Lower Bound Techniques for QBF Proof Systems Meena Mahajan The - - PowerPoint PPT Presentation

lower bound techniques for qbf proof systems
SMART_READER_LITE
LIVE PREVIEW

Lower Bound Techniques for QBF Proof Systems Meena Mahajan The - - PowerPoint PPT Presentation

Lower Bound Techniques for QBF Proof Systems Meena Mahajan The Institute of Mathematical Sciences, HBNI, Chennai. Complexity, Algorithms, Automata, Logic Meeting CAALM 2019, CMI, India. 21-25 Jan 2019 24 Jan 2019 Meena Mahajan, IMSc Credits


slide-1
SLIDE 1

Lower Bound Techniques for QBF Proof Systems

Meena Mahajan

The Institute of Mathematical Sciences, HBNI, Chennai.

Complexity, Algorithms, Automata, Logic Meeting CAALM 2019, CMI, India. 21-25 Jan 2019

24 Jan 2019 Meena Mahajan, IMSc

slide-2
SLIDE 2

Credits

My work on QBF proof complexity – partially supported by the EU Marie Curie IRSES grant CORCON. joint work with Olaf Beyersdorff Firedrich-Schiller Univ, Jena, Germany Leroy Chew Univ of Leeds, UK Anil Shukla formerly at IMSc, Chennai now at IIT Ropar

24 Jan 2019 Meena Mahajan, IMSc

slide-3
SLIDE 3

QBF Proof systems

What are they? Why do we study them?

24 Jan 2019 Meena Mahajan, IMSc

slide-4
SLIDE 4

Satisfiability

SAT: Satisfiability.

  • eg. Is there an assignment to x, y, z satisfying all the clauses

(x ∨ y ∨ z), (x ∨ ¬ y ∨ ¬ z), (¬ x ∨ y ∨ ¬ z), (¬ x ∨ ¬ y ∨ z)? Quintessential NP-complete problem. Very hard – in theory.

24 Jan 2019 Meena Mahajan, IMSc

slide-5
SLIDE 5

Satisfiability

SAT: Satisfiability.

  • eg. Is there an assignment to x, y, z satisfying all the clauses

(x ∨ y ∨ z), (x ∨ ¬ y ∨ ¬ z), (¬ x ∨ y ∨ ¬ z), (¬ x ∨ ¬ y ∨ z)? Quintessential NP-complete problem. Very hard – in theory. In practice – a solved problem! Many good SAT solvers around.

24 Jan 2019 Meena Mahajan, IMSc

slide-6
SLIDE 6

Satisfiability

SAT: Satisfiability.

  • eg. Is there an assignment to x, y, z satisfying all the clauses

(x ∨ y ∨ z), (x ∨ ¬ y ∨ ¬ z), (¬ x ∨ y ∨ ¬ z), (¬ x ∨ ¬ y ∨ z)? Quintessential NP-complete problem. Very hard – in theory. In practice – a solved problem! Many good SAT solvers around. Ambitious programs to design good solvers for problems harder than SAT.

24 Jan 2019 Meena Mahajan, IMSc

slide-7
SLIDE 7

QBF solvers

QBF: Quantified Boolean Formula Subsumes SAT. eg. Is this QBF true? ∃x∃y∃z(x ∨ y ∨ z), (x ∨ ¬ y ∨ ¬ z), (¬ x ∨ y ∨ ¬ z), (¬ x ∨ ¬ y ∨ z) PSPACE-complete, so much more expressive than SAT.

  • eg. Is this formula true?

∃e ∀u ∃c ∃d (¬ e ∨ c)(e ∨ d)(¬ u ∨ c)(u ∨ d)(¬ c ∨ ¬ d) Quite a few QBF solvers developed in the last couple of decades.

24 Jan 2019 Meena Mahajan, IMSc

slide-8
SLIDE 8

Improving solvers

How to improve the performance of a solver? Understand where it flounders.

24 Jan 2019 Meena Mahajan, IMSc

slide-9
SLIDE 9

Improving solvers

How to improve the performance of a solver? Understand where it flounders. Underlying solver heuristics are formal proof systems: Runs of SAT/QBF solver provide proofs of unsatisfiability/falsity. Lower bounds in formal proof system (no short proof of unsat/falsity) ⇓ no short runs.

24 Jan 2019 Meena Mahajan, IMSc

slide-10
SLIDE 10

Improving solvers

How to improve the performance of a solver? Understand where it flounders. Underlying solver heuristics are formal proof systems: Runs of SAT/QBF solver provide proofs of unsatisfiability/falsity. Lower bounds in formal proof system (no short proof of unsat/falsity) ⇓ no short runs. Proving lower bounds – back to theory!

24 Jan 2019 Meena Mahajan, IMSc

slide-11
SLIDE 11

The Resolution Proof System for UNSAT

C: bag of clauses. ˜ a: assignment to the variables. If ˜ a satisfies Then ˜ a satisfies C = . . . A ∨ x B ∨ ¬ x . . . C′ = . . . A ∨ x B ∨ ¬ x . . . A ∨ B

24 Jan 2019 Meena Mahajan, IMSc

slide-12
SLIDE 12

The Resolution Proof System for UNSAT

C: bag of clauses. ˜ a: assignment to the variables. If ˜ a satisfies Then ˜ a satisfies C = . . . A ∨ x B ∨ ¬ x . . . C′ = . . . A ∨ x B ∨ ¬ x . . . A ∨ B C0 ∈ SAT = ⇒ C1 ∈ SAT = ⇒ . . . = ⇒ Ct−1 ∈ SAT = ⇒ Ct ∈ SAT C0 ∈ SAT ⇐ . . . ⇐ Ci ∈ SAT ⇐ . . . ⇐ Ct ∈ SAT ⇐ ∈ Ct

24 Jan 2019 Meena Mahajan, IMSc

slide-13
SLIDE 13

Extending Resolution to QBFs

QBFs: Quantified Boolean Formulas W.l.o.g., QBF in prenex CNF: Q x · F( x); F a set of clauses. Resolution is sound: If Q x · F(x) is true, and we add a clause C to F through resolution to get F ′, then Q x · F ′(x) is also true.

