
108: spine order parameter of Bollob\'{a}s et al. \cite{scaling:window:2sat}
109: to random constraint satisfaction problems, rigorously showing that for
110: such problems a discontinuity of the spine is associated with a
111: $2^{\Omega(n)}$ resolution complexity  (and thus a $2^{\Omega(n)}$
112: complexity of DPLL algorithms) on random instances. The two phenomena
113: have a common underlying cause: the emergence of ``large'' (linear size)
114: minimally unsatisfiable subformulas of a random formula at the
115: satisfiability phase transition. 
116: 
117: We present several further results that add weight to the intuition that
118: random constraint satisfaction problems with a sharp threshold and a
119: continuous spine are ``qualitatively similar to random 2-SAT''. Finally,
120: we argue that it is the spine rather than the backbone parameter whose
121: continuity has implications for the decision complexity of combinatorial
122: problems, and we provide experimental evidence that the two parameters
123: can behave in a different manner.  
124: \end{abstract} 
125: 
126: {\bf Keywords:} constraint satisfaction problems, phase transitions, spine, resolution complexity. \\
127: 
128: {\bf MR Categories:} Primary 68Q25, Secondary 82B27. 
129: 
130: %\end{opening} 
131: 
132: 
133: \section{Introduction}
134: 
135: The major promise of {\em phase transitions in combinatorial problems}
136: has been to shed light on the ``practical'' algorithmic complexity of
137: combinatorial problems.  A possible connection has been highlighted by
138: results of Monasson et al.~\cite{2+p:nature, 2+p:rsa} that are based on
139: experimental evidence and nonrigorous arguments from statistical
140: mechanics.  Studying a version of random satisfiability that
141: ``interpolates'' between 2-SAT and 3-SAT, they suggested that the order
142: of the phase transition, combinatorially expressed by continuity of an
143: order parameter called the {\em backbone}, might have implications for
144: the problem's typical-case complexity.  A discontinuous or first-order
145: transition appeared to be symptomatic of exponential complexity, whereas
146: a continuous or second-order transition was correlated with polynomial
147: complexity.
148: 
149: It is understood by now that this connection is limited.  For instance,
150: $k$-XOR-SAT is a problem believed, based on arguments from statistical
151: mechanics~\cite{zecchina:kxorsat}, to have a first-order phase
152: transition. But it is easily solved by a polynomial algorithm, Gaussian
153: elimination. So, if any connection exists between first-order phase
154: transitions and the complexity of a given problem, it cannot involve
155: {\em all} polynomial time algorithms for the problem. Fortunately, this
156: does not end all hopes for a connection with computational complexity:
157: {\em descriptive complexity}~\cite{descriptive} provides a principled
158: way to measure the complexity of problems with respect to more limited
159: classes of algorithms, those expressible in a given framework.   Here we
160: focus on the {\em Davis-Putnam-Longman-Loveland} (DPLL) class of
161: algorithms~\cite{beame:dp}.
162: 
163: One way to identify the connection between phase transitions and
164: computational complexity is to formalize the underlying intuition
165: connecting the two notions in a purely combinatorial way, devoid of any
166: physics considerations.
167: % Indeed, there exists a nonrigorous argument of this sort:
168: First-order phase transitions amount to a discontinuity in the (suitably
169: rescaled) size of the backbone.
170: % But, at least
171: For random $k$-SAT~\cite{monasson:zecchina}, and more specifically for
172: the optimization problem MAX-$k$-SAT, the backbone has a combinatorial
173: interpretation: it is the set of literals that are ``frozen'', or assume
174: the same value, in all {\em optimal} assignments.  Intuitively, a large
175: backbone size has implications for the complexity of finding such
176: assignments: all literals in the backbone require specific values in
177: order to satisfy the formula optimally, but an algorithm assigning
178: variables in an iterative fashion has very few ways to know what those
179: ``right'' values to assign are. In the case of a first-order phase
180: transition, the backbone of formulas just above the transition contains,
181: with high probability, a fraction of the literals that is bounded away
182: from zero. An algorithm such as DPLL that assigns values to variables
183: iteratively
184: % fashion has probably no way to consistently identify these
185: % ``frozen'' variables, and will
186: may misassign a backbone variable whose height, in a binary tree
187: characterizing the behavior of the algorithm, is $\Omega(n)$ where $n$
188: is the number of variables. This would force a backtrack on the tree.
189: Assuming the algorithm cannot significantly ``reduce'' the size of the
190: explored portion of this tree, a first-order phase transition would then
191: w.h.p.\ imply a $2^{\Omega(n)}$ lower bound for the running time of DPLL
192: on random instances located slightly above the transition.
193: 
194: There exists, however, a significant flaw in the heuristic argument
195: above: the backbone is defined with respect to {\em optimal} assignments
196: for the given formula,  meaning assignments that satisfy the largest
197: possible number of clauses (or all of them, in the case where the
198: formula is satisfiable).  The argument suggests that a discontinuity in
199: the backbone size will make it difficult for algorithms that assign
200: variables in an iterative manner to find {\em optimal} solutions.  The
201: complexity of the optimization problem is, however, often different from
202: that of the corresponding decision problem. For instance, that is the
203: case in XOR-SAT, where the decision problem is easy but the optimization
204: problem is hard. As mentioned above, XOR-SAT is presumed to have a
205: first-order phase transition, so it is not clear at all that the
206: continuity or discontinuity of the backbone should be the relevant
207: predictor for the complexity of the {\em decision} problem as well. 
208: 
209: Fortunately, it turns out that the intuition of the previous argument
210: also holds for a different order parameter, a ``weaker'' version of the
211: backbone called the {\em spine}, introduced
212: in~\cite{scaling:window:2sat} in order to prove that random 2-SAT has a
213: second-order phase transition. Unlike the backbone, the spine is defined
214: in terms of the {\em decision} problem, hence it could conceivably have
215: a larger impact on the complexity of these problems. Of course, the same
216: caveat applies as for the backbone: any connection with computational
217: complexity can only involve complexity classes that have weaker
218: expressive power than the class of polynomial time algorithms. 
219: 
220: We aim in this paper to provide evidence that for random constraint
221: satisfaction problems it is the behavior of the spine, rather than the
222: backbone, impacts the complexity of the underlying decision problem. To
223: accomplish this:
224: 
225: \begin{enumerate}
226: \item We discuss the proper definitions of the backbone and spine for
227: random constraint satisfaction problems (CSP).
228: 
229: \item We formally establish a simple connection between a discontinuity
230: in the relative size of the spine at the threshold and the resolution
231: complexity of random satisfiability problems. In a nutshell, a necessary
232: and sufficient condition for the existence of a discontinuity is the
233: existence of an $\Omega(n)$ lower bound (w.h.p.)\ on the size of
234: minimally unsatisfiable subformulas of a random (unsatisfiable)
235: subformula. But standard methods from proof
236: complexity~\cite{ben-sasson:resolution:width} imply that for all
237: problems where we can prove such an $\Omega(n)$ lower bound, there is a
238: $2^{\Omega(n)}$ lower bound on their resolution complexity and hence on
239: the complexity of DPLL algorithms as well~\cite{beame:dp}.  This
240: property arises from the expansion of the underlying formula's
241: hypergraph, and is {\em independent} of the precise definition of the
242: problem at hand.  Conversely we show (Theorem~\ref{second:order}) that
243: for {\em any generalized satisfiability problem}, a second-order phase
244: transition implies, in the region where most formulas are unsatisfiable,
245: an upper bound on resolution complexity that is smaller than any
246: exponential: $O(2^{\epsilon n})$ for every $\epsilon>0$.
247: 
248: \item We give a sufficient condition
249: (Theorem~\ref{sufficient:first-order}) for the existence of a
250: discontinuous jump in the size of the spine. We then show
251: (Theorem~\ref{implicates:first-order}) that this condition is fulfilled
252: by all problems whose constraints have no implicates of size two or
253: less.  Qualitatively, our results suggest that all satisfiability
254: problems with a continuous phase transition in the spine are
255: ``2-SAT-like''.
256: 
257: \item Finally, we present experimental results that attempt to clarify
258: whether the backbone and the spine can behave differently at the phase
259: transition. The {\em graph bipartition problem} (GBP) is one case where
260: this seems to happen. In contrast, for random {\em 3-coloring} (3-COL),
261: the backbone and spine appear to have similar behavior.
262: 
263: \end{enumerate}
264: 
265: A note on the significance of our results: a first-order transition or
266: discontinuity in the size of the spine is weaker than a discontinuity in
267: the size of the backbone. In the last section of the paper we give a
268: numerical demonstration of an example where the backbone and spine
269: behave differently.  And unlike for the backbone, we do not have a
270: physical interpretation for the spine.  But this is not our intention.
271: The argument connecting the continuity of the backbone order parameter
272: with the complexity of decision problems is problematic, and what we
273: rigorously show is that --- with no physical considerations in mind ---
274: {\em the intuitive connection holds instead for the spine}.
275: % Also, in 
276: % the last section of the paper presents experimental work suggesting
277: % that {\em the backbone and the spine can behave differently}.
278: % For reasons of space proofs are deferred to the Appendix.
279: 
280: \section{Preliminaries}
281: 
282: Throughout this paper we assume a general familiarity with the concepts
283: of phase transitions in combinatorial
284: problems~\cite{martin:monasson:zecchina}, random
285: structures~\cite{bol:b:random-graphs}, and proof
286: complexity~\cite{beame:proof:survey}. We assume more detailed
287: familiarity with certain fundamental results on sharp thresholds
288: \cite{friedgut:k:sat,chvatal:szemeredi:resolution,ben-sasson:resolution:width},
289: and we make use of some of the methods associated with those results.
