Quantified Boolean Formulas: Complexity and Expressiveness

Transcription

1 Dagstuhl SAT Interactions 2012 Quantified Boolean Formulas: Complexity and Expressiveness Uwe Bubeck University of Paderborn

2 Outline Introduction Free Variables and Equivalence Complexity Expressiveness Conclusion Uwe Bubeck QBF: Complexity and Expressiveness 2

3 Section 1 Introduction

4 Quantified Boolean Formulas 1/2 QBF extends propositional logic by allowing universal and existential quantifiers over propositional variables. Inductive definition: 1. Every propositional formula is a QBF. 2. If Φ is a QBF then xφ and yφ are also QBFs. 3. If Φ 1 and Φ 2 are QBFs then Φ 1, Φ 1 Φ 2 and Φ 1 Φ 2 are also QBFs. Uwe Bubeck QBF: Complexity and Expressiveness 4

5 Quantified Boolean Formulas 2/2 In a closed QBF, every variable is quantified. Semantics definition for closed QBF: y Φ y is true if and only if Φ y/0 is true or Φ y/1 is true. x Φ x is true if and only if Φ x/0 is true and Φ x/1 is true. A closed QBF is either true or false. Uwe Bubeck QBF: Complexity and Expressiveness 5

6 Prenexing Closed QBFs Φ and Ψ are logically equivalent (Φ Ψ) they are satisfiability equivalent they both evaluate to the same truth value. Every QBF can be transformed in linear time into a logically equivalent prenex formula Q 1 v 1 Q k v k φ by: 1. renaming quantified variables, 2. transformation into negation normal form (NNF), 3. moving quantifiers to the front by maxiscoping: Qv Φ Ψ Qv (Φ Ψ) for {, } and Q,. Uwe Bubeck QBF: Complexity and Expressiveness 6

7 Tree Models: Definition A closed prenex QBF Q 1 v 1 Q k v k φ is true if and only if there exists a tree such that: [Samulowitz et. al., 2006] 1. Each inner node is labeled with a variable v i, 1 i k, its outgoing edges are labeled with v i = 0 or v i = Inner nodes labeled with v i have two children if and only if Q i =, and one child otherwise. 3. For each path from root to leaf labeled with (v i1,, v i j ), we have 1 i 1 < < i j k (i.e. order as in the prefix). 4. On each path from root to leaf, the edge labels are a satisfying assignment to the propositional matrix φ. Uwe Bubeck QBF: Complexity and Expressiveness 7

8 Tree Models: Definition A closed prenex QBF Q 1 v 1 Q k v k φ is true if and only if there exists a tree such that: [Samulowitz et. al., 2006] 1. Each inner node is labeled with a variable v i, 1 i k, its outgoing edges are labeled with v i = 0 or v i = Inner nodes labeled with v i have two children if and only if Q i =, and one child otherwise. 3. For each path from root to leaf labeled with (v i1,, v i j ), we have 1 i 1 < < i j k (i.e. order as in the prefix). 4. On each path from root to leaf, the edge labels are a satisfying assignment to the propositional matrix φ. Uwe Bubeck QBF: Complexity and Expressiveness 8

9 Tree Models: Example Example: x 1 y 1 x 2 y 2 x 1 y 2 x 1 y 2 y 1 x 2 x 1 x 1 = 0 x 1 = 1 y 1 y 1 y 1 =? y 1 =? x 2 x 2 = 0 x 2 = 1 x 2 = 0 x 2 = 1 x 2 y 2 y 2 y 2 y 2 y 2 =? y 2 =? y 2 =? y 2 =? Uwe Bubeck QBF: Complexity and Expressiveness 9

10 Tree Models: Example Example: x 1 y 1 x 2 y 2 x 1 y 2 x 1 y 2 y 1 x 2 x 1 x 1 = 0 x 1 = 1 y 1 y 1 y 1 = 1 y 1 =? x 2 x 2 = 0 x 2 = 1 x 2 = 0 x 2 = 1 x 2 y 2 y 2 y 2 y 2 y 2 =? y 2 =? y 2 =? y 2 =? Uwe Bubeck QBF: Complexity and Expressiveness 10