24 Jan 2019 Meena Mahajan, IMSc

slide-14
SLIDE 14

Extending Resolution to QBFs

QBFs: Quantified Boolean Formulas W.l.o.g., QBF in prenex CNF: Q x · F( x); F a set of clauses. Resolution is sound: If Q x · F(x) is true, and we add a clause C to F through resolution to get F ′, then Q x · F ′(x) is also true. But Resolution alone is not enough. Consider ∃x ∀u (x ∨ ¬ u) (¬ x ∨ u). Resolution can add (x ∨ ¬ x) or (u ∨ ¬ u). Useless. Universal variable u has to be handled differently. Two ways to proceed, modelling

  • CDCL-based solvers
  • expansion-based solvers

24 Jan 2019 Meena Mahajan, IMSc

slide-15
SLIDE 15

The Evaluation Game on QBFs

QBF Q x · F(x) Two players, Red and Blue, step through quantifier prefix left-to-right. Red picks values for ∃ variables, Blue for ∀ variables. Assignment constructed: ˜ a. Red wins a run of the game if F(˜ a) true. Otherwise Blue wins.

24 Jan 2019 Meena Mahajan, IMSc

slide-16
SLIDE 16

The Evaluation Game on QBFs

QBF Q x · F(x) Two players, Red and Blue, step through quantifier prefix left-to-right. Red picks values for ∃ variables, Blue for ∀ variables. Assignment constructed: ˜ a. Red wins a run of the game if F(˜ a) true. Otherwise Blue wins. example: ∃x ∀u (x ∨ ¬ u) (¬ x ∨ u).

24 Jan 2019 Meena Mahajan, IMSc

slide-17
SLIDE 17

The Evaluation Game on QBFs

QBF Q x · F(x) Two players, Red and Blue, step through quantifier prefix left-to-right. Red picks values for ∃ variables, Blue for ∀ variables. Assignment constructed: ˜ a. Red wins a run of the game if F(˜ a) true. Otherwise Blue wins. example: ∃x ∀u (x ∨ ¬ u) (¬ x ∨ u). Red: x = 1, Blue: u = 1: Red wins Red: x = 1, Blue: u = 0: Blue wins Red: x = 0, Blue: u = 1: Blue wins

24 Jan 2019 Meena Mahajan, IMSc

slide-18
SLIDE 18

The Evaluation Game on QBFs

QBF Q x · F(x) Two players, Red and Blue, step through quantifier prefix left-to-right. Red picks values for ∃ variables, Blue for ∀ variables. Assignment constructed: ˜ a. Red wins a run of the game if F(˜ a) true. Otherwise Blue wins. example: ∃x ∀u (x ∨ ¬ u) (¬ x ∨ u). Red: x = 1, Blue: u = 1: Red wins Red: x = 1, Blue: u = 0: Blue wins Red: x = 0, Blue: u = 1: Blue wins Blue can always win: set u = x.

24 Jan 2019 Meena Mahajan, IMSc

slide-19
SLIDE 19

The Evaluation Game on QBFs

QBF Q x · F(x) Two players, Red and Blue, step through quantifier prefix left-to-right. Red picks values for ∃ variables, Blue for ∀ variables. Assignment constructed: ˜ a. Red wins a run of the game if F(˜ a) true. Otherwise Blue wins. example: ∃x ∀u (x ∨ ¬ u) (¬ x ∨ u). Red: x = 1, Blue: u = 1: Red wins Red: x = 1, Blue: u = 0: Blue wins Red: x = 0, Blue: u = 1: Blue wins Blue can always win: set u = x. Q x · F(x) false if and only if Blue has a winning strategy. Use this to extend Resolution.

24 Jan 2019 Meena Mahajan, IMSc

slide-20
SLIDE 20

The ∀ reduction rule

Consider this scenario: Q x · F(x) is true. So Red has a winning strategy. F(x) has a clause C in which the rightmost variable (as per Q x) is a universal variable u. i.e. C = A ∨ ℓ; ℓ ∈ {u, ¬ u}; all variables in A are left of u.

24 Jan 2019 Meena Mahajan, IMSc

slide-21
SLIDE 21

The ∀ reduction rule

Consider this scenario: Q x · F(x) is true. So Red has a winning strategy. F(x) has a clause C in which the rightmost variable (as per Q x) is a universal variable u. i.e. C = A ∨ ℓ; ℓ ∈ {u, ¬ u}; all variables in A are left of u. Then, by the time Blue has to fix u, Red’s strategy must ensure that sub-clause A is already satisfied. That is, Red has a winning strategy on Q x · [F(x) ∧ A]. So Q x · [F(x) ∧ A] is also true.