290: 
291: Two models arising in the theory of random structures are:
292: \begin{itemize}
293: \item The {\em constant probability model} $\Gamma(n,p)$. A random
294: string of bits $X\in\Gamma(n,p)$ is obtained by independently
295: setting each bit of $X$ to 1 with probability $p$, and the rest to 0.
296: \item The {\em counting model} $\Gamma(n,m)$. A random string
297: $X\in\Gamma(n,m)$ is obtained by setting $m$ bits of $X$, chosen
298: uniformly at random, to 1 and the rest to 0.
299: \end{itemize}
300: 
301: For the following purposes, let us work within the constant probability
302: model.
303: Consider a property $A$ that is monotonically increasing, in that
304: if $A$ holds for a given string of bits $X$, then changing any of these
305: bits from 0 to 1 preserves property $A$.  
306: For any $\epsilon >0$, let $p_{\epsilon}= p_{\epsilon}(n)$
307: be the canonical probability such that $\Pr_{X \in
308: \Gamma(n,p_{\epsilon}(n))}[X \mbox{ satisfies } A]= \epsilon$, where
309: $p_\epsilon$ increases monotonically with $\epsilon$.
310: {\em Sharp thresholds} are those for which the function has a
311: ``sudden jump'' from value 0 to 1:
312: 
313: \begin{definition} \label{sharp}
314: Property $A$ has a {\em sharp threshold} iff for every $0<\epsilon
315: < 1/2$, we have $\lim_{n\goesto \infty} \frac{p_{1-\epsilon}(n)-
316: p_{\epsilon}(n)}{p_{1/2}(n)} = 0$. $A$ has {\em a coarse
317: threshold} if for some $\epsilon > 0$ it holds that
318: $\lim\inf_{n\goesto \infty} \frac{p_{1-\epsilon}(n)-
319: p_{\epsilon}(n)}{p_{1/2}(n)} > 0$.
320: \end{definition}
321: 
322: We will use the model of random
323: constraint satisfaction from Molloy~\cite{molloy-stoc2002}:
324: 
325: \begin{definition}\label{model} Let ${\cal D} = \{0,1,\ldots, t-1\}$,
326: $t\geq 2$ be a fixed set. Consider all $2^{t^{k}}-1$ possible 
327: nonempty sets of $k$-ary constraint templates (relations) 
328:  with values taken from ${\cal D}$. Let
329: ${\cal C}$ be such a nonempty set of constraint templates.
330: 
331: A random formula $\phi\in\CSP({\cal C})$ is a set of constraints 
332: specified under the counting model by the following procedure:
333: 
334: \begin{enumerate}
335: % \item $n$ is the number of variables.
336: % 
337: % \item $m$ is the number of clauses, chosen by the following
338: % procedure: first
339: \item Select, uniformly at random and with replacement,
340: $m$ hyperedges of the complete $k$-uniform hypergraph on $n$ variables.
341: 
342: \item For each hyperedge, choose a random ordering of the variables
343: involved in it. Choose a random constraint template from  ${\cal C}$ and 
344: apply it to the list of (ordered) variables.
345: \end{enumerate}
346: 
347: We use the notation $\SAT({\cal C})$ (instead of $\CSP({\cal C})$)
348: when t=2.
349: % Also, for $\Phi$ an instance of $\CSP({\cal C})$ we
350: % denote by $Opt(\Phi)$ the set of {\em optimal assignments} for $\Phi$, and by $opt(\Phi)$ the number of constraints left
351: % unsatisfied by an optimal assignment. 
352: \end{definition}
353: 
354: For an instance $\Phi\in\CSP({\cal C})$, we denote by $\Var(\Phi)$ the
355: set of variables that actually appear in $\Phi$, and by $\opt(\Phi)$ the
356: number of constraints left unsatisfied by an {\em optimal assignment}
357: for $\Phi$.
358: % Finally, we denote by $\Opt(\Phi)$ the set of
359: % {\em optimal assignments} for $\Phi$, and by $\opt(\Phi)$ the number of
360: % constraints left unsatisfied by an optimal assignment. 
361: 
362: % We will denote by $Var$ the set of variables a particular instance 
363: % of $\CSP({\cal C})$ is {\em defined on}, and by $Var(\Phi)$ the variables 
364: % that {\em effectively appear in $\Phi$}. 
365: % When dealing with {\em boolean} constraint satisfaction problems, 
366: % we will also use $Lit(\Phi)$ to denote the set of {\em literals} 
367: % (variables and negated variables) the formula is defined on.
368: 
369: Just as in random graphs~\cite{bol:b:random-graphs}, under fairly
370: liberal conditions one can use the constant probability model
371: instead of the counting model from the previous definition. The
372: interesting range of the parameter $m$ is when the ratio $m/n$ is a
373: constant, $c$, called the {\em constraint density}. The original
374: investigation of the order of the phase transition in $k$-SAT used an
375: order parameter called {\em the backbone},
376: \begin{equation}\label{bb:initial}
377: B(\Phi) = \{ x\in \Var(\Phi)\ |\ \exists\ W\in \{x,\overline{x}\} : 
378: \opt(\Phi \cup W) >\opt(\Phi) \},
379: % B(\Phi) = \{ x\in \Var(\Phi)\ |\ \exists\ \lambda \in \{0,1\} : 
380: % \forall\ X \in \Opt(\Phi), X(x)=\lambda \},
381: \end{equation}
382: or more precisely {\em the backbone fraction} 
383: \begin{equation}\label{fb}
384: f_{B}(\Phi)=\frac{|B(\Phi)|}{n}.
385: \end{equation} 
386: 
387: % One could define the backbone in term of variables, rather than literals. 
388: % However, to compute the backbone fraction we would have to normalize by $|Var(\Phi)|$ instead of $|Var(\Phi)|$, and the definition would not change the behavior of the parameter.  
389: 
390: Bollob\'{a}s et al.~\cite{scaling:window:2sat} have investigated
391: the order of the
392: phase transition in $k$-SAT (for $k=2$) under a different order
393: parameter, a ``monotonic version'' of the backbone called {\em the
394: spine}
395: \begin{equation}\label{spine:initial}
396: S(\Phi) = \{ x\in \Var(\Phi) \ |\ \exists\  W\in\{x,\overline{x}\},
397: \Xi \subseteq \Phi : \Xi \in
398: \SAT, \Xi \AND W \in \overline{\SAT}\}.
399: \end{equation}
400: Here, ``$\in\SAT$'' means ``is satisfiable'' and
401: ``$\in\overline{\SAT}$'' means ``is unsatisfiable''.
402: 
403: The corresponding version of \eqref{fb} is
404: 
405: \begin{equation}\label{fspine}
406: f_{S}(\Phi)=\frac{|S(\Phi)|}{n}.
407: \end{equation} 
408: 
409: They showed that random 2-SAT has a continuous (second-order)
410: phase transition: the size of $f_{S}$ approaches zero w.h.p.\
411: (as $n\goesto
412: \infty$) for constraint density $c<c_\mathrm{2-SAT}=1$, and is continuous at
413: $c=c_\mathrm{2-SAT}$.
414: By contrast, nonrigorous arguments from statistical
415: mechanics~\cite{monasson:zecchina} imply that for 3-SAT the parameter 
416: $f_{B}$ jumps discontinuously from zero to a positive value at the
417: transition point $c=c_\mathrm{3-SAT}$ (a first-order phase transition).
418: 
419: \section{How to define the backbone/spine for random CSP (and beyond)}
420: 
421: We would like to extend the concepts of backbone and spine to
422: general constraint satisfaction problems.
423: The extended definitions
424: must preserve as many of the properties of the
425: backbone/spine as possible. 
426: 
427:  Certain
428: differences between the case of random $k$-SAT and the general
429: case force us to employ an alternative definition of the
430: backbone/spine. The most obvious is that
431: \eqref{spine:initial} involves negations of variables, unlike Molloy's
432: model. Also, these definitions are inadequate
433: for problems whose solution space presents a relabeling symmetry,
434: such as the case of {\em graph coloring} where
435: the set of (optimal) colorings is closed under permutations of the
436: colors. Due to this symmetry, no variable can be ``frozen'' to a fixed
437: value $\lambda$ as in \eqref{bb:initial}.
438: 
439: We therefore define the backbone/spine of a
440: random instance of $\CSP({\cal C})$ in a slightly different manner. 
441: Let $\hat{C}$ be the set of constraints obtained by applying the
442: constraint templates in ${\cal C}$ to all ordered lists of $k$ variables
443: chosen from the set of all $n$ variables.
444: 
445: \begin{definition}\label{spine-first} 
446: \[
447: B(\Phi) = \{x\in \Var(\Phi)\ |\ \exists\ C \in \hat{C} : x\in C,
448: \opt(\Phi \cup C) >\opt(\Phi) \},
449: \]
450: \[
451: S(\Phi) = \{ x\in \Var(\Phi)\ |\ \exists\ C\in \hat{C}, \Xi \subseteq \Phi : x\in C, \Xi \in \CSP,
452: \Xi \cup C \in \overline{\CSP}\}.
453: \]
454: 
455: \end{definition}
456: 
457: For $k$-CNF formulas whose (original)
458: backbone/spine contains at least three literals, a variable $x$ is
459: in the (new version of the) backbone/spine if and only if either
460: $x$ or $\overline{x}$ were present in the old version. In
461: particular the new definition does not change the order of the
462: phase transition of random $k$-SAT.
