A Faster Clause-Shortening Algorithm for SAT with No Restriction on Clause Length



Similar documents
Asymptotic Growth of Functions

In nite Sequences. Dr. Philippe B. Laval Kennesaw State University. October 9, 2008

Soving Recurrence Relations

Running Time ( 3.1) Analysis of Algorithms. Experimental Studies ( 3.1.1) Limitations of Experiments. Pseudocode ( 3.1.2) Theoretical Analysis

5 Boolean Decision Trees (February 11)

I. Chi-squared Distributions

Incremental calculation of weighted mean and variance


SECTION 1.5 : SUMMATION NOTATION + WORK WITH SEQUENCES

Department of Computer Science, University of Otago

A probabilistic proof of a binomial identity

Properties of MLE: consistency, asymptotic normality. Fisher information.

Chapter 6: Variance, the law of large numbers and the Monte-Carlo method

Convexity, Inequalities, and Norms

Lecture 4: Cauchy sequences, Bolzano-Weierstrass, and the Squeeze theorem

Infinite Sequences and Series

Chapter 7 Methods of Finding Estimators

CS103A Handout 23 Winter 2002 February 22, 2002 Solving Recurrence Relations

Section 11.3: The Integral Test

Chair for Network Architectures and Services Institute of Informatics TU München Prof. Carle. Network Security. Chapter 2 Basics

Approximating Area under a curve with rectangles. To find the area under a curve we approximate the area using rectangles and then use limits to find

PROCEEDINGS OF THE YEREVAN STATE UNIVERSITY AN ALTERNATIVE MODEL FOR BONUS-MALUS SYSTEM

Overview of some probability distributions.

Annuities Under Random Rates of Interest II By Abraham Zaks. Technion I.I.T. Haifa ISRAEL and Haifa University Haifa ISRAEL.

Lecture 2: Karger s Min Cut Algorithm


SAMPLE QUESTIONS FOR FINAL EXAM. (1) (2) (3) (4) Find the following using the definition of the Riemann integral: (2x + 1)dx

Modified Line Search Method for Global Optimization

Vladimir N. Burkov, Dmitri A. Novikov MODELS AND METHODS OF MULTIPROJECTS MANAGEMENT

Maximum Likelihood Estimators.

ON AN INTEGRAL OPERATOR WHICH PRESERVE THE UNIVALENCE

where: T = number of years of cash flow in investment's life n = the year in which the cash flow X n i = IRR = the internal rate of return

Chapter 5: Inner Product Spaces

CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 8

University of California, Los Angeles Department of Statistics. Distributions related to the normal distribution

CS103X: Discrete Structures Homework 4 Solutions

Taking DCOP to the Real World: Efficient Complete Solutions for Distributed Multi-Event Scheduling

Output Analysis (2, Chapters 10 &11 Law)

THE ABRACADABRA PROBLEM

Integer Factorization Algorithms

Cooley-Tukey. Tukey FFT Algorithms. FFT Algorithms. Cooley

Week 3 Conditional probabilities, Bayes formula, WEEK 3 page 1 Expected value of a random variable

Lecture 4: Cheeger s Inequality

Measures of Spread and Boxplots Discrete Math, Section 9.4

THIN SEQUENCES AND THE GRAM MATRIX PAMELA GORKIN, JOHN E. MCCARTHY, SANDRA POTT, AND BRETT D. WICK

Basic Elements of Arithmetic Sequences and Series

The Stable Marriage Problem

UC Berkeley Department of Electrical Engineering and Computer Science. EE 126: Probablity and Random Processes. Solutions 9 Spring 2006

Sequences and Series

Our aim is to show that under reasonable assumptions a given 2π-periodic function f can be represented as convergent series

Your organization has a Class B IP address of Before you implement subnetting, the Network ID and Host ID are divided as follows:

1. MATHEMATICAL INDUCTION

Chapter 14 Nonparametric Statistics

Recovery time guaranteed heuristic routing for improving computation complexity in survivable WDM networks

A Note on Sums of Greatest (Least) Prime Factors

A Combined Continuous/Binary Genetic Algorithm for Microstrip Antenna Design

Example 2 Find the square root of 0. The only square root of 0 is 0 (since 0 is not positive or negative, so those choices don t exist here).