24 Jan 2019 Meena Mahajan, IMSc

slide-22
SLIDE 22

The proof system QU-Res = Res+∀Red

Q x · C Grow the bag of clauses C using Resolution: If A ∨ x and B ∨ ¬ x are in the bag, can add A ∨ B (provided not a tautology), ∀-Reduction: If A ∨ ℓ(u) in the bag, and all variables in A left of u, can add A, until the empty clause is added.

24 Jan 2019 Meena Mahajan, IMSc

slide-23
SLIDE 23

The proof system QU-Res (cont’d)

Sound: A derivation of reveals a winning strategy for Blue.

[vanGelder 2012]

Complete: Use a winning strategy of Blue to decide which clauses to derive.

24 Jan 2019 Meena Mahajan, IMSc

slide-24
SLIDE 24

The proof system QU-Res (cont’d)

Sound: A derivation of reveals a winning strategy for Blue.

[vanGelder 2012]

Complete: Use a winning strategy of Blue to decide which clauses to derive.

Suffices to resolve with existential pivots only (Q-Res, [KleineB¨ uningKarpinskiFl¨

  • gel 1995] )

Suffices to eliminate variables in right-to-left order of quantification blocks (Level-ordered Q-Res)

24 Jan 2019 Meena Mahajan, IMSc

slide-25
SLIDE 25

A derivation in Q-Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)] e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1

24 Jan 2019 Meena Mahajan, IMSc

slide-26
SLIDE 26

A derivation in Q-Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)] e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 ¯ u ∨ c e ∨ ¯ c ¯ e ∨ c u ∨ ¯ c

24 Jan 2019 Meena Mahajan, IMSc

slide-27
SLIDE 27

A derivation in Q-Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)] e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 ¯ u ∨ c e ∨ ¯ c ¯ e ∨ c u ∨ ¯ c e ∨ ¯ u ¯ e ∨ u

24 Jan 2019 Meena Mahajan, IMSc

slide-28
SLIDE 28

A derivation in Q-Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)] e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 ¯ u ∨ c e ∨ ¯ c ¯ e ∨ c u ∨ ¯ c e ∨ ¯ u ¯ e ∨ u e ¯ e

24 Jan 2019 Meena Mahajan, IMSc

slide-29
SLIDE 29

A derivation in Q-Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)] e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 ¯ u ∨ c e ∨ ¯ c ¯ e ∨ c u ∨ ¯ c e ∨ ¯ u ¯ e ∨ u e ¯ e

  • 24 Jan 2019

Meena Mahajan, IMSc

slide-30
SLIDE 30

Adding the ∀-reduction rule

[Beyersdorff,Bonacina,Chew ITCS 2016]

P: Any sound and complete line-based proof system for UNSAT

eg Cutting Planes, Polynomial Calculus, Frege, restrictions of Frege (AC0-Frege, AC0[p]-Frege, TC0-Frege ...)

⇓ P+∀Red: a sound and complete proof system for QBF

24 Jan 2019 Meena Mahajan, IMSc

slide-31
SLIDE 31

The CP+∀Red proof system

CP+∀Red: Cutting Planes + ∀ Reduction.

[Beyersdorff,Chew,M,Shukla FSTTCS 2016]

Cutting Planes: Encode clauses as integer inequalities. x ∨ y ∨ z → x + y + z ≥ 1 x ∨ ¬ y ∨ z → x + (1 − y) + z ≥ 1 (x − y + z ≥ 0) x ∨ ¬ y ∨ ¬ z → x + (1 − y) + (1 − z) ≥ 1 (x − y − z ≥ −1) Bags of inequalities, not clauses. Evaluation game: Red tries to satisfy all inequalities. Blue tries to falsify some inequality.

24 Jan 2019 Meena Mahajan, IMSc

slide-32
SLIDE 32

The rules in Cutting Planes

If Red (∃) can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (for k ∈ Z>0) (a · x means a1x1 + a2x2 + . . . + anxn.)

24 Jan 2019 Meena Mahajan, IMSc

slide-33
SLIDE 33

The rules in Cutting Planes

If Red (∃) can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (for k ∈ Z>0) Then Red can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (a + b) · x ≥ A + B the + rule

24 Jan 2019 Meena Mahajan, IMSc

slide-34
SLIDE 34

The rules in Cutting Planes

If Red (∃) can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (for k ∈ Z>0) Then Red can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (a + b) · x ≥ A + B the + rule ka · x ≥ kA the × rule

24 Jan 2019 Meena Mahajan, IMSc

slide-35
SLIDE 35

The rules in Cutting Planes

If Red (∃) can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (for k ∈ Z>0) Then Red can win . . . a · x ≥ A b · x ≥ B kc · x ≥ C . . . (a + b) · x ≥ A + B the + rule ka · x ≥ kA the × rule c · x ≥ C

k

  • the ÷ rule

24 Jan 2019 Meena Mahajan, IMSc

slide-36
SLIDE 36

∀ reduction in CP+∀Red

If Red can win with I containing a · x ≥ A where the rightmost non-zero coefficient in a is blue, a = a’ b 00...0, (ie a universal variable, u) then Red can win with I ∪ {a′000...00 · x ≥ A − b} ↑ (Set u = 1) ∪ {a′000...0 · x ≥ A} ↑ (Set u = 0) . This Blue-elimination is the ∀-Reduction rule.