463: 
464: Alternatively, in studying 3-colorability (3-COL) of random graphs
465: $G=(V,E)$, Culberson and Gent~\cite{frozen:development} defined the
466: spine of a colorable graph $G$ to be the set of vertex pairs $(x,y)\in
467: V^{2}$ that get assigned the same color in {\em all} colorings of $G$. 
468: 
469: Following up on the idea of defining the backbone and spine in terms of {\em constraints} rather than {\em variables}, and by analogy with the definition in~\cite{scaling:window:2sat},  
470: one can extend the definition of $S(G)$ to general graphs\footnote{Culberson and Gent employ an ``effective'' version of the spine they call {\em frozen development} that is more amenable to 
471: experimental analysis. Frozen development is a subset of the spine, as
472: defined in \eqref{spine-3col}.} by 
473: 
474: \begin{equation}\label{spine-3col}  
475: S(G) = \{ (x,y)\in V^{2}\ |\ \exists\ H \subseteq G : H \in
476: \mbox{3-COL}, H \cup (x,y) \in \overline{\mbox{3-COL}}\}.
477: \end{equation}
478: 
479: We can further extend these definitions to all random constraint satisfaction 
480: problems $\CSP({\cal C})$: 
481: 
482: \begin{definition}\label{spine-two} 
483: \[
484: B_{C}(\Phi) = \{C \in \hat{C}\ |\ \opt(\Phi \cup C) > \opt(\Phi) \},
485: \]
486: \[
487: S_{C}(\Phi) = \{ C\in \hat{C}\ |\ \exists\ \Xi \subseteq \Phi : \Xi \in
488: \CSP, \Xi \cup C \in \overline{\CSP}\}.
489: \]
490: \end{definition} 
491: Similarly,
492: % if we denote by ${\cal C}(\Phi)$ the set of constraints obtained 
493: % by applying the constraint templates in ${\cal C}$ to all ordered lists 
494: % of $k$ variables chosen from $Var$, 
495: one can define the {\em backbone/spine fraction} by 
496: \[
497: f_{B_C}(\Phi)= \frac{|B_{C}(\Phi)|}{|\hat{C}|},  
498: \]
499: and 
500: \[
501: f_{S_C}(\Phi)= \frac{|S_{C}(\Phi)|}{|\hat{C}|}.
502: \]
503: 
504: We will refer to these concepts as the {\em constraint-based}
505: backbone/spine (fractions), as opposed to the previously defined {\em
506: variable-based} quantities. The two are clearly related.  For instance 
507: one can easily show that 
508: \[
509: B(\Phi)=\cup_{C\in B_{C}(\Phi)} \Var(C),
510: \]
511: where $\Var(C)$ represents all variables appearing in constraint $C$
512: alone.  It is also clear that $|B_{C}(\Phi)|=O(|B(\Phi)|^k)$ and
513: similarly for the spine.  Since $|\hat{C}|=\Theta(n^k)$, it follows that
514: the continuity of $f_{B}$ or $f_{S}$ implies
515: the continuity of $f_{B_C}$ or $f_{S_C}$.  However, the converse is
516: not in general true, and so
517: % 
518: % (a similar identity holds for the two definitions of the spine). 
519: % 
520: % If the variable-based backbone/spine of an 
521: % instance of $\CSP({\cal C})$ has size $u$, then the constraint-based 
522: % backbone/spine has size $O(u^{k})$. It follows readily 
523: % that continuity of $f_{B}$ ($f_{S}$) will imply 
524: % the continuity of $f_{B_C}$ ($f_{S_C}$). 
525: % 
526: % On the other hand, while the size of the variable-based 
527: % backbone is at most $k$ times the size the relation of the constraint-based 
528: % backbone (and similarly for the spine), this relation is {\em not} powerful 
529: % enough to show that
530: the two backbone/spine fractions do not necessarily behave in the same 
531: way.
532: % The reason for that is that we are using different normalization factors 
533: % ($|Var|$ in one case, $\theta(|Var|^{k})$ in the other) 
534: % when defining the two versions of the order parameters. 
535: 
536: Given the two types of definitions, which
537: should we choose? The answer depends on the problem, as well as 
538: on the issue we wish to address. For instance,
539: in the statistical mechanics analysis of
540: combinatorial problems, the presumably ``correct'' definition of the backbone
541: emerges from the analysis
542: undertaken in \cite{monasson:zecchina} for random $k$-SAT. 
543: But since we are interested
544: in a combinatorial definition, with no physics considerations in
545: mind, the only principled way to choose between the two types of
546: order parameters (one based on variables, the other based on
547: constraints) is based on the class of algorithms we are concerned
548: with. In the case of random constraint satisfaction problems and DPLL
549: algorithms, it is variables that get assigned values, so
550: Definition~\ref{spine-first} is preferred. On the other hand, 
551: constraint-based definitions can make sense for problems that
552: share some characteristics with random 3-COL (i.e., binary
553: constraint satisfaction problems, and problems with built-in
554: symmetries of the solution space). In a later section we will see an
555: example, the case of {\em graph bipartition}, where the
556: constraint-based backbone and spine seem to behave differently.
557: (Whether one can come with a natural example of this phenomenon
558: for the variable-based backbone is an interesting open problem.)
559: 
560: 
561:  
562: 
563:  
564: 
565: 
566: 
567: \section{Spine discontinuity and resolution complexity of random CSP}\label{main:section}
568: 
569: In this section we will study the continuity of the spine-based order
570: parameter  $f_{S}$ for {\em boolean} random constraint satisfaction,
571: or satisfiability, problems. The kind of continuous/discontinuous
572: behavior we are looking for is formalized by the following definition (a
573: similar one can be given for the constraint-based versions of the order
574: parameter):   
575: 
576: \begin{definition}\label{cont} 
577:  Let ${\cal C}$ be such that $\SAT({\cal C})$ has a
578: sharp threshold. Problem $\SAT({\cal C})$ has a {\em discontinuous spine}
579: if there exists $\eta > 0$ such that for every sequence $m = m(n)$ we have
580: \begin{equation}\label{jump:general}
581: \lim_{n\goesto \infty} \Pr_{m=m(n)}[\Phi \in \SAT] = 0 \implies
582: \lim_{n\goesto \infty}\Pr_{m=m(n)}[ f_{S}(\Phi)\geq
583: \eta]= 1.
584: \end{equation}
585: If, on the other hand, for every $\epsilon >0$ there exists a constant 
586: $c=c({\epsilon})$ such that the map $\epsilon \goesto c({\epsilon})$ is monotonically increasing and 
587: \begin{equation}\label{jump:continuous}
588: \lim_{n\goesto \infty} \Pr_{m=c({\epsilon})n}[\Phi \in \SAT] = 0\
589: \mathrm{and}\ \lim_{n\goesto \infty}\Pr_{m=c({\epsilon})n}[
590: f_{S}(\Phi)\geq \epsilon]= 0
591: \end{equation}
592: we say that $\SAT({\cal C})$ has a {\em continuous spine}.
593: \end{definition}
594: 
595: We now give a simple observation that will be the basis for identifying 
596: discontinuities of the spine: 
597: 
598: \begin{proposition}\label{spine:unsat}
599: Let $\Phi$ be a minimally unsatisfiable formula, and let $x$ be a
600: literal that appears in $\Phi$. Then, by Definition~\ref{spine-first}, $x\in S(\Phi)$.
601: \end{proposition}
602: 
603: \begin{proof}
604: There exists $C\in \Phi$ such that $x\in C$. But $\Phi \setminus C$ is satisfiable and 
605: $(\Phi \setminus C)\cup C$ is not, 
606: thus $x\in S(\Phi)$.   
607: \end{proof} 
608: \qed 
609: 
610: 
611: \begin{cor}\label{3sat:first-order}
612: $k$-SAT, $k\geq 3$ has a discontinuous spine.
613: \end{cor}
614: 
615: \begin{proof}
616: To show a discontinuous spine it is sufficient to show
617: that a random unsatisfiable formula contains w.h.p.\ a minimally
618: unsatisfiable subformula involving a linear number of literals.
619: % A way to accomplish this is by directly employing the
620: In the Chv\'{a}tal-Szemer\'{e}di proof
621: \cite{chvatal:szemeredi:resolution} that w.h.p.\ random $k$-SAT has
622: exponential resolution size for $k\geq 3$, % where the above claim
623: the claim is implicitly proved. 
624: \end{proof}
625: \qed
626: 
627: \begin{definition} The {\em width} of a resolution proof $P$ of the unsatisfiability of a CNF-formula $F$ is defined to be the maximum number of 
628: literals in any clause that appears in the proof $P$.   
629: 
630: If $\Phi$ is an instance of $\SAT({\cal C})$, denote by $\Cl(\Phi)$ the
631: CNF formula obtained by expressing each constraint of $\Phi$ as a
632: conjunction of {\em clauses} (i.e., expressing $\Phi$ in conjunctive
633: form).  
634: 
635: The {\em resolution complexity} of an instance $\Phi$ of
636: $\SAT({\cal C})$ is defined as the length of the smallest resolution 
637: proof  of $\Cl(\Phi)$. 
638: \end{definition} 
639: 
640:  A simple observation is that a continuous spine
641: has implications for resolution
642: complexity:
643: 
644: \begin{theorem}\label{second:order}
645: Let ${\cal C}$ be a set of constraint templates such that $\SAT({\cal C})$ has a
646: continuous spine.