11 Tree Models: Example Example: x 1 y 1 x 2 y 2 x 1 y 2 x 1 y 2 y 1 x 2 x 1 x 1 = 0 x 1 = 1 y 1 y 1 y 1 = 1 y 1 =? x 2 x 2 = 0 x 2 = 1 x 2 = 0 x 2 = 1 x 2 y 2 y 2 y 2 y 2 y 2 = 0 y 2 = 0 y 2 =? y 2 =? Uwe Bubeck QBF: Complexity and Expressiveness 11

12 Tree Models: Example Example: x 1 y 1 x 2 y 2 x 1 y 2 x 1 y 2 y 1 x 2 x 1 x 1 = 0 x 1 = 1 y 1 y 1 y 1 = 1 y 1 = 1 x 2 x 2 = 0 x 2 = 1 x 2 = 0 x 2 = 1 x 2 y 2 y 2 y 2 y 2 y 2 = 0 y 2 = 0 y 2 = 1 y 2 = 1 Uwe Bubeck QBF: Complexity and Expressiveness 12

13 Tree Models: Example Example: x 1 y 1 x 2 y 2 x 1 y 2 x 1 y 2 y 1 x 2 x 1 x 1 = 0 x 1 = 1 y 1 y 1 y 1 = 1 y 1 = 1 y 2 y 2 y 2 = 0 y 2 = 1 Universal Reduction Uwe Bubeck QBF: Complexity and Expressiveness 13

14 Tree Models: Generalization The requirement that variables must appear in the same order as in the prefix (rule 3) can be further relaxed: by grouping similar quantifiers into quantifier blocks. v 1 v n1 v n1 +1 v n2 v n2 +1 v n3 φ S 1 S 2 S 3 by considering variable dependency schemes. Uwe Bubeck QBF: Complexity and Expressiveness 14

15 Dependency Schemes 1/2 Consider the following formula: x y 1 u y 2 y 3 [ x u y 3 y 1 y 3 y 1 y 3 u y 3 ( y 1 y 2 )] Which variables must be above y 2 in the tree without altering the truth value of the formula? In general: dependency decision problem is PSPACE-complete. [Samer / Szeider, 2007] Uwe Bubeck QBF: Complexity and Expressiveness 15

16 Dependency Schemes 2/2 x y 1 u y 2 y 3 [ x u y 3 y 1 y 3 y 1 y 3 u y 3 ( y 1 y 2 )] One simple heuristic: recovering non-prenex structure: x y 1 [[ u y 3 x u y 3 y 1 y 3 y 1 y 3 u y 3 ] [ y 2 y 1 y 2 ]] Identified informally by [Biere, 2004], Formalized in [Samer / Szeider, 2007], [Bubeck / Kleine Büning, 2007], Currently tightest scheme by [van Gelder, 2011]. Still no practicable algorithm for computing tight schemes! Uwe Bubeck QBF: Complexity and Expressiveness 16

17 Section 2 Free Variables and Equivalence

18 Semantics of Free Variables 1/2 A focus of this talk is to allow free variables. Notation: Φ z 1,, z n for formula Φ with free variables z 1,, z n. QBF* is the class of quantified Boolean formulas with free variables. The valuation of a QBF* depends on the values of the free variables: Φ z 1,, z n t: z 1,, z n QBF is satisfied by a truth assignment 0,1 if and only if Φ t(z 1 ),, t(z n ) Φ[z 1 /t z 1,, z n /t z n ] is true. Uwe Bubeck QBF: Complexity and Expressiveness 18

19 Semantics of Free Variables 2/2 Example: Φ a, b, c, d : = y a y b y y c d Then Φ 0,1,0,1 Φ 0,1,0,0 etc. = y 0 y 1 y ( y 0 1) is true, = y 0 y 1 y ( y 0 0) is false, y a y b y y c d is true if and only if a c d b c d is true. Every QBF* is equivalent to a propositional formula. Uwe Bubeck QBF: Complexity and Expressiveness 19

20 Logical Equivalence Propositional or quantified Boolean formulas α and β with (free) variables z 1,, z n are logically equivalent, written as α z 1,, z n β(z 1,, z n ), if and only if α t(z 1 ),, t(z n ) = β(t(z 1 ),, t(z n )) for each truth assignment t to z 1,, z n. Quantified variables are not directly considered for logical equivalence. They can be seen as local within the respective formula. Uwe Bubeck QBF: Complexity and Expressiveness 20

21 Satisfiability Equivalence a c d b c d y a y b y y c d Without quantifier: a c d b c d a y b y y c d Problem: a = b = 1, c = d = 0, y = 1 satisfies left, but not right side. Relaxation: satisfiability equivalence a c d b c d SAT a y b y y c d Existence of a satisfying assignment for one side implies there is some satisfying assignment for the other side. Uwe Bubeck QBF: Complexity and Expressiveness 21

22 Equivalence and Rewriting For propositional formulas α, β, γ, it holds that β γ α α β/γ. Parts of a formula can be replaced with logically equivalent formulas. But: β SAT γ α SAT α β/γ. Satisfiability equivalence is sometimes too weak. Uwe Bubeck QBF: Complexity and Expressiveness 22

23 Section 3 Complexity

24 QBF* Complexity 1/2 Well known: the decision problem for QBF and the satisfiability problem for QBF* are PSPACE-complete. [Meyer / Stockmeyer, 1973] QBF* consequence and equivalence problems are also PSPACE-complete. QDNF* remains PSPACE-complete. Some verification problems are also PSPACE-complete: Propositional LTL satisfiability [Sistla / Clarke, 1985] Symbolic reachability in sequential circuits [Savitch, 1970] Uwe Bubeck QBF: Complexity and Expressiveness 24

25 QBF* Complexity 2/2 Typical approaches to reduce the complexity: Restriction of the matrix to special classes of propositional formulas Bounding of quantifier alternations in the prefix: φ φ k quantifier blocks, outermost existential: k quantifier blocks, outermost universal: prefix type Σ k prefix type Π k Uwe Bubeck QBF: Complexity and Expressiveness 25

26 Polynomial-time Hierarchy Well-known relationship with polynomial-time hierarchy: For fixed k 1, the satisfiability problem for QBF* with prefix type Σ k is Σ P k -complete, and Π P k -complete for prefix type Π k. [Stockmeyer, 1976] P Σ k+1 NP Σ k P P, Π k+1 Δ 0 P Σ 0 P Π 0 P P P P co Σ k+1, Δ k+1 P Σ k P Uwe Bubeck QBF: Complexity and Expressiveness 26

27 Schaefer s Dichotomy Theorem Consider a quantified constraint expression Q 1 y 1 Q n y n f 1 x 1,1,, x 1,m1 f k x k,1,, x k,mk with Q 1,, Q n {, }, Boolean functions f 1,, f k C and arguments x i,j y 1,, y n {0,1}. Dichotomy Theorem: Let C be a finite set of constraints. If C is Horn, anti-horn, bijunctive / Krom (equiv. to 2-CNF) or affine (equiv. to XOR-CNF) then QSAT c C is in P. Otherwise, QSAT c C is PSPACE-complete. [Schaefer, 1978], [Dalmau, 1997] [Creignou / Khanna et al., 2001] Uwe Bubeck QBF: Complexity and Expressiveness 27

28 QHORN* and Q2-CNF* QHORN* is the class of QCNF* formulas with at most one positive literal per clause. QHORN* satisfiability is decidable in quadratic time O( Φ ), where is the number of universal variables and Φ the formula length. Algorithm: by unit propagation [Flögel et al. 1995] or by universal expansion [Bubeck / Kleine Büning, 2008] Q2-CNF* satisfiability is decidable in linear time. Algorithm: by strongly connected components [Aspvall et al., 1979] Uwe Bubeck QBF: Complexity and Expressiveness 28

29 Section 4 Expressiveness

30 Other Representations We can also define logical equivalence with other representations of Boolean functions: a c d b a c d b c d y a y b y y c d For each assignment to a, b, c, d, all three representations evaluate to the same truth value. Uwe Bubeck QBF: Complexity and Expressiveness 30

31 Quantified Encodings 1/2 Main Theme: How compact are encodings in QBF* (or subclasses) versus other logically equivalent representations? Encoding Techniques: 1. Abbreviate exact repetitions by existential variables: A B C D A B C E A B C F y y A B C y D y E y F Simple implication if term occurs only in one polarity: y y A B C y D y E y F (quantified versions of [Tseitin, 1970] and [Plaisted, 1986]) Uwe Bubeck QBF: Complexity and Expressiveness 31

32 Quantified Encodings 2/2 2. Compress conjunctions of renamings / instantiations by universal variables: φ A 1, B 1 φ A 2, B 2 φ A 3, B 3 u v u A i v B i i=1 3 φ u, v [Dershowitz et al., 2005], [Meyer/Stockmeyer, 1973] 3. Iterative Squaring (Extension of 2.): Φ x 0, x n = x 1 x n 1 φ x 0, x 1 φ x 1, x 2 φ x n 1, x n Encoding: Φ n x 0, x n y Φ n/2 x 0, y Φ n/2 y, x n y u v u x 0 v y u y v x n Φ n/2 u, v [Meyer/Stockmeyer, 1973] Uwe Bubeck QBF: Complexity and Expressiveness 32

33 Nested Instantiations 1/3 Instantiations might also be nested: A 1 A 2 A 1 A 2 B 1 B 2 B 1 B 2 is of the form ψ(φ A 1, A 2, φ B 1, B 2 ). Other example: Parity of n Boolean variables f 0 p 1, p 2 p 1 p 2 p 1 p 2 f 1 p 1, p 2, p 3, p 4 f 0 f 0 p 1, p 2, f 0 p 3, p 4 f 2 p 1,, p 16 f 1 (f 1 p 1,, p 4,, f 1 p 13,, p 16 ) log 2 log 2 n + 1 definitions of size O(n). Uwe Bubeck QBF: Complexity and Expressiveness 33

34 Nested Instantiations 2/3 General Definition: A Nested Boolean Function (NBF) is a sequence of functions D f k = (f 0,, f k ) with initial functions f 0,, f t defined by f i (x i ) α i (x i ) for a propositional formula α i over x i (x i,1,, x i,n i) compound functions f t+1,, f k of the form f i x i f j0 (f j1 (x i 1 ),, f jr (x i r )) with previously defined functions f j0,, f jr and matching tuples x i i 1,, x r over variables from x i or Boolean constants. [Bubeck / Kleine Büning, 2012], [Cook / Soltys, 1999] Uwe Bubeck QBF: Complexity and Expressiveness 34

35 Nested Instantiations 3/3 By clever combination of previously presented encoding techniques, every NBF can be transformed in linear time into a logically equivalent prenex QBF*. [Bubeck / Kleine Büning, 2012] The inverse direction is very simple by simulating quantifier expansion, but length increases slightly to O( v Φ ) for a QBF* Φ with v quantified variables. Application: Configuration Problems ( Talk by Hans) Future Work: Solvers, interesting NBF subclasses Uwe Bubeck QBF: Complexity and Expressiveness 35

36 Circuits and Existential Quantifiers There is a close connection between fan-out in Boolean circuits and existential quantification: 1. Transformation from circuit to formula a c d b y 1 y 8 y 1 y 2 y 3 y 4 a y 1 c y 2 d y 3 b y 4 y 5 y 2 y 3 y 5 y 1 y 5 y 6 y 6 y 7 y 5 y 4 y 7 y 6 y 7 y 8 y 8 y 8 [Bauer / Brand et al., 1973] [Anderaa / Börger, 1981] Uwe Bubeck QBF: Complexity and Expressiveness 36

37 Circuits and Existential Quantifiers There is a close connection between fan-out in Boolean circuits and existential quantification: 1. Transformation from circuit to formula a c d b y 1 y 8 y 1 y 2 y 3 y 4 a y 1 c y 2 d y 3 b y 4 y 5 y 2 y 3 y 5 y 1 y 5 y 6 y 6 y 7 y 5 y 4 y 7 y 6 y 8 ) ( y 7 y 8 y 8 y 8 y a y b y y c d Uwe Bubeck QBF: Complexity and Expressiveness 37

38 The Class HORN b The previous linear transformation produces existentially quantified formulas in CNF with at most one positive bound variable per clause (i.e. the bound variables respect the Horn property). We call such formulas HORN b. General: for QK QCNF, QK b is the class of formulas Φ z 1,, z n = Q 1 v 1 Q m v m i φ i b (v 1,, v m ) φ i f (z 1,, z n ) where Q 1 v 1 Q m v m i φ b i (v 1,, v m ) is a formula in QK, and φ f i (z 1,, z n ) an arbitrary clause over free variables. Uwe Bubeck QBF: Complexity and Expressiveness 38

39 Circuits and Existential Quantifiers 2. Transformation from formula to circuit Any HORN b formula can be transformed in polynomial time into a logically equivalent Boolean circuit. [Anderaa / Börger, 1981] [Kleine Büning / Zhao / Bubeck, 2009] HORN b and circuits have similar expressiveness. Uwe Bubeck QBF: Complexity and Expressiveness 39

40 HORN b vs Propositional Formulas There are HORN b (even HORN*) formulas for which there is no logically equivalent propositional CNF of polynomial length. [Kleine Büning / Lettmann, 1999] Is a polynomial-length transformation from HORN b to arbitrary propositional formulas possible? Equivalent: Are circuits with unrestricted fan-out more expressive than with fan-out 1? Uwe Bubeck QBF: Complexity and Expressiveness 40

41 HORN b vs QHORN b Can the expressive power of HORN b be enhanced by universal quantification? Not significantly. For every Φ QHORN b with universal quantifiers, there exists a logically equivalent HORN b formula of quadratic length O( Φ ) which can be computed also in time O( Φ ). [Bubeck, 2010] That means QHORN b satisfiability is NP-complete. Uwe Bubeck QBF: Complexity and Expressiveness 41

42 Minimal Falsity and Quantification A closed QCNF formula is minimal false (MF) iff it is false and removing an arbitrary clause makes it true. Example: x y 0 y 1 y 0 z 0 y 0 z 1 y 0 z 2 y 1 z 3 x z 4 Bound parts: x y 0 y 1 (y 0 ) ( y 0 ) ( y 0 ) (y 1 ) (x) MF subformulas: 1. y 0 (y 0 ) ( y 0 ) 2. y 0 (y 0 ) ( y o ) 3. x (x) Uwe Bubeck QBF: Complexity and Expressiveness 42

43 Minimal Falsity and Quantification x y 0 y 1 y 0 z 0 y 0 z 1 y 0 z 2 y 1 z 3 x z 4 Bound parts: x y 0 y 1 (y 0 ) ( y 0 ) ( y 0 ) (y 1 ) (x) MF subformulas: 1. y 0 (y 0 ) ( y 0 ) 2. y 0 (y 0 ) ( y o ) 3. x (x) Corresponding bound parts: 1. z 0, z 1 2. z 0, z 2 3. z 4 For each MF subformula, one of the corresponding free parts must be satisfied. So the formula is equivalent to z 0 z 1 z 0 z 2 z 4. Uwe Bubeck QBF: Complexity and Expressiveness 43