24 Jan 2019 Meena Mahajan, IMSc

slide-37
SLIDE 37

Proofs in the CP+∀Red proof system

Keep using the +, ×, ÷ and ∀Reduction rules. Red can win with I = I0 ⇓ Red can win with I1 ⇓ Red can win with I2 ⇓ . . . ⇓ Red can win with It. If It contains 0 ≥ 1, then Red can’t win with It, and so Red can’t win with I.

24 Jan 2019 Meena Mahajan, IMSc

slide-38
SLIDE 38

Expansion-Based Systems

∀uQ x · F(u, x) is true

  • [Q

x · F(0, x)] ∧ [Q x · F(1, x)] is true

  • Q

xu/0Q xu/1 ·

  • F(0,

xu/0) ∧ F(1, xu/1)

  • is true

24 Jan 2019 Meena Mahajan, IMSc

slide-39
SLIDE 39

Expansion-Based Systems

∀uQ x · F(u, x) is true

  • [Q

x · F(0, x)] ∧ [Q x · F(1, x)] is true

  • Q

xu/0Q xu/1 ·

  • F(0,

xu/0) ∧ F(1, xu/1)

  • is true

Expand the initial formula judiciously, on the fly. Then use standard resolution. Expansion-based systems: ∀Exp+Res [Janota,Marques-Silva 2015], IR [Beyersdorff,Chew,Janota 2014].

24 Jan 2019 Meena Mahajan, IMSc

slide-40
SLIDE 40

Part of a derivation in ∀Exp+Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)]

e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 24 Jan 2019 Meena Mahajan, IMSc

slide-41
SLIDE 41

Part of a derivation in ∀Exp+Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)]

e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 cu/1 cu/1 e ∨ du/1 ¯ cu/1 ∨ ¯ du/1 24 Jan 2019 Meena Mahajan, IMSc

slide-42
SLIDE 42

Part of a derivation in ∀Exp+Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)]

e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 cu/1 cu/1 e ∨ du/1 ¯ cu/1 ∨ ¯ du/1 cu/1 e ∨ ¯ cu/1 24 Jan 2019 Meena Mahajan, IMSc

slide-43
SLIDE 43

Part of a derivation in ∀Exp+Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)]

e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 cu/1 cu/1 e ∨ du/1 ¯ cu/1 ∨ ¯ du/1 cu/1 e ∨ ¯ cu/1 e e 24 Jan 2019 Meena Mahajan, IMSc

slide-44
SLIDE 44

Part of a derivation in ∀Exp+Res

∃e ∀u ∃c ∃d [(¬ e ∨ c), (¬ u ∨ c), (e ∨ d), (u ∨ d), (¬ c ∨ ¬ d)]

e u c d (¯ u ∨ c) (e ∨ d) (¯ u ∨ c) d (e ∨ d) (¯ c ∨ ¯ d) 1 u c d (¯ e ∨ c) (u ∨ d) (¯ e ∨ c) d (u ∨ d) (¯ c ∨ ¯ d) 1 cu/1 cu/1 e ∨ du/1 ¯ cu/1 ∨ ¯ du/1 cu/1 e ∨ ¯ cu/1 e e ¯ e ∨ cu/0 ¯ e ∨ cu/0 du/0 ¯ cu/0 ∨ ¯ du/ ¯ e ∨ cu/0 ¯ cu/0 ¯ e ¯ e

  • 24 Jan 2019

Meena Mahajan, IMSc

slide-45
SLIDE 45

Merging complementary literals

Consider ∃x ∀u (x ∨ ¬ u)(¬ x ∨ u).

24 Jan 2019 Meena Mahajan, IMSc

slide-46
SLIDE 46

Merging complementary literals

Consider ∃x ∀u (x ∨ ¬ u)(¬ x ∨ u). Resolve on x; instead of tautology u ∨ ¬ u, merge u and ¬ u into u∗. Intended meaning: Blue’s winning strategy for u is not dictated by this clause, but will be decided by the setting to x.

24 Jan 2019 Meena Mahajan, IMSc

slide-47
SLIDE 47

Merging complementary literals

Consider ∃x ∀u (x ∨ ¬ u)(¬ x ∨ u). Resolve on x; instead of tautology u ∨ ¬ u, merge u and ¬ u into u∗. Intended meaning: Blue’s winning strategy for u is not dictated by this clause, but will be decided by the setting to x. Proof Systems that use merging: LD-Q-Res (Long-Distance QRes), [Balabanov,Jiang 2012] LQU+-Res, [Balabonav,Widl,Jiang 2014] IRM (Instantiation, Resolution, Merge) [Beyersdorff,Chew,Janota 2014].

24 Jan 2019 Meena Mahajan, IMSc

slide-48
SLIDE 48

The relative power of some QBF proof systems:

Q-Res ∀Exp+Res LD-Q-Res QU-Res LQU+-Res IR IRM CP+∀Red PC+∀Red Frege+∀Red

24 Jan 2019 Meena Mahajan, IMSc

slide-49
SLIDE 49

Lower Bounds for QBF proof systems

from propositional hardness. not useful for understanding QBF solvers

24 Jan 2019 Meena Mahajan, IMSc

slide-50
SLIDE 50

Lower Bounds for QBF proof systems

from propositional hardness. not useful for understanding QBF solvers by adapting techniques for propositional hardness. let’s review

24 Jan 2019 Meena Mahajan, IMSc

slide-51
SLIDE 51

The Size-Width Relation in Resolution

In Resolution, Short proofs are narrow.

(Size of proof: number of steps. Width of proof: max width of clause in proof.)