647: Then for every constraint density $c>lim_{\epsilon \goesto
648: 0} c({\epsilon})$, and {\em every} $\epsilon>0$, random
649: formulas of constraint density $c$ have w.h.p.\ resolution
650: complexity $O(2^{\epsilon n})$.
651: \end{theorem}
652: 
653: \begin{proof}
654: 
655: Because of Proposition~\ref{spine:unsat} and the fact that  
656: $\SAT({\cal C})$ has a continuous spine, for every $\epsilon>0$, minimally
657: unsatisfiable subformulas of a random formula $\Phi$ with
658: constraint density $c({\epsilon})$ contain w.h.p.\ at most $\epsilon n$ 
659: variables. Consider the backtrack tree of the natural DPLL algorithm that
660: tries to satisfy constraints one at a time on such a minimally
661: unsatisfiable subformula $F$. By the usual correspondence between
662: DPLL refutations and resolution complexity (e.g., \cite{beame:dp})
663: this yields a resolution proof of the unsatisfiability of $\Phi$
664: having size at most $2^{\epsilon n}$.
665: 
666: Taking $\epsilon$ to be small enough that $c({\epsilon})<c$, and using the 
667: fact that resolution complexity of a random formula is a monotonically
668: decreasing function of the constraint density, we get the desired result.  
669: \end{proof}
670: \qed
671: 
672: Let us observe that we have stated the preceding theorem using condition
673: $c>lim_{\epsilon \goesto 0} c({\epsilon})$ since we cannot be sure, even
674: for $k$-SAT, that the phase transition takes place at a constant value
675: of the constraint density $c$. In practice one would of course expect
676: that, for a problem with a continuous spine, there
677: exists a sequence $c({\epsilon})$ as in Definition~\ref{cont} having the
678: constraint density at the phase transition as its limit.   
679: 
680: 
681: 
682: \begin{definition}
683: Denote by $|F|$ the
684: number of constraints that appear in formula $F$. Define  
685: 
686: \[ c^{*}(F)= \max\left\{
687: \frac{|H|}{|\Var(H)|}: \emptyset \neq H \subseteq
688: F\right\}.
689: \]
690: \end{definition}
691: 
692: The next result gives a sufficient condition for a generalized
693: satisfiability problem to have a discontinuous spine. Interestingly, it is one condition
694: studied in \cite{molloy-stoc2002}.
695: 
696: \begin{theorem} \label{sufficient:first-order}
697: Let ${\cal C}$ be such that $\SAT({\cal C})$ has a
698: sharp threshold. If there exists $\epsilon > 0$ such that for
699: every minimally unsatisfiable formula $F$ it holds that
700: $c^{*}(F) > \frac{1+\epsilon}{k-1}$, 
701: then $\SAT({\cal C})$ has a discontinuous spine.
702: 
703: \end{theorem}
704: 
705: \begin{proof}
706: The proof is similary to that of Corollary~\ref{3sat:first-order}: we will show that 
707: w.h.p.\ a random formula contains a minimally unsatisfiable subformula containing a linear number of variables, and apply Proposition~\ref{spine:unsat}. 
708: 
709: To accomplish that, we first recall the following concept from
710: \cite{chvatal:szemeredi:resolution}:
711: 
712: \begin{definition}
713: Let $x,y>0$. A $k$-uniform hypergraph with $n$ vertices is {\em
714: ($x$,$y$)-sparse} if every set of $s\leq xn$ vertices contains at
715: most $ys$ edges.
716: \end{definition}
717: 
718: We also recall Lemma 1 from the same paper.
719: \begin{lemma}\label{sparsity:hypergraph}
720: Let $k,c>0$ and $y>1/(k-1)$. Then w.h.p.\ the
721: random $k$-uniform hypergraph with $n$ vertices and $cn$ edges is
722: $(x,y)$-sparse, where
723: \begin{equation}\label{x-sparse}
724: x = \left( \frac{1}{2e}\left(\frac{y}{ce}\right)^{y}\right)^{\frac{1}{y(k-1)-1}}.
725: \end{equation}
726: \end{lemma}
727: 
728: Let $y=\frac{1+\epsilon}{k-1}$.  Directly applying
729: Lemma~\ref{sparsity:hypergraph}, w.h.p.\ a random
730: $k$-uniform hypergraph with $cn$ edges is $(x_{0},y)$ sparse, for
731: $x_{0}=(\frac{1}{2e}(\frac{y}{ce})^{y})^{\frac{1}{\epsilon}}$.
732: The critical observation is then that the existence of a minimally
733: unsatisfiable formula with $xn$ variables and with $c^{*}(F) >
734: \frac{1+\epsilon}{k-1}$ implies that the $k$-uniform hypergraph
735: associated with the given formula is {\em  not} $(x,y)$-sparse.
736: It follows that any formula
737: with fewer than $x_{0}n/k$ constraints (and thus fewer than 
738:  $x_{0}n$ variables) is satisfiable. Therefore, any 
739: minimally unsatisfiable subformula of random formula $\Phi$ has more
740: than $x_{0}n/k$ constraints. 
741: 
742: To show that such formulas have many variables, we again employ the expansion 
743: of the formula hypergraph given by Lemma~\ref{sparsity:hypergraph}, and 
744: infer that {\em all} subformulas of size less than $xn$ of $\Phi$ (in particular those that are also subformulas of a minimally unsatisfiable subformula of 
745: $\Phi$) have a linear number of variables. 
746: \end{proof}
747: \qed
748: 
749: 
750: One can give an explicitly defined class of satisfiability
751: problems for which the previous result applies:
752: 
753: \begin{theorem}\label{implicates:first-order}
754: Let $k\geq 2$ and let ${\cal C}$ be such that $\SAT({\cal C})$ has a
755: sharp threshold.  If  {\em no} clause template $C\in {\cal C}$ has (when
756: expressed as a CNF-formula) an implicate of length 2 or 1 then
757: \begin{enumerate}
758: \item For every minimally unsatisfiable formula $F$, 
759: $
760: c^{*}(F)\geq \frac{2}{2k-3}$. Therefore $\SAT({\cal C})$ satisfies the conditions of the previous
761: theorem, i.e., it has a discontinuous spine.
762: \item Moreover, there exists a constant $\eta >0$ such that w.h.p.\ 
763: random instances of $\SAT({\cal C})$ have $\Omega(2^{\eta n})$ resolution
764: complexity\footnote{This result subsumes some of the results in
765: \cite{mitchell:cp02}. While a preliminary version of this paper was under consideration (and publicly available \cite{istrate:allthat}) related and technically more sophisticated results have been given independently in  \cite{molloy:focs2003}.}.
766: \end{enumerate}
767: \end{theorem}
768: 
769: The condition in the theorem is violated, as expected, by random 2-SAT.
770: It is also violated by the random version of the NP-complete problem
771: 1-in-$k$-SAT. This can be seen as follows.  The problem can be
772: represented as $\CSP({\cal C})$, for ${\cal C}$ a set of $2^{k}$
773: constraints corresponding to all ways to negate some of the variables,
774: and has a rigorously determined ``2-SAT-like'' location of the
775: transition point~\cite{istrate:1ink:sat}. However, the formula 
776: \[
777: C(x_{1}, x_{2}, \ldots, x_{k-1},
778: x_{k})\AND C(\overline{x_{k}}, x_{k+1},  \ldots , x_{2k-2},
779: x_{1})
780: \]
781: 
782: \[
783: \AND C(\overline{x_{1}}, x_{2k-1}, \ldots, x_{3k-3},
784: \overline{x_{k}}) \AND C(x_{k}, x_{3k-2}, \ldots, x_{4k-4},
785: \overline{x_{1}}),
786: \]
787: where $C$ is the constraint ``1-in-$k$'', is minimally
788: unsatisfiable but has clause/variable ratio $1/(k-1)$ and
789: implicates $\overline{x_{1}} \OR \overline{x_{k}}$ and $x_{1}\OR
790: x_{k}$.
791: 
792: \vspace{5mm}
793: 
794: \begin{proof}
795: 
796: \begin{enumerate}
797: \item For any real $r \geq 1$, formula $F$ and set of clauses
798: $G\subseteq F$,
799:  define the {\em $r$-deficiency of $G$}, $\delta_{r}(G)=
800: r|G|-2|\Var(G)|$. Also define
801: \begin{equation} \label{max}
802: \delta^{*}_{r}(F)= \max\{\delta_{r}(G): \emptyset \neq G \subseteq
803: F\}
804: \end{equation}
805: 
806: \begin{definition}
807: Let $F$ be a formula, and let $C_{1}, C_{2}, \ldots, C_{i}, \ldots, C_{m}$ be 
808: a listing of the constraints in $F$. 
809: \item A variable $v$ is {\em private} for constraint $C_{i}$ if $v$
810: appears in $C_{i}$ but in no other constraint.
811: \item Variable $v$ is {\em free in $C_{i}$} if $v$ appears in $C_{i}$ but in no $C_{j}$, $j<i$. Otherwise we say that $v$ is {\em bound in $C_{i}$}.  
812: \end{definition} 
813: 
814: 
815: We claim that for any minimally unsatisfiable $F$,
816: $\delta^{*}_{2k-3}(F)\geq 0$. Indeed, assume this was not true.
817: Then there exists such $F$ such that:
818: \begin{equation}\label{deff}
819: \delta_{2k-3}(G)\leq -1\mbox{ for all }\emptyset \neq G\subseteq
820: F.
821: \end{equation}
822: \begin{lemma} \label{1-transversal}
823: Let $F$ be a formula for which condition~\ref{deff} holds. Then
824: there exists an ordering $C_{1}, \ldots, C_{|F|}$ of constraints
825: in $F$ such that each constraint $C_{i}$ contains at least $k-2$
826: variables that are free in $C_{i}$. 
