LongDistance QResolution with Dependency Schemes
Abstract
Resolution proof systems for quantified Boolean formulas (QBFs) provide a formal model for studying the limitations of stateoftheart searchbased QBF solvers that use these systems to generate proofs. We study a combination of two proof systems supported by the solver DepQBF: Qresolution with generalized universal reduction according to a dependency scheme and long distance Qresolution. We show that the resulting proof system—which we call longdistance Q(D)resolution—is sound for the reflexive resolutionpath dependency scheme. In fact, we prove that it admits strategy extraction in polynomial time. This comes as an application of a general result, by which we identify a whole class of dependency schemes for which longdistance Q(D)resolution admits polynomialtime strategy extraction. As a special case, we obtain soundness and polynomialtime strategy extraction for long distance Q(D)resolution with the standard dependency scheme. We further show that searchbased QBF solvers using a dependency scheme D and learning with longdistance Qresolution generate longdistance Q(D)resolution proofs. The above soundness results thus translate to partial soundness results for such solvers: they declare an input QBF to be false only if it is indeed false. Finally, we report on experiments with a configuration of DepQBF that uses the standard dependency scheme and learning based on longdistance Qresolution.
Keywords
QBF Qresolution Dependency schemes Strategy extraction1 Introduction
Quantified Boolean formulas (QBFs) offer succinct encodings for problems from domains such as formal verification, synthesis, and planning [5, 13, 16, 30, 38, 43]. Although the combination of (more verbose) propositional encodings with SAT solvers is still the stateoftheart approach to many of these problems, QBF solvers are gaining ground. An arsenal of new techniques has been introduced over the past few years [10, 11, 14, 22, 23, 25, 26, 29, 32, 33, 35], and these advances in solver technology have been accompanied by the development of a better understanding of the underlying QBF proof systems and their limitations [4, 7, 8, 9, 12, 18, 27, 42].
Searchbased solvers implementing the QCDCL algorithm [15, 46] represent one of the principal stateoftheart approaches in QBF solving. Akin to modern SAT solvers, these solvers rely on successive variable assignments in combination with fast constraint propagation and learning. Unlike SAT solvers, however, searchbased QBF solvers are constrained by the variable dependencies induced by the quantifier prefix:^{1} while SAT solvers can assign variables in any order, searchbased QBF solvers can only assign variables from the leftmost quantifier block that contains unassigned variables, since the assignment of a variable further to the right might depend on the variable assignment to this block. In the most extreme case, this forces solvers into a fixed order of variable assignments, rendering decision variable heuristics ineffective.
The searchbased solver DepQBF uses dependency schemes to partially bypass this restriction [10, 31]. Dependency schemes can sometimes identify pairs of variables as independent, allowing the solver to assign them in any order. This gives decision heuristics more freedom and results in increased performance [10].
 (a)
Longdistance Qresolution permits the derivation of tautological clauses in certain cases [2, 45, 47]. This system can be used in constraint learning as an alternative to Qresolution, leading to fewer backtracks during search and, sometimes, reduced runtime [19]. In addition, clause learning based on longdistance Qresolution is substantially easier to implement. However, it is open whether QCDCL with learning based on longdistance Qresolution is sound when combined with known dependency schemes.
 (b)
For applications in verification and synthesis, it is not enough for solvers to decide whether an input QBF is true or false—they also have to generate a certificate. Such certificates can be efficiently constructed from Qresolution [2] and even longdistance Qresolution proofs [3]. However, it is not clear whether this is possible for proofs generated by QCDCL with the standard dependency scheme.
Our proof relies on a familiar interpretation of Qresolution refutations as winning strategies for the universal player in the evaluation game [24]. We identify a natural property of dependency schemes D that not only allows for the interpretation of an LDQ(D)refutation as a winning strategy for the universal player, but even implies polynomialtime certificate extraction from an LDQ(D)refutation. We then show that the reflexive resolution path dependency scheme in fact has this property.
One of our motivations for studying the combination of longdistance Qresolution and dependency schemes is that it is already supported by DepQBF. To complement our theoretical results, and to provide further motivation for resolving the question of soundness of “true” answers, we performed experiments using a configuration of DepQBF with both features activated. These experiments show that performance with learning based on LDQ(D)resolution is on par with and—in some cases—even slightly better than the performance of DepQBF with other configurations of constraint learning.
1.1 Organization
Section 2 establishes basic notions used throughout this paper. In Sect. 3, we review dependency schemes and introduce the LDQ(D) proof system. In Sect. 4, we present a version of QCDCL which combines dependency schemes with learning based on longdistance Qresolution and argue that this algorithm generates LDQ(D) proofs. Section 5 is split into two parts: in the first part, we define a property of dependency schemes D and prove that it is sufficient for soundness of LDQ(D); in the second part, we show that the reflexive resolutionpath dependency scheme has this property. In Sect. 6, we report on experiments with a version of DepQBF that generates LDQ(D\(^\text {std}\))proofs. In Sect. 7, we briefly discuss recently published related work. We conclude in Sect. 8 with some open questions.
2 Preliminaries
2.1 Formulas and Assignments
A literal is a negated or unnegated variable. If x is a variable, we write \(\overline{x} = \lnot x\) and \(\overline{\lnot x} = x\), and let \( var (x) = var (\lnot x) = x\). We sometimes call literals x and \(\lnot x\) the positive and negative polarity of variable x. If X is a set of literals, we write \(\overline{X}\) for the set \(\{\,\overline{x} \;{:}\;x \in X \,\}\). A clause is a finite disjunction of literals, and a term is a finite conjunction of literals. We call a clause tautological if it contains the same variable negated as well as unnegated. A CNF formula is a finite conjunction of nontautological clauses. Whenever convenient, we treat clauses and terms as sets of literals, and CNF formulas as sets of sets of literals. We write \( var (S)\) for the set of variables occurring (negated or unnegated) in a clause or term S, that is, \( var (S) = \{\, var (\ell ) \;{:}\;\ell \in S \,\}\). Moreover, we let \( var (\varphi ) = \bigcup _{C \in \varphi } var (C)\) denote the set of variables occurring in a CNF formula \(\varphi \).
A truth assignment (or simply assignment) to a set X of variables is a mapping \(\tau : X \rightarrow \{0,1\}\). We write [X] for the set of truth assignments to X, and extend \(\tau : X \rightarrow \{0, 1\}\) to literals by letting \(\tau (\lnot x) = 1  \tau (x)\) for \(x \in X\). Let \(\tau : X \rightarrow \{0, 1\}\) be a truth assignment. The restriction \(C[\tau ]\) of a clause (term) S by \(\tau \) is defined as follows: if there is a literal \(\ell \in S \cap (X \cup \overline{X})\) such that \(\tau (\ell ) = 1\) (\(\tau (\ell ) = 0)\) then \(S[\tau ] = 1\) (\(S[\tau ] = 0\)). Otherwise, \(S[\tau ] = S {\setminus } (X \cup \overline{X})\). The restriction \(\varphi [\tau ]\) of a CNF formula \(\varphi \) by the assignment \(\tau \) is defined \(\varphi [\tau ] = \{\,C[\tau ] \;{:}\;C \in \varphi , C[\tau ] \ne 1 \,\}\).
2.2 PCNF Formulas
A PCNF formula is denoted by \(\varPhi = {\mathcal {Q}}.\varphi \), where \(\varphi \) is a CNF formula and \(\mathcal {Q}= Q_1 X_1 \dots Q_n X_n\) is a sequence such that \(Q_i \in \{\forall , \exists \}\), \(Q_i \ne Q_{i+1}\) for \(1 \le i < n\), and the \(X_i\) are pairwise disjoint sets of variables. We call \(\varphi \) the matrix of \(\varPhi \) and \(\mathcal {Q}\) the (quantifier) prefix of \(\varPhi \), and refer to the \(X_i\) as quantifier blocks. We require that \( var (\varphi ) = X_1 \cup \dots \cup X_n\) and write \( var (\varPhi ) = var (\varphi )\). We define a partial order \(<_{\varPhi }\) on \( var (\varphi )\) as \(x<_{\varPhi } y \Leftrightarrow x \in X_i, y \in X_j, i < j\). We extend \(<_{\varPhi }\) to a relation on literals in the obvious way and drop the subscript whenever \(\varPhi \) is understood. For \(x \in var (\varPhi )\) we let \(R_{\varPhi }(x) = \{\,y \in var (\varPhi ) \;{:}\;x <_{\varPhi } y \,\}\) and \(L_{\varPhi }(x) = \{\,y \in var (\varPhi ) \;{:}\;y <_{\varPhi } x \,\}\) denote the sets of variables to the right and to the left of x in \(\varPhi \), respectively. Relative to the PCNF formula \(\varPhi \), variable x is called existential (universal) if \(x \in X_i\) and \(Q_i = \exists \) (\(Q_i = \forall \)). The set of existential (universal) variables occurring in \(\varPhi \) is denoted \( var _{\exists }(\varPhi )\) (\( var _{\forall }(\varPhi )\)). The size of a PCNF formula \(\varPhi = {\mathcal {Q}}.\varphi \) is defined as \(\varPhi  = \sum _{C \in \varphi }C\). If \(\tau \) is an assignment, then \(\varPhi [\tau ]\) denotes the PCNF formula \(\mathcal {Q}'.\varphi [\tau ]\), where \(\mathcal {Q}'\) is the quantifier prefix obtained from \(\mathcal {Q}\) by deleting variables that do not occur in \(\varphi [\tau ]\). True and false PCNF formulas are defined in the usual way.
2.3 Countermodels
Let \(\varPhi = \mathcal {Q}.\varphi \) be a PCNF formula. A countermodel of \(\varPhi \) is an indexed family \(\{f_{u}\}_{u \in var _{\forall }(\varPhi )}\) of functions \(f_u: [L_{\varPhi }(u)] \rightarrow \{0, 1\}\) such that \(\varphi [\tau ] = \{\emptyset \}\) for every assignment \(\tau : var (\varPhi ) \rightarrow \{0, 1\}\) satisfying \(\tau (u) = f_u(\tau _{L_{\varPhi }(u)})\) for \(u \in var _{\forall }(\varPhi )\).
Proposition 1
(Folklore) A PCNF formula is false if, and only if, it has a countermodel.
3 Dependency Schemes and LDQ(D)Resolution
In this section, we introduce the proof system LDQ(D), which combines Q(D)resolution [42] with longdistance Qresolution [2]. Qresolution is a generalization of propositional resolution to PCNF formulas [28]. Qresolution is of practical interest due to its relation to search based QBF solvers that implement Quantified Conflict Driven Constraint Learning (QCDCL) [15, 46]: the traces of QCDCL solvers correspond to Qresolution proofs [19, 21]. QCDCL—which will be described in more detail in Sect. 4—generalizes the wellknown DPLL procedure [17] from SAT to QSAT. In a nutshell, DPLL searches for a satisfying assignment of an input formula by propagating unit clauses and assigning pure literals until the formula cannot be simplified any further, at which point it picks an unassigned variable and branches on the assignment of this variable. Any of the remaining variables can be chosen for assignment, but the order of assignment can have significant effects on the runtime. Modern conflict driven clause learning (CDCL) SAT solvers derived from the DPLL algorithm use sophisticated heuristics to determine what variable to assign next [34].
In QCDCL, the quantifier prefix imposes constraints on the order of variable assignments: a variable may be assigned only if it occurs in the leftmost quantifier block with unassigned variables. Often, this is more restrictive than necessary. For instance, variables from disjoint subformulas may be assigned in any order. Intuitively, a variable can be assigned as long as it does not depend on any unassigned variable. This is the intuition underlying a generalization of QCDCL implemented in the solver DepQBF [10, 31]. DepQBF uses a dependency scheme [39] to compute an overapproximation of variable dependencies. Dependency schemes are mappings that associate every PCNF formula with a binary relation on its variables that refines the order of variables in the quantifier prefix.^{2}
Definition 1
(Dependency Scheme) A dependency scheme is a mapping D that associates each PCNF formula \(\varPhi \) with a relation \(D_{\varPhi } \subseteq \{\,(x, y) \;{:}\;x <_{\varPhi } y \,\}\) called the dependency relation of \(\varPhi \) with respect to D.
The mapping which simply returns the prefix ordering of an input formula can be thought of as a baseline dependency scheme:
Definition 2
(Trivial Dependency Scheme) The trivial dependency scheme \(\text {D}^{\text { trv}}\) associates each PCNF formula \(\varPhi \) with the relation \(\text {D}^{\text { trv}}_{\varPhi } = \{\,(x, y) \;{:}\;x <_{\varPhi } y \,\}\).
DepQBF uses a dependency relation to determine the order in which variables can be assigned: if y is a variable and there is no unassigned variable x such that (x, y) is in the dependency relation, then y is considered ready for assignment. DepQBF also uses the dependency relation to generalize the \(\forall \)reduction rule used in clause learning [10]. As a result of its use of dependency schemes, DepQBF generates proofs in a generalization of Qresolution called Q(D)resolution [42], a proof system that takes a dependency scheme D as a parameter.
Dependency schemes can be partially ordered based on their dependency relations: if the dependency relation computed by a dependency scheme \(D_1\) is a subset of the dependency relation computed by a dependency scheme \(D_2\) for each PCNF formula, then \(D_1\) is more general than \(D_2\). The more general a dependency scheme, the more freedom DepQBF has in choosing decision variables. Currently, (aside from the trivial dependency scheme) DepQBF supports the socalled standard dependency scheme [39].^{3} We will work with the more general reflexive resolutionpath dependency scheme [42], a variant of the resolutionpath dependency scheme [41, 44]. This dependency scheme computes an overapproximation of variable dependencies based on whether two variables are connected by a (pair of) resolution path(s).
Definition 3
 1.
For all \(i \in [k]\), there is a \(C_i \in \varphi \) such that \(\ell _{2i1}, \ell _{2i} \in C_i\).
 2.
For all \(i \in [k]\), \( var (\ell _{2i1}) \ne var (\ell _{2i})\).
 3.
For all \(i \in [k1]\), \(\{\ell _{2i}, \ell _{2i+1}\} \subseteq X \cup \overline{X}\).
 4.
For all \(i \in [k1]\), \(\overline{\ell _{2i}} = \ell _{2i+1}\).
One can think of a resolution path as a potential chain of implications: if each clause \(C_i\) contains exactly two literals, then assigning \(\ell _1\) to 0 requires setting \(\ell _{2k}\) to 1. If, in addition, there is such a path from \(\overline{\ell _1}\) to \(\overline{\ell _{2k}}\), then \(\ell _1\) and \(\ell _{2k}\) have to be assigned opposite values. Accordingly, the resolution path dependency scheme identifies variables connected by a pair of resolution paths as potentially dependent on each other.
Definition 4

x and y, as well as \(\lnot x\) and \(\lnot y\), are connected in \(\varPhi \) with respect to X.

x and \(\lnot y\), as well as \(\lnot x\) and y, are connected in \(\varPhi \) with respect to X.
Definition 5
The reflexive resolutionpath dependency scheme is the mapping \(\text { D }^{\text { rrs}}\) that assigns to each PCNF formula \(\varPhi = {\mathcal {Q}}.\varphi \) the relation \(\text { D }^{\text { rrs}}_{\varPhi } = \{\,x <_{\varPhi } y \;{:}\;\{x,y\}\) is a resolutionpath dependency pair in \(\varPhi \) with respect to \(R_{\varPhi }(x) {\setminus } var _{\forall }(\varPhi ) \,\}\).
Both Qresolution and Q(D)resolution only allow for the derivation of nontautological clauses, that is, clauses that do not contain a literal negated as well as unnegated. Longdistance Qresolution is a variant of Qresolution that admits tautological clauses in certain cases [2]. Variants of QCDCL that allow for learned clauses to be tautological [45, 47] have been shown to generate proofs in longdistance Qresolution [19].
 1.
If \(\ell \) is a literal and \(\mathcal {P}_1\) and \(\mathcal {P}_2\) are proof DAGs with distinct sinks \(v_1\) and \(v_2\), then \(\mathcal {P}_1 \odot _{\ell } \mathcal {P}_2\) is the proof DAG consisting of the union of \(\mathcal {P}_1\) and \(\mathcal {P}_2\) along with a new sink v that has two incoming edges, the first one from \(v_1\) and the second one from \(v_2\). Moreover, if \(C_1\) is the label of \(v_1\) in \(\mathcal {P}_1\) and \(C_2\) is the label of \(v_2\) in \(\mathcal {P}_2\), then v is labeled with the clause \((C_1 {\setminus } \{\ell \}) \cup (C_2 {\setminus } \{\overline{\ell }\})\).
 2.
If u is a variable and \(\mathcal {P}\) is a proof DAG with a sink w labeled with C, then \(\mathcal {P} u\) denotes the proof DAG obtained from \(\mathcal {P}\) by adding an edge from w to a new node v such that v is labeled with \(C {\setminus } \{u, \lnot u\}\).
Definition 6

Source nodes are labeled with input clauses from \(\varphi \).

If a node with label C has parents labeled \(C_1\) and \(C_2\) then C can be derived from \(C_1\) and \(C_2\) by (longdistance) resolution.

If a node labeled with a clause C has a single parent with label \(C'\) then C can be derived from \(C'\) by \(\forall \)reduction with respect to the dependency scheme D.
Let \(\mathcal {P}\) be an LDQ(D)derivation from a PCNF formula \(\varPhi \). The (clause) label of the sink node is called the conclusion of \(\mathcal {P}\), denoted \(Cl(\mathcal {P})\). If the conclusion of \(\mathcal {P}\) is the empty clause then we refer to \(\mathcal {P}\) as an LDQ(D)refutation of \(\varPhi \). For a node v of \(\mathcal {P}\), the subderivation (of \(\mathcal {P}\)) rooted at v is the proof DAG induced by v and its ancestors in \(\mathcal {P}\). It is straightforward to verify that the resulting proof DAG is again an LDQ(D)derivation from \(\varPhi \). For convenience, we will identify (sub)derivations with their sinks. The size of \(\mathcal {P}\), denoted \(\mathcal {P}\), is the total number of literal occurrences in clause labels of \(\mathcal {P}\).
4 QCDCL with Dependency Schemes Generates LDQ(D)Proofs
In this section, we present a version of the QCDCL algorithm that uses dependency schemes [10, 31] and performs constraint learning based on longdistance Qresolution [19, 46].^{5} Generalizing an argument by Egly et al. [19], we will show that this algorithm produces LDQ(D)proofs when using a dependency scheme D.
Starting from an input PCNF formula \(\varPhi \), QCDCL generates (“learns”) constraints—clauses and terms—until it produces an empty constraint, at which point it returns true (if the empty term is learned) or false (if the empty clause is learned).
One can think of QCDCL solving as proceeding in rounds. Along with a set of clauses and terms, the solver maintains an assignment \(\sigma \) (the trail, which we will represent by a sequence of literals in the order of their assignment). During each round, this assignment is extended by quantified Boolean constraint propagation (QBCP) and—possibly—branching.
Quantified Boolean constraint propagation (with dependency scheme D) consists in the exhaustive application of universal and existential reduction (relative to \(D_{\varPhi }\)) in combination with unit assignments.^{6} More specifically, QBCP reports a clause C as falsified if \(C[\sigma ] \ne 1\) and universal reduction can be applied to \(C[\sigma ]\) to obtain the empty clause. Dually, a term T is considered satisfied if \(T[\sigma ] \ne 0\) and \(T[\sigma ]\) can be reduced to the empty term. A clause C is unit under \(\sigma \) if \(C[\sigma ] \ne 1\) and universal reduction yields a clause \(C' = (\ell )\), for some existential literal \(\ell \) such that \( var (\ell )\) is unassigned. Dually, a term T is unit under \(\sigma \) if \(T[\sigma ] \ne 0\) and existential reduction can be applied to obtain a term \(T' = (\ell )\) containing a single universal literal \(\ell \). If \(C = (\ell )\) is a unit clause, then the assignment \(\sigma \) has to be extended by \(\ell \) in order not to falsify C, and if \(T = (\ell )\) is a unit term, then \(\sigma \) has to be extended by \(\overline{\ell }\) in order not to satisfy T. If several clauses or terms are unit under \(\sigma \), QBCP nondeterministically picks one and extends the assignment accordingly. This is repeated until a constraint is empty or no unit constraints remain.
If QBCP does not lead to an empty constraint, the assignment \(\sigma \) is extended by branching. That is, the solver chooses an unassigned variable y such that every variable x with \((x, y) \in D_{\varPhi }\) is assigned, and extends the assignment \(\sigma \) by y or \(\lnot y\).
The resulting assignment can be partitioned into socalled decision levels. The decision level of an assignment \(\sigma \) is the number of literals in \(\sigma \) that were assigned by branching. The decision level of a literal \(\ell \) in \(\sigma \) is the decision level of the prefix of \(\sigma \) that ends with \(\ell \). Note that each decision level greater than 0 can be associated with a unique variable assigned by branching.
Eventually, the assignment maintained by QCDCL must falsify a clause or satisfy a term. When this happens (we call this a conflict), the solver proceeds to conflict analysis to derive a learned constraint C. Initially, C is the falsified clause (satisfied term), called the conflict clause (term). The solver finds the existential (universal) literal in C that was assigned last by QBCP, and the antecedent clause (term) R responsible for this assignment. A new constraint is derived by resolving C and R and applying universal (existential) reduction (again, relative to \(D_{\varPhi }\)). This is done repeatedly until the resulting constraint C is asserting. A clause (term) S is asserting if there is a unique existential (universal) literal \(\ell \in S\) with maximum decision level (greater than zero) among literals in S, the corresponding decision variable is existential (universal), and every universal (existential) variable \(y \in var (S)\) such that \((y, var (\ell )) \in D_{\varPhi }\) is assigned at a lower decision level (an asserting constraint becomes unit after backtracking). Once an asserting constraint has been found, it is added to the solver’s set of constraints. Finally, QCDCL backtracks, undoing variable assignments until reaching a decision level computed from the learned constraint.
Pseudocode for the main QCDCL loop is shown as Algorithm 1, and pseudocode for conflict analysis is shown as Algorithm 2. We now formally state and prove a correspondence between clauses learned by QCDCL and LDQ(D)resolution.
Proposition 2
Every clause learned by Algorithm 1 given an input PCNF \(\varPhi \) can be derived from \(\varPhi \) by LDQ(D)resolution.
Proof
We will show that each learned clause constructed during conflict analysis can be derived by LDQ(D)resolution from input clauses or previously learned clauses. In addition, we prove an invariant saying that each such clause C can be reduced to the empty clause under the trail assignment restricted to variables up to and including the existential variable assigned last among those in C. Formally, if \(\ell _1\dots \ell _k\) is the trail and \(i = \max \{\,j \;{:}\;\overline{\ell _j} \in C, 1 \le j \le k \,\}\), then the restricted trail is \(\ell _1\dots \ell _i\). If such an i does not exist—which can happen only if C contains no existential variable—the restricted trail is empty. Let \(\tau \) denotes the assignment corresponding to the restricted trail. We want to show that \(C[\tau ]\) simplifies to the empty clause by \(\forall \)reduction. The proof is by induction on the number of resolution operations performed by Algorithm 2.
The base case with C being the conflict clause is trivial. For the inductive step, suppose that C can be reduced to the empty clause under the restricted trail assignment \(\tau = \ell _1\dots \ell _i\). Thus \(\overline{\ell _i} \in C\) is the existential literal falsified last among literals in C and conflict analysis would resolve C with the antecedent R of \(\ell _i\) next. We have to show the resolvent satisfies the above invariant and that this resolution operation is permissible in LDQ(D)resolution. Let \(\tau ' = \ell _1\dots \ell _{i1}\) denote the trail at the time when \(\ell _i\) was propagated. Clause R is unit under \(\tau '\), that is, \(R[\tau ']\) reduces to \((\ell _i)\). In particular, we have \(\tau '(\ell ) = 0\) for every existential literal \(\ell \in R {\setminus } \{\ell _i\}\) and every universal literal \(\ell \in R\) assigned by \(\tau '\). Since \(C[\tau ]\) reduces to the empty clause by assumption, we further must have \(\tau (\ell ) = 0\) for every existential literal and every assigned universal literal of C. We conclude that \(\tau '(\ell ) = 0\) for every existential literal and every assigned universal literal in the resolvent \(C' = (C \cup R) {\setminus } \{\ell _i, \overline{\ell _i}\}\). This property is clearly not affected by unassigning universal variables or unassigning existential variables not occurring in \(C'\), so \(C'\) reduces to the empty clause under its corresponding restricted trail assignment, proving that the invariant is preserved. Furthermore, it entails that any literal \(\ell \in R\) such that \(\overline{\ell } \in C\) must be universal and unassigned by \(\tau '\). Since \(R[\tau ']\) can be reduced to \((\ell _i)\) by \(\forall \)reduction, we must have \(( var (\ell ), var (\ell _i)) \notin D_{\varPhi }\) for each such literal \(\ell \), so \(C'\) can be derived from C and R in LDQ(D)resolution. \(\square \)
As in the case of QCDCL without dependency schemes [19, 21] an analogue of this result can be proved for learned terms and a dual proof system (“Qconsensus”) that operates on terms instead of clauses.
The proof of Proposition 2 uses the fact that two clauses \(C_1 \vee u \vee e\) and \(C_2 \vee \lnot u \vee \lnot e\) can be resolved on variable e even if \(u < e\) as long as \((u, e) \notin D_{\varPhi }\). The following example illustrates that this generalization of the resolution rule is necessary for LDQ(D)resolution to trace QCDCL with dependency schemes and longdistance Qresolution.
Example 1
We now construct a possible trace of QCDCL on \(\varPhi \) with \(\text { D }^{\text { rrs}}\) and learning based on longdistance Qresolution. At decision level 0 the unit clauses in \(\psi \) are propagated, setting \(a = 1\) and \(b = 1\). This does not lead to further propagation and QCDCL proceeds with the decision \(y \overset{d}{=} 0\). Note that y can be assigned before x because \((x, y) \notin \text { D }^{\text { rrs}}_{\varPhi }\). This assignment does not lead to any literals being propagated, so the algorithm makes another decision \(z_2 \overset{d}{=} 0\). Now clause \(C_2\) simplifies to the unit clause \((z_1)\) and \(z_1 = 1\) is propagated. Clause \(C_1\) only contains x under the resulting assignment and we have a conflict. Conflict analysis first resolves clauses \(C_1\) and \(C_2\) to obtain the clause \(C_{12} = (x \vee y \vee z_2 \vee \lnot a)\). Variable a depends on x, so \(\forall \)reduction cannot be applied. Since \(z_2\) is the only variable from the second decision level in clause \(C_{12}\) and \(z_2\) does not depend on x, \(C_{12}\) is asserting and the clause is learned by QCDCL. Backtracking undoes the decision involving \(z_2\) and propagates \(z_2 = 1\) instead. As a result, \(C_4\) simplifies to \((z_0)\), unit propagation assigns \(z_0 = 1\), and clause \(C_5\) is falsified. Conflict analysis resolves \(C_4\) and \(C_5\) to derive the learned (unit) clause \(C_{45} = (\overline{z_2})\), which causes QCDCL to backtrack to decision level 0 and propagate \(z_2 = 0\). Now clause \(C_{12}\) simplifies to \((x \vee y)\). Since y is independent of x we can apply \(\forall \)reduction to obtain the unit clause (y) which propagates the assignment \(y = 1\). Clause \(C_6\) in turn becomes unit and propagates \(z_3 = 0\). As a result, clause \(C_3\) simplifies to \((\overline{x})\) and reduces to the empty clause by \(\forall \)reduction. Conflict analysis resolves \(C_3\) and \(C_6\) so as to obtain the clause \((\overline{x} \vee \overline{y} \vee \overline{b})\). Variable b depends on x, so \(\forall \)reduction cannot be applied. Next, clause \((\overline{x} \vee \overline{y} \vee \overline{b})\) is resolved with \(C_{12}\) to derive \((x \vee \overline{x} \vee z_2 \vee \overline{a} \vee \overline{b})\). Note that this resolution step is permissible since \((x, y) \notin \text { D }^{\text { rrs}}_{\varPhi }\). Further resolution steps involving unit clauses yield the clause \((x \vee \overline{x})\), which can be reduced to the empty clause, so that QCDCL terminates with return value FALSE.
5 Soundness of and Strategy Extraction for LDQ(D\(^\text {rrs}\))
5.1 PolynomialTime Strategy Extraction from LDQ(D)Refutations
A PCNF formula can be associated with an evaluation game played between an existential and a universal player. These players take turns assigning quantifier blocks in the order of the prefix. The existential player wins if the matrix evaluates to 1 under the resulting variable assignment, while the universal player wins if the matrix evaluates to 0. One can show that the formula is true (false) if and only if the existential (universal) player has a winning strategy in this game, and this winning strategy is a (counter)model.
Goultiaeva et al. [24] proved that a Qresolution refutation can be used to compute winning moves for the universal player in the evaluation game. The idea is that universal maintains a “restriction” of the refutation by the assignment constructed in the evaluation game, which is a refutation of the restricted formula.
For assignments made by the existential player, the universal player only needs to consider each instance of resolution whose pivot variable is assigned: one of the premises is not satisfied and can be used to (re)construct a refutation.
When it is universal’s turn, the quantifier block for which she needs to pick an assignment is leftmost in the restricted formula. This means that \(\forall \)reduction of these variables is blocked by any of the remaining existential variables and can only be applied to a purely universal clause. In a Qresolution refutation, these variables must therefore be reduced at the very end, and because Qresolution does not permit tautological clauses, only one polarity of each universal variable from the leftmost block can appear in a refutation. It follows that universal can maintain a Qresolution refutation by assigning variables from the leftmost block in such a way as to map the associated literals to 0, effectively deleting them from the remaining Qresolution refutation.
In this manner, the universal player can maintain a refutation until the end of the game, when all variables have been assigned. At that point, a refutation consists only of the empty clause, which means that the assignment chosen by the two players falsifies a clause of the original matrix and universal has won the game.
Egly et al. [19] observed that this argument goes through even in the case of longdistance Qresolution, since a clause containing both u and \(\lnot u\) for a universal variable u can only be derived by resolving on an existential variable to the left of u, but no such existential variable exists if u is from the leftmost block.
In this section, we will prove that this argument can be generalized to LDQ(D\(^\text {rrs}\))refutations. We illustrate this correspondence with an example:
Example 2
The key property that allows universal to maintain a refutation in the above example is that universal variables from the leftmost quantifier block may appear in at most one polarity. We now show that this property is in fact sufficient for soundness of LDQ(D) when combined with a natural monotonicity property of dependency schemes.
Definition 7
A dependency scheme D is monotone if \(D_{\varPhi [\tau ]}\subseteq D_{\varPhi }\) for every PCNF formula \(\varPhi \) and every assignment \(\tau \) to a subset of \( var (\varPhi )\). We say that D is simple if, for every PCNF formula \(\varPhi =\forall X\mathcal {Q}.\varphi \), every LDQ(D)derivation \(\mathcal {P}\) from \(\varPhi \), and every universal variable \(u \in X\), u or \(\overline{u}\) does not appear in \(\mathcal {P}\). A dependency scheme D is normal if it is both monotone and simple.
As in the case of Qresolution, universal’s move for a particular quantifier block can be computed from the assignment corresponding to the previous moves and the refutation in polynomial time. Since every polynomialtime algorithm can be implemented by a family of polynomiallysized circuits, and because these circuits can even be computed in polynomial time [1, p. 109], it follows that LDQ(D) admits polynomialtime strategy extraction when D is normal (in “Appendix Appendix:”, we present an explicit construction with a more specific bound on the runtime).
Theorem 1
Let D be a normal dependency scheme. There is a polynomialtime algorithm that, given a PCNF formula \(\varPhi \) and an LDQ(D)refutation of \(\varPhi \), computes a countermodel of \(\varPhi \).
As an application of this general result, we will prove that the reflexive resolutionpath dependency scheme is normal in Sect. 5.2 below.
Theorem 2
\(\mathrm {\text { D }^{\text { rrs}}}\) is normal.
Corollary 1
There is a polynomialtime algorithm that, given a PCNF formula \(\varPhi \) and an LDQ(D\(^\text {rrs}\))refutation of \(\varPhi \), computes a countermodel of \(\varPhi \).
This result immediately carries over to the less general standard dependency scheme.
Corollary 2
There is a polynomialtime algorithm that, given a PCNF formula \(\varPhi \) and an LDQ(D\(^\text {std}\))refutation of \(\varPhi \), computes a countermodel of \(\varPhi \).
In combination with Proposition 1, these results imply soundness of both proof systems.
Corollary 3
The systems LDQ(D\(^\text {std}\)) and LDQ(D\(^\text {rrs}\)) are sound.
5.2 The Reflexive ResolutionPath Dependency Scheme is Normal
In order to prove Theorem 2 and show that \(\text { D }^{\text { rrs}}\) is normal, we will need some insight into the relationship between resolution paths and LDQ(D\(^\text {rrs}\))derivations. For a formula \(\varPhi \) and a universal literal u, we will denote by \(T_u(\varPhi )\) the set of existential literals e such that \(u < e\) and such that e is reachable from u by a resolution path via existential variables to the right of u in \(\varPhi \).
Lemma 1
Let \(\varPhi = \forall X \mathcal {Q}.\varphi \) be a PCNF formula and let u be a universal literal with \( var (u) \in X\). Let \(C_1, C_2\in \varphi \) be clauses such that for some existential literal x, \(x\in C_1\) and \(\overline{x}\in C_2\), and let \(C=C_1\cup C_2{\setminus }\{x,\overline{x}\}\). Then \(T_u(\varPhi )=T_u(\mathcal {Q}.\varphi \cup \{C\} )\).
Proof
Let \(\varPhi '= \mathcal {Q}.\varphi \cup \{C\}\). Of course, by adding clauses to a formula, we preserve all existing resolution paths, so \(T_u(\varPhi )\subseteq T_u(\varPhi ')\). We will prove that the opposite inclusion holds as well. Let \(e\in T_u(\varPhi ')\) and let \(\pi \) be a resolution path in \(\varPhi '\) certifying this. If \(\pi \) is also a resolution path in \(\varPhi \), we are done. If it is not, it must be because it performs a Ctransition, namely it contains two consecutive literals \(l_1, l_2\) such that \(var(l_1)\ne var(l_2)\), \(l_1,l_2\in C\), but \(\{l_1,l_2\}\not \subseteq C_1\) and \(\{l_1,l_2\}\not \subseteq C_2\). In this case, without loss of generality, we have \(l_1\in C_1\) and \(l_2\in C_2\). Let \(\pi _1\) be the prefix of \(\pi \) up to and including \(l_1\) and \(\pi _2\) be the suffix of \(\pi \) starting with \(l_2\). Let \(\pi '\) be the concatenation of \(\pi _1\), x, \(\overline{x}\), and \(\pi _2\). It is clearly a valid resolution path and it uses one fewer Ctransitions than \(\pi \). Iterating this process, we can remove all Ctransitions from \(\pi \) to obtain a resolution path in \(\varPhi \). The resulting resolution path has the same endpoints and therefore certifies that \(e\in T_u(\varPhi )\). \(\square \)
The previous lemma implies that when considering reachability from an outermost universal literal in a formula \(\varPhi \), we can use clauses derived from \(\varPhi \) by LDQ(D\(^\text {rrs}\))resolution as well. Indeed, adding clauses produced by the resolution rule does not change the set of reachable literals by Lemma 1, and adding clauses produced by universal reduction clearly does not even create new resolution paths. Particularly, if two literals ever appear together in a derived clause, there is a resolution path between them. This is summarized by the following corollary.
Corollary 4
Let \(\mathcal {P}\) be an LDQ(D\(^\text {rrs}\))derivation from a PCNF formula \(\varPhi = \forall X \mathcal {Q}.\varphi \) and let \(u\in X, u\in Cl(\mathcal {P})\). Then for all existential literals \(e\in Cl(\mathcal {P})\), there is a resolution path from u to e in \(\varPhi \).
As a first step towards proving Theorem 2, we will prove that both polarities of an outermost universal literal cannot appear together in a single clause of a derivation.
Lemma 2
Let \(\mathcal {P}\) be an LDQ(D\(^\text {rrs}\))derivation from a PCNF formula \(\varPhi = \forall X \mathcal {Q}.\varphi \) and let \(u\in X\). Then \(u\notin Cl(\mathcal {P})\) or \(\lnot u\notin Cl(\mathcal {P})\).
Proof
Towards a contradiction, suppose \(u,\lnot u\in Cl(\mathcal {P})\). Since input clauses do not contain both polarities of any literal, there must be a resolution step inside the derivation, which merges u and \(\lnot u\) into one clause. Let \(\mathcal {P}'=\mathcal {P}_1\odot _x \mathcal {P}_2\) be such a step. Then, without loss of generality, \(x,u\in Cl(\mathcal {P}_1)\) and \(\lnot x,\lnot u\in Cl(\mathcal {P}_2)\) and by Corollary 4, there is a resolution path from u to x and from \(\lnot u\) to \(\lnot x\), i.e. \((u,x)\in \text { D }^{\text { rrs}}_\varPhi \). However, if x depends on u, opposite polarities of u cannot be merged in a resolution step with the pivot x, a contradiction. \(\square \)
We will further assume that the derivation considered in the proof of Theorem 2 is in the normal form given by the following lemma.
Lemma 3
Let \(\mathcal {P}\) be an LDQ(D\(^\text {rrs}\))derivation from a PCNF formula \(\varPhi = \forall X \mathcal {Q}.\varphi \), let \(u \in X\) be a universal variable such that both u and \(\lnot u\) appear in \(\mathcal {P}\), and let \(Y = var _{\exists }(\varPhi )\). There exists an LDQ(D\(^\text {rrs}\))derivation \(\mathcal {P}'\) from a PCNF formula \(\varPhi ' = \forall u \exists Y.\varphi '\) such that \(\mathcal {P}'\) contains both u and \(\lnot u\) but only \(\forall \)reduction steps with respect to one polarity of u.
Proof
Let \(\varphi '\) be the result of removing all universal variables except u from \(\varphi \). Removing universal variables does not introduce new resolution paths, so \(\text { D }^{\text { rrs}}_{\varPhi '} \subseteq \text { D }^{\text { rrs}}_{\varPhi }\). The derivation \(\mathcal {P}'\) can be obtained from \(\mathcal {P}\) in the following way. We first delete all occurrences of universal variables other than u from \(\mathcal {P}\), along with \(\forall \)reduction steps involving such variables. The result is an LDQ(D\(^\text {rrs}\))derivation from \(\varPhi '\), and it still contains both u and \(\lnot u\). Next, we choose a subderivation containing both u and \(\lnot u\) such that none of its proper subderivations contains both literals. By Lemma 2, the conclusion C of this derivation can contain at most one of u and \(\lnot u\). If \(u \in C\) we omit all \(\forall \)reduction steps involving u. Otherwise, we omit all \(\forall \)reduction steps involving \(\lnot u\). This yields the desired LDQ(D\(^\text {rrs}\))derivation \(\mathcal {P}'\) from \(\varPhi '\). \(\square \)
Using Lemmas 2 and 3, we can proceed to finish the proof of Theorem 2.
Proof of Theorem 2
Towards a contradiction, consider an LDQ(D\(^\text {rrs}\))derivation from a formula \(\varPhi = \forall X \mathcal {Q}.\varphi \) and let \(u\in X\) be such that both polarities of u occur in this derivation. Let \(Y = var _{\exists }(\varPhi )\) and let \(\mathcal {P}\) denote the simplified derivation given by Lemma 3. Assume without loss of generality that \(\mathcal {P}\) is a tree (any derivation can be turned into a treelike derivation by copying proof nodes), and that \(\mathcal {P}\) does not contain \(\forall \)reduction steps involving \(\lnot u\). Since \(\lnot u\) occurs in \(\mathcal {P}\) but is not reduced, \(\lnot u\) must occur in the conclusion of \(\mathcal {P}\). Thus u cannot occur in the conclusion by Lemma 2. Since u is present in the derivation \(\mathcal {P}\), this means there must be a reduction step on u somewhere in \(\mathcal {P}\). As u is the only universal variable and we omitted all reduction steps on \(\lnot u\), all reduction steps in \(\mathcal {P}\) are on u and \(\mathcal {P}\) must have the form depicted in Fig. 3, where \(\mathcal {P}_n=\mathcal {P}_{n+1}u\) is a lowermost reduction step on u and the subsequent resolutions are on pivots \(x_n,\ldots , x_1\). Let \(C_0=Cl(\mathcal {P}), C_i=Cl(\mathcal {P}_i), C_i'=Cl(\mathcal {P}_i')\). The clauses \(C_1',\ldots ,C_n',C_{n+1}\) are derived by LDQ(D\(^\text {rrs}\))resolution and by Lemma 1 we know that we can use them to show resolutionpath connections as if they were input clauses. By the transformations we considered we know that starting from an arbitrary LDQ(D\(^\text {rrs}\))derivation we can obtain a valid LDQ(D\(^\text {rrs}\))derivation in this form, so any contradiction we derive from here means a contradiction with the assumption that an LDQ(D\(^\text {rrs}\))derivation contains both polarities of a universal variable from the outermost block, thus proving Theorem 2. With that, we are ready to finish the proof.
We will prove that there is a resolution path from \(\lnot u\) to u going through an existential literal in \(C_{n+1}\), which is in contradiction with the soundness of reduction of u from \(C_{n+1}\). Let us consider open resolution paths, i.e. resolution paths without their final literal. If an open resolution path ends in a literal \(\ell \) of clause C, we say that the path leads to the clause C. By induction on n, we will prove that there is an open resolution path from \(\lnot u\) which leads to the clause \(C_n\). If \(n=1\), we have the path \(\lnot u, \lnot x_1, x_1\). For \(n>1\), let \(\pi \) be the open path leading to \(C_{n1}\) and let \(\ell \) be its last literal. Then either \(\ell \in C_n\), in which case we have an open path leading to \(C_n\), or \(\ell \in C_n'\), in which case we have the open path \(\pi , \lnot x_n, x_n\) leading to \(C_n\). An open path that leads to \(C_n\) also leads to \(C_{n+1}\), because those two clauses only differ in the presence of u and therefore can be closed by the literal u to obtain the required resolution path. \(\square \)
6 Experiments
To gauge the potential of clause learning based on LDQ(D\(^\text {std}\)), we ran experiments with the searchbased solver DepQBF^{7} in version 5.0. By default, DepQBF supports proof generation only in combination with the trivial dependency scheme—in that case, it generates Qresolution or longdistance Qresolution proofs (depending on whether longdistance resolution is enabled). However, by uncommenting a few lines in the source code, proof generation can also be enabled with the standard dependency scheme, and this option can even be combined with longdistance resolution. This leads to the solver generating Q(D\(^\text {std}\))resolution or LDQ(D\(^\text {std}\))resolution proofs (see Sect. 4).
 1.
Longdistance Qresolution with \(\forall /\exists \)reduction according to D\(^\text {std}\) (LDQD).
 2.
Longdistance Qresolution with ordinary \(\forall /\exists \)reduction (LDQ).
 3.
Qresolution with \(\forall /\exists \)reduction according to D\(^\text {std}\) (QD).
 4.
Ordinary Qresolution (Q).
Solved instances, solved true instances, solved false instances, and total runtime in seconds (including timeouts) with preprocessing (but without QBCE)
Configuration  Solved  True  False  Time 

Application track  
LDQD  377  186  191  343,455 
LDQ  377  186  191  345,459 
QD  377  183  194  343,928 
Q  376  182  194  345,914 
QBFLib track  
LDQD  130  69  61  140,743 
LDQ  131  69  62  141,646 
QD  129  67  62  140,975 
Q  127  65  62  142,679 
Results with preprocessing and dynamic QBCE
Configuration  Solved  True  False  Time 

Application track  
LDQD  385  195  190  339,143 
LDQ  388  201  187  336,739 
QD  392  201  191  334,965 
Q  389  198  191  337,141 
QBFLib track  
LDQD  145  75  70  132,567 
LDQ  133  64  69  141,682 
QD  137  70  67  134,150 
Q  129  62  67  142,399 
Results for the application track with QBCE (but without preprocessing)
Configuration  Solved  True  False  Time 

LDQD  440  223  217  287,012 
LDQ  435  223  212  291,574 
QD  440  225  215  291,661 
Q  437  221  216  337,141 
7 Related Work
QCDCL with learning based on longdistance Qresolution was first described by Zhang and Malik [45]. They presented an argument for the soundness of using tautological clauses (respectively, contradictory terms) within their algorithm but did not study longdistance Qresolution as a proof system. Lacking a sound theoretical foundation, the use of tautological clauses in QCDCL was abandoned in favour of more complicated methods for constraint learning that avoid their generation [20, 21, 33].
Interest in longdistance Qresolution was renewed when Balabanov and Jiang [2] introduced the proof system we presented in Sect. 3 (restricted to the trivial dependency scheme) and proved its soundness. Egly et al. [19] showed that a family of formulas known to be hard for Qresolution [28] admits short longdistance Qresolution proofs. They also demonstrated that QCDCL with learning based on longdistance resolution generates longdistance Qresolution proofs and presented a new version of DepQBF that implements this algorithm. Finally, they showed that a longdistance Qresolution proof can be interpreted as a winning strategy in the evaluation game associated with a QBF, generalizing a result by Goultiaeva et al. [24]. While these results established a solid theoretical framework for the use of longdistance Qresolution within QCDCL they did not provide an intuitive account of the semantics of individual tautological clauses. Such an account was subsequently presented by Balabanov et al. [3], who showed that tautological literals \(u, \lnot u \in C\) can be interpreted as proxies for “phase functions” that determine whether a variable or its negation is present in clause C based on the values assigned to pivot variables appearing in the derivation of C. The authors used this interpretation to generalize the lineartime strategy extraction algorithm of Balabanov and Jiang [2] to longdistance Qresolution proofs.
Recently and independently of this work, Beyersdorff and Blinkhorn investigated the soundness of Qresolution proof systems parameterized by dependency schemes [6]. They define a property of dependency schemes D—full exhibition—which ensures that a certain version of longdistance Q(D)resolution is sound, and show that the reflexive resolutionpath dependency scheme has that property.
In a nutshell, a dependency scheme D is fully exhibited if every true QBF \(\varPhi \) has a model \(\{ f_e \}_{e \in var _{\exists }(\varPhi )}\) such that \(f_e\) may only depend on a universal variable u if \((u, e) \in D_{\varPhi }\) (such models have elsewhere been referred to as Dmodels [40]). It is fairly straightforward to show that Q(D)resolution is sound if D has this property, but generalizing this result to proof systems with longdistance resolution presents a challenge. Beyersdorff and Blinkhorn show that full exhibition is sufficient for soundness of a restricted version of LDQ(D)resolution, where complementary universal literals that are “merged” by resolution must be annotated with the (existential) pivot variable, and universal reduction can be applied only if every existential variable occurring in the premise or the annotation of a universal variable is independent of the universal variable to be reduced. However, it is uncertain whether proofs generated by DepQBF with LDQ(D)learning satisfy this additional restriction.
How full exhibition relates to our normality property is not entirely clear. Beyersdorff and Blinkhorn prove that full exhibition is not sufficient for soundness of LDQ(D)resolution as defined here. In combination with Theorem 1, this shows that dependency schemes that are fully exhibited need not be normal. Whether there are normal dependency schemes that are not fully exhibited, on the other hand, remains open. Indeed, there is some evidence to the effect that normality entails full exhibition: consider a dependency scheme D that is not fully exhibited, and let \(\varPhi = \forall u \mathcal {Q}.\varphi \) be a true QBF that does not have a Dmodel. Suppose u is the only universal variable of \(\varPhi \). In this restricted case, the (non)existence of a Dmodel can be expressed as a QBF \(\varPsi \) by simply shifting existentials independent of u to the left. Because \(\varPhi \) does not have a Dmodel, \(\varPsi \) must be false and admit a Qresolution refutation \(\mathcal {P}\), which is also an LDQ(D)refutation of \(\varPhi \). Because \(\varPhi \) is true, LDQ(D)resolution must be unsound and so D cannot be normal by Theorem 1. Obviously, the assumption that u is the only universal variable of \(\varPhi \) is very restrictive, but since we can suppose that D is monotone (recall that a dependency scheme is normal if it is both simple and monotone), there is hope that the argument for an arbitrary QBF can be reduced to this case by instantiating with a suitable variable assignment.
8 Discussion
The results of Sects. 4 and 5 establish a partial soundness proof of QCDCL with learning based on LDQ(D\(^\text {std}\)): we now know that we can trust such a solver when it outputs “false”. To prove that “true” answers can be trusted as well, one has to show soundness of quantified term resolution (Qconsensus) when combined with the standard dependency scheme and longdistance resolution. The reason this does not follow from the results proved here is that they rely on a correspondence of Qresolution derivations with dependencyinducing resolution paths that is not immediate for terms generated from an input PCNF: there is a correspondence of Qconsensus derivations with dual “resolution paths” that connect such terms, but these “resolution paths” do not induce resolutionpath dependencies in the input PCNF.
The experiments in Sect. 6 indicate that we should not expect significant performance gains when switching from learning with Q(D\(^\text {std}\))resolution to LDQ(D\(^\text {std}\)). This is in spite of the fact that, from a purely theoretical perspective, LDQ(D\(^\text {std}\)) is a stronger proof system: a wellstudied class of QBFs introduced by Kleine Büning, Karpinski, and Flögel requires exponentiallysized Qresolution proofs [28] but admits short longdistance Qresolution refutations [19], and since the standard dependency scheme does not offer any improvement over trivial dependencies on these formulas (see [12]) we obtain an exponential separation of LDQ(D\(^\text {std}\))resolution from Q(D\(^\text {std}\))resolution. From a practical point of view, the main benefit of using LDQ(D\(^\text {std}\))resolution over Q(D\(^\text {std}\))resolution is that conflict analysis is much simpler (cf. [19]). A learned constraint can be obtained from a conflict simply by resolving variables in the reverse order of their propagation (see Sect. 4). Methods that avoid the generation of tautological clauses (contradictory terms) during learning are significantly more involved [20, 21, 33].
We have shown that LDQ(D\(^\text {rrs}\))refutations allow for polynomialtime strategy extraction. In practice, the corresponding algorithm generates circuits that are frequently larger by an order of magnitude than the refutations provided as input. It is unclear whether this increase in size can be avoided by careful engineering alone or only by using a different approach. Faster (linear time) strategy extraction algorithms are known for “ordinary” Qresolution and longdistance Qresolution [2, 3]. Unfortunately, their underlying idea of setting universal variables so as to falsify the premise of some \(\forall \)reduction step no longer works when dependency schemes enter the mix: generalized \(\forall \)reduction may remove a universal variable u even in the presence (in the premise) of an existential variable e such that \(u < e\) and the universal player can only be sure to falsify the premise if the eliteral is false, but she does not know the value of e at the time of assigning u. We believe that developing lineartime strategy extraction algorithms for Q(D)resolution or LDQ(D)resolution is going to require a better understanding of the power of these proof systems visàvis Qresolution and longdistance Qresolution [12].
Footnotes
 1.
We consider QBFs in prenex normal form.
 2.
The original definition of dependency schemes [39] is more restrictive than the one given here, but the additional requirements are irrelevant for the purposes of this paper.
 3.
Strictly speaking, it uses a refined version of the standard dependency scheme [31, p. 49].
 4.
The resolution rule as defined here is more general than the one considered in an earlier version of this paper [36], in that we admit complementary universal literals to be “merged” as long as the pivot is independent according to D (rather than \(\text {D}^{\text { trv}}\)). This definition—which we is required to capture proofs generated by DepQBF (see Example 1 in Sect. 4)—was proposed in (independent) work by Beyersdorff and Blinkhorn [6].
 5.
The presentation is adapted from a recent paper on QCDCL with dependency learning [37].
 6.
For simplicity, we do not consider the pure literal rule as part of QBCP.
 7.
 8.
As a sanity check, we verified that all configurations that were able to solve a particular instance returned the same result.
 9.
 10.
Our definition slightly differs from the original for the resolution rule: if restriction removes the pivot variable from both premises, we simply pick the (restriction of the) first premise as the result (rather than the clause containing fewer literals). This simplifies the certificate extraction argument given below.
Notes
Acknowledgements
Open access funding provided by Austrian Science Fund (FWF). We would like to thank Florian Lonsing for helpful discussions and for pointing out how to configure DepQBF so that it generates LDQ(D\(^\text {std}\)) proofs. This research was partially supported by Austrian Science Fund (FWF) Grants P27721 and W1255N23.
References
 1.Arora, S., Barak, B.: Computational Complexity–A Modern Approach. Cambridge University Press, Cambridge (2009)CrossRefzbMATHGoogle Scholar
 2.Balabanov, V., Jiang, J.R.: Unified QBF certification and its applications. Formal Methods Syst. Des. 41(1), 45–65 (2012)CrossRefzbMATHGoogle Scholar
 3.Balabanov, V., Jiang, J.R., Janota, M., Widl, M.: Efficient extraction of QBF (counter) models from longdistance resolution proofs. In: Bonet, B., Koenig, S. (eds.) Proceedings of the TwentyNinth AAAI Conference on Artificial Intelligence, AAAI 2015, pp. 3694–3701. AAAI Press (2015)Google Scholar
 4.Balabanov, V., Widl, M., Jiang, J.R.: QBF resolution systems and their proof complexities. In: Sinz, C., Egly, U. (eds.) Theory and Applications of Satisfiability Testing–SAT 2014, Lecture Notes in Computer Science, vol. 8561, pp. 154–169. Springer, Berlin (2014)Google Scholar
 5.Benedetti, M., Mangassarian, H.: QBFbased formal verification: experience and perspectives. J. Satisf. Boolean Model. Comput. 5(1–4), 133–191 (2008)MathSciNetzbMATHGoogle Scholar
 6.Beyersdorff, O., Blinkhorn, J.: Dependency schemes in QBF calculi: semantics and soundness. In: Rueher, M. (ed.) Principles and Practice of Constraint Programming22nd International Conference, CP 2016, Lecture Notes in Computer Science, vol. 9892, pp. 96–112. Springer, Berlin (2016)Google Scholar
 7.Beyersdorff, O., Chew, L., Janota, M.: On unification of QBF resolutionbased calculi. In: CsuhajVarjú, E., Dietzfelbinger, M. Ésik, Z. (eds.) Mathematical Foundations of Computer Science 2014–39th International Symposium, MFCS 2014, Lecture Notes in Computer Science, vol. 8635, pp. 81–93. Springer, Berlin (2014)Google Scholar
 8.Beyersdorff, O., Chew, L., Janota, M.: Proof complexity of resolutionbased QBF calculi. In: Mayr, E.W., Ollinger, N. (eds.) 32nd International Symposium on Theoretical Aspects of Computer Science, STACS 2015, LIPIcs, vol. 30, pp. 76–89. Schloss Dagstuhl–LeibnizZentrum fuer Informatik (2015)Google Scholar
 9.Beyersdorff, O., Chew, L., Mahajan, M., Shukla, A.: Feasible interpolation for QBF resolution calculi. In: Halldórsson, M.M., Iwama, K., Kobayashi, N., Speckmann, B. (eds.) Automata, Languages, and Programming–42nd International Colloquium, ICALP 2015, Lecture Notes in Computer Science, vol. 9134, pp. 180–192. Springer, Berlin (2015)Google Scholar
 10.Biere, A., Lonsing, F.: Integrating dependency schemes in searchbased QBF solvers. In: Strichman, O., Szeider, S. (eds.) Theory and Applications of Satisfiability Testing–SAT 2010, Lecture Notes in Computer Science, vol. 6175, pp. 158–171. Springer, Berlin (2010)Google Scholar
 11.Biere, A., Lonsing, F., Seidl, M.: Blocked clause elimination for QBF. In: Bjørner, N., SofronieStokkermans, V. (eds.) International Conference on Automated Deduction–CADE 23, Lecture Notes in Computer Science, vol. 6803, pp. 101–115. Springer, Berlin (2011)Google Scholar
 12.Blinkhorn, J., Beyersdorff, O.: Shortening QBF proofs with dependency schemes. In: Theory and Applications of Satisfiability Testing–SAT 2017, Lecture Notes in Computer Science, vol. 10491, pp. 263–280. Springer, Berlin (2017)Google Scholar
 13.Bloem, R., Könighofer, R., Seidl, M.: SATbased synthesis methods for safety specs. In: McMillan, K.L., Rival, X. (eds.) Verification, Model Checking, and Abstract Interpretation–VMCAI 2014, Lecture Notes in Computer Science, vol. 8318, pp. 1–20. Springer, Berlin (2014)Google Scholar
 14.Bubeck, U.: Modelbased transformations for quantified Boolean formulas. Ph.D. thesis, University of Paderborn (2010)Google Scholar
 15.Cadoli, M., Schaerf, M., Giovanardi, A., Giovanardi, M.: An algorithm to evaluate Quantified Boolean Formulae and its experimental evaluation. J. Autom. Reason. 28(2), 101–142 (2002)MathSciNetCrossRefzbMATHGoogle Scholar
 16.Cashmore, M., Fox, M., Giunchiglia, E.: Partially grounded planning as Quantified Boolean Formula. In: Borrajo, D., Kambhampati, S., Oddi, A., Fratini, S. (eds.) 23rd International Conference on Automated Planning and Scheduling, ICAPS 2013. AAAI (2013)Google Scholar
 17.Davis, M., Logemann, G., Loveland, D.: A machine program for theoremproving. Commun. ACM 5, 394–397 (1962)MathSciNetCrossRefzbMATHGoogle Scholar
 18.Egly, U.: On sequent systems and resolution for QBFs. In: Cimatti, A., Sebastiani, R. (eds.) Theory and Applications of Satisfiability Testing–SAT 2012, Lecture Notes in Computer Science, vol. 7317, pp. 100–113. Springer, Berlin (2012)Google Scholar
 19.Egly, U., Lonsing, F., Widl, M.: Longdistance resolution: proof generation and strategy extraction in searchbased QBF solving. In: McMillan, K.L., Middeldorp, A., Voronkov, A. (eds.) Logic for Programming, Artificial Intelligence, and Reasoning–LPAR 2013, Lecture Notes in Computer Science, vol. 8312, pp. 291–308. Springer, Berlin (2013)Google Scholar
 20.Gelder, A.V.: Contributions to the theory of practical quantified boolean formula solving. In: Milano, M. (ed.) Principles and Practice of Constraint Programming–18th International Conference, CP 2012, Lecture Notes in Computer Science, vol. 7514, pp. 647–663. Springer, Berlin (2012)Google Scholar
 21.Giunchiglia, E., Narizzano, M., Tacchella, A.: Clause/term resolution and learning in the evaluation of quantified Boolean formulas. J. Artif. Intell. Res. 26, 371–416 (2006)MathSciNetzbMATHGoogle Scholar
 22.Goultiaeva, A., Bacchus, F.: Exploiting QBF duality on a circuit representation. In: Fox, M., Poole, D. (eds.) Proceedings of the TwentyFourth AAAI Conference on Artificial Intelligence, AAAI 2010. AAAI Press (2010)Google Scholar
 23.Goultiaeva, A., Seidl, M., Biere, A.: Bridging the gap between dual propagation and CNFbased QBF solving. In: Macii, E. (ed.) Design, Automation and Test in Europe, DATE 13, pp. 811–814. EDA Consortium San Jose, CA, USA/ACM DL (2013)Google Scholar
 24.Goultiaeva, A., Van Gelder, A., Bacchus, F.: A uniform approach for generating proofs and strategies for both true and false QBF formulas. In: Walsh, T. (ed.) Proceedings of IJCAI 2011, pp. 546–553. IJCAI/AAAI (2011)Google Scholar
 25.Heule, M., Seidl, M., Biere, A.: A unified proof system for QBF preprocessing. In: Demri, S., Kapur, D., Weidenbach, C. (eds.) Automated Reasoning–7th International Joint Conference, IJCAR 2014, Lecture Notes in Computer Science, vol. 8562, pp. 91–106. Springer, Berlin (2014)Google Scholar
 26.Janota, M., Klieber, W., MarquesSilva, J., Clarke, E.M.: Solving QBF with counterexample guided refinement. In: Cimatti, A., Sebastiani, R. (eds.) Theory and Applications of Satisfiability Testing–SAT 2012, Lecture Notes in Computer Science, vol. 7317, pp. 114–128. Springer, Berlin (2012)Google Scholar
 27.Janota, M., MarquesSilva, J.: On propositional QBF expansions and Qresolution. In: Järvisalo, M., Van Gelder, A. (eds.) Theory and Applications of Satisfiability Testing–SAT 2013, Lecture Notes in Computer Science, vol. 7962, pp. 67–82. Springer, Berlin (2013)Google Scholar
 28.Kleine Büning, H., Karpinski, M., Flögel, A.: Resolution for quantified Boolean formulas. Inf. Comput. 117(1), 12–18 (1995)MathSciNetCrossRefzbMATHGoogle Scholar
 29.Klieber, W., Sapra, S., Gao, S., Clarke, E.M.: A nonprenex, nonclausal QBF solver with gamestate learning. In: Strichman, O., Szeider, S. (eds.) Theory and Applications of Satisfiability Testing–SAT 2010, Lecture Notes in Computer Science, vol. 6175, pp. 128–142. Springer, Berlin (2010)Google Scholar
 30.Kronegger, M., Pfandler, A., Pichler, R.: Conformant planning as benchmark for QBFsolvers. In: International Workshop on Quantified Boolean Formulas–QBF 2013 (2013). http://fmv.jku.at/qbf2013/
 31.Lonsing, F.: Dependency schemes and searchbased QBF solving: theory and practice. Ph.D. thesis, Johannes Kepler University, Linz, Austria (2012)Google Scholar
 32.Lonsing, F., Bacchus, F., Biere, A., Egly, U., Seidl, M.: Enhancing searchbased QBF solving by dynamic blocked clause elimination. In: Davis, M., Fehnker, A., McIver, A., Voronkov, A. (eds.) Logic for Programming, Artificial Intelligence, and Reasoning20th International Conference, LPAR20 2015. Lecture Notes in Computer Science, vol. 9450, pp. 418–433. Springer, Berlin (2015)Google Scholar
 33.Lonsing, F., Egly, U., Van Gelder, A.: Efficient clause learning for quantified Boolean formulas via QBF pseudo unit propagation. In: Järvisalo, M., Van Gelder, A. (eds.) Theory and Applications of Satisfiability Testing–SAT 2013, Lecture Notes in Computer Science, vol. 7962, pp. 100–115. Springer, Berlin (2013)Google Scholar
 34.MarquesSilva, J.P.: The impact of branching heuristics in propositional satisfiability algorithms. In: Barahona, P., Alferes, J.J. (eds.) Progress in Artificial Intelligence, 9th Portuguese Conference on Artificial Intelligence, EPIA ’99, Lecture Notes in Computer Science, vol. 1695, pp. 62–74. Springer, Berlin (1999)Google Scholar
 35.Niemetz, A., Preiner, M., Lonsing, F., Seidl, M., Biere, A.: Resolutionbased certificate extraction for QBF. In: Cimatti, A., Sebastiani, R. (eds.) Theory and Applications of Satisfiability Testing–SAT 2012, Lecture Notes in Computer Science, vol. 7317, pp. 430–435. Springer, Berlin (2012)Google Scholar
 36.Peitl, T., Slivovsky, F., Szeider, S.: Long distance Qresolution with dependency schemes. In: Creignou, N., Berre, D.L. (eds.) Theory and Applications of Satisfiability Testing–SAT 2016, Lecture Notes in Computer Science, vol. 9710, pp. 500–518. Springer, Berlin (2016)Google Scholar
 37.Peitl, T., Slivovsky, F., Szeider, S.: Dependency learning for QBF. In: Gaspers, S., Walsh, T. (eds.) Theory and Applications of Satisfiability Testing–SAT 2017, Lecture Notes in Computer Science, vol. 10491, pp. 298–313. Springer, Berlin(2017)Google Scholar
 38.Rintanen, J.: Asymptotically optimal encodings of conformant planning in QBF. In: 22nd AAAI Conference on Artificial Intelligence, AAAI 2007, pp. 1045–1050. AAAI (2007)Google Scholar
 39.Samer, M., Szeider, S.: Backdoor sets of quantified Boolean formulas. J. Autom. Reason. 42(1), 77–97 (2009)MathSciNetCrossRefzbMATHGoogle Scholar
 40.Slivovsky, F.: Structure in #SAT and QBF. Ph.D. thesis, TU Wien (2015)Google Scholar
 41.Slivovsky, F., Szeider, S.: Computing resolutionpath dependencies in linear time. In: Cimatti, A., Sebastiani, R. (eds.) Theory and Applications of Satisfiability Testing–SAT 2012, Lecture Notes in Computer Science, vol. 7317, pp. 58–71. Springer, Berlin (2012)Google Scholar
 42.Slivovsky, F., Szeider, S.: Soundness of Qresolution with dependency schemes. Theoret. Comput. Sci. 612, 83–101 (2016)MathSciNetCrossRefzbMATHGoogle Scholar
 43.Staber, S., Bloem, R.: Fault localization and correction with QBF. In: MarquesSilva, J., Sakallah, K.A. (eds.) Theory and Applications of Satisfiability Testing–SAT 2007, Lecture Notes in Computer Science, vol. 4501, pp. 355–368. Springer, Berlin(2007)Google Scholar
 44.Van Gelder, A.: Variable independence and resolution paths for quantified Boolean formulas. In: Lee, J. (ed.) Principles and Practice of Constraint Programming–CP 2011, Lecture Notes in Computer Science, vol. 6876, pp. 789–803. Springer, Berlin (2011)Google Scholar
 45.Zhang, L., Malik, S.: Conflict driven learning in a quantified Boolean satisfiability solver. In: Pileggi, L.T. Kuehlmann, A. (eds.) Proceedings of the 2002 IEEE/ACM International Conference on Computeraided Design, ICCAD 2002, pp. 442–449. ACM/IEEE Computer Society (2002)Google Scholar
 46.Zhang, L., Malik, S.: The quest for efficient boolean satisfiability solvers. In: Brinksma, D., Larsen, K.G. (eds.) Computer Aided Verification: 14th International Conference (CAV 2002), Lecture Notes in Computer Science, vol. 2404, pp. 17–36 (2002)Google Scholar
 47.Zhang, L., Malik, S.: Towards a symmetric treatment of satisfaction and conflicts in quantified Boolean formula evaluation. In: Hentenryck, P.V. (ed.) Principles and Practice of Constraint Programming–8th International Conference, CP 2002, Lecture Notes in Computer Science, vol. 2470, pp. 200–215. Springer, Berlin (2002)Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.