Theorem ([Ben-Sasson,Wigderson 2001])

For all unsatisfiable CNFs F in n variables: S( ResT F) ≥ exp

  • w
  • Res F
  • − w(F)
  • .

(tree-like proofs; no reusing clauses)

S( Res F) = exp

  • w
  • Res F
  • − w(F)

2 n

  • .

24 Jan 2019 Meena Mahajan, IMSc

slide-52
SLIDE 52

The Size-Width relation in Q-Res

In Q-Res, this fails completely!

[Beyersdorff,Chew,M,Shukla STACS 2016, ACM Trans. Comp. Logic 2018]

24 Jan 2019 Meena Mahajan, IMSc

slide-53
SLIDE 53

The Size-Width relation in Q-Res

In Q-Res, this fails completely!

[Beyersdorff,Chew,M,Shukla STACS 2016, ACM Trans. Comp. Logic 2018]

∀ u1u2 . . . un ∃ e0e1 . . . en   (e0) (¬ ei−1 ∨ ui ∨ ei) for i ∈ [n] (¬ en)   Using Resolution, derive u1 ∨ . . . ∨ un. (n + 1 steps) Then using ∀Red, derive . (n steps) So proof of size O(n). Even tree-like. We show: Any proof must derive u1 ∨ . . . ∨ un. So width of any proof Ω(n).

24 Jan 2019 Meena Mahajan, IMSc

slide-54
SLIDE 54

Re-defining Width For QBFs

Problem: accumulation of universal variables. Possible fix: Redefine Width∃. Count only existential variables. Now does an analogue of the short-proofs-are-narrow hold?

24 Jan 2019 Meena Mahajan, IMSc

slide-55
SLIDE 55

Re-defining Width For QBFs

Problem: accumulation of universal variables. Possible fix: Redefine Width∃. Count only existential variables. Now does an analogue of the short-proofs-are-narrow hold? No! Completion Principle: clausal encoding of ∃ X ∈ {0, 1}n×n ∀ z (z ∨ ∃ all-1s row) ∧ (¬ z ∨ ∃all-0s column)

24 Jan 2019 Meena Mahajan, IMSc

slide-56
SLIDE 56

Re-defining Width For QBFs

Problem: accumulation of universal variables. Possible fix: Redefine Width∃. Count only existential variables. Now does an analogue of the short-proofs-are-narrow hold? No! Completion Principle: clausal encoding of ∃ X ∈ {0, 1}n×n ∀ z (z ∨ ∃ all-1s row) ∧ (¬ z ∨ ∃all-0s column) Under appropriate short clausal encoding, proof of size O(n2). Even tree-like proof: no reusing derived clauses. We show: Any proof must have width∃ Ω(n).

24 Jan 2019 Meena Mahajan, IMSc

slide-57
SLIDE 57

Size-Width∃ relation for non-tree-like proofs

∃e1∀u1∃c1∃d1 ∃e2∀u2∃c2∃d2 . . . ∃en∀un∃cn∃dn for i ∈ [n], (¬ ei ∨ ci) (ei ∨ di) (¬ ui ∨ ci) (ui ∨ di) ¬ c1 ∨ ¬ d1 ∨ ¬ c2 ∨ ¬ d2 ∨ . . . ∨ ¬ cn ∨ ¬ dn Winning strategy for universal player: ui = ¬ ei.

24 Jan 2019 Meena Mahajan, IMSc

slide-58
SLIDE 58

Size-Width∃ relation for non-tree-like proofs

∃e1∀u1∃c1∃d1 ∃e2∀u2∃c2∃d2 . . . ∃en∀un∃cn∃dn for i ∈ [n], (¬ ei ∨ ci) (ei ∨ di) (¬ ui ∨ ci) (ui ∨ di) ¬ c1 ∨ ¬ d1 ∨ ¬ c2 ∨ ¬ d2 ∨ . . . ∨ ¬ cn ∨ ¬ dn Winning strategy for universal player: ui = ¬ ei. Encode last clause with additional ∃ variables as short clauses.

24 Jan 2019 Meena Mahajan, IMSc

slide-59
SLIDE 59

Size-Width∃ relation for non-tree-like proofs

∃e1∀u1∃c1∃d1 ∃e2∀u2∃c2∃d2 . . . ∃en∀un∃cn∃dn for i ∈ [n], (¬ ei ∨ ci) (ei ∨ di) (¬ ui ∨ ci) (ui ∨ di) ¬ c1 ∨ ¬ d1 ∨ ¬ c2 ∨ ¬ d2 ∨ . . . ∨ ¬ cn ∨ ¬ dn Winning strategy for universal player: ui = ¬ ei. Encode last clause with additional ∃ variables as short clauses. Short proofs in Q-Res, size nO(1). We show: Width∃ of any Q-Res proof Ω(n).

24 Jan 2019 Meena Mahajan, IMSc

slide-60
SLIDE 60

Size-Width∃ relation for non-tree-like proofs

∃e1∀u1∃c1∃d1 ∃e2∀u2∃c2∃d2 . . . ∃en∀un∃cn∃dn for i ∈ [n], (¬ ei ∨ ci) (ei ∨ di) (¬ ui ∨ ci) (ui ∨ di) ¬ c1 ∨ ¬ d1 ∨ ¬ c2 ∨ ¬ d2 ∨ . . . ∨ ¬ cn ∨ ¬ dn Winning strategy for universal player: ui = ¬ ei. Encode last clause with additional ∃ variables as short clauses. Short proofs in Q-Res, size nO(1). We show: Width∃ of any Q-Res proof Ω(n). Large width requirement does not give size lower bound.

24 Jan 2019 Meena Mahajan, IMSc

slide-61
SLIDE 61

Lower Bounds for QBF proof systems

from propositional hardness. not useful for understanding QBF solvers by adapting techniques for propositional hardness. let’s review: size-width fails for Q-Res

24 Jan 2019 Meena Mahajan, IMSc

slide-62
SLIDE 62

Lower Bounds for QBF proof systems

from propositional hardness. not useful for understanding QBF solvers by adapting techniques for propositional hardness. let’s review: size-width fails for Q-Res interpolation?

24 Jan 2019 Meena Mahajan, IMSc

slide-63
SLIDE 63

Feasible Interpolation – the propositional case

F = A( p, q) ∧ B( p, r) in UNSAT

  • for all assignments

a to p, either A( a, q) or B( a, r) in UNSAT. A( a, q) in SAT = ⇒ B( a, r) in UNSAT.

24 Jan 2019 Meena Mahajan, IMSc

slide-64
SLIDE 64

Feasible Interpolation – the propositional case

F = A( p, q) ∧ B( p, r) in UNSAT

  • for all assignments

a to p, either A( a, q) or B( a, r) in UNSAT. A( a, q) in SAT = ⇒ B( a, r) in UNSAT. GIven a, can we tell which is in UNSAT?

24 Jan 2019 Meena Mahajan, IMSc

slide-65
SLIDE 65

Feasible Interpolation – the propositional case

F = A( p, q) ∧ B( p, r) in UNSAT

  • for all assignments

a to p, either A( a, q) or B( a, r) in UNSAT. A( a, q) in SAT = ⇒ B( a, r) in UNSAT. GIven a, can we tell which is in UNSAT? We want an interpolant circuit C in p variables: C( a) = 0 = ⇒ A( a, q) is in UNSAT, and C( a) = 1 = ⇒ B( a, r) is in UNSAT.

24 Jan 2019 Meena Mahajan, IMSc

slide-66
SLIDE 66

Feasible Interpolation – the propositional case

F = A( p, q) ∧ B( p, r) in UNSAT

  • for all assignments

a to p, either A( a, q) or B( a, r) in UNSAT. A( a, q) in SAT = ⇒ B( a, r) in UNSAT. GIven a, can we tell which is in UNSAT? We want an interpolant circuit C in p variables: C( a) = 0 = ⇒ A( a, q) is in UNSAT, and C( a) = 1 = ⇒ B( a, r) is in UNSAT. A( a, q) is in SAT = ⇒ C( a) = 1 = ⇒ B( a, r) is in UNSAT.

24 Jan 2019 Meena Mahajan, IMSc

slide-67
SLIDE 67

Feasible Interpolation – the propositional case (cont’d)

Theorem ([Kraj´

ıˇ cek 1997],[Pudl´ ak 1997])

Resolution proofs of size s give Boolean circuits of size sO(1) computing interpolants. Cutting Planes proofs of size s give real arithmetic circuits of size sO(1) computing interpolants.

24 Jan 2019 Meena Mahajan, IMSc

slide-68
SLIDE 68

Feasible Interpolation – the propositional case (cont’d)

Theorem ([Kraj´

ıˇ cek 1997],[Pudl´ ak 1997])

Resolution proofs of size s give Boolean circuits of size sO(1) computing interpolants. Cutting Planes proofs of size s give real arithmetic circuits of size sO(1) computing interpolants. If p variables appears only positively in A( p, q) or only negatively in B( p, r), then interpolant circuit is (real-) monotone.

24 Jan 2019 Meena Mahajan, IMSc

slide-69
SLIDE 69

Feasible Interpolation – the propositional case (cont’d)

Theorem ([Kraj´

ıˇ cek 1997],[Pudl´ ak 1997])

Resolution proofs of size s give Boolean circuits of size sO(1) computing interpolants. Cutting Planes proofs of size s give real arithmetic circuits of size sO(1) computing interpolants. If p variables appears only positively in A( p, q) or only negatively in B( p, r), then interpolant circuit is (real-) monotone. All resolution / cutting-plane proofs of the clique-colour formulas are

  • f exponential size.

(Clique-colour formulas: CNF encodings of “∃ a graph that is (k − 1)-colourable and has a k-clique.”)

24 Jan 2019 Meena Mahajan, IMSc

slide-70
SLIDE 70

Feasible Interpolation for QBFs

∃ p Q q Q r [ A( p, q) ∧ B( p, r) ] is false

p [Q qA( p, q) ∧ Q rB( p, r)] is false ⇓ for all assignments a to p, either Q q A( a, q) or Q r B( a, r) is false. Interpolant circuit: C( a) = 0 = ⇒ Q q A( a, q) is false, and C( a) = 1 = ⇒ Q r B( a, r) is false.

24 Jan 2019 Meena Mahajan, IMSc

slide-71
SLIDE 71

Feasible Interpolation works for many QBF proof systems

The Clique-coClique formulas: CNF encodings of ∃ an n-vertex graph G, ∀ u, u implies G has a k-clique, ¬u implies G has no k-clique. (Note: To express no clique, universal quantifiers used. Not succinctly expressible as UNSAT instance.)