827: \end{lemma}
828: 
829: \begin{proof}
830: Denote by $v_{i}$ the number of variables that appear in {\em exactly}
831: $i$ constraints of $F$. We have $ \sum_{i\geq 1} i v_{i} = k |F|$,
832: therefore $2|\Var(F)|-v_{1}\leq k |F|$. This can be rewritten as
833: $v_{1}\geq 2|\Var(F)|-k|F|> |F| (2k-3 - k)= (k-3) |F|$, where we use
834: \eqref{deff}. Therefore there exists at least one constraint $C$ in $F$
835: with at least $k-2$ variables that are private in $F$, hence necessarily
836: free in $F$.  We set $C_{|F|}=C$ and apply this argument recursively to
837: $F\setminus C$.
838: \end{proof}
839: \qed
840: 
841: Let us show now that $F$ cannot be minimally unsatisfiable. Construct a
842: satisfying assignment for $F$ incrementally, so that the partial assignment 
843: constructed up to stage $j$ will satisfy constraints $C_{1}, \ldots, C_{j}$. 
844: 
845: Indeed, suppose we have constructed a partial assignment that satisfies 
846: $C_{1}, \ldots, C_{j-1}$, and consider now constraint
847: $C_{j}$. At most two of the variables in $C_{j}$ are bound in
848: $C_{j}$. Since $C_{j}$ has no implicates of length two or less, no matter 
849: what the assignment to these two variables might have been in the previous 
850: stages, one can set the variables that are free in $C_{j}$ in a way that 
851: satisfies this clause. Iteratively performing this construction 
852: yields a satisfying assignment for $F$, in contradiction with our
853: assumption that $F$ was minimally unsatisfiable.
854: 
855: Therefore $\delta^{*}_{2k-3}(F)\geq 0$, a statement equivalent to
856: our conclusion.
857: 
858: \item To prove the resolution complexity lower bound we use the
859: size-width connection for resolution complexity obtained in
860: \cite{ben-sasson:resolution:width}: it is sufficient to prove that 
861: there exists $\eta >0$ such that w.h.p.\ random instances of $\SAT({\cal C})$
862: having constraint density $c$ have resolution width at least $\eta n$. 
863: 
864: To accomplish this, we use the same strategy 
865: as in \cite{ben-sasson:resolution:width}: define for a unsatisfiable formula $\Phi$ a measure $\mu:Clauses \goesto {\bf N}$ (where $Clauses$ is the set of 
866: all possible disjunctions of literals from $\Var(\Phi)$, including 
867: the contradictory clause $\Box$) such that 
868: \begin{enumerate}
869: \item for every clause $C$ that appears in $\Cl(\Phi)$, $\mu(C)\leq 1$,
870: \item w.h.p.\ $\mu(\Box)$ is ``large''.
871: \item Infer that in any refutation there exists a clause $C$ with ``medium'' 
872: $\mu(C)$, and 
873: \item prove that if $\mu(C)$ is ``medium'' than the width of $C$ is
874: ``large''. 
875: \end{enumerate}
876: 
877: As in \cite{ben-sasson:resolution:width}, define 
878: 
879: \[
880: \mu(C)=\min\{|\Xi|: \Xi\subseteq  \Phi, \Xi \models C\},
881: \]
882: 
883: where $\models$ is the logical entailment relation. In particular 
884: $\mu(\Box)$ is the size of the smallest unsatisfiable subformula of $\Phi$. 
885:  $\mu$ is subadditive, 
886: that is, for every clauses $C_{1}$ and $C_{2}$ that share a variable $x$ 
887: appearing with opposite signs in the two clauses, 
888: \[
889: \mu(res_{x}(C_{1},C_{2}))\leq \mu(C_{1})+\mu(C_{2}). 
890: \]
891: 
892: where $res_{x}(C_{1},C_{2})$ denotes the clause obtained by applying resolution to clauses 
893: $C_{1}$, $C_{2}$ with respect to variable $x$. 
894: It is clear that condition a) is satisfied.
895: As to b), the following is true: 
896: 
897: \begin{lemma}
898: There exists $\eta_{1}>0$ such that for any $c>0$, w.h.p.\
899: $\mu(\Box)\geq \eta_{1} n$, where  $\Phi$ is a random
900: instance of $\SAT({\cal C})$ having constraint density $c$.
901: \end{lemma}
902: 
903: %\iffalse
904: %%%%%%%%
905: 
906: \begin{proof}
907: In the proof of Theorem~\ref{sufficient:first-order} we have shown
908: that there exists $\eta_{0}>0$ such that w.h.p.\ any unsatisfiable
909: subformula of a given formula has at least $\eta_{0} n$
910: constraints. Therefore {\em any} formula $F$ made up of {\em clauses} from
911: the CNF-representation of constraints in $\Phi$, and which has
912: fewer than $\eta_{0} n$ clauses is satisfiable (since it is less tight 
913: than the conjunction of those constraints). 
914: 
915: The claim now follows by taking $\eta_{1} = \eta_{0}$.
916: \end{proof}
917: \qed
918: %\fi
919: 
920: The only (slightly) nontrivial step of the proof, which critically
921: uses the fact that constraints in ${\cal C}$ do not have
922: implicates of length one or two, is to prove that clause
923: implicates of subformulas of ``medium'' size have ``many''
924: variables. 
925: 
926: \begin{lemma}\label{expansion}
927: There exists $d>0$ and $\eta_{2}>0$ such that w.h.p.\ (when  $\Phi$ is a random
928: instance of $\SAT({\cal C})$ having constraint density $c$)
929:  every clause $C$ present in a refutation of $\Cl(\Phi)$ 
930: that satisfies $\frac{dn}{2}<\mu(C)\leq dn$ also satisfies  $|C|\geq
931: \eta_{2} n$.
932: \end{lemma}
933: 
934: 
935: \begin{proof}
936: 
937: Given a clause $C$, let $\Xi$ be a subformula of $\Phi$, having minimal size, such
938: that $\Xi \models C$. We claim:
939: 
940: \begin{lemma}\label{appear} 
941: For every constraint $P$ of $\Xi$ that contains $k-2$ private variables, 
942: at least one of these variables appears in $C$.
943: \end{lemma}
944: 
945: \begin{proof} 
946: Suppose there exists a constraint $D$ of $\Xi$ with at least $k-2$ private 
947: variables such that none of its private variables appears in $C$. Because of the minimality of $\Xi$ there exists an assignment $F$
948: that satisfies $\Xi \setminus \{D\}$ but does not satisfy $D$ or
949: $C$. Since $D$ has no implicates of size two, there exists an
950: assignment $G$, that differs from $F$ only on the private
951: variables of $D$, that satisfies $\Xi$. But since $C$ does not
952: contain any of the private variables of $D$, $F$ coincides with
953: $G$ on variables in $C$. The conclusion is that $G$ does not
954: satisfy $C$, contradicting the fact that $\Xi\models C$.
955: \end{proof} 
956: \qed
957: 
958: 
959: Now define $x(\cdot, \cdot)$ to be the function from \eqref{x-sparse} that 
960: describes the dependence of $x$ on $y$ and $c$. For a constant $\epsilon >0$ 
961: to be determined later, define 
962: 
963: \[
964: d = min(inf\{x(2/(2k-3+\epsilon),c)| c\geq c_{\SAT({\cal C})}\}, \eta_{1}).
965: \]
966: 
967: Since $\SAT({\cal C})$ has a sharp threshold, the first term of the minimum 
968: expression is, like $\eta_{1}$, strictly greater than zero. Therefore,
969: $d>0$.
970: 
971: \begin{lemma}\label{private}  
972: There exists constant $\eta_{2}>0$ such that w.h.p., when  $\Phi$ is a random
973: instance of $\SAT({\cal C})$ having constraint density $c$ and $W\subseteq \Phi$
974:  is a formula with at most $dn$ constraints,  $W$ contains at least $\eta_{2} n$ constraints each of which has at least $k-2$ private variables.  
975: \end{lemma} 
976: \begin{proof} 
977: 
978: To prove Lemma~\ref{private} we first need: 
979: 
980: \begin{lemma} \label{pr} 
981: Let $\epsilon>0$ be a constant. If $F$ is a formula with 
982: $c^{*}(F)\leq \frac{2}{2k-3+\epsilon}$ then for every
983: subformula $G$ of $F$, at least $(\epsilon/3)
984: |G|$ constraints of $G$ have  at
985: least $k-2$ private variables.
986: \end{lemma} 
987: 
988: \begin{proof} 
989:  Indeed,
990: since $c^{*}(G)\leq \frac{2}{2k-3+\epsilon}$, by an argument similar
991: to the one used in the proof of Lemma~\ref{1-transversal},   $v_{1}(G)\geq
992: (k-3+\epsilon)|G|$. Since constraints in $G$ have
993: arity $k$, at least $(\epsilon/3) |G|$ have
994: more than $k-3$ (i.e., at least $k-2$) private variables. 
995: \end{proof} 
996: \qed
997: 
998: Returning to the proof of Lemma~\ref{private}, 
999: choose $y=\frac{2}{2k-3+\epsilon}$ in
1000: Lemma~\ref{sparsity:hypergraph} for $\epsilon>0$ a small enough
1001: constant. Because of the definition of $d$, when $\Phi$ is a random
1002: instance of $\SAT({\cal C})$ having constraint density $c$, w.h.p.\ formula   
1003: $\Phi$ is $(d,y)$ sparse. Since $|W|\leq dn$, this easily implies the fact that 
1004: \[
1005: c^{*}(W) \leq \frac{2}{2k-3+\epsilon}. 