The Power of Free Branching in a General Model of Backtracking and Dynamic Programming Algorithms

Project Deliverables. CS 361, Lecture 28. Outline. Project Deliverables. Administrative. Project Comments

Hypothesis testing. Null and alternative hypotheses

Systems Design Project: Indoor Location of Wireless Devices

Confidence Intervals for One Mean

THE HEIGHT OF q-binary SEARCH TREES

WHEN IS THE (CO)SINE OF A RATIONAL ANGLE EQUAL TO A RATIONAL NUMBER?

CHAPTER 3 THE TIME VALUE OF MONEY

Time Value of Money, NPV and IRR equation solving with the TI-86

Chapter 7 - Sampling Distributions. 1 Introduction. What is statistics? It consist of three major areas:

Research Article Sign Data Derivative Recovery

Partial Di erential Equations

Theorems About Power Series

A Recursive Formula for Moments of a Binomial Distribution

Plug-in martingales for testing exchangeability on-line

Lecture 5: Span, linear independence, bases, and dimension

1. C. The formula for the confidence interval for a population mean is: x t, which was

Chapter 7: Confidence Interval and Sample Size

Confidence Intervals. CI for a population mean (σ is known and n > 30 or the variable is normally distributed in the.

Notes on exponential generating functions and structures.

Class Meeting # 16: The Fourier Transform on R n

Ekkehart Schlicht: Economic Surplus and Derived Demand

Estimating Probability Distributions by Observing Betting Practices

CHAPTER 7: Central Limit Theorem: CLT for Averages (Means)

CS85: You Can t Do That (Lower Bounds in Computer Science) Lecture Notes, Spring Amit Chakrabarti Dartmouth College

Multi-server Optimal Bandwidth Monitoring for QoS based Multimedia Delivery Anup Basu, Irene Cheng and Yinzhe Yu

Data Analysis and Statistical Behaviors of Stock Market Fluctuations

Determining the sample size

Present Value Factor To bring one dollar in the future back to present, one uses the Present Value Factor (PVF): Concept 9: Present Value

Lesson 15 ANOVA (analysis of variance)

Nr. 2. Interpolation of Discount Factors. Heinz Cremers Willi Schwarz. Mai 1996

1 Computing the Standard Deviation of Sample Means

Discrete Mathematics and Probability Theory Spring 2014 Anant Sahai Note 13

Math C067 Sampling Distributions

5: Introduction to Estimation

Chapter 5 Unit 1. IET 350 Engineering Economics. Learning Objectives Chapter 5. Learning Objectives Unit 1. Annual Amount and Gradient Functions

Transcription:

Joural o Satisfiability, Boolea Modelig ad Computatio 1 2005) 49-60 A Faster Clause-Shorteig Algorithm for SAT with No Restrictio o Clause Legth Evgey Datsi Alexader Wolpert Departmet of Computer Sciece Roosevelt Uiversity 430 S. Michiga Av. Chicago, IL 60605, USA edatsi@roosevelt.edu awolpert@roosevelt.edu Abstract We give a radomized algorithm for testig satisfiability of Boolea formulas i cojuctive ormal form with o restrictio o clause legth. This algorithm uses the clauseshorteig approach proposed by Schuler [14]. The ruig time of the algorithm is O 2 1 1/α)) where α = lm/) + Ol l m) ad, m are respectively the umber of variables ad the umber of clauses i the iput formula. This boud is asymptotically better tha the previously best kow 2 1 1/ log2m)) boud for SAT. 1. Keywords: SAT with o restrictio o clause legth, upper boud, clause shorteig Submitted July 2005; revised Novber 2005; published Novber 2005 1. Itroductio Durig the past few years there has bee cosiderable progress i obtaiig upper bouds o the complexity of solvig the Boolea satisfiability probl. This lie of research has produced ew algorithms for k-sat. They were further used to prove otrivial upper bouds for SAT o restrictio o clause legth). Upper bouds for k-sat. The best kow upper bouds for k-sat are based o two approaches: the satisfiability codig lma [9, 8] ad multistart radom walk [12, 13]; both give close upper bouds o the time of solvig k-sat. Schöig s radomized algorithm i [12] has the 2 2/k) boud where is the umber of variables i the iput formula. The radomized algorithm i [8] has a slightly better boud this algorithm is ofte referred to as the PPSZ algorithm). The multistart-radom-walk approach is deradomized usig coverig codes i [3], which gives the best kow 2 2/k + 1)) boud for determiistic k-sat algorithms. The PPSZ algorithm is deradomized with the same boud for Uiquek-SAT [11]. For small values of k, these bouds are improved: for example, 3-SAT ca be solved by a radomized algorithm with the 1.324 boud [7] ad by a determiistic algorithm with the 1.473 boud [2]. 1. The results of this paper without proofs) appeared i the proceedigs of SAT 2005 [6] c 2005 Delft Uiversity of Techology ad the authors.

E. Datsi ad A. Wolpert Upper bouds for SAT o restrictio o clause legth). The first otrivial upper boud for SAT is give i [10]: the 2 1 1/2 ) boud for a radomized algorithm that uses the PPSZ algorithm as a subroutie. A close boud for a determiistic algorithm is proved i [4]. A much better boud for SAT is the 2 1 1/ log2m)) boud, where m is the umber of clauses i the iput formula. This boud is due to Schuler [14]. His radomized algorithm is based o a approach that ca be called clause shorteig: the algorithm repeatedly applies a k-sat subroutie to formulas obtaied by shorteig iput clauses to legth k. Schuler s algorithm is deradomized i [5]; the deradomizatio gives a determiistic algorithm that solves SAT with the same boud. I this paper we improve the 2 1 1/ log2m)) boud: we give a radomized algorithm that solves SAT with the followig upper boud o the ruig time: «2 1 1 l m )+Ol l m) Idea of the algorithm. Our algorithm for SAT is a repetitio of a polyomial-time procedure P that tests satisfiability of a iput formula F. If F is satisfied by a truth assigmet A, the procedure P fids A with probability at least p. As usual, repeatig P o the iput formula O1/p) times, we ca fid A with a costat probability. Whe describig P, we view clauses as sequeces rather tha sets) of literals. Each clause l 1, l 2,..., l s such that s > k is divided ito two segmets as follows: First segmet: l 1,..., l k, Secod segmet: l k+1,..., l s We say that the first segmet is true uder a assigmet A if at least oe of its literals is true uder A; otherwise we say that the first segmet is false uder A. If A is clear from the cotext, we omit the words uder A. Let F cosist of clauses C 1,..., C m ad let A be a fixed satisfyig assigmet to F. The procedure P is based o the followig dichotomy: Case 1. The iput formula F has may clauses i which the first segmet is false. We call such clauses bad. Suppose that the umber of bad clauses is greater tha or equal to some d. The if we choose a clause C i from C 1,..., C m at radom, the first segmet i C i is false with probability at least d/m. Assumig that C i is bad, we ca simplify F by assigig false to all literals occurrig i the first segmet of C i. Case 2. The iput formula F has few less tha d) bad clauses. If we cosider a formula made up of the first segmets of all iput clauses, we get a k-cnf formula i which almost all clauses cotai true literals. Oly the first segmets of bad clauses prevet us from fidig A by applyig a k-sat algorithm to this formula. To fix it, we guess all bad clauses ad replace their first segmets by clauses of legth k that cotai true literals. More exactly, for each bad clause, we replace its first segmet by k literals chose at radom from this clause. The probability that the resultig k-cnf formula is satisfied by A is easily estimated. The procedure P first assumes that Case 2 holds. To process this case, P ivokes a subroutie deoted S) that trasforms F ito a k-cnf formula ad applies a k-sat 1) 50

A Faster Clause-Shorteig Algorithm algorithm. If o satisfyig assigmet is foud, Case 1 is cosidered to hold. The P simplifies the iput formula ad recursively ivokes itself o the simplified formula. The procedure P ad the subroutie S) uses k ad d as parameters. Clearly, the success probability of P depeds o values of the parameters. What values of k ad d maximize the success probability? I Sect. 4 we fid values of the parameters k logm/) + Olog logm)) ad d / log 3 m) that give the followig lower boud o the success probability: «2 1 1 l m )+Ol l m) Sect.2 cotais basic defiitios ad otatio. I Sect. 3 we describe the procedure P ad the subroutie S. I Sect. 4 we prove a lower boud o the success probability of P. I Sect. 5 we defie our algorithm for SAT ad prove the claimed upper boud o its ruig time. 2. Defiitios ad Notatio We deal with Boolea formulas i cojuctive ormal form CNF). By a variable we mea a Boolea variable that takes truth values t true) or f false). A literal is a variable x or its egatio x. If l is a literal the l deotes its complet, i.e. if l is x the l deotes x, ad if l is x the l deotes x. Similarly, if v deotes oe of the truth values t or f, we write v to deote its complet. A clause C is a set of literals such that C cotais o completary literals. A formula F is a set of clauses; ad m deote, respectively, the umber of variables ad the umber of clauses i F. If each clause i F cotais at most k literals, we say that F is a k-cnf formula. A assigmet to variables x 1,..., x is a mappig from {x 1,..., x } to {t, f}. This mappig is exteded to literals: each literal x i is mapped to the complet of the truth value assiged to x i. We say that a clause C is satisfied by a assigmet A or, C is true uder A) if A assigs t to at least oe literal i C. Otherwise, we say that C is falsified by A or, C is false uder A). The formula F is satisfied by A if every clause i F is satisfied by A. I this case, A is called a satisfyig assigmet for F. Let F be a formula ad l 1,..., l s be literals. We write F [l 1 =f,..., l s =f] to deote the formula obtaied from F by assigig the value f to all of l 1,..., l s. This formula is obtaied from F as follows: the clauses that cotai ay literal from l 1,..., l s are deleted from F, ad the literals l 1,..., l s are deleted from the other clauses. Note that F [l 1 = f,..., l s = f] may cotai the pty clause or may be the pty formula. Let A ad A be two assigmets that differ oly i the values assiged to a literal l. The we say that A is obtaied from A by flippig the value of l. By SAT we mea the followig computatioal probl: Give a formula F i CNF, decide whether F is satisfiable. The k-sat probl is the restricted versio of SAT that allows oly clauses cosistig of at most k literals. We write log x to deote log 2 x. 51

E. Datsi ad A. Wolpert 3. Procedure P ad Subroutie S Both procedures S ad P use the parameters k ad d; their values are determied i the ext sectio. 3.1 Descriptio of S Suppose that a iput formula F has a satisfyig assigmet A such that F has oly few < d) clauses i which the first segmet is false uder A, i.e. Case 2 of the dichotomy holds. The the subroutie S fids such a assigmet with the probability estimated i Sect. 4). The subroutie takes two steps: 1. Reductio of F to a k-cnf formula F such that ay satisfyig assigmet to F satisfies F. 2. Use of a k-sat algorithm to fid a satisfyig assigmet to F. First, S guesses all bad clauses, i.e. clauses i which the first segmet is false uder A. More exactly, the subroutie guesses a set {B 1,..., B d 1 } of clauses such that all bad clauses are cotaied i this set. For each clause B i, the subroutie guesses k literals i B i such that this guessed set cotais a true literal. The the subroutie replaces each clause i F by a set of k literals chose from this clause either the guessed set for B 1,..., B d 1 or the first segmets for the other clauses i F. The resultig formula is deoted by F. It is obvious that if we guessed right the A satisfies F. To test satisfiability of k-cnf formulas at the secod step, S uses a polyomial-time radomized algorithm that fids a satisfyig assigmet with a expoetially small probability. We choose Schöig s algorithm [12] to perform this testig we could use ay algorithm that has at least the same success probability, for example the PPSZ algorithm [8]). More exactly, we use oe radom walk of Schöig s algorithm, which has the success probability at least 2 2/k) up to a costat [13]. If S fids a satisfyig assigmet the the procedure returs it ad halts. Otherwise, S should retur o but for techical reasos, istead of o, we require S retur F which is used i the exteral procedure P. Note that if d = 1, i.e. there is o bad clause, the S simply fids a satisfyig assigmet to the k-cnf formula made up of the m first segmets. Also ote that the smaller d, the higher the probability of guessig F. Subroutie S Iput: Formula F with m clauses over variables, itegers k ad d. Output: Satisfyig assigmet or k-cnf formula. 1. Reduce F to a k-cnf formula F as follows: a) Choose d 1 clauses B 1,..., B d 1 i F at radom. b) Replace each clause that does ot belog to {B 1,..., B d 1 } by its first segmet. c) For each B i, choose k literals i B i at radom ad replace B i by the chose set. 52

A Faster Clause-Shorteig Algorithm 2. Test satisfiability of F usig oe radom walk of legth 3 see [12] for details): a) Choose a iitial assigmet a uiformly at radom; b) Repeat 3 times: i. If F is satisfied by the assigmet a the retur a ad halt; ii. Pick ay clause C i F such that C is falsified by a. Choose a literal l i C uiformly at radom. Modify a by flippig the value of l. c) Retur F. 3.2 Descriptio of P Suppose that a iput formula F has a satisfyimg assigmet A. The procedure first calls Subroutie S. The result of S depeds o which case of the dichotomy holds. 1. The iput formula F has may d) bad clauses. The S ever fids A. 2. The iput formula F has few < d) bad clauses. The S returs A with some probability). If S does ot retur a satisfyig assigmet, the procedure P simplifies F ad recursively ivokes itself o the simplified formula, i.e. P processes Case 1 of the dichotomy. To simplify the formula, P uses the fact that A falsifies at least oe clause i F. Note that such a clause has bee obtaied from a log clause loger tha k) i F because each short clause must cotai a true literal. Therefore, P chooses a clause at radom from those clauses i F that are obtaied from log clauses i F. Let l 1,..., l k be all literals of the chose clause. The P assigs f to these literals ad reduces F to F [l 1 =f,..., l k =f]. Procedure P Iput: Formula F with m clauses over variables, itegers k ad d. Output: Satisfyig assigmet or o. 1. Ivoke S o F, k, ad d. 2. Choose a clause at radom from those clauses i F that are obtaied from log loger tha k) clauses of F. Let the chose clause cosist of literals l 1,..., l k. 3. Simplify F by assigig f to all l 1,..., l k, i.e. reduce F to F [l 1 =f,..., l k =f]. 4. Recursively ivoke P o F [l 1 =f,..., l k =f]. 5. Retur o. Note that Schuler s algorithm [14] is a special case of P: take d = 1, k = log2m), ad use the PPSZ algorithm istead of Schöig s algorithm at step 2 i S. 53

E. Datsi ad A. Wolpert 4. Success Probability of Procedure P Give a iput formula F ad some values of the parameters k ad d, Procedure P fids a fixed satisfyig assigmet A or returs o. What is the probability of fidig A? I this sectio, we give a lower boud o the success probability of P ad choose values of k ad d that maximize this boud. We defie the miimum success probabilities for S ad P as follows: s, m, k, d) is the miimum success probability of S where the miimum is take over all pairs F, A such that i) A satisfies F ; ii) F has variables ad m clauses; iii) F has less tha d bad clauses. p, m, k, d) is the miimum success probability of P where the miimum is take over all pairs F, A such that i) A satisfies F ; ii) F has variables ad m clauses. I all ext lmas we assume that m > d 2. Lma 1. Subroutie S has the followig lower boud o the success probability: s, m, k, d) > ) d 1) log e 1 2 k ) Proof. Let A be a fixed satisfyig assigmet to F. The probability of fidig A is the product of three probabilities s 1, s 2, ad s 3. The probability s 1 is the probability that the set {B 1,..., B d 1 } of chose clauses cotais all bad clauses step 1a i S). The probability s 2 is the probability of guessig sets with true literals for all B 1,..., B d 1 step 1b i S). The probability s 3 is the probability of fidig A by oe radom walk step 2 i S). Sice B 1,..., B d 1 are chose at radom from m iput clauses, we have usig Stirlig s approximatio as i [1, page 4]): s 1 1 m d 1) 2πm d 1 m ) ) 1 d 1 m 2 H d 1 m ) m > 3 2 For further estimatio we use the iequality log1 + x) < x log e: s 1 > 3 2 d 1 2 H m ) m = 3 m 2 log 2 d 1) d 1) m d+1) log1+ m d+1) d 1 > 3 2 = 3 2 2 d 1) log m d 1) m d+1) d 1 m d+1) log e d 1 ) d 1) d 1 2 H m ) m To estimate s 2, we estimate the probability of guessig a set with a true literal for each B i. Obviously, B i has at most k) subsets of literals of size k. For a fixed true literal i Bi, there are at most 1 k 1) of such subsets that cotai this literal. Hece s 2 ) 1 d 1 k 1) = ) k d 1 k) 54

A Faster Clause-Shorteig Algorithm A lower boud o s 3 is proved i [13]: 2 ) 3 2 2. k Usig the iequality 1 x 2 x log e, we have s 3 2 ) 3 2 2 k = 2 3 2 1 1 ) k 2 3 2 2 log e k = 2 log e 3 2 1 k ) The product of the above bouds for s 1, s 2, ad s 3 gives the followig boud: s 1 s 2 s 3 > 3 2 = kd 1) d 1) d 1) k ) d 1) 2 1 log e k ) Fially, we relax the above boud to ) d 1) log e 1 2 k ). ) d 1) 2/3) 2 1 log e k ) Our goal is to choose values of the parameters k ad d viewed as fuctios of ad m) so as to maximize the success probability of Procedure P. We choose the fuctios =, m) ad d 0 = d 0, m) i two steps. First, we fid a lower boud o the success probability of P. The maximum of this boud is attaied whe k is a certai fuctio of, m, ad d. The, substitutig this fuctio i the boud, we fid d 0, m) that maximizes the boud. Lma 2. For all k such that k 1 + log ) d d 1) log e Procedure P has the followig lower boud o the success probability: p, m, k, d) > ) d 1) log e 1 2 k ) Proof. Let A be a fixed satisfyig assigmet to F. Suppose that Procedure P fids A with t recursive calls, where t is some iteger i the iterval 0 t /k. The P simplifies formulas t times. For each simplificatio, the probability of guessig a clause i F that cosists of false literals is at least d/m. Therefore, with probability at least d/m) t, we get a formula G such that G is still satisfied by A; G has less tha d bad clauses; G has at most kt variables. Usig Lma 1, we get the followig lower boud o the probability of fidig A with t recursive calls: p, m, k, d) > ) d t m kt)) d 1) log e kt)1 2 k ) = ) d t m ) d 1) 1 kt ) d 1) 2 1 log e k )+kt1 log e k ) = β, m, k, d) d m ) t 1 kt = β, m, k, d) 2 t log m d ) d 1) 2 kt1 log e k ) ) d 1) log1 kt ) + kt1 log e k ) 2) 55

E. Datsi ad A. Wolpert where β deotes the lower boud i the claim: β, m, k, d) = ) d 1) log e 1 2 k ) Note that the β boud is the same as the boud give by Lma 1. Sice log1 x) < x log e for all 0 x < 1, we have p, m, k, d) > β, m, k, d) 2 t log m kd 1) )+ log e+k log e d = β, m, k, d) 2 t k 1+ mi t [β, m, k, d) 2 t k d 1) log e 1+ log d ) d 1) log e log d ) ] where the miimum is take over all t such that 0 t /k. The secod iequality i 2) is equivalet to ) d 1) log e k 1 + log ) d 0. Therefore, the miimum is attaied whe t = 0, which proves the claim. It follows from Lma 2 that Procedure P has the followig lower boud o the success probability: p, m, k, d) > mi [β, m, k, d) 2 t d 1) log e k 1+ log d ) ] 3) t What value of k maximizes this boud for give, m, ad d? A more thorough aalysis shows that this boud is maximum whe the iequality o k is replaced by the equality: k = 1 + log ) d d 1) log e Cosider the boud obtaied from 3) by substitutig this value for k. Differetiatio by d shows that this boud is maximum whe d is close to / log 3 ). However, k ad d must be itegers, so we take the ceiligs of those values of k ad d. Lma 3. For all ad m such that 10 < 2 3 /2, take the followig values of k ad d: = log ) + 3 log log) 1 + log e d 0 = log 3 ) log 3 4) ) The p, m,, d 0 ) > 2 1 log 2 +2 56

A Faster Clause-Shorteig Algorithm Proof. First, we show that Lma 2 ca be applied for ad d 0, i.e. we show that 2) holds. It is straightforward to check the first iequality i 2). For the secod iequality we have ) ) 1 + d 0 1) log e log d 0 = ) 1 + d 0 1) log e log) + log d 0 log ) ) + 3 log log) log) + log = 0. log 3 ) Next, we substitute ad d 0 i the boud give by Lma 2. Note that d 0 = / log 3 )+ δ where 0 δ < 1. The we have usig < ) p, m,, d 0 ) > ) d0 1) 2 1 log e = ) δ+1 ) log 3 ) 2 1 log e > ) log 3 ) 2 1 log e = 2 log 2 ) > 2 = 2 log 2 ) log) 1 log e log) log ) 1 log e log) 2 log 2 1 log e ) It rais to prove that 2 log 2 ) + log e 1 log e ) + 2 This iequality is equivalet to log e log e + 2 2 log 2 ) which, i tur, is equivalet to ) 2 ) k0 1 + 2k0 log) log e 5) Our assumptio 2 3 /2 implies < log). Ideed, ) < log + 3 log log) + 1 [ ) ] = log) log loglog 3 )) 2 log) Therefore, the first factor i 5) is less tha 1 uder the assumptio. To make sure that the secod factor is less tha log e 1.44, we otice that 5. Ideed, if 10 the > 3 log log) > 3 log log10) > 5. 57

E. Datsi ad A. Wolpert 5. Mai Result The mai algorithm Algorithm A) is a repetitio of Procedure P. Lma 3 shows that if we substitute ad d 0 defied i 4) for the parameters of P the the success probability of P is at least 2 1 log 2 +2 The umber r of repetitios is take to be the iverse to this probability: r = 2 1 log e +2 which gives a costat success probability of Algorithm A. Algorithm A Iput: Formula F i CNF with m clauses over variables. Output: Satisfyig assigmet or o. 1. Compute ad d 0 as i 4) ad r as i 6). 2. Repeat the followig r times: a) Ru PF,, d 0 ); b) If a satisfyig assigmet is foud, retur it ad halt. 3. Retur o. Theor 1. Algorithm A rus i time O 2 m) 2 1 log e +2 For ay satisfiable iput formula such that 10 < 2 3 /2, Algorithm A fids a satisfyig assigmet with probability greater tha 1/2. Proof. To prove the first claim, we eed to estimate the ruig time of Procedure P. The procedure recursively ivokes itself at most / times. Each call scas the iput formula ad elimiates at most variables. Therefore, the procedure performs less tha scas. Algorithm A repeats the procedure at most r times, which gives the boud i the claim. Let p A, m) be the success probability of Algorithm A. The p A, m) 1 1 p, m,, d 0 )) r Usig Lma 3 ad the iequality 1 x) r e xr, we have p A, m) 1 e 1. 6) Rark 1. Note that we ca write the boud i Theor 1 as O 2 m) 2 «1 1 l m )+Ol l m). 58

A Faster Clause-Shorteig Algorithm Rark 2. For formulas with a costat clause desity, i.e. if m is liear i, our boud is better tha i the geeral case: 2 1 1 Ol l m) Rark 3. Our algorithm is based o the clause-shorteig approach proposed by Schuler [14]. His algorithm is deradomized with the same upper boud o the ruig time [5]. It would be atural to try to apply the same method of deradomizatio to our algorithm. However, the direct applicatio gives a determiistic algorithm with a much worse upper boud. Is it possible to deradomize Algorithm A with the same or a slightly worse upper boud? Ackowledgmets We thak Edward A. Hirsch ad Daiel Rolf for helpful discussios. We also thak the aoymous reviewers for their help i improvig the paper. Refereces [1] B. Bollobás. Radom Graphs. Cambridge Uiversity Press, 2d editio, 2001. [2] T. Bruegga ad W. Ker. A improved local search algorithm for 3-SAT. Theoretical Computer Sciece, 329 1-3):303 313, Decber 2004. [3] E. Datsi, A. Goerdt, E. A. Hirsch, R. Kaa, J. Kleiberg, C. Papadimitriou, P. Raghava, ad U. Schöig. A determiistic 2 2/k + 1)) algorithm for k-sat based o local search. Theoretical Computer Sciece, 289 1):69 83, 2002. [4] E. Datsi, E. A. Hirsch, ad A. Wolpert. Algorithms for SAT based o search i Hammig balls. I Proceedigs of the 21st Aual Symposium o Theoretical Aspects of Computer Sciece, STACS 2004, volume 2996 of Lecture Notes i Computer Sciece, pages 141 151. Spriger, March 2004. [5] E. Datsi ad A. Wolpert. Deradomizatio of Schuler s algorithm for SAT. I Proceedigs of the 7th Iteratioal Coferece o Theory ad Applicatios of Satisfiability Testig, SAT 2004, pages 69 75, May 2004. [6] E. Datsi ad A. Wolpert. A improved upper boud for SAT. I Proceedigs of the 8th Iteratioal Coferece o Theory ad Applicatios o Satisfiability Testig, SAT 2005, volume 3569 of Lecture Notes i Computer Sciece, pages 400 407. Spriger, Jue 2005. [7] K. Iwama ad S. Tamaki. Improved upper bouds for 3-SAT. I Proceedigs of the 15th Aual ACM-SIAM Symposium o Discrete Algorithms, SODA 2004, page 328, Jauary 2004. A prelimiary versio appeared i Electroic Colloquium o Computatioal Complexity, Report No. 53, July 2003. 59

E. Datsi ad A. Wolpert [8] R. Paturi, P. Pudlák, M. E. Saks, ad F. Zae. A improved expoetial-time algorithm for k-sat. I Proceedigs of the 39th Aual IEEE Symposium o Foudatios of Computer Sciece, FOCS 98, pages 628 637, 1998. [9] R. Paturi, P. Pudlák, ad F. Zae. Satisfiability codig lma. I Proceedigs of the 38th Aual IEEE Symposium o Foudatios of Computer Sciece, FOCS 97, pages 566 574, 1997. [10] P. Pudlák. Satisfiability algorithms ad logic. I Proceedigs of the 23rd Iteratioal Symposium o Mathatical Foudatios of Computer Sciece MFCS 98), volume 1450 of Lecture Notes i Computer Sciece, pages 129 141. Spriger-Verlag, 1998. [11] D. Rolf. Deradomizatio of PPSZ for Uique-k-SAT. I Proceedigs of the 8th Iteratioal Coferece o Theory ad Applicatios o Satisfiability Testig, SAT 2005, volume 3569 of Lecture Notes i Computer Sciece, pages 216 225. Spriger, Jue 2005. [12] U. Schöig. A probabilistic algorithm for k-sat ad costrait satisfactio probls. I Proceedigs of the 40th Aual IEEE Symposium o Foudatios of Computer Sciece, FOCS 99, pages 410 414, 1999. [13] U. Schöig. A probabilistic algorithm for k-sat based o limited local search ad restart. Algorithmica, 32 4):615 623, 2002. [14] R. Schuler. A algorithm for the satisfiability probl of formulas i cojuctive ormal form. Joural of Algorithms, 54 1):40 44, Jauary 2005. A prelimiary versio appeared as a techical report i 2003. 60