24 Jan 2019 Meena Mahajan, IMSc

slide-72
SLIDE 72

Feasible Interpolation works for many QBF proof systems

The Clique-coClique formulas: CNF encodings of ∃ an n-vertex graph G, ∀ u, u implies G has a k-clique, ¬u implies G has no k-clique. (Note: To express no clique, universal quantifiers used. Not succinctly expressible as UNSAT instance.)

Theorem ([Beyersdorff,Chew,M,Shukla ICALP15, LMCS17, FSTTCS16])

All the resolution-based QBF proof systems Q-Res, QU-Res, LD-Q-Res, LQU+-Res, ∀Exp+Res, IR, IRM as well as the proof system CP+∀Red, admit feasible monotone interpolation. All Clique-coClique formulas need exponential-sized proofs in all these proof systems.

24 Jan 2019 Meena Mahajan, IMSc

slide-73
SLIDE 73

Lower Bounds for QBFproof systems

from propositional hardness. not useful for understanding QBF solvers by adapting techniques for propositional hardness. let’s review:

Size lower bounds from Width lower bounds does not work with the simplest extension of Resolution, Q-Res. Feasible Interpolation works for all Resolution based systems and for CP+∀Red.

24 Jan 2019 Meena Mahajan, IMSc

slide-74
SLIDE 74

Lower Bounds for QBFproof systems

from propositional hardness. not useful for understanding QBF solvers by adapting techniques for propositional hardness. let’s review:

Size lower bounds from Width lower bounds does not work with the simplest extension of Resolution, Q-Res. Feasible Interpolation works for all Resolution based systems and for CP+∀Red.

from strategy extraction. all-new; specific to QBFs The winning strategy of the universal player in the evaluation game leads to new lower bound techniques.

24 Jan 2019 Meena Mahajan, IMSc

slide-75
SLIDE 75

Strategy Extraction

Main idea: A proof reveals information about a winning strategy. Examine a proof. Construct a circuit of a special type for computing the winning strategy. Circuit type: depends on the proof system Circuit size: depends on the proof size If the winning strategy is hard to compute in the relevant circuit model, then all proofs in the proof system must be large.

24 Jan 2019 Meena Mahajan, IMSc

slide-76
SLIDE 76

From Proof to Decision List for Winning Strategy

Blue has to choose the value of a variable u. Blue knows values of all variables left of u; partial assignment a. Proof lines L1, L2, . . . , Lm. ∀Red on u at (1 <) i1 < i2 < . . . < ik (≤ m). Lir : eliminate u from Ljr , jr < ir.

24 Jan 2019 Meena Mahajan, IMSc

slide-77
SLIDE 77

From Proof to Decision List for Winning Strategy

Blue has to choose the value of a variable u. Blue knows values of all variables left of u; partial assignment a. Proof lines L1, L2, . . . , Lm. ∀Red on u at (1 <) i1 < i2 < . . . < ik (≤ m). Lir : eliminate u from Ljr , jr < ir. if Li1( a) false then set u to make Lj1( a) false elseif Li2( a) false then set u to make Lj2( a) false . . . . . . elseif Lik( a) false then set u to make Ljk( a) false else set u = 0.

24 Jan 2019 Meena Mahajan, IMSc

slide-78
SLIDE 78

From Proof to Decision List for Winning Strategy

Blue has to choose the value of a variable u. Blue knows values of all variables left of u; partial assignment a. Proof lines L1, L2, . . . , Lm. ∀Red on u at (1 <) i1 < i2 < . . . < ik (≤ m). Lir : eliminate u from Ljr , jr < ir. if Li1( a) false then set u to make Lj1( a) false elseif Li2( a) false then set u to make Lj2( a) false . . . . . . elseif Lik( a) false then set u to make Ljk( a) false else set u = 0.

[Beyersdorff,Chew,Janota 2015], [Beyersdorff,Bonacina,Chew 2016]:

This strategy is a winning strategy for Blue. Strategy description: A Decision List for each universal variable.

24 Jan 2019 Meena Mahajan, IMSc

slide-79
SLIDE 79

Strategy Extraction from P+∀Red proofs

Proof with s ∀Reduction steps ⇓ Winning strategy can be computed by a Decision List with s steps. QU-Res: Each condition is a clause. (is a1 ∨ a2 ∨ . . . ∨ an true?) CP+∀Red: Each condition is a linear threshold function. (is c1a1 + c2a2 + . . . + cnan ≥ b?)

24 Jan 2019 Meena Mahajan, IMSc

slide-80
SLIDE 80

Genuine QBF bounds for P+∀Red proofs

Proof with s ∀Reduction steps ⇓ Winning strategy can be computed by a Decision List with s steps. Contrapositive gives lower bounds on number of ∀Reduction steps, not just on proof size.

24 Jan 2019 Meena Mahajan, IMSc

slide-81
SLIDE 81

Genuine QBF bounds for P+∀Red proofs

Proof with s ∀Reduction steps ⇓ Winning strategy can be computed by a Decision List with s steps. Contrapositive gives lower bounds on number of ∀Reduction steps, not just on proof size. i.e. Proving the QBF false will require large size even with a SAT oracle (appropriately formalised).

24 Jan 2019 Meena Mahajan, IMSc

slide-82
SLIDE 82

Winning Strategies compute desired functions

Fix Boolean function f . Fix small circuit C computing f (size m). Define QBF Qf ,C: ∃x1x2 . . . xn∀w∃z1z2 . . . zm

  • (w = zm)

(zi = value of ith gate of C(x)) : i ∈ [m]

  • (zi clauses enforce zm = f (x).)