1006: \]
1007: 
1008: Lemma~\ref{private} follows by applying Lemma~\ref{pr} to formula $W$
1009: with $\eta_{2}=\epsilon/3$. Applying this result and Lemma~\ref{appear} to
1010: formula $\Xi$ also concludes the proof of Lemma~\ref{expansion}. 
1011: 
1012: \end{proof} 
1013: \end{proof} 
1014: \qed
1015: 
1016: The proof of item 2. of
1017: Theorem~\ref{implicates:first-order} now follows: since for any
1018: clause $K$ in $\Cl(\Phi)$ we have $\mu(K)=1$, since
1019: $\mu(\Box)>\eta_{1} n$ and since $0<d\leq \eta_{1}$, 
1020: there indeed exists a
1021: clause $C$ such that
1022: \begin{equation}\label{intermediate}
1023: \mu(C)\in [dn/2, dn].
1024: \end{equation}
1025: 
1026: Indeed,
1027:  let $C^{\prime}$ be a clause in the resolution refutation of $\Phi$ minimal with
1028: the property that $\mu(C^{\prime})> dn$. Then at least one clause
1029: $C$ of the two involved in deriving $C^{\prime}$ satisfies
1030: \eqref{intermediate}.
1031: Applying Lemma~\ref{expansion} we infer that the width of $C$ is at least 
1032: $\eta_{2} n$. Using the size-width connection from\cite{ben-sasson:resolution:width} completes the proof of item 2. of Theorem~\ref{implicates:first-order}. 
1033: 
1034: \end{enumerate}
1035: 
1036: \end{proof}
1037: \qedbox
1038: 
1039: \subsection{Threshold location and discontinuous spines}
1040: 
1041: Molloy \cite{molloy-stoc2002} has studied threshold properties of 
1042: random constraint satisfaction problems, describing a technical 
1043: property of the constraint set (called {\em very well-behavedness})
1044: that is necessary for the existence of a sharp threshold. In 
1045: \cite{istrate:sharp} we have shown that Molloy's {\em well-behavedness} condition is actually necessary and sufficient for boolean constraints (this has 
1046: been independently proved by Creignou and Daud\'{e} \cite{creignou-daude-thresholds}). Thus we have 
1047: completely characterized sets $C$ for which $\SAT({\cal C})$ has 
1048: a sharp threshold. 
1049: 
1050: The well-behavedness condition has implication for the clause/variable 
1051: ratio of minimally unsatisfiable formulas: it has to be larger than $1/(k-1)$. 
1052: Furthermore, Molloy has shown that if the density of minimally unsatisfiable 
1053: formulas is {\em bounded away from  $1/(k-1)$} (i.e., it satisfies 
1054: the conditions of Theorem~\ref{sufficient:first-order}) then the location of 
1055: the transition is {\em strictly larger than $\frac{1}{k(k-1)}$}. 
1056: 
1057: We have seen that the same density condition is sufficient to guarantee 
1058: the discontinuity of the spine and exponential resolution complexity. 
1059: A natural question therefore arises: is it
1060: possible to relate the continuity (or discontinuity) of the spine to the
1061: {\em location} of the phase transition ?
1062: 
1063: At first this does not seem to be possible: we have already encountered  
1064: two problems that fail to satisfy the sufficient condition for a
1065: discontinuous spine, random 2-SAT, for which the transition has been
1066: proven to be of second order~\cite{scaling:window:2sat}, and random
1067: 1-in-$k$-SAT, for which a similar result
1068: holds~\cite{istrate:1ink:sat}. Both have a threshold location 
1069: strictly higher than Molloy's lower bound of $\frac{1}{k(k-1)}$. However, 
1070: the most natural specification of the random model for the two problems 
1071: involves applying constraints on both variables and their negations. 
1072: For both problems the actual 
1073: location of the threshold is {\em twice} the value given by Theorem 3 in
1074: \cite{molloy-stoc2002}, at clause/variable ratio $\frac{2}{k(k-1)}$. This 
1075: suggests that the following tempting intuitive picture might be accurate, 
1076: at least in a more restricted setting:  
1077: 
1078: \begin{enumerate}
1079: 
1080: \item Problems with a continuous spine are ``2-SAT-like'', and have a
1081: phase transition at constraint density $c^\cont_k= \frac{2}{k(k-1)}$. 
1082: \item  Problems with a discontinuous spine
1083: have a phase transition located at constraint density $c> c^\cont_k$. 
1084: \end{enumerate} 
1085: 
1086: To obtain results that partly support the intuition above, we have to
1087: modify the random model from Definition~\ref{model} to allow negated
1088: variables. 
1089: 
1090: \begin{definition}
1091: Let ${\cal C}$ be a set of constraint templates. The {\em closure of ${\cal C}$}, denoted $\overline{\cal C}$ is the set of 
1092: constraints 
1093: \begin{equation} \label{closure} 
1094: \overline{\cal C}= \{C(x_{1}^{\epsilon_{1}}, \ldots, x_{k}^{\epsilon_{k}})\mbox{ }|\mbox{ }C\in {\cal C}\mbox{ and }\epsilon_{1}, \ldots, \epsilon_{k}\in \{\pm 1\}\}, 
1095: \end{equation}   
1096: where for a variable $x_i$ we define $x_i^{1}:=x_i$, $x_i^{-1}:=
1097: \overline{x_i}$. 
1098: 
1099: Set ${\cal C}$ is {\em good} if $|\overline{\cal C}|=|{\cal C}| 2^{k}$, that is all elements on the right hand side of 
1100: \eqref{closure} are distinct. 
1101: 
1102: \end{definition} 
1103: 
1104: 
1105: 
1106: \begin{definition}\label{neg} 
1107: Let ${\cal C}$ be a good set of constraint templates. 
1108: Denote by $\SATneg({\cal C})$ the version of $\SAT({\cal C})$ that
1109: generates a random formula by the following process: 
1110: 
1111: \begin{enumerate}
1112: % \item $n$ is the number of variables.
1113: % 
1114: % \item $m$ is the number of constraints, chosen by the following
1115: % procedure: first
1116: \item Select, uniformly at random and with replacement,
1117: $m$ hyperedges of the complete $k$-uniform hypergraph on $n$ variables.
1118: 
1119: \item For each hyperedge $e$, choose a random ordering $o_{e}$ of the variables
1120: involved in it. 
1121: 
1122: \item Independently with probability 1/2 negate each variable appearing in $o_{e}$. 
1123: \item Choose a random constraint template from  ${\cal C}$ and 
1124: apply it to the ordered list of literals in $o_{e}$.
1125: \end{enumerate}
1126: 
1127: % Again, when there is no need to be that specific we will drop the parameters 
1128: % $n,m$ from our notation. 
1129: % 
1130: \end{definition} 
1131: 
1132: It is easy to see that problems such as $k$-SAT and 1-in-$k$-SAT can be
1133: expressed using the framework of Definition~\ref{neg}.  The following
1134: result shows that the intuition connecting the discontinuity of the
1135: spine, resolution complexity and the location of the phase transition
1136: does indeed have merit: a strenghtening of the condition guaranteeing
1137: the existence of a discontinuous spine and exponential resolution
1138: complexity also implies that the satisfiability threshold is
1139: located at a value higher than $c^\cont_k$:
1140:  
1141: 
1142: \begin{theorem}
1143: Let ${\cal C}$ be a good set such that 
1144: \begin{enumerate}
1145: \item $\SATneg({\cal C})$ has a sharp threshold (the 
1146: result in \cite{istrate:sharp} can be easily adapted to completely 
1147: characterize such sets ${\cal C}$).   
1148: \item  There exists $\epsilon > 0$ such that,  
1149: for every minimally unsatisfiable formula $F$ whose constraints are drawn 
1150: from template set ${\cal C}$, the ratio of the number of constraints in $F$ to the number of 
1151: distinct literals (variables and negated variables) appearing in $F$ is 
1152: at least $\frac{1+\epsilon}{k-1}$.                        
1153: \end{enumerate} 
1154: Then 
1155: 
1156: \begin{enumerate} 
1157: \item There is a
1158: constant $\delta >0$ such that random instances of $\SATneg({\cal C})$ 
1159: with $m=cn$, where $c\leq \frac{2}{k(k-1)} (1+\delta)$, are 
1160: satisfiable with probability $1-o(1)$. 
1161: \item Problem $\SATneg({\cal C})$ has a discontinuous spine and exponential resolution complexity. 
1162: \end{enumerate} 
1163: \end{theorem}
1164: 
1165: \begin{proof} 
1166: \begin{enumerate} 
1167: \item 
1168: Since  $\SATneg({\cal C})$ has a sharp threshold, it is sufficient to show 
1169: that there exists a fixed constant $\eta >0$ such that the probability 
1170: that a random formula is satisfiable is at least $\eta$. 
1171: 
1172: Suppose $m= cn$ with $c= \frac{2}{k(k-1)} (1+\delta)$, with $\delta > 0$ 
1173: small enough. Define random model $\SATneg_2({\cal C})$ that is a
1174: variant of $\SATneg({\cal C})$ as follows: 
1175: 
1176: \begin{enumerate} 
1177: \item Choose a random $k$-uniform hypergraph $H$ with $m$ edges on the  vertex set (of cardinality $2n$) consisting of {\em variables and negated variables}.
1178: \item For every edge $e\in H$ create a random permutation $o_{e}$ of its elements.  