44 Minimal Falsity and Quantification MF subformulas of the bound parts determine the role of the free parts in QCNF* formulas: Q φ i b φ i f φ i1 f φ ir f 1 i q Q φ b i φ b 1 i r MF [Bubeck / Kleine Büning, 2010] How do restrictions on the structure of the MF skeleton influence the expressiveness? all MF subformulas of the bound parts Uwe Bubeck QBF: Complexity and Expressiveness 44

45 Section 5 Conclusion

46 Conclusion Free variables are nothing to be afraid of. Compared to propositional logic, there is a very rich set of QBF* subclasses. How do these subclasses differ in expressiveness? Is there a hierarchy PROP < poly len HORN b < poly len CNF* < poly len CNF*...? How are these classes related to other representations, e.g. circuits? Uwe Bubeck QBF: Complexity and Expressiveness 46

47 References 1/6 S. Anderaa and E. Börger. The Equivalence of Horn and Network Complexity for Boolean Functions B. Aspvall, M. Plass and R. Tarjan. A Linear-Time Algorithm for Testing the Truth of Certain Quantified Boolean Formulas M. Bauer, D. Brand, M. Fischer, A. Meyer and M. Paterson. A Note on Disjunctive Form Tautologies A. Biere. Resolve and Expand U. Bubeck. Model-Based Transformations for Quantified Boolean Formulas Uwe Bubeck QBF: Complexity and Expressiveness 51

48 References 2/6 U. Bubeck and H. Kleine Büning. Bounded Universal Expansion for Preprocessing QBF U. Bubeck and H. Kleine Büning. Models and Quantifier Elimination for Quantified Horn Formulas U. Bubeck and H. Kleine Büning. The Power of Auxiliary Variables for Propositional and Quantified Boolean Formulas U. Bubeck and H. Kleine Büning. Encoding Nested Boolean Functions as Quantified Boolean Formulas S. Cook and M. Soltys. Boolean Programs and Quantified Propositional Proof Systems Uwe Bubeck QBF: Complexity and Expressiveness 52

49 References 3/6 N. Creignou, S. Khanna and M. Sudan. Complexity Classifications of Boolean Constraint Satisfaction Problems V. Dalmau Some Dichotomy Theorems on Constant-free Boolean Formulas N. Dershowitz, Z. Hanna and J. Katz. Bounded Model Checking with QBF U. Egly, M. Seidl, H. Tompits, S. Woltran and M. Zolda. Comparing Different Prenexing Strategies for Quantified Boolean Formulas A. Flögel, M. Karpinski and H. Kleine Büning. Resolution for Quantified Boolean Formulas Uwe Bubeck QBF: Complexity and Expressiveness 53

50 References 4/6 H. Kleine Büning and U. Bubeck. Theory of Quantified Boolean Formulas H. Kleine Büning and T. Lettmann. Propositional Logic: Deduction and Algorithms H. Kleine Büning, X. Zhao and U. Bubeck. Resolution and Expressiveness of Subclasses of Quantified Boolean Formulas and Circuits A. Meyer and L. Stockmeyer. Word Problems Requiring Exponential Time D. Plaisted and S. Greenbaum. A Structure-preserving Clause Form Translation Uwe Bubeck QBF: Complexity and Expressiveness 54

51 References 5/6 M. Samer and S. Szeider. Backdoor Sets of Quantified Boolean Formulas H. Samulowitz, J. Davies, F. Bacchus. Preprocessing QBF W. Savitch. Relationships Between Nondeterministic and Deterministic Tape Complexities T. Schaefer. The Complexity of Satisfiability Problems A. Sistla and E. Clarke. The Complexity of Propositional Linear Time Logics Uwe Bubeck QBF: Complexity and Expressiveness 55

52 References 6/6 L. Stockmeyer. The Polynomial-Time Hierarchy G. Tseitin. On the Complexity of Derivation in Propositional Calculus A. van Gelder. Variable Independence and Resolution Paths Uwe Bubeck QBF: Complexity and Expressiveness 56