24 Jan 2019 Meena Mahajan, IMSc

slide-83
SLIDE 83

Winning Strategies compute desired functions

Fix Boolean function f . Fix small circuit C computing f (size m). Define QBF Qf ,C: ∃x1x2 . . . xn∀w∃z1z2 . . . zm

  • (w = zm)

(zi = value of ith gate of C(x)) : i ∈ [m]

  • (zi clauses enforce zm = f (x).)

Blue can choose w = f (x) and force a win. No other way for Blue to win. f (x) not hard to compute – it has a small circuit. If no small decision list, then no small proof.

24 Jan 2019 Meena Mahajan, IMSc

slide-84
SLIDE 84

Winning Strategies Hard for Decision Lists

The Parity function has an O(n) size circuit. The Parity function requires exponentially long decision lists of

  • clauses. ([H˚

astad ]: ⊕ ∈ AC0.) Hence

Theorem ([Beyersdorff,Chew,Janota 2015])

Any Q-Res or QU-Res proof for Q-Parity must be of exponential size.

24 Jan 2019 Meena Mahajan, IMSc

slide-85
SLIDE 85

Winning Strategies Hard for Decision Lists

The Parity function has an O(n) size circuit. The Parity function requires exponentially long decision lists of

  • clauses. ([H˚

astad ]: ⊕ ∈ AC0.) Hence

Theorem ([Beyersdorff,Chew,Janota 2015])

Any Q-Res or QU-Res proof for Q-Parity must be of exponential size. The Inner Product function has an O(n) size circuit. IP(x, y) x · y mod 2. The IP function needs > 2n/2 − 1 steps in a decision list of linear threshold functions. ([Tur´

an,Vatan 1997]) Hence

Theorem ([Beyersdorff,Chew,M,Shukla 2016])

Any CP+∀Red proof for Q-IP must be of exponential size.

24 Jan 2019 Meena Mahajan, IMSc

slide-86
SLIDE 86

Other Sources of Hardness

[Beyersdorff,Pich LICS 2016]

Every lower bound in Frege+∀Red stems from

either propositional hardness,

  • r a circuit lower bound.

No other source of hardness.

24 Jan 2019 Meena Mahajan, IMSc

slide-87
SLIDE 87

Other Sources of Hardness

[Beyersdorff,Pich LICS 2016]

Every lower bound in Frege+∀Red stems from

either propositional hardness,

  • r a circuit lower bound.

No other source of hardness. Not true for weaker systems.

24 Jan 2019 Meena Mahajan, IMSc

slide-88
SLIDE 88

Other Sources of Hardness (cont’d)

∃x1 · · · xn∀u1 · · · un∃t1 · · · tn (xi ∨ ui ∨ ti) i ∈ [n] (¬ xi ∨ ¬ ui ∨ ti) i ∈ [n] (¬ t1 ∨ · · · ∨ ¬ tn) Blue has a trivial winning strategy: ui = xi. But the formula is still hard to prove false in QU-Res. Why?

24 Jan 2019 Meena Mahajan, IMSc

slide-89
SLIDE 89

Winning Strategies needing Varied Responses (cont’d)

Many responses needed in winning strategy. 2n possible values for u1 · · · un, all necessary. cost of formula high. Each line can contribute only so much: capacity small. Hence proof size must be large.

Theorem ([Beyersdorff,Blinkhorn,Hinde ITCS2018])

Size-Cost-Capacity Theorem: For any proof π of a QBF φ in a system P+∀Red, Size(π) × Capacity(π) ≥ Cost(φ)

24 Jan 2019 Meena Mahajan, IMSc

slide-90
SLIDE 90

Winning Strategies needing Varied Responses (cont’d)

Many responses needed in winning strategy. 2n possible values for u1 · · · un, all necessary. cost of formula high. Each line can contribute only so much: capacity small. Hence proof size must be large.

Theorem ([Beyersdorff,Blinkhorn,Hinde ITCS2018])

Size-Cost-Capacity Theorem: For any proof π of a QBF φ in a system P+∀Red, Size(π) × Capacity(π) ≥ Cost(φ) Similar result for expansion-based systems; [Beyersdorff,Blinkhorn STACS 18].

24 Jan 2019 Meena Mahajan, IMSc

slide-91
SLIDE 91

To conclude ...

QBF Proof systems What are they? Formal systems for proving false QBFs false. Why do we study them? Lower bounds can help guide development of better solvers. (Also, strong lower bounds will separate complexity classes.)

24 Jan 2019 Meena Mahajan, IMSc

slide-92
SLIDE 92

To conclude ...

QBF Proof systems What are they? Formal systems for proving false QBFs false. Why do we study them? Lower bounds can help guide development of better solvers. (Also, strong lower bounds will separate complexity classes.) What do we know? Extracting strategies from proofs leads to lower bounds.

24 Jan 2019 Meena Mahajan, IMSc

slide-93
SLIDE 93

To conclude ...

QBF Proof systems What are they? Formal systems for proving false QBFs false. Why do we study them? Lower bounds can help guide development of better solvers. (Also, strong lower bounds will separate complexity classes.) What do we know? Extracting strategies from proofs leads to lower bounds. What next? Continue the cycle QBF solver QBF proof system Lower bound

formalise underlying system prove improve

24 Jan 2019 Meena Mahajan, IMSc

slide-94
SLIDE 94

Thank you

24 Jan 2019 Meena Mahajan, IMSc