1179: \item Apply a random constraint template in ${\cal C}$ to variables in the 
1180: ordered list $o_{e}$. 
1181: 
1182: \end{enumerate}
1183: 
1184: This model differs from the random model $\SATneg({\cal C})$ in that 
1185: it allows for constraints that include a  $k$-tuple of literals
1186: involving two opposite literals. 
1187: 
1188: Define $W$ to be the event that formula $\Phi$ contains some clause
1189: involving two opposite literals. It is easy to see that the expected
1190: number of such clauses in a random formula $\Phi$ is constant.
1191: Therefore, with positive probability $\lambda >0$ in a random formula
1192: generated according to  $\SATneg_{2}({\cal C})$, the bad event $W$
1193: will {\em not} happen. 
1194: 
1195: Let $Z$ denote the event that a random formula with $m= cn$ clauses
1196: generated according to the random model $\SATneg_2({\cal C})$ is satisfiable. 
1197: 
1198: Then the probability that a random formula in  $\SATneg({\cal C})$ is 
1199: satisfiable is equal to $Pr[Z|\overline{W}]$. To show that this is bounded 
1200: away from zero it is enough to prove that $Pr[Z]=1-o(1)$. 
1201: 
1202: The $k$-uniform hypergraph on the $2n$ nodes (variables 
1203: and their negations) corresponding to choosing 
1204: a random instance of $\SATneg_2({\cal C})$ is a random $k$-uniform hypergraph. Thus we want to show that 
1205: a formula generated by first choosing such a random $k$-uniform hypergraph $H$, and then 
1206: applying a random constraint template from ${\cal C}$ on the given
1207: literals is w.h.p.\ satisfiable. 
1208: 
1209: The proof of this is entirely similar to a step in the proof of Molloy's Theorem 3 in \cite{molloy-stoc2002}, and 
1210: amounts to showing that w.h.p.\ the hypergraph $H$ does not contain any hypergraph of high density, 
1211: corresponding to the fact that minimally unsatisfiable subformulas have clause/variable density at least $\frac{1}{k-1} (1+\epsilon)$. 
1212: Rather than repeating an argument that is presented in detail in that paper, we refer the reader to  \cite{molloy-stoc2002}. 
1213: 
1214: \item Since ${\cal C}$ is good, one can simply apply
1215: Theorem~\ref{sufficient:first-order} to $\SAT(\overline{{\cal C}})$,
1216: which is equivalent to problem $\SATneg({\cal C})$.
1217: \end{enumerate} 
1218: \end{proof} 
1219: \qed
1220: 
1221: 
1222: 
1223: \section{Beyond random satisfiability: comparing the behavior of the backbone and spine}
1224: 
1225: In this section we investigate empirically the continuity of
1226: the backbone for two graph problems, random three coloring (3-COL)
1227: and the graph bipartition problem (GBP).  Both can be phrased as
1228: decision or as optimization problems, in the same manner as $k$-SAT and
1229: MAX-$k$-SAT.
1230: % The latter problem is the decision counterpart 
1231: % of the graph bisection problem in the same manner the $k$-SAT problem is 
1232: % the decision counterpart of MAX-$k$-SAT. 
1233: 
1234: We consider a large number of instances of random graphs, of sizes up to
1235: $n=1024$ and over a range of mean degree values near the threshold.
1236: For each instance we determine the backbone fraction $f$.
1237: 
1238: Culberson and Gent~\cite{frozen:development} have shown experimentally
1239: that the 3-COL spine fraction $f_{S_C}$, as defined in
1240: Definition~\ref{spine-two}, exhibits a discontinuous transition. To be
1241: consistent with this study, we use the backbone fraction $f_{B_C}$ from
1242: the same definition. We employ a rapid heuristic called {\em extremal
1243: optimization\/}~\cite{BoPe1}.  Although an incomplete procedure,
1244: numerical studies~\cite{BoPe2} as well as testbed comparisons with an
1245: exact algorithm ~\cite{Trick}, have shown that extremal optimization
1246: yields an excellent approximation of $f_{B_C}$ around the critical
1247: region (see~\cite{BoPe1} for further discussions, that, we believe,
1248: convincingly support this assertion).  Fig.~1 shows $f_{B_C}$
1249: as a function of mean degree. 
1250: 
1251: Culberson and Gent have speculated that at the 3-COL threshold, although
1252: their spine is discontinuous, the backbone might be {\em continuous\/}.
1253: The results in Fig.~1a suggest otherwise.  For 3-COL, $f_{B_C}$
1254: does not appear to vanish above the threshold, indicating a
1255: discontinuous large $n$ backbone~\cite{BoPe2}.
1256: 
1257: We next study the graph bipartion problem (GBP):
1258: 
1259: \begin{definition}
1260: GBP is the following decision problem.  Given a (not necessarily
1261: connected) graph $G$ with $n$ vertices, $n$ being an even number,
1262: determine whether it can be partitioned into two edge-disjoint sets
1263: having $n/2$ vertices each.
1264: \end{definition} 
1265: This problem cannot, strictly speaking, be cast in the setup of random
1266: constraint satisfaction problems from Definition~\ref{model}, since not
1267: every partition of vertices of $G$ is allowed. It can be cast to a
1268: satisfiability problem (with variables associated to nodes, values
1269: associated to each partition and constraint ``$x = y$'' associated to
1270: the edge between the corresponding vertices) but we must add the
1271: additional requirement that {\em all satisfying assignments contain an
1272: equal number of ones and zeros}.  Thus the complexity-theoretic
1273: observations of Section~\ref{main:section} do not automatically apply to
1274: it.  We can, however, give a ``DPLL-like'' class of algorithms for GBP,
1275: so the the hope of obtaining results similar to the previous ones is not
1276: so far-fetched.
1277: 
1278: Let us investigate the continuity of the backbone/spine under the model
1279: in Definition~\ref{spine-two}.  It is easy to see that the
1280: constraint-based spine $S_C(G)$ of a GBP instance $G$ contains all edges
1281: belonging to a connected component of size larger than $n/2$. Since the
1282: GBP threshold takes place where the giant component becomes larger than
1283: $n/2$, $f_{S_C}$ is discontinuous there.  On the other hand, the
1284: backbone fraction $f_{B_C}$ (Fig.~1b) appears to remain
1285: continuous, vanishing at large $n$ on both sides of the threshold. 
1286: 
1287: We have noted earlier that the discontinuity of $f_{B_C}$ is a stronger
1288: property than the discontinuity of $f_B$.  Thus for 3-COL it follows
1289: that the variable-based backbone is discontinuous as well.  By contrast,
1290: it is not clear for GBP whether the variable-based backbone is
1291: continuous: our preliminary experimental evidence is as yet
1292: inconclusive. 
1293: 
1294: The results in Fig.~1b suggest that the spine and the backbone
1295: can behave differently at the threshold, though they do not yet address
1296: the question of whether the spine's discontinuity really has
1297: computational implications for the decision problem's complexity.  After
1298: all, unlike 3-COL, GBP can easily be solved in polynomial time by
1299: dynamic programming. This situation is similar to that of XOR-SAT, where
1300: a polynomial algorithm exists but the complexity of {\em resolution
1301: proofs/DPLL algorithms} is exponential.  The class of ``DPLL-like''
1302: algorithms that can solve GBP can no longer be simulated in a
1303: straightforward manner by {\em resolution proofs}, however it can be
1304: simulated using proof systems $Res(k)$ that are extensions of
1305: resolution~\cite{krajicek:resk}. Some of the hardness results for
1306: resolution extend to these more powerful proof systems, and in
1307: \cite{istrate:descriptive} we investigate the extent to which our
1308: present results apply to this class of proof systems.  These
1309: preliminary results imply that, indeed, the discontinuity of the spine
1310: {\em does} have computational implications for GBP.
1311: 
1312: \begin{figure}[H]
1313: %\tabcapfont
1314: \centerline{%
1315: \begin{tabular}{c@{\hspace{2pc}}c}
1316: \includegraphics[angle=-90, width = .45\linewidth]{3col} & 
1317: \includegraphics[angle=-90, width = .45\linewidth]{gbp}  \\
1318: (a) 
1319: 3-COL~\cite{BoPe2} & (b) GBP. 
1320: \end{tabular}}
1321: 
1322: \caption{
1323: Plot of the estimated constraint-based backbone fraction $f_{B_C}$
1324: on random graphs, as a function of mean degree $c$.  For
1325: 3-COL, the systematic error based on benchmark comparisons with random
1326: graphs is negligible compared to the statistical error bars;
1327: for GBP, $f_{B_C}$ is found by exact enumeration.
1328: % For 3-COL we average over 2300, 500, and 280 instances for $n=32$, 64, and
1329: % 128 at each value of $pn$.  For the GBP, we average over 3000, 350, and
1330: % 100 instances for $n=50$, 100, and 200.
1331: The thresholds
1332: $c\approx4.70$ for 3-COL and $c=2\ln2$ for GBP are shown by dashed
1333: lines.\label{mass}}
1334: \end{figure}
1335: 
1336: \section{Discussion}
1337: 
1338: We have shown that the existence of a discontinuous spine
1339: in a random satisfiability problem is often correlated with a
1340: $2^{\Omega(n)}$ peak in the complexity of resolution/DPLL
1341: algorithms at the transition point. The underlying reason is that the two phenomena
1342: (the jump in the order parameter and the resolution complexity
1343: lower bound) have common causes.
1344: 
1345: The example of random $k$-XOR-SAT
1346: shows that a general connection between a first-order phase transition and
1347: the complexity of the underlying decision problems is hopeless:
1348: Ricci-Tersenghi et al.~\cite{zecchina:kxorsat}
1349: have presented a non-rigorous argument using the replica method
1350: that shows that this problem has a first-order phase transition, and
1351: the following weaker result is a direct consequence of
1352: Theorem~\ref{implicates:first-order}: 
1353:  
1354: \begin{proposition}
1355: Random $k$-XOR-SAT, $k\geq 3$, has a discontinuous spine.
1356: \end{proposition}
1357: 
1358: However, our results, as well as work in progress mentioned above,
1359: suggest that the continuity/discontinuity of the spine is a predictor
1360: for the complexity of the {\em restricted classes} of decision
1361: algorithms that can be simulated by ``resolution-like'' proof systems.
1362: Furthermore, experimental evidence in the previous section suggests that the
1363: backbone and the spine do not always behave similarly.
1364: Our analysis indicates that the spine, rather than the backbone, is the
1365: order parameter to consider in studying the complexity of combinatorial
1366: problems.
1367: 
1368: \section{Acknowledgments}
1369: 
1370: This work has been supported by the U.S.\ Department of Energy under
1371: contract W-705-ENG-36, through the LANL LDRD program, and by grant
1372: 0312510 from the Division of Materials Research at the National Science
1373: Foundation.
1374: 
1375: \begin{thebibliography}{99}
1376: 
1377: \bibitem{2+p:nature}
1378: R.~Monasson, R.~Zecchina, S.~Kirkpatrick, B.~Selman, and L.~Troyansky.
1379: \newblock Determining computational complexity from characteristic phase
1380:   transitions.
1381: \newblock {\em Nature}, 400(8):133--137, 1999.
1382: 
1383: \bibitem{2+p:rsa}
1384: R.~Monasson, R.~Zecchina, S.~Kirkpatrick, B.~Selman, and L.~Troyansky.
1385: \newblock $2+p$-{SAT}: Relation of typical-case complexity to the nature of the
1386:   phase transition.
1387: \newblock {\em Random Structures and Algorithms}, 15(3--4):414--435, 1999.
1388: 
1389: \bibitem{zecchina:kxorsat}
1390: F.~Ricci-Tersenghi, M.~Weight, and R.~Zecchina.
1391: \newblock Simplest random $k$-satisfiability problem.
1392: \newblock {\em Physical Reviews E}, 63:026702, 2001.
1393: 
1394: \bibitem{descriptive} 
1395: N.~Immerman. 
1396: \newblock {\em Descriptive Complexity}. 
1397: \newblock Springer Graduate Texts in Computer Science, 1999. 
1398: 
1399: \bibitem{beame:dp}
1400: P.~Beame, R.~Karp, T.~Pitassi, and M.~Saks.
1401: \newblock The efficiency of resolution and {D}avis-{P}utnam procedures.
1402: \newblock {\em SIAM Journal of Computing}, 31(4):1048--1075, 2002.
1403: 
1404: \bibitem{monasson:zecchina}
1405: R.~Monasson and R.~Zecchina.
1406: \newblock Statistical mechanics of the random $k$-{SAT} model.
1407: \newblock {\em Physical Review E}, 56:1357, 1997.
1408: 
1409: \bibitem{scaling:window:2sat}
1410: B.~Bollob\'{a}s, C.~Borgs, J.T. Chayes, J.~H. Kim, and D.~B. Wilson.
1411: \newblock The scaling window of the 2-{SAT} transition.
1412: \newblock Technical report, Los Alamos e-print server,
1413:   http://xxx.lanl.gov/ps/math.CO/9909031, 1999.
1414: 
1415: \bibitem{ben-sasson:resolution:width}
1416: E.~Ben-{S}asson and A.~Wigderson.
1417: \newblock Short {P}roofs are {N}arrow:{R}esolution made {S}imple.
1418: \newblock {\em Journal of the ACM}, 48(2), 2001.
1419: 
1420: \bibitem{martin:monasson:zecchina}
1421: O.~Martin, R.~Monasson, and R.~Zecchina.
1422: \newblock Statistical mechanics methods and phase transitions in combinatorial
1423:   optimization problems.
1424: \newblock {\em Theoretical Computer Science}, 265(1-2):3--67, 2001.
1425: 
1426: \bibitem{bol:b:random-graphs}
1427: B.~Bollob\'{a}s.
1428: \newblock {\em Random Graphs}.
1429: \newblock Academic Press, 1985.
1430: 
1431: \bibitem{beame:proof:survey}
1432: P.~Beame and T.~Pitassi.
1433: \newblock Propositional proof complexity: Past present and future.
1434: \newblock In {\em Current Trends in {T}heoretical {C}omputer {S}cience}, pages
1435:   42--70. 2001.
1436: 
1437: \bibitem{friedgut:k:sat}
1438: E.~Friedgut.
1439: \newblock Necessary and sufficient conditions for sharp thresholds of graph
1440:   properties, and the k-{SAT} problem. with an appendix by {J}. {B}ourgain.
1441: \newblock {\em Journal of the A.M.S.}, 12:1017--1054, 1999.
1442: 
1443: \bibitem{chvatal:szemeredi:resolution}
1444: V.~Chv\'{a}tal and E.~Szemer\'{e}di.
1445: \newblock Many hard examples for resolution.
1446: \newblock {\em Journal of the ACM}, 35(4):759--768, 1988.
1447: 
1448: \bibitem{molloy-stoc2002}
1449: M.~Molloy.
1450: \newblock Models for random constraint satisfaction problems.
1451: \newblock In {\em Proceedings of the 32nd ACM Symposium on Theory of
1452:   Computing}, 2002.
1453: 
1454: \bibitem{achlioptas:friedgut:kcol}
1455: D.~Achlioptas and E.~Friedgut.
1456: \newblock A sharp threshold for $k$-colorability.
1457: \newblock {\em Random Structures and Algorithms}, 14(1):63--70, 1999.
1458: 
1459: \bibitem{frozen:development}
1460: J.~Culberson and I.~Gent.
1461: \newblock Frozen development in Graph Coloring. 
1462: \newblock {\em Theoretical Computer Science}, 265(1-2):227--264, 2001.
1463: 
1464: \bibitem{mitchell:cp02}
1465: D.~Mitchell.
1466: \newblock Resolution complexity of {R}andom {C}onstraints.
1467: \newblock In {\em Eigth International Conference on Principles and Practice of
1468:   Constraint Programming}, 2002.
1469: 
1470: \bibitem{molloy:focs2003}
1471: M.~Molloy and M.~Salavatipour.
1472: \newblock The resolution complexity of random constraint satisfaction problems.
1473: \newblock (to appear in FOCS'2003).
1474: 
1475: \bibitem{istrate:1ink:sat}
1476: D.~Achlioptas, A.~Chtcherba, G.~Istrate, and C.~Moore.
1477: \newblock The phase transition in random 1-in-k {SAT} and {NAE }3{SAT}.
1478: \newblock In {\em Proceedings of the 13th ACM-SIAM Symposium on Discrete
1479:   Algorithms}, 2001. Journal version in preparation. 
1480: 
1481: \bibitem{BoPe1}
1482: S.~Boettcher and A.~Percus.
1483: \newblock Nature's way of optimizing.
1484: \newblock {\em Artificial Intelligence}, 119:275--286, 2000.
1485: 
1486: \bibitem{BoPe2}
1487: S.~Boettcher and A.G.~Percus.
1488: \newblock Extremal Optimization at the Phase Transition of the
1489: 3-Coloring Problem.
1490: \newblock {\em Physical Review E}, vol. 69,066703, 2004.
1491: 
1492: 
1493: \bibitem{EOperc}
1494: S.~Boettcher.
1495: \newblock Extremal optimization of graph partition at the percolation
1496:   threshold.
1497: \newblock {\em J. Phys. A: Math. Gen.}, 32:5201--5211, 1999.
1498: 
1499: \bibitem{Trick}
1500: M.A. Trick.
1501: \newblock color.c graph coloring code.
1502: \newblock {A}vailable at http://mat.gsia.cmu.edu/COLOR/solvers/trick.c.
1503: 
1504: \bibitem{krajicek:resk}
1505: J.~Krajicek.
1506: \newblock On the weak pigeonhole principle.
1507: \newblock {\em Fundamenta Matematicae}, 170(1--3):123--140, 2001.
1508: 
1509: \bibitem{istrate:allthat}
1510: G.~Istrate.
1511: \newblock Phase transitions and {\em all that}. 
1512: \newblock {P}reprint CS.CC/0211012, ACM Computer Repository at arXiv.org. 
1513: 
1514: \bibitem{istrate:descriptive}
1515: G.~Istrate.
1516: \newblock Descriptive complexity and first-order phase transitions.
1517: \newblock (in progress).
1518: 
1519: \bibitem{istrate:sharp} 
1520: G.~Istrate. 
1521: \newblock Threshold Properties of Random Constraint Satisfaction Problems. 
1522: \newblock Accepted to a special volume of {\em Discrete Applied Mathematics} on Typical-case complexity and phase transitions. 
1523: 
1524: \bibitem{creignou-daude-thresholds}
1525: N.~Creignou and H. Daud\'{e}.
1526: \newblock Combinatorial sharpness criterion and phase transition classification for random CSPs
1527: \newblock. {\em Information and Computation}, vol. 190, no.2 (2004), pp. 220-238.
1528: 
1529: \end{thebibliography}
1530: 
1531: %\end{article}
1532: \end{document}
1533: 
1534: 
1535: