1 Introduction

Checking the satisfiability of formulas (Garey and Johnson 1990), along with checking whether formulas are satisfied in a given model (i.e. model checking Emerson and Clarke 1980), is an important computational task associated with formalisms used for specifying software systems (Huth and Ryan 2004). In the case of multiagent systems (Wooldridge 2009; Shoham and Leyton-Brown 2008), the satisfiability problem underlies the following two tasks, appearing during system development. The first is the task of specification verification. Checking whether there exists a system that satisfies the given specification is essentially the satisfiability problem. The second task is related to implementation of individual agents (Shoham 1993). Often such implementations are based on logical formalisms and execution of programs of agents involves reasoning tasks related to that formalism. These tasks are based on checking the satisfiability of formulas.

The problem with checking the satisfiability of formulas of modal logics for multiagent systems is its high computational complexity (Halpern and Moses 1992). Such logics usually combine fix-point modalities, like common beliefs and mutual intentions, with axioms interconnecting different types of modalities, like introspection axioms and realism axioms (Levesque et al. 1990; Rao and Georgeff 1991; Wooldridge 2000; Dunin-Kęplicz and Verbrugge 2010). Richness of the formalisms leads to high complexity of computational tasks such as checking the satisfiability. One way of addressing this problem is restricting the language of the formalism (Halpern 1995). Knowing possible language restrictions and the associated complexity of computational tasks can help designers of programming languages or the graphical languages for systems design to enforce the chosen restrictions and ensure lower cost of computational tasks like reasoning or checking system validity. A classic example of this approach is the Horn fragment of first order logic that is adopted in logic programming languages, like Prolog.

The rest of the paper is organized as follows. We start with discussing the related literature and the state of the art in Sect. 2. Then we introduce the logical framework in Sect. 3. In Sect. 4 we present the general idea of modal context restriction together with two restrictions for logic \({\textsc {Team}}{\textsc {Log}}.\)We discuss the restrictions and provide examples in Sect. 5. Section 6 contains the complexity results. Section 7 contains the conclusions.

2 Related literature

The main focus of this paper is the complexity of the satisfiability problem for modal BDI logics. Modal logics of agency based on BDI model (Bratman 1987) are formalisms used to specify individual agents in terms of their beliefs, goals/desires and intentions. In the context of multi-agent systems this formalism is extended with fix-point modalities for common beliefs and collective (or joint) intentions and commitments. These extensions where first introduced in the seminal work of Cohen, Levesque et al. (1990). This was followed by a number of well known early formalism including KARO (van Linder et al. 1994, 1998; Meyer et al. 1999; van der Hoek et al. 1999; Aldewereld et al. 2004), \(\mathcal {LORA}\) (Wooldridge 2000), and \({\textsc {Team}}{\textsc {Log}}\) (Dunin-Kęplicz and Verbrugge 1996, 2002, 2004, 2010). More recent works on BDI logics and related formalisms focus on extending the formalism and for obtaining better formalisation of concepts like desires, intentions, or coalitions. Dubois et al. (2017) consider the problem of desires revision and propose a formalism for reasoning about desires based on possibilistic logic. Wobcke (2015) proposed an agent dynamic logic (ADL) that allows for reasoning about intentions and action. (Bauters et al. 2014a, b, 2017) develop a BDI formalism that allows for modelling and reasoning under uncertainty. Ågotnes and Alechina (2018) study axiomatisation and complexity of epistemic coalition logic, a formalism combining modalities for expressing knowledge and common knowledge with the modality expressing that a group of agents is effective to make a formula true. Lorini and Sartor (2016) propose a logic for reasoning about social influence based on Belnap et al. (2001) logic of ‘seeing to it that’, STIT. The BDI model of agency remains an actively studied and used model up to this day. Most recent applications include multiagent system organizations (Keogh and Sonenberg 2020), supply chain quality inspection (Yan et al. 2020), and traffic simulations (Rüb and Dunin-Kęplicz 2020).

A common characteristic of multi-agent BDI logics is adopting, along with standard modal systems \(\text {K}_{n}\), \(\text {KD}_{n}\) or \(\text {KD45}_{n}\), mixed axioms that interrelate modalities representing different aspects of agent description. Well-known examples of such axioms are the realism axioms (Cohen and Levesque 1990; Rao and Georgeff 1998; Wooldridge 2000) and the introspection axioms (Rao and Georgeff 1991; Dunin-Kęplicz and Verbrugge 2002, 2004). In the case of basic BDI logics for a single agent (without the temporal or dynamic component), addition of these axioms does not change the complexity of the satisfiability problem, and they all remain PSPACE complete (Rao and Georgeff 1998; Dziubiński et al. 2007). In the multiagent case such logics are extended by lifting individual modalities representing beliefs, goals or intentions to the group level by introducing fixpoint modalities representing common beliefs, mutual goals or mutual intentions (Levesque et al. 1990; Wooldridge 2000; Aldewereld et al. 2004; Dunin-Kęplicz and Verbrugge 2010; Ågotnes and Alechina 2018). Adding such modalities leads to EXPTIME hard satisfiability problem (Halpern and Moses 1992), and presence of mixed axioms does not affect this result (Dziubiński et al. 2007).

One of the ways of dealing with high complexity of logical formalisms is restricting their language, so that the complexity of the satisfiability problem is reduced.Footnote 1 Restricting modal depth of formulas by a constant may lead to NPTIME complete satisfiability problem, while combining this with restricting the number of propositional symbols leads to linear time solvability of the problem. This, however, is with a constant that depends exponentially on the number of these symbols (Halpern 1995). In the case of modal logics with fixpoint modalities the restrictions mentioned above are not that promising, as the satisfiability problem remains EXPTIME hard, even when modal depth of formulas is bounded by 2 (Halpern and Moses 1992; Dziubiński et al. 2007). Motivated by these results, Dziubiński (2013) proposed a new kind of language restriction called modal context restriction and applied restrictions of this kind to standard systems of multimodal logics (generated by different combinations of axioms K, T, D, 4 and 5) enriched with fixpoint modalities. This leads to PSPACE completeness and, when combined with modal depth restriction, to NPTIME completeness of the satisfiability problem.

Another type of restrictions considered in the literature are sub-propositional fragments of modal logics. In Nguyen (2005) it is shown that the Horn fragment of some of modal systems has NPTIME complete satisfiability problem, which becomes PTIME complete when combined with modal depth restriction. More recently, Bresolin et al. (2016) and (Bresolin et al. 2018) considered the Horn and the Krom fragments of modal basic logics K, T, K4, and S4, also in combination with allowing box or diamond operators in positive literals only. They show that all these fragments are PTIME solvable. Wałęga (2019) considered the core fragment (i.e. an intersection of the Horn and Krom fragments) and showed that if only box operators in positive literals are allowed, the fragment is NL-complete. In Bauland et al. (2006) restrictions on propositional operators used in formulas are considered. To study different sets of propositional operators used in formulas, Post lattice (Post 1941), which has been successfully used to classify the complexity problems for propositional calculus, is used. It is shown that in the case of basic normal modal system \(\text {K}\) there is a trichotomy: depending on the boolean operators used the satisfiability problem is either PSPACE complete, coNPTIME complete or PTIME solvable). In the case of normal modal system \(\text {KD}\) there is a dichotomy: the satisfiability problem is either PSPACE complete or PTIME solvable. Almost complete characterization was also obtained for modal systems \(\text {T}\), \(\text {S}4\) and \(\text {S}5\). Similar approach was also applied to LTL in Bauland et al. (2009) and to \(\hbox {CTL}^{*}\) and CTL in Meier et al. (2008).

In this paper, we build on the idea of modal context restrictions from Dziubiński (2013) and present modal context restrictions for BDI logics with two types of mixed axioms, realism axioms and introspection axioms, interrelating modalities of different basic multimodal logics. Presence of mixed axioms results in a richer setup which makes the problem of designing the right modal context restriction much more difficult than in Dziubiński (2013) and results in more complex restrictions. In particular, we propose two novel modal context restrictions, called \({\mathbf{R}}_1\) and \({\mathbf{R}}_2\). We show that the basic modal context restriction, \({\mathbf{R}}_1\), leads to PSPACE completeness of the satisfiability problem. However, existence of the introspection axioms results in PSPACE hardness of the problem, even if modal depth of formulas is bounded by 2. An interesting feature of the considered logic is the fact that models satisfying the formulas may be exponentially deep (with respect to their size). For this reason a standard tableau based algorithm has to be extended so that it uses polynomial space to decide the satisfiability. Two of the restrictions proposed in the paper, \(\hbox {R}_2\) and \(\hbox {R}_1(c)\), lead to NPTIME solvability of the satisfiability problem, when combined with modal depth restriction. As working formalism we chose \({\textsc {Team}}{\textsc {Log}}\) (Dunin-Kęplicz and Verbrugge 2010), a well-known BDI formalism designed to formalise teamwork in multiagent systems.

3 Logical framework

\({\textsc {Team}}{\textsc {Log}},\) developed by Dunin-Kęplicz and Verbrugge in a series of papers (Dunin-Kęplicz and Verbrugge 1996, 2002, 2003, 2004) and a book (Dunin-Kęplicz and Verbrugge 2010), is a logical framework proposed to formalize individual and group aspects of BDI systems. The full \({\textsc {Team}}{\textsc {Log}}\) is a very reach formalism allowing for expressing and reasoning about various aspects of individual agents and multiagent systems relevant to cooperative problem solving. Moreover, it is intended to be suitable for possible enrichments that could be designed for chosen classes of multiagent systems. In this paper we focus on the core \({\textsc {Team}}{\textsc {Log}}\) (which will be simply called \({\textsc {Team}}{\textsc {Log}}\)) presented below. For the full framework see (Dunin-Kęplicz and Verbrugge 2010).

\({\textsc {Team}}{\textsc {Log}}\) is a propositional multimodal logic that introduces five sets of modal operators based on a finite and non-empty set of agents \({\mathcal {A}}\). Three sets of operators, \({\Omega }^{\mathrm {B}}= \{[\mathrm {B}]_{j}: j \in {\mathcal {A}}\}\), \({\Omega }^{\mathrm {G}}= \{[\mathrm {G}]_{j}: j \in {\mathcal {A}}\}\) and \({\Omega }^{\mathrm {I}}= \{[\mathrm {I}]_{j}: j \in {\mathcal {A}}\}\), are used for representing beliefs, goals and intentions, respectively, of individual agents. We call them individual modalities, for short, and the set of these modalities is denoted by \({\Omega }^{\mathrm {ind}}\). Additional two sets of operators, \({{\Omega }^{\mathrm {B}}}^{+}= \{[\mathrm {B}]^{+}_{G}: G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \} \}\) and \({{\Omega }^{\mathrm {I}}}^{+}= \{[\mathrm {I}]^{+}_{G}: G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\}\) are used for representing common beliefs and mutual intentions of groups of agents. We call them fixpoint modalities. The set of all modal operators of \({\textsc {Team}}{\textsc {Log}}\) is denoted by \({\Omega }^{\mathrm {T}}\).

The language of \({\textsc {Team}}{\textsc {Log}},\) denoted by \({{\mathcal {L}}}^{\mathrm {T}}\), is based on a countable set of propositional variables \({\mathcal {P}}\) and on a set of modal operators \({\Omega }^{\mathrm {T}}\). It is a minimal set of formulas satisfying the following properties

  • \({\mathcal {P}}\subseteq {{\mathcal {L}}}^{\mathrm {T}}\),

  • If \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\), then \(\lnot \varphi \in {{\mathcal {L}}}^{\mathrm {T}}\),

  • If \(\varphi _1 \in {{\mathcal {L}}}^{\mathrm {T}}\) and \(\varphi _2 \in {{\mathcal {L}}}^{\mathrm {T}}\), then \(\varphi _1 \wedge \varphi _2 \in {{\mathcal {L}}}^{\mathrm {T}}\),

  • If \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) and \(\Box \in {\Omega }^{\mathrm {T}}\), then \(\Box \varphi \in {{\mathcal {L}}}^{\mathrm {T}}\).

We will also use the following standard abbreviations for propositional constants and operators:

  • \(\top {\mathop {=}\limits ^{\text {\tiny def}}}p \wedge \lnot p\), where \(p \in {\mathcal {P}}\),

  • \(\top {\mathop {=}\limits ^{\text {\tiny def}}}\lnot \bot \),

  • \(\varphi _1 \vee \varphi _2 {\mathop {=}\limits ^{\text {\tiny def}}}\lnot (\lnot \varphi _1 \wedge \lnot \varphi _2)\),

  • \(\varphi _1 \rightarrow \varphi _2 {\mathop {=}\limits ^{\text {\tiny def}}}\lnot (\varphi _1 \wedge \lnot \varphi _2)\),

  • \(\varphi _1 \leftrightarrow \varphi _2 {\mathop {=}\limits ^{\text {\tiny def}}}(\varphi _2 \rightarrow \varphi _2) \wedge (\varphi _2 \rightarrow \varphi _1)\),

as well as the following abbreviations for general beliefs and general intentions of a non-empty group of agents G

  • \([\mathrm {B}]_{G}\varphi {\mathop {=}\limits ^{\text {\tiny def}}}\bigwedge _{j \in G} [\mathrm {B}]_{j}\varphi \),

  • \([\mathrm {I}]_{G}\varphi {\mathop {=}\limits ^{\text {\tiny def}}}\bigwedge _{j \in G} [\mathrm {I}]_{j}\varphi \).

Given a finite set of formulas \(\Phi \), we will use \(\bigwedge \Phi \) to denote the conjunction of all formulas in the set, and \(\bigvee \Phi \) to denote the disjunction of all formulas in the set. We will also use the conventions that \(\bigwedge \varnothing = \top \) and \(\bigvee \varnothing = \bot \).

Throughout the paper we will refer to the notion of single negation. Given a formula \(\varphi \),

$$\begin{aligned} \sim \!\varphi = \left\{ \begin{array}{ll} \psi, \hfill & \text { if }\varphi = \lnot \psi \text { for some formula }\psi , \hfill \\ \lnot \varphi, &\text { otherwise}. \end{array} \right. \end{aligned}$$

A set of formulas \(\Phi \) is closed under single negation iff for all \(\varphi \in \Phi \), it holds that \(\sim \!\varphi \in \Phi \). Given a set of formulas \(\Phi \) we will use \(\lnot \Phi \) to denote the smallest set containing \(\Phi \) and closed under single negation.

We will use \(|\varphi |\) to denote the length of a formulaFootnote 2 and \(\mathrm {dep}(\varphi )\) to denote the modal depth of a formula. Both notions have standard meaning and we omit the definition here. Given a finite set of formulas \(\Phi \),

$$\begin{aligned} \mathrm {dep}(\Phi ) = \left\{ \begin{array}{l l} 0, \hfill & \text { if }\Phi = \varnothing \hfill \\ \max \{\mathrm {dep}(\varphi ) : \varphi \in \Phi \}, \hfill & \text { otherwise}. \hfill \end{array} \right. \end{aligned}$$

3.1 Deduction system

\({\textsc {Team}}{\textsc {Log}}\) combines axiom systems \(\text {KD45}_{n}\), associated with modal operators from \({\Omega }^{\mathrm {B}}\) representing beliefs, \(\text {K}_{n}\), associated with modal operators from \({\Omega }^{\mathrm {G}}\) representing goals, and \(\text {KD}_{n}\), associated with modal operators from \({\Omega }^{\mathrm {I}}\) representing intentions. Additionally, axioms interrelating different modalities, called mixed axioms, as well as axioms related to fixpoint modalities from \({{\Omega }^{\mathrm {B}}}^{+}\) and \({{\Omega }^{\mathrm {I}}}^{+}\), representing common beliefs and mutual intentions, are introduced. All these are presented below.

For each of the modal operators \(\Box \in {\Omega }^{\mathrm {ind}}\), the following axioms and deduction rules of the basic modal system K are adopted:

$$\begin{aligned} \begin{array}{lll} \mathbf{P } \hfill & \text {All instances of propositional tautologies} & \hfill \\ \mathbf{K } \hfill & \Box \varphi \wedge \Box (\varphi \rightarrow \psi ) \rightarrow \Box \psi &\hfill \\ \mathbf{MP} \hfill & \text {From } \varphi \text { and } \varphi \rightarrow \psi \text { infer }\psi \hfill & \text {(Modus ponens)} \hfill \\ \mathbf{GEN} \hfill & \text {From }\varphi \text { infer }\Box \varphi \hfill & \text {(Generalization).} \hfill \\ \end{array} \end{aligned}$$

Additionally, depending on which of the sets \({\Omega }^{\mathrm {B}}\), \({\Omega }^{\mathrm {G}}\) or \({\Omega }^{\mathrm {I}}\), \(\Box \) belongs to, a subset of the following axioms is adopted:

$$\begin{aligned} \begin{array}{ll} \mathbf{D } \hfill &{} \lnot \Box \bot \hfill \\ \mathbf{4 } \hfill &{} \Box \varphi \rightarrow \Box \Box \varphi \hfill \\ \mathbf{5 } \hfill &{} \lnot \Box \varphi \rightarrow \Box \lnot \Box \varphi . \hfill \\ \end{array} \end{aligned}$$

The intended meaning of a formula \([\mathrm {B}]_{j}\varphi \) is that agent j believes that \(\varphi \) and axioms and inference rules of the standard doxastic modal logic (c.f. Meyer and van der Hoek 1995; Fagin et al. 2003) forming the \(\text {KD45}_{n}\) system are adopted for modal operators from \({\Omega }^{\mathrm {B}}\). The interpretations of the axioms are in this case as follows: K (distribution of beliefs), D (consistency of beliefs), 4 (positive introspection) and 5 (negative introspection).

The intended meaning of a formula \([\mathrm {G}]_{j}\varphi \) is that agent j has goal \(\varphi \) and axioms and inference rules of the system \(\text {K}_{n}\) are adopted for modal operators from \({\Omega }^{\mathrm {G}}\). Axiom K is interpreted as distribution of goals. Note that goals can be inconsistent.

The intended meaning of a formula \([\mathrm {I}]_{j}\varphi \) is that agent j intends \(\varphi \) and axioms and inference rules of the system \(\text {KD}_{n}\) are adopted for modal operators from \({\Omega }^{\mathrm {I}}\). The interpretations of the axioms are in this case as follows: K (distribution of intentions) and D (consistency of intentions).

Mixed axioms, interrelating different modalities, are as follows:

$$\begin{aligned} \begin{array}{lll} \mathbf{BG4} \hfill &{} [\mathrm {G}]_{j}\varphi \rightarrow [\mathrm {B}]_{j}[\mathrm {G}]_{j}\varphi &{} \text {(Positive introspection of goals)} \hfill \\ \mathbf{BG5} \hfill &{} \lnot [\mathrm {G}]_{j}\varphi \rightarrow [\mathrm {B}]_{j}\lnot [\mathrm {G}]_{j}\varphi \hfill &{} \text {(Negative introspection of goals)} \hfill \\ \mathbf{BI4} \hfill &{} [\mathrm {I}]_{j}\varphi \rightarrow [\mathrm {B}]_{j}[\mathrm {I}]_{j}\varphi \hfill &{} \text {(Positive introspection of intentions)} \hfill \\ \mathbf{BI5} \hfill &{} \lnot [\mathrm {I}]_{j}\varphi \rightarrow [\mathrm {B}]_{j}\lnot [\mathrm {I}]_{j}\varphi \hfill &{} \text {(Negative introspection of intentions)} \hfill \\ \mathbf{IG} \hfill &{} [\mathrm {I}]_{j}\varphi \rightarrow [\mathrm {G}]_{j}\varphi \hfill &{} \text {(Goals and intentions compatibility).} \hfill \end{array} \end{aligned}$$

Axioms BG4 and BG5 correspond to positive and negative introspection of goals: an agent is aware of the goals it has and of the goals it does not have. Analogous axioms, BI4 and BI5, are adopted for intentions. Axiom IG corresponds to goals and intentions compatibility: intentions of an agent are a subset of its goals.

The axioms of positive and negative introspection of goals and intentions were also adopted in the first version of the formalism of Rao and Georgeff (1991). The axiom of goals and intentions compatibility is discussed in Rao and Georgeff (1998) as one of the realism axioms, extending the realism axiom of Cohen and Levesque (1990) for beliefs and goals compatibility to compatibility of goals and intentions. It was also adopted by Wooldridge in his \(\mathcal {LORA}\) formalism (Wooldridge 2000).

For fixpoint modalities \([O]^{+}_{G} \in {{\Omega }^{\mathrm {B}}}^{+}\cup {{\Omega }^{\mathrm {I}}}^{+}\) the following axiom and a rule of inference are adopted:

$$\begin{aligned} \begin{array}{lll} \mathbf{C } \hfill &{} [O]^{+}_{G}\varphi \leftrightarrow [O]_{G}(\varphi \wedge [O]^{+}_{G}\varphi ) &{} \hfill \\ \mathbf{RC} \hfill &{} \text {From }\varphi \rightarrow [O]_{G}(\psi \wedge \varphi ) \text { infer }\varphi \rightarrow [O]^{+}_{G}\psi \hfill &{} \text {(Induction).} \hfill \end{array} \end{aligned}$$

The intended meaning of a formula \([\mathrm {B}]^{+}_{G}\varphi \) is that group G has a common belief that \(\varphi \). This notion is defined in terms of general beliefs, \([\mathrm {B}]_{G}\varphi \), meaning that every agent in G believes that \(\varphi \). Thus there is a common belief that \(\varphi \) in group G if and only if every agent in G believes that \(\varphi \), every agent in G believes that every agent in G believes that \(\varphi \), etc., ad infinitum.

The intended meaning of a formula \([\mathrm {I}]^{+}_{G}\varphi \) is that group G has a mutual intention that \(\varphi \). This notion is defined in terms of general intentions, \([\mathrm {I}]_{G}\varphi \), meaning that every agent in G intends that \(\varphi \), in analogous way to how common beliefs are defined.

Fixpoint modalities such as common beliefs or mutual intentions are widely used in formalism for multiagent systems such as the formalism of Levesque et al. (1990), \(\mathcal {LORA}\) of Wooldridge (2000) and the extension of KARO by Aldewereld, Hoek and Meyer and Bratman (1987).

As we wrote above, operators of general beliefs and general intentions, \([\mathrm {B}]_{G}\) and \([\mathrm {I}]_{G}\), are defined in terms of individual beliefs and individual intentions and a formula with these operators can be translated to a formula without them in linear time, which increases the size of the formula be a linear factor \(\le |{\mathcal {A}}|\). For that reason we will omit them from now on, as they are not relevant for the complexity issues considered in this paper.

3.2 Semantics

Formulas from \({{\mathcal {L}}}^{\mathrm {T}}\) are interpreted in Kripke models with accessibility relations corresponding to modalities from \({\Omega }^{\mathrm {T}}\). Since accessibility relations corresponding to operators from \({{\Omega }^{\mathrm {B}}}^{+}\) and \({{\Omega }^{\mathrm {I}}}^{+}\) can be defined in terms of accessibility relations corresponding to individual modalities \({\Omega }^{\mathrm {ind}}\), the definition of Kripke models is based on relations corresponding to these operators only.

Definition 1

(Kripke frame) A Kripke frame is a tuple \({\mathcal {F}} =\left( W, \left\{ O_j : [O]_{j} \in {\Omega }^{\mathrm {ind}}\right\} \right) \), where

  • \(W \ne \varnothing \) is the set of possible worlds.

  • For all \([O]_{j} \in {\Omega }^{\mathrm {ind}}\), \(O_j \subseteq W \times W\). Each relation \(O_j\) stands for the accessibility relation corresponding to the operator \([O]_{j}\).

Definition 2

(Kripke model) A Kripke model is a pair \({\mathcal {M}} = \left( {\mathcal {F}},Val\right) \), where \({\mathcal {F}}\) is a Kripke frame and

  • \(Val : {\mathcal {P}}\times W \rightarrow \{0,1\}\) is a valuation function that assigns the truth values to atomic propositions in worlds.

Given a binary relation \(R \subseteq W \times W\) and \(w \in W\) we will use R(w) to denote the set of worlds accessible from w, that is \(R(w) = \{v \in W : (w,v) \in R\}\). Moreover, we will use \(R^{+}\) to denote the transitive closure of R. Additionally, given a family of relations \(\{R_j : j \in {\mathcal {A}}\}\) and a set of agents \(G\subseteq {\mathcal {A}}\) , relation \(R_G = \bigcup _{j \in G} R_j\). The relation corresponding to a modal operator \([O]^{+}_{G} \in {\Omega }^{\mathrm {T}}\) is \(O^{+}_G\).

Definition 3

(Satisfaction) Let \({\mathcal {M}}\) be a Kripke model, w be a world in \({\mathcal {M}}\) and \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) be a formula. The notion of \(\varphi \) being satisfied (or being true or holding) in \({\mathcal {M}}\) at w is defined inductively as follows:

$$\begin{aligned} \begin{array}{lll} ({\mathcal {M}},w) \vDash p \hfill & \text {iff } \hfill & Val(p,w) = 1, \hfill \\ ({\mathcal {M}},w) \vDash \lnot \varphi \hfill & \text {iff } \hfill & ({\mathcal {M}},w) \nvDash \varphi , \hfill \\ ({\mathcal {M}},w) \vDash \varphi _1 \wedge \varphi _2 \hfill & \text {iff } \hfill & ({\mathcal {M}},w) \vDash \varphi _1\text { and }({\mathcal {M}},w) \vDash \varphi _2, \hfill \\ ({\mathcal {M}},w) \vDash [O]_{j}\varphi \hfill & \text {iff } \hfill & ({\mathcal {M}},v) \vDash \varphi \text {, for all }v \in O_j(w), \hfill \\ ({\mathcal {M}},w) \vDash [O]^{+}_{G}\varphi \hfill & \text {iff } \hfill & ({\mathcal {M}},v) \vDash \varphi \text {, for all }v \in O^{+}_G(w). \hfill \\ \end{array} \end{aligned}$$

Let \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) be a formula. We say that \(\varphi \) is valid in a Kripke model \({\mathcal {M}}\) if for every world w in \({\mathcal {M}}\), \(({\mathcal {M}},w) \vDash \varphi \). We denote this fact by \({\mathcal {M}} \vDash \varphi \). We say that \(\varphi \) is satisfiable in \({\mathcal {M}}\) if there exists a world w in \({\mathcal {M}}\) such that \(({\mathcal {M}},w) \vDash \varphi \). Let \({\mathcal {C}}\) be a class of Kripke models. We say that \(\varphi \) is valid in \({\mathcal {C}}\) if \({\mathcal {M}} \vDash \varphi \), for every \({\mathcal {M}} \in {\mathcal {C}}\). We denote this fact by \({\mathcal {C}} \vDash \varphi \). We say that \(\varphi \) is satisfiable in \({\mathcal {C}}\) if there exists \({\mathcal {M}} \in {\mathcal {C}}\) such that \(\varphi \) is satisfiable in \({\mathcal {M}}\).

Axioms of modal systems \(\text {K}_{n}\), \(\text {KD}_{n}\) and \(\text {KD45}_{n}\), as far as they do not hold on all frames like K, correspond to well-known structural properties on Kripke frames, in the sense that they hold on all frames having certain structural properties (c.f. van Benthem 1984). Axiom D (adopted for \([\mathrm {B}]_{j}\) and \([\mathrm {I}]_{j}\)) corresponds to seriality of \(O_j\) (where \(O \in \{\mathrm {B},\mathrm {I}\}\))

$$\begin{aligned} \forall s (O_j(s) \ne \varnothing ), \end{aligned}$$

axiom 4 (adopted for \([\mathrm {B}]_{j}\)) corresponds to transitivity of \(\mathrm {B}_j\)

$$\begin{aligned} \forall s,t (t \in \mathrm {B}_j(s) \rightarrow \mathrm {B}_j(t) \subseteq \mathrm {B}_j(s)), \end{aligned}$$

and axiom 5 (adopted for \([\mathrm {B}]_{j}\)) corresponds to Euclidity of \(\mathrm {B}_j\)

$$\begin{aligned} \forall s,t (t \in \mathrm {B}_j(s) \rightarrow \mathrm {B}_j(s) \subseteq \mathrm {B}_j(t)). \end{aligned}$$

Similarly, the mixed axioms correspond to certain properties of Kripke frames. Axioms of positive introspection, B O4 with \(O \in \{\mathrm {G},\mathrm {I}\}\), correspond to the following property

$$\begin{aligned} \forall s,t (t \in \mathrm {B}_j(s) \rightarrow O_j(t) \subseteq O_j(s)). \end{aligned}$$

We will call this property generalized transitivity. Axioms of negative introspection, B O5, with \(O \in \{\mathrm {G},\mathrm {I}\}\), correspond to the property

$$\begin{aligned} \forall s,t (t \in \mathrm {B}_j(s) \rightarrow O_j(s) \subseteq O_j(t)). \end{aligned}$$

We will call this property generalized Euclidity. Finally, axiom IG corresponds to the property

$$\begin{aligned} \mathrm {G}_j \subseteq \mathrm {I}_j. \end{aligned}$$

Proofs of these correspondences are given in Dunin-Kęplicz and Verbrugge (2004). Also, they follow directly from Sahlqvist theorem (c.f. Blackburn et al. 2002, for example). The class of all Kripke frames with accessibility relations satisfying the properties above will be called \({\textsc {Team}}{\textsc {Log}}\) frames. Analogously \({\textsc {Team}}{\textsc {Log}}\) models are defined. We will say that \(\varphi \) is \({\textsc {Team}}{\textsc {Log}}\) provable, denoted by \(\vdash _T \varphi \), if there exists a proof of \(\varphi \) that includes axioms from \({\textsc {Team}}{\textsc {Log}}.\) The deduction system of \({\textsc {Team}}{\textsc {Log}}\) is sound and complete with respect to the class of \({\textsc {Team}}{\textsc {Log}}\) models, as was shown in Dunin-Kęplicz and Verbrugge (2002).

Theorem 1

Let \({\mathcal {T}}\) be the class of \({\textsc {Team}}{\textsc {Log}}\) models. Then for any \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\)

$$\begin{aligned} {\mathcal {T}}\vDash \varphi \text { iff } \vdash _{T} \varphi \end{aligned}$$

4 Modal context restriction

In this section we introduce a family of language restrictions for modal logics called modal context restrictions. The idea was already presented in Dziubiński (2013). After introducing the concept, we propose two restrictions of this kind, called \({\mathbf{R}}_1\) and \({\mathbf{R}}_2\), as well as a refinement of one of them, called \({\mathbf{R}}_{1(c)}\), which will be studied in the remaining part of the paper.

We start by defining the notion of modal context restriction for a general language of multimodal logic.Footnote 3 First we need a notion of modal context of a formula within a formula. Consider a case of a formula appearing only once as a subformula of some given formula. This subformula is in a scope of a sequence (possibly empty) of modal operators. This sequence is the modal context of the subformula in the given formula. In general a formula may appear more than once as a subformula of another formula. In this case there is a set of sequence of modal operators, one for each appearance of the subformula. To get an intuition of this concept, assume a modal language over the set of two modal operators \({\Omega }= \{\Box _0,\Box _1\}\). Every formula has an expression-tree representation associated with it. Leafs of such a tree are labelled with propositional symbols and internal nodes are labelled with operators. Each subtree of the expression-tree corresponds to a subformula of the formula. For example, a formula

$$\begin{aligned} \varphi \equiv \Box _1 (p \wedge \Box _0 (q \vee \Box _1 p)) \end{aligned}$$

is represented by a tree in Fig. 1.

Fig. 1
figure 1

Expression tree for formula \(\Box _1 (p \wedge \Box _0 (q \vee \Box _1 p)\)

Expression-tree representation is unique (up to isomorphism). Modal context of a subformula of a formula is the set of sequences of modal operators on the paths from the root of the expression-tree representing the formula to the roots of expression-trees representing the subformula. In the example above, modal context of q in \(\varphi \) is \(\{\Box _1 \Box _0\}\), modal context of p in \(\varphi \) is \(\{\Box _1,\Box _1 \Box _0 \Box _1\}\), modal context of \(p \wedge \Box _0 (q \vee \Box _1 p)\) in \(\varphi \) is \(\{\Box _1\}\), and modal context of \(\varphi \) in \(\varphi \) is the empty sequence, \(\varepsilon \). If a formula is not a subformula of the other formula, then its modal context is \(\varnothing \).

Formally, let \({{\mathcal {L}}}[{\mathcal {P}},{\Omega }]\) be a multimodal language over a set of propositional variables \({\mathcal {P}}\) and a set of (unary) modal operators \({\Omega }\). Given \(\varphi \in {{\mathcal {L}}}[{\mathcal {P}},{\Omega }]\), let

$$\begin{aligned} \mathrm {Sub}(\varphi ) = \{\psi : \psi \text { is a subformula of } \varphi \} \end{aligned}$$

be the set of all subformulas of \(\varphi \).

Definition 4

(Modal context of a formula within a formula) Let \(\{\varphi ,\xi \} \in {{\mathcal {L}}}[{\mathcal {P}},{\Omega }]\). The modal context of formula \(\xi \) within formula \(\varphi \) is a set of finite sequences over \({\Omega }\), \(\mathrm {cont}\left( \xi ,\varphi \right) \subseteq {\Omega }^{*}\), defined inductively as follows:

  • \(\mathrm {cont}\left( \xi ,\varphi \right) = \varnothing \), if \(\xi \notin \mathrm {Sub}(\varphi )\),

  • \(\mathrm {cont}\left( \varphi ,\varphi \right) = \{\varepsilon \}\),

  • \(\mathrm {cont}\left( \xi ,\lnot \psi \right) = \mathrm {cont}\left( \xi ,\psi \right) \), if \(\xi \ne \lnot \psi \),

  • \(\mathrm {cont}\left( \xi ,\psi _1 \wedge \psi _2\right) = \mathrm {cont}\left( \xi ,\psi _1\right) \cup \mathrm {cont}\left( \xi ,\psi _2\right) \), if \(\xi \ne \psi _1 \wedge \psi _2\),

  • \(\mathrm {cont}\left( \xi ,\Box \psi \right) = \Box \cdot \mathrm {cont}\left( \xi ,\psi \right) \), if \(\xi \ne \Box \psi \) and \(\Box \in {\Omega }\),

where \(\Box \cdot S = \{\Box \cdot s : s \in S \}\), for \(\Box \in {\Omega }\) and \(S \subseteq {\Omega }^{*}\).

Definition 5

(Modal context of a formula) Modal context of a formula \(\varphi \in {{\mathcal {L}}}[{\mathcal {P}},{\Omega }]\), \(\mathrm {cont}\left( \varphi \right) \), is the sum of modal contexts of all its subformulas, that is

$$\begin{aligned} \mathrm {cont}\left( \varphi \right) = \bigcup _{\xi \in \mathrm {Sub}(\varphi )} \mathrm {cont}\left( \xi ,\varphi \right) . \end{aligned}$$

Having defined modal context of a formula, we are ready to define the notion of modal context restriction. The restriction is the set of sequences of modal operators that are allowed in modal contexts of formulas.

Definition 6

(Modal context restriction) A modal context restriction is a set of sequences over \({\Omega }\), \(R \subseteq {\Omega }^{*}\), constraining possible modal contexts of subformulas within formulas. We say that \(\varphi \in {{\mathcal {L}}}[{\mathcal {P}},{\Omega }]\) satisfies a modal context restriction \(R \subseteq {\Omega }^{*}\) iff \(\mathrm {cont}\left( \varphi \right) \subseteq R\).

To see an example of modal context restriction, consider the multi-modal language over the set of modal operators \({\Omega }= \{\Box _0, \Box _1\}\). Suppose that we would like to allow only the formulas were modal operators (if present) are alternating in modal contexts of subformulas. Such a restriction can be defined by a regular expression \(\left( (\varepsilon \cup \Box _1) \cdot (\Box _0 \Box _1)^{*}\right) \cup \left( (\varepsilon \cup \Box _0) \cdot (\Box _1 \Box _0)^{*}\right) \). Alternatively, we could define it by specifying that subsequences \(\Box _1 \Box _1\) and \(\Box _0 \Box _0\) are forbidden: \({\Omega }^{*} \setminus ({\Omega }^{*} \cdot (\Box _1 \Box _1 \cup \Box _0 \Box _0) \cdot {\Omega }^{*})\) (that is, as a complement of the language containing all the sequences that we want to forbid).

Notice that modal context restriction is a generalisation of restricting modal depth of formulas by a constant c. Indeed, we could define it as a set of sequences of modal operators of length at most c.

4.1 Restricting modal context of \({\textsc {Team}}{\textsc {Log}}\)

In this paper we study two modal context restrictions for the language of \({\textsc {Team}}{\textsc {Log}}\) that lead to PSPACE completeness of the satisfiability problem. The restrictions are presented below.

The first of the restrictions, \({\mathbf{R}}_1\), is motivated by the formula used to show that the satisfiability problem for basic multimodal logics with fixpoint modalities is EXPTIME hard, even if depth of formulas is bounded by 2 (c.f. Dziubiński et al. 2007). The modal context of this formula contains sequences of the form \([O]^{+}_{G}[O]_{j}\) with \(j \in G\). Restriction \({\mathbf{R}}_1\) forbids such sequences in modal contexts of formulas. It forbids also sequences \([O]^{+}_{G}[O]^{+}_{H}\) with \(G \cap H \ne \varnothing \). This is motivated by the fact that in \({\textsc {Team}}{\textsc {Log}}\) a formula \([O]^{+}_{G}[O]^{+}_{H}\psi \) is equivalent to \([O]^{+}_{G}\bigwedge _{j \in H} ([O]_{j}\psi \wedge [O]_{j}[O]^{+}_{H}\psi )\), the modal context of which contains the sequence \([O]^{+}_{G}[O]_{j}\) that we want to forbid. In the definition below these forbidden sequences are grouped in the sets \(S_O(G)\). Additionally the restriction forbids subsequences \([\mathrm {I}]^{+}_{G}[\mathrm {B}]_{j}[\mathrm {I}]_{j}\) with \(j \in G\). This is motivated by the fact that, due to mixed axioms BI4 and BI5, the formula \([\mathrm {I}]^{+}_{G}[\mathrm {B}]_{j}[\mathrm {I}]_{j}\varphi \) is equivalent to \([\mathrm {I}]^{+}_{G}[\mathrm {I}]_{j}\varphi \), which is the sequence that we initially wanted to forbid. For similar reasons sequences \([\mathrm {I}]^{+}_{G}[\mathrm {B}]^{+}_{H}[\mathrm {I}]_{j}\), \([\mathrm {I}]^{+}_{G}[\mathrm {B}]_{j}[\mathrm {I}]^{+}_{F}\) and \([\mathrm {I}]^{+}_{G}[\mathrm {B}]^{+}_{H}[\mathrm {I}]^{+}_{F}\) are forbidden for if \(j \in G \cap H\) or \(j \in G \cap F\) or \(F \cap G \cap H \ne \varnothing \), respectively. In the definition below these forbidden sequences are grouped in the sets \(S_{\mathrm {IB}(G)}\).

Definition 7

(Restriction \({\mathbf{R}}_1\)) Let

$$\begin{aligned} \mathbf {R}_\mathbf {1} = {\Omega }^{*} \setminus \left( {\Omega }^{*} \cdot \left[ \bigcup _{G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}} \left( S_{\mathrm {I}}(G) \cup S_{\mathrm {IB}}(G)\right) \cup \bigcup _{G \in \mathrm {P}({\mathcal {A}}), |G| \ge 2} S_{\mathrm {B}}(G) \right] \cdot {\Omega }^{*} \right) , \end{aligned}$$

where

$$\begin{aligned} S_{\mathrm {IB}}(G)&= \bigcup _{j \in G}[\mathrm {I}]^{+}_{G} \cdot \left\{ [\mathrm {B}]_{j}, [\mathrm {B}]^{+}_{\{j\}}\right\} ^* \cdot T_{\mathrm {B}}(\{j\}) \cdot T_{\mathrm {I}}(\{j\}), \text { and}\\ S_O(G)&= [O]^{+}_{G} \cdot T_O(G), \\ T_O(G)&= \{[O]_{j} : j \in G \} \cup \{[O]^{+}_{H} : H \in \mathrm {P}({\mathcal {A}}), H \cap G \ne \varnothing \}, \end{aligned}$$

for \(O \in \{\mathrm {B},\mathrm {I}\}\). The set of formulas in \({{\mathcal {L}}}^{\mathrm {T}}\) satisfying restriction \({\mathbf{R}}_1\) will be denoted by \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\).

Roughly speaking, restriction \({\mathbf{R}}_1\) forbids any operator \([O]_{j}\) or \([O]^{+}_{H}\), with \(O \in \{\mathrm {B},\mathrm {I}\}\) in scope of operator \([O]^{+}_{G}\), if \(j \in G\) or \(G \cap H \ne \varnothing \). The following formulas satisfy restriction \({\mathbf{R}}_1\):

$$\begin{aligned}&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}p,&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {G}]_{1}p. \end{aligned}$$

Indeed, the modal context of the first formula is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}}, [\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}\right\} \) and none of these sequences contains any of the forbidden ones. The modal context of the second formula is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}}, [\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {G}]_{1}\right\} \) and, again, none of these sequences contains any of the forbidden ones. The following formulas violate restriction \({\mathbf{R}}_1\):

$$\begin{aligned}&[\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}[\mathrm {I}]_{1}q,&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}p. \end{aligned}$$

The modal context of the first one is \(\left\{ \varepsilon , [\mathrm {I}]^{+}_{\{1,2\}}, [\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}, [\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}[\mathrm {I}]_{1}\right\} \), and it contains a forbidden sequence \([\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}[\mathrm {I}]_{1}\). The modal context of the second one is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}},[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}\right\} \), and it contains a forbidden sequence \([\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}\).

As we show in Sect. 6 (Proposition 1), the satisfiability problem for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is PSPACE hard, even if modal depth of formulas is bounded by 2. This result is shown using a formula which contains, in its modal context, sequences of the form \([\mathrm {B}]^{+}_{G}[\mathrm {G}]_{j}\) and \([\mathrm {B}]^{+}_{G}[\mathrm {I}]_{j}\), with \(j \in G\). Motivated by that, we study another modal context restriction, \({\mathbf{R}}_2\), which refines \({\mathbf{R}}_1\) by forbidding such sequences. Additionally, for the reasons similar to those explained in the case of restriction \({\mathbf{R}}_1\), subsequences forbidden by restriction \({\mathbf{R}}_2\) contain sequences of the form \([\mathrm {B}]^{+}_{G}[\mathrm {I}]^{+}_{H}\) with \(G \cap H \ne \varnothing \). Restriction \({\mathbf{R}}_2\), when combined with restricting modal depth of formulas by a constant, makes the \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem NPTIME solvable.

Definition 8

(Restriction \({\mathbf{R}}_2\)) Let

$$\begin{aligned} \mathbf {R}_\mathbf {2} = {\Omega }^{*} \setminus \left( {\Omega }^{*} \cdot \left[ \bigcup _{G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}} \left( S_{\mathrm {I}}(G) \cup S_{\mathrm {IB}}(G)\right) \cup \bigcup _{G \in \mathrm {P}({\mathcal {A}}), |G| \ge 2} {\tilde{S}}_{\mathrm {B}}(G) \right] \cdot {\Omega }^{*} \right) , \end{aligned}$$

where

$$\begin{aligned} {\tilde{S}}_{\mathrm {B}}(G)&= [\mathrm {B}]^{+}_{G} \cdot \left( \{[\mathrm {G}]_{j} : j \in G \} \cup \bigcup _{O \in \{\mathrm {B},\mathrm {I}\}} T_{O}(G)\right) \end{aligned}$$

and \(S_{\mathrm {IB}}\), \(S_{\mathrm {I}}\) and \(T_O\), for \(O \in \{\mathrm {B},\mathrm {I}\}\), are defined like in the case of restriction \({\mathbf{R}}_1\). The set of formulas in \({{\mathcal {L}}}^{\mathrm {T}}\) satisfying restriction \({\mathbf{R}}_2\) will be denoted by \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\).

Restriction \({\mathbf{R}}_2\) extends \({\mathbf{R}}_1\) by forbidding any operator \([O]_{j}\) or \([O]^{+}_{H}\), with \(O \in \{\mathrm {G},\mathrm {I}\}\), in the context of \([\mathrm {B}]^{+}_{G}\), if \(j \in G\) or \(H \cap G \ne \varnothing \). Thus any formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) satisfying restriction \({\mathbf{R}}_2\), satisfies restriction \({\mathbf{R}}_1\) as well, that is \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\subseteq {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\). Notice that if \(|{\mathcal {A}}| = 1\), then \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}= {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\).

The following formulas satisfy restriction \({\mathbf{R}}_2\).

$$\begin{aligned}&[\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1}p,&[\mathrm {B}]^{+}_{\{1,2\}}\left( q \vee [\mathrm {B}]_{3}p\right) . \end{aligned}$$

Indeed, the modal context of the first formula is \(\left\{ \varepsilon , [\mathrm {I}]^{+}_{\{1,2\}}, [\mathrm {I}]^{+}_{\{1,2\}}[\mathrm {B}]_{1} \right\} \) and none of these sequences contains any of the forbidden ones. The modal context of the second formula is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}}, [\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {B}]_{3} \right\} \) and, again, none of these sequences contains any of the forbidden ones.

The following formulas violate restriction \({\mathbf{R}}_2\)

$$\begin{aligned}&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}p,&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {G}]_{1}p. \end{aligned}$$

Modal context of the first formula is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}}, [\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}\right\} \), and it contains a forbidden sequence \([\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}\). Modal context of the second formula is \(\left\{ \varepsilon , [\mathrm {B}]^{+}_{\{1,2\}}, [\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {G}]_{1}\right\} \), and it contains a forbidden sequence \([\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {G}]_{1}\). Notice that the modal contexts above do not contain sequences forbidden by restriction \({\mathbf{R}}_1\), and so the formulas belong to \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\).

4.2 Restriction \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\)

As we show in Sect. 5, restriction \({\mathbf{R}}_2\) is too strong, as it forbids formulas such as collective intentions, important for specifying cooperating teams of agents (Dunin-Kęplicz and Verbrugge 2010). Therefore we consider a refinement of restriction \({\mathbf{R}}_1\) that, when combined with restricting modal depth, makes the \([{\textsc {Team}}{\textsc {Log}}\) satisfiability problem NPTIME solvable.

The restriction is motivated by the formula used in proof of Proposition 1. One of the conjuncts in this formula is a formula of the form \([\mathrm {B}]^{+}_{G}\psi \) where \(\psi \) is a propositional formula built of 2N atoms of the form \([\mathrm {I}]_{j}\xi \) (alternatively \([\mathrm {G}]_{j}\xi \)), with \(j \in G\). These atoms are used to implement a counter that can enforce a path of length \({\mathcal {O}}(2^{N})\) in the model for the formula. To prevent such a construction, restriction \({\mathbf{R}}_{1(c)}\) bounds, by a constant c, the number of subformulas of the form \([\mathrm {I}]_{j}\xi \), \([\mathrm {G}]_{j}\xi \) and \([\mathrm {I}]^{+}_{H}\xi \), within the direct context of modal operators \([\mathrm {B}]^{+}_{G}\) with \(j \in G\) or \(H \cap G \ne \varnothing \). That is, whenever a formula from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) has a subformula which violates modal context restriction \({\mathbf{R}}_2\), then this formula must satisfy this additional restriction.

To define the restriction, we need to define the set of subformulas of a formula taken with respect to propositional operators only. Let \(\mathrm {PT}(\varphi )\) be defined inductively as follows:

  1. 1.

    \(\mathrm {PT}(p) = \{p\}\), where \(p \in {\mathcal {P}}\),

  2. 2.

    \(\mathrm {PT}(\lnot \psi ) = \{\lnot \psi \} \cup \mathrm {PT}(\psi )\),

  3. 3.

    \(\mathrm {PT}(\psi _1 \wedge \psi _2) = \mathrm {PT}(\psi _1) \cup \mathrm {PT}(\psi _2)\).

  4. 4.

    \(\mathrm {PT}(\Box \psi ) = \{\Box \psi \}\), where \(\Box \in {\Omega }^{\mathrm {T}}\).

Given a formula \(\Box \varphi \), the set \(\mathrm {PT}(\varphi )\) contains the subformulas in the direct context of a modal operator \(\Box \). So for example in the case of \(\Box (p \wedge \lnot q \vee \Box (\Box p \wedge r) \vee \Box r)\), \(\mathrm {PT}(p \wedge \lnot q \vee \Box (\Box p \wedge r) \vee \Box r ) = \{p, q, \lnot q, \Box (\Box p \wedge r), \Box r\}\) contains the subformulas in the direct context of the most external operator \(\Box \).

The restriction is defined as follows.

Definition 9

(Restriction \({\mathbf{R}}_{1(c)}\)) Let \(c \ge 0\). A formula \(\varphi \) satisfies the restriction \({\mathbf{R}}_{1(c)}\) if \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) and either \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) or one of the following holds:

  • \(\varphi \) is of the form \(\lnot \psi \) and \(\psi \) satisfies restriction \({\mathbf{R}}_{1(c)}\),

  • \(\varphi \) is of the form \(\psi _1 \wedge \psi _2\) and \(\psi _1\) and \(\psi _2\) satisfy restriction \({\mathbf{R}}_{1(c)}\),

  • \(\varphi \) is of the form \([O]_{j}\psi \), with \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\) and \(j \in {\mathcal {A}}\), and \(\psi \) satisfies restriction \({\mathbf{R}}_{1(c)}\),

  • \(\varphi \) is of the form \([\mathrm {I}]^{+}_{G}\psi \), with \(G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\), and \(\psi \) satisfies restriction \({\mathbf{R}}_{1(c)}\),

  • \(\varphi \) is of the form \([\mathrm {B}]^{+}_{G}\psi \), with \(G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\), \(\psi \) satisfies restriction \({\mathbf{R}}_{1(c)}\) and \(\left| \{[O]_{j}\xi : [O]_{j}\xi \in \lnot \mathrm {PT}(\psi ) \text { and } j\in G \} \cup \{[\mathrm {I}]^{+}_{H}\xi : [\mathrm {I}]^{+}_{H}\xi \in \lnot \mathrm {PT}(\psi ) \text { and } H \cap G \ne \varnothing \}\right| \le c\).

The set of formulas in \({{\mathcal {L}}}^{\mathrm {T}}\) satisfying restriction \({\mathbf{R}}_{1(c)}\) will be denoted by \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\).

The following formulas satisfy restriction \({\mathbf{R}}_{1(1)}\).

$$\begin{aligned}&[\mathrm {B}]^{+}_{\{1,2\}}[\mathrm {I}]^{+}_{\{1,2\}}p,&[\mathrm {B}]^{+}_{\{1,2\}}\left( [\mathrm {I}]_{2}p \vee q\right) . \end{aligned}$$

Firstly, both the formulas satisfy restriction \({\mathbf{R}}_1\). The set of formulas in the direct context of \([\mathrm {B}]^{+}_{\{1,2\}}\) in the first formula, \(\mathrm {PT}([\mathrm {I}]^{+}_{\{1,2\}}p) = \{[\mathrm {I}]^{+}_{\{1,2\}}p\}\) and since this set contains only one formula, restriction \({\mathbf{R}}_{1(1)}\) is satisfied. The set of formulas in the direct context of \([\mathrm {B}]^{+}_{\{1,2\}}\) in the second formula, \(\mathrm {PT}([\mathrm {I}]_{2}p \vee q) = \{[\mathrm {I}]_{2}p,q\}\) and since this set contains only one formula of the form \([\mathrm {I}]_{j}\xi \), \([\mathrm {G}]_{j}\xi \) or \([\mathrm {I}]^{+}_{H}\xi \) with \(j \in \{1,2\}\) or \(H \cap \{1,2\} \ne \varnothing \), namely \([\mathrm {I}]_{2}p\), so restriction \({\mathbf{R}}_{1(1)}\) is satisfied as well.

The following formulas violate restriction \({\mathbf{R}}_{1(1)}\) and satisfy restriction \({\mathbf{R}}_{1(2)}\)

$$\begin{aligned}&[\mathrm {B}]^{+}_{\{1,2\}}\left( [\mathrm {I}]_{1}p \wedge [\mathrm {I}]_{2}q\right) ,&[\mathrm {B}]^{+}_{\{1,2\}}\left( [\mathrm {G}]_{2}p \vee [\mathrm {I}]^{+}_{\{2,3\}}q\right) . \end{aligned}$$

Both the formulas satisfy restriction \({\mathbf{R}}_1\). The set of formulas in the direct context of \([\mathrm {B}]^{+}_{\{1,2\}}\) in the first formula, \(\mathrm {PT}([\mathrm {I}]_{1}p \wedge [\mathrm {I}]_{2}q) = \{[\mathrm {I}]_{1}p,[\mathrm {I}]_{2}q)\) and since this set contains two formulas of the form \([\mathrm {I}]_{j}\xi \) with \(j \in \{1,2\}\), so restriction \({\mathbf{R}}_{1(c)}\) with \(c = 1\) is violated, but it is satisfied with \(c = 2\). The set of formulas in the direct context of \([\mathrm {B}]^{+}_{\{1,2\}}\) in the second formula, \(\mathrm {PT}([\mathrm {G}]_{2}p \vee [\mathrm {I}]^{+}_{\{2,3\}}q) = \{[\mathrm {G}]_{2}p,[\mathrm {I}]^{+}_{\{2,3\}}q\}\) and since this set contains two formulas, \([\mathrm {G}]_{j}p\) with \(j \in \{1,2\}\), and \([\mathrm {I}]^{+}_{\{H\}}q\) with \(H \cap \{1,2\} \ne \varnothing \), so restriction \({\mathbf{R}}_{1(c)}\) with \(c = 1\) is violated, but it is satisfied with \(c = 2\).

5 Discussion

Let us start the discussion of the two restrictions with formulas specifying beliefs of groups of agents. When interpreted in the context of BDI agents, \({\mathbf{R}}_1\) can be seen as forbidding common introspection of beliefs within a group of agents. In other words, it forbids any formula of the form \([\mathrm {B}]^{+}_{G}\varphi \) where \(\varphi \) contains, within the scope of propositional operators, any formulas referring to beliefs of agents from G. For example the following formula specifies that group G commonly beliefs that some agent j believes that \(\varphi \) holds

$$\begin{aligned}{}[\mathrm {B}]^{+}_{G}[\mathrm {B}]_{j}\varphi \end{aligned}$$

If \(j \notin G\) (and if \(\varphi \) satisfies \({\mathbf{R}}_1\)), then this formula satisfies restriction \({\mathbf{R}}_1\). However, if \(j \in G\), then the formula does not satisfy the restriction. Similarly, the following formula specifies that group G commonly believes that some other group H commonly believes that \(\varphi \) holds

$$\begin{aligned}{}[\mathrm {B}]^{+}_{G}[\mathrm {B}]^{+}_{H}\varphi \end{aligned}$$

If \(G \cap H = \varnothing \) (and if \(\varphi \) satisfies \({\mathbf{R}}_1\)), then this formula satisfies restriction \({\mathbf{R}}_1\) and it violates it if \(G \cap H \ne \varnothing \). The second case could be seen as a consequence of the first one, given that the formula \([\mathrm {B}]^{+}_{G}[\mathrm {B}]^{+}_{H}\varphi \leftrightarrow [\mathrm {B}]^{+}_{G}\left( [\mathrm {B}]^{+}_{H}\varphi \wedge [\mathrm {B}]_{j}\varphi \right) \) is provable in \({\textsc {Team}}{\textsc {Log}}.\) Summarizing, as long as the objects of common beliefs of a group of agents are ‘external’ for that group, i.e. do not concern beliefs of the agents in the group, restriction \({\mathbf{R}}_1\) is not violated.

Let us look at restriction \({\mathbf{R}}_2\) now. It forbids, in addition to what is forbidden by \({\mathbf{R}}_1\), common introspection of goals and intentions within a group of agents. Hence the following formula, specifying that group G of agents commonly believes that agent j has intention \(\varphi \):

$$\begin{aligned}{}[\mathrm {B}]^{+}_{G}[\mathrm {I}]_{j}\varphi \end{aligned}$$

is allowed by \({\mathbf{R}}_2\) if \(j \notin G\) (and \(\varphi \) satisfies \({\mathbf{R}}_2\)) and is forbidden otherwise. Similarly with the formula \([\mathrm {B}]^{+}_{G}[\mathrm {G}]_{j}\varphi \), specifying that group G of agents has common belief that agent j has goal \(\varphi \), as well as with the formula

$$\begin{aligned}{}[\mathrm {B}]^{+}_{G}[\mathrm {I}]^{+}_{H}\varphi , \end{aligned}$$

specifying that group G of agents has common belief that group H of agents has mutual intention \(\varphi \). In this case the formula satisfies condition \({\mathbf{R}}_2\) as long as \(G \cap H = \varnothing \) (and \(\varphi \) satisfies \({\mathbf{R}}_2\)) and violates it otherwise.

Summarizing, similarly like in the case of restriction \({\mathbf{R}}_1\), restriction \({\mathbf{R}}_2\) is not violated as long as the objects of common beliefs of a group of agents are ‘external’ for that group, where external means this time that it does not concern any informational or motivational attitudes of the agents in the group.

Interpretation of the restrictions in the case of mutual intentions of groups of agents is similar to that of common beliefs. There is one addition, however. Both restrictions, \({\mathbf{R}}_1\) and \({\mathbf{R}}_2\), forbid formulas of the form

$$\begin{aligned}{}[\mathrm {I}]^{+}_{G}[\mathrm {B}]_{j}[\mathrm {I}]_{j}\varphi , \end{aligned}$$

with \(j \in G\), specifying that group G of agents has mutual intention that agent j believes that it has intention \(\varphi \). This is needed because, due to awareness axioms, formulas of this form are equivalent to formulas of the form \([\mathrm {I}]^{+}_{G}[\mathrm {I}]_{j}\varphi \) (which are forbidden like the analogous ones for common beliefs). For similar reasons formulas \([\mathrm {I}]^{+}_{G}[\mathrm {B}]^{+}_{H}[\mathrm {I}]_{j}\varphi \), \([\mathrm {I}]^{+}_{G}[\mathrm {B}]_{j}[\mathrm {I}]^{+}_{F}\varphi \) and \([\mathrm {I}]^{+}_{G}[\mathrm {B}]^{+}_{H}[\mathrm {I}]^{+}_{F}\varphi \) are forbidden by \({\mathbf{R}}_1\) and \({\mathbf{R}}_2\), if \(j \in G \cap H\) or \(j \in G \cap F\) or \(F \cap G \cap H \ne \varnothing \), respectively.

Now let us turn to formulas specifying important properties of multiagent systems. For example one of the fundamental notions underlying teamwork is that of collective intention, defined as follows (Dunin-Kęplicz and Verbrugge 2010):

$$\begin{aligned} \mathrm{C \text{- } INT}_{G}\left( \varphi \right) \equiv [\mathrm {I}]^{+}_{G}\varphi \wedge [\mathrm {B}]^{+}_{G}[\mathrm {I}]^{+}_{G}\varphi \end{aligned}$$

does not satisfy \({\mathbf{R}}_2\), while it satisfies \({\mathbf{R}}_1\) (as long as it is satisfied by formulas \(\varphi \) and \([\mathrm {I}]^{+}_{G}\varphi \)).

Another notion fundamental for specifying teamwork is (bilateral) social commitments between agents in a team. For example social commitment of agent i towards agent j with respect to some action \(\alpha \) is defined as follows (Dunin-Kęplicz and Verbrugge 2010):Footnote 4

$$\begin{aligned} \mathrm{COMM}\left( i, j, \alpha \right) \equiv [\mathrm {I}]_{j}\alpha \wedge [\mathrm {G}]_{j}done(i,\alpha ) \wedge [\mathrm {B}]^{+}_{\{i,j\}}\left( [\mathrm {I}]_{j}done(i,\alpha ) \wedge [\mathrm {G}]_{j}done(i,\alpha )\right) . \end{aligned}$$

This formula does not satisfy \({\mathbf{R}}_2\), because of presence of formulas \([\mathrm {I}]_{j}done(i,a)\) and \([\mathrm {G}]_{j}done(i,a)\) in the direct context of operator \([\mathrm {B}]^{+}_{\{i,j\}}\). However, it satisfies restriction \({\mathbf{R}}_1\). Notice also that the formula satisfies \({\mathbf{R}}_1(c)\) with \(c = 2\).

The third important notion is collective commitment. Several variants of it can be defined, corresponding to different strength of motivational and informational interdependencies within a team of agents. In this case even restriction \({\mathbf{R}}_1\) can be too strong. Consider, for example, the strongest two forms of collective commitment, robust and strong collective commitment:

$$\begin{aligned} \mathrm{R \text{-} COMM}_{G,P} \left( \varphi \right) \equiv\,&\mathrm{C\text{-}INT}_{G}\left( \varphi \right) \wedge constitutes(P,\varphi ) \wedge [\mathrm {B}]^{+}_{G}constitutes(P,\varphi ) \\& \wedge \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} [\mathrm {B}]^{+}_{G}\mathrm{COMM}\left( i, j, \alpha \right) \\ \mathrm{S \text{-}COMM}_{G,P} \left( \varphi \right) \equiv\,&\mathrm{C\text{-}INT}_{G}\left( \varphi \right) \wedge constitutes(P,\varphi ) \wedge [\mathrm {B}]^{+}_{G}constitutes(P,\varphi ) \\&\wedge[\mathrm {B}]^{+}_{G}\left( \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} \mathrm{COMM}\left( i, j, \alpha \right) \right) \end{aligned}$$

In both cases the last component, expressing team awareness about the distribution of social commitments within the group involves common beliefs within the group about beliefs of agents from this group (which are contained in the definition of \(\mathrm{COMM}\left( i, j, \alpha \right) \)). Hence, both definitions of commitments do not satisfy the restriction \({\mathbf{R}}_1\). The formula defining the robust commitment is not a problem, because it can be replaced by an equivalent formula that satisfies restriction \({\mathbf{R}}_1\). Let responsibility of agent i towards agent j with respect to action \(\alpha \) be defined as follows:

$$\begin{aligned} \mathrm{RESP}\left( i, j, \alpha \right) \equiv [\mathrm {I}]_{j}\alpha \wedge [\mathrm {G}]_{j}done(i,a) \end{aligned}$$

Notice that social commitment of one agent towards another is responsibility plus awareness about this responsibility. Consider now the following definition of robust commitment:

$$\begin{aligned} \mathrm{R} \text{-} {\text{COMM}}^{\prime }_{G,P} \left( \varphi \right) \equiv\,&\mathrm{C} \text{-}\text{INT}_{G}\left( \varphi \right) \wedge constitutes(P,\varphi ) \wedge [\mathrm {B}]^{+}_{G}constitutes(P,\varphi ) \\& \wedge \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} [\mathrm {B}]^{+}_{G}\mathrm{RESP}\left( i, j, \alpha \right) \end{aligned}$$

It can be easily seen that \(\mathrm{R \text{-} COMM}_{G,P} \left( \varphi \right) \) is equivalent to \(\mathrm{R} \text{-} \text{COMM}^{\prime }_{G,P} \left( \varphi \right) \), as the formula \([\mathrm {B}]^{+}_{G}\left( \psi \wedge [\mathrm {B}]_{j}\psi \right) \leftrightarrow [\mathrm {B}]^{+}_{G}\psi \) is provable in \({\textsc {Team}}{\textsc {Log}}\) for any \(\psi \), \(G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\) and \(j \in G\).

In the case of strong commitment one could deal with the problem by lowering the level of awareness about the distribution of the social commitments, replacing subformula \([\mathrm {B}]^{+}_{G}\left( \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} \mathrm{COMM}\left( i, j, \alpha \right) \right) \) with \([\mathrm {B}]_{G}\left( \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} \mathrm{COMM}\left( i, j, \alpha \right) \right) \).Footnote 5 Another possibility is to consider an alternative definition of bilateral commitment, where social commitments are replaced by bilateral responsibilities

$$\begin{aligned} \mathrm{S} \text{-} \text{COMM}^{\prime }_{G,P} \left( \varphi \right) \leftrightarrow&\mathrm{C} \text{-} \text{INT}_{G}\left( \varphi \right) \wedge constitutes(P,\varphi ) \wedge [\mathrm {B}]^{+}_{G}constitutes(P,\varphi ) \\& \wedge [\mathrm {B}]^{+}_{G}\left( \bigwedge _{\alpha \in P} \bigvee _{i,j \in G} \mathrm{RESP}\left( i, j, \alpha \right) \right) \\&\wedge\bigwedge _{\alpha \in P} \bigwedge _{i,j \in G} \mathrm{RESP}\left( i, j, \alpha \right) \rightarrow [\mathrm {B}]^{+}_{\{i,j\}}\left( \mathrm{RESP}\left( i, j, \alpha \right) \right) \end{aligned}$$

In this version of the notion of strong commitment the group is fully aware of bilateral responsibilities within the group with regard to the actions of the plan P for achieving the goal \(\varphi \), but the awareness about existence of bilateral awareness about these responsibilities is not required.

As the examples above show, restriction \({\mathbf{R}}_1\) is sufficiently weak to allow for expressing the key properties of cooperating teams of agents, such as collective intentions and social (bilateral commitments). What the restriction forbids, are the formulas where groups of agents have common beliefs about beliefs of agents from the group or have mutual intentions with regards to intentions of agents from the group. Essentially, the objects of common beliefs must be ‘external’ with respect to the group. In particular, they may be beliefs of agents from outside the group about the beliefs of agents in the group. Similarly in the case of mutual intentions. This restriction may be a problem when some kinds of collective commitments are concerned. The problem there is the awareness part, where common beliefs about social (bilateral) commitments within a group are expressed. It can be overcome by restating the formula expressing the collective commitment (like in the case of robust commitment). If this is impossible (like in the case of social commitment), alternative forms of collective commitments, where social commitments are replaced with responsibilities, can be considered. In these variants, the group as a whole holds a common belief about all the bilateral responsibilities, but it does not hold common belief about the bilateral awareness about these responsibilities. We would like to note that other, weaker, forms of collective commitments, like the team commitment and the distributed commitment (c.f. Dunin-Kęplicz and Verbrugge 2010) are expressible with restriction \({\mathbf{R}}_1\).

Restriction \({\mathbf{R}}_2\) is too strong to allow for expressing even collective intentions. However it is still of use. Methodologies of agent oriented modelling and design, like the one proposed by Kinny et al. (1996); Kinny and Georgeff (1997); Kinny (1998), often divide the process of agents modelling by separating construction of belief model, goal model and plan model. Similarly, in specification of multiagent systems using formalisms like \({\textsc {Team}}{\textsc {Log}},\) separate parts could be distinguished, where purely informational and purely motivational aspects of individual agents and groups of agents are specified and parts where interrelations between these parts are specified. In such cases restriction \({\mathbf{R}}_2\) can be applied to the purely informational or purely motivational parts, while restriction \({\mathbf{R}}_1\) can be applied to the mixed parts.

In the discussion above we restricted attention to \({\textsc {Team}}{\textsc {Log}}\) formalism. However, similar formulas, where agents in a group hold common beliefs about intentions or goals of group members appear also in other theories of teamwork. See for example (Levesque et al. 1990; Wooldridge and Jennings 1999; Aldewereld et al. 2004).

6 Complexity of the satisfiability problem

In this section we study the complexity of checking the satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) and \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\). We focus on presenting the algorithms and general ideas. For that reason we moved most of the proofs to the Appendix.

For checking the satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) we will use the method based on pre-tableau construction presented in Halpern and Moses (1992). However, adopting a similar algorithm for \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) would not work. This is because, as we show below, formulas of \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) may require an exponentially deep model with respect to the size of input formula, while all the algorithms based on the pre-tableau method perform a depth first search constructing sequences of nodes that constitute the tree-like structure of the pre-tableau for a given input.

Proposition 1

Let \(|{\mathcal {A}}| \ge 2\). Then there exists a satisfiable formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) such that any \({\textsc {Team}}{\textsc {Log}}\) model \({\mathcal {M}}\) in which it is satisfied contains a sequence of pairwise different worlds of length exponential with respect to \(|\varphi |\).

Proposition 1 is shown by writing a formula that implements a binary counter. That is, the formula enforces a path in its model, where formulas of the form \([\mathrm {I}]_{j}\xi _1,\ldots ,[\mathrm {I}]_{\xi _N}\) (alternatively \([\mathrm {G}]_{j}\xi _1,\ldots ,[\mathrm {G}]_{\xi _N}\)) have different valuations, implementing a counter over N bits.

6.1 Complexity of \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\)

Algorithm 1 for checking satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) presented in this section is based on the standard tableau method for modal logics. Our presentation follows (Halpern and Moses 1992) in the general methodology, which can be summarized as consisting of the following steps:

  1. 1.

    Define the notion of modal tableau for the logic in question. A modal tableau is a Kripke frame with worlds labelled with sets of formulas and accessibility relations satisfying additional properties associated with axioms generating the logic considered.

  2. 2.

    Show that any formula of the logic is satisfiable iff there is a tableau for it.

  3. 3.

    Give an algorithm for checking the satisfiability of a formula. The algorithm constructs a tree-like structure called a pre-tableau which forms a basis for the tableau for the formula.

  4. 4.

    Show that the algorithm has a termination property and is valid.

  5. 5.

    Analyse the computational complexity of the algorithm.

In steps 1 and 2 we follow (Dziubiński et al. 2007) where modal tableau for \({\textsc {Team}}{\textsc {Log}}\) is defined. The main difficulty are steps 3 and 4, particularly the termination property. In step 3 we extend the algorithm from (Dziubiński et al. 2007) so that fixpoint modalities can be dealt with. For step 4 we find the property of sets of formulas processed by the algorithm that ‘decreases’ during execution of the algorithm. In these steps we extend the ideas used in Dziubiński (2013) for standard systems of multimodal logics enriched with fixpoint modalities.

The notion of modal tableau is based on the notion of model graph, which we define below.

Definition 10

(Model graph) A model graph \({\mathcal {T}}\) is a tuple \({\mathcal {T}}= \left( W,\left\{ O_j : [O]_{j} \in {\Omega }^{\mathrm {ind}}\right\} ,L\right) \), where W and \(O_j\) are defined like in a Kripke frame and L is a labelling function associating with each state \(w \in W\) a set L(w) of formulas.

The modal tableau for \({\textsc {Team}}{\textsc {Log}}\) is a model graph with labelling sets of formulas satisfying additional properties, which we define below. Firstly, all labels of states are closed propositional tableaux. We define these notions below.

Definition 11

(Closed set of formulas) A set of formulas \(\Phi \) is closed if it satisfies the following condition, for all \(G \subseteq {\mathcal {A}}\) and \(O \in \{\mathrm {B},\mathrm {I}\}\):

Cl:

If \([O]^{+}_{G}\varphi \in \Phi \), then \(\{[O]_{j}[O]^{+}_{G}\varphi , [O]_{j}\varphi : j \in G\} \subseteq \Phi \),

Given a formula \(\varphi \), we will use \(\mathrm {Cl}(\varphi )\) to denote the smallest closed set of formulas containing \(\varphi \). Similarly, given a set of formulas \(\Phi \) we will use \(\mathrm {Cl}(\Phi )\) to denote the smallest closed set of formulas having \(\Phi \) as a subset.

Definition 12

(Propositional tableau) A propositional tableau is a set \({\mathcal {T}}\) of formulas such that \({\mathcal {T}}\) is not trivially inconsistentFootnote 6 and:

  1. 1.

    If \(\lnot \lnot \psi \in {\mathcal {T}}\) then \(\psi \in {\mathcal {T}}\).

  2. 2.

    If \(\varphi \wedge \psi \in {\mathcal {T}}\) then \(\varphi \in {\mathcal {T}}\) and \(\psi \in {\mathcal {T}}\).

  3. 3.

    If \(\lnot (\varphi \wedge \psi ) \in {\mathcal {T}}\) then either \(\{\sim \!\varphi ,\psi \} \subseteq T\) or \(\{\varphi ,\sim \!\psi \} \subseteq {\mathcal {T}}\) or \(\{\sim \!\varphi ,\sim \!\psi \} \subseteq {\mathcal {T}}\).

A propositional tableau for a formula \(\varphi \) is a minimal propositional tableau \({\mathcal {T}}\) such that \(\varphi \in {\mathcal {T}}\). It is easy to see that every closed propositional tableau for \(\varphi \) is a maximal consistent subset of \(\lnot \mathrm {Cl}(\mathrm {PT}(\varphi ))\).Footnote 7 Notice that, by definition, a propositional tableau cannot be trivially inconsistent.

figure a

A modal tableau for \({\textsc {Team}}{\textsc {Log}},\) called a \({\textsc {Team}}{\textsc {Log}}\) tableau is defined as follows.

Definition 13

(\({\textsc {Team}}{\textsc {Log}}\) tableau) A modal tableau is a model graph \({\mathcal {T}}= \left( W,\left\{ O_j : [O]_{j} \in {\Omega }^{\mathrm {ind}}\right\} ,L\right) \) such that for all \(w \in W\), L(w) is a closed propositional tableau. Moreover, for any \([O]_{j} \in {\Omega }^{\mathrm {T}}\) and any \([O]^{+}_{G} \in {\Omega }^{\mathrm {T}}\) the following conditions are satisfied, for all \(w \in W\):

T1:

If \([O]_{j}\varphi \in L(w)\) and \(v \in O_j(w)\), then \(\varphi \in L(v)\). If \([O]^{+}_{G}\varphi \in L(w)\) and \(v \in O^{+}_G(w)\), then \(\varphi \in L(v)\).

T2:

If \(\lnot [O]_{j}\varphi \in L(w)\), then there exists \(v \in O_j(w)\) such that \(\sim \!\varphi \in L(v)\). If \(\lnot [O]^{+}_{G}\varphi \in L(w)\), then there exists \(v \in O^{+}_G(w)\) such that \(\sim \!\varphi \in L(v)\).

The following conditions are satisfied if \([O]_{j} \in {\Omega }^{\mathrm {T}}\) is associated with additional axioms from D5 (c.f. Halpern and Moses 1992):

  • If \([O]_{j}\) is associated with axiom D, then the following condition is satisfied, for any \(w \in W\):

    • TD If \([O]_{j}\varphi \in L(w)\), then either \(\varphi \in L(w)\) or \(O_j(w) \ne \varnothing \).

  • If \([O]_{j}\) is associated with axiom 4, then the following condition is satisfied, for any \(w \in W\):

    • T4 If \(v \in O_j(w)\) and \([O]_{j}\varphi \in L(w)\), then \([O]_{j}\varphi \in L(v)\).

  • If \([O]_{j}\) is associated with axiom 5, then the following condition is satisfied, for any \(w \in W\):

    • T5 If \(v \in O_j(w)\) and \([O]_{j}\varphi \in L(v)\), then \([O]_{j}\varphi \in L(w)\).

The following additional conditions, associated with axioms B O4, B O5 (with \(O \in \{\mathrm {G},\mathrm {I}\}\)) and IG are satisfied for all \(j \in {\mathcal {A}}\) and \(w \in W\):

  • TB O4 If \(v \in \mathrm {B}_j(w)\) and \([O]_{j}\varphi \in L(w)\), then \([O]_{j}\varphi \in L(v)\).

  • TB O5 If \(v \in \mathrm {B}_j(w)\) and \([O]_{j}\varphi \in L(v)\), then \([O]_{j}\varphi \in L(w)\).

  • TIG If \(v \in \mathrm {G}_j(w)\) and \([\mathrm {I}]_{j}\varphi \in L(w)\), then \(\varphi \in L(v)\).

A \({\textsc {Team}}{\textsc {Log}}\) tableau is a modal tableau satisfying conditions T1 and T2 (for all \([O]_{j}\) with \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\) and \(j \in {\mathcal {A}}\)), condition TD (for all \([O]_{j}\) with \(O \in \{\mathrm {B},\mathrm {I}\}\) and \(j \in {\mathcal {A}}\)), conditions T4 and T5 (for all \([\mathrm {B}]_{j}\) with \(j \in {\mathcal {A}}\)) and conditions TBG4, TBG5, TBI4, TBI5 and TIG. Given a formula \(\varphi \), we say that \({\mathcal {T}}\) is a tableau for \(\varphi \) if there exists a state \(w \in W\) such that \(\varphi \in L(w)\).

The following proposition links existence of \({\textsc {Team}}{\textsc {Log}}\) tableau for a formula with its satisfiability.

Proposition 2

A formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) is satisfiable iff there is a \({\textsc {Team}}{\textsc {Log}}\) tableau for \(\varphi \).

6.1.1 Algorithm for \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\)

Algorithm 1 tries to construct a pre-tableau – a tree-like structure that forms the basis for a \({\textsc {Team}}{\textsc {Log}}\) tableau for an input formula \(\varphi \). A pre-tableau consists of nodes connected with a successor relation. Each node can have zero or more successors and each of them has zero or one predecessor. There is at most one node in the pre-tableau that has no predecessors and it is called the root. Each node is labelled with a set of formulas. The root is labelled with the set containing the input formula only. The nodes of a pre-tableau can be divided into two groups: internal nodes and states. Successors of states correspond to accessibility relations and are created for formulas of the form \(\lnot [O]_{j}\psi \) in the labels of states and, in the case of modal operators with which axiom D is associated, for formulas of the form \([O]_{j}\psi \) in the labels of states. To construct a \({\textsc {Team}}{\textsc {Log}}\) tableau based on a given pre-tableau, a subset of states of the pre-tableau is selected and the accessibility relations are constructed on the basis of successor relations for states. Labels of states must be closed propositional tableaux satisfying additional requirements given below. Internal nodes correspond to subsequent steps of constructing labels of states. For the more detailed explanation of the notion of pre-tableaux see (Halpern and Moses 1992).

Now we turn to defining additional properties that will be satisfied by states of pre-tableaux constructed by Algorithm 1. Firstly, the notion of \([O]^{+}\)-expanded set of formulas, for \(O \in \{\mathrm {B},\mathrm {I}\}\), related to fixpoint modalities is needed.

Definition 14

(\([O]^{+}\)-expanded set of formulas) A set of formulas \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}\) is \([O]^{+}\)-expanded, with \(O \in \{\mathrm {B},\mathrm {I}\}\) and \(G \subseteq {\mathcal {A}}\), if the following condition is satisfied:

CE:

If \(\lnot [O]^{+}_{G}\psi \in \Phi \), then for all \(j \in G\), \(\{[O]_{j}\psi , [O]_{j}[O]^{+}_{G}\psi \} \subseteq \lnot \Phi \) and there exists \(j \in G\) such that either \(\lnot [O]_{j}\psi \in \Phi \) or \(\lnot [O]_{j}[O]^{+}_{G}\psi \in \Phi \).

Secondly, the notion of \([\mathrm {B}]\)-expanded set of formulas is needed

Definition 15

(\([\mathrm {B}]\)-expanded tableau) A \([\mathrm {B}]\)-expanded tableau is a \([\mathrm {B}]^{+}\)-expanded and \([\mathrm {I}]^{+}\)-expanded closed propositional tableau \({\mathcal {T}}\) such that for all \(j \in {\mathcal {A}}\):

  1. 1.

    If \([\mathrm {B}]_{j}\varphi \in \lnot {\mathcal {T}}\) and \([O]_{j}\psi \in \lnot \mathrm {PT}(\varphi )\), then \([O]_{j}\psi \in \lnot {\mathcal {T}}\), where \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\).

  2. 2.

    If \([\mathrm {B}]_{j}\varphi \in \lnot {\mathcal {T}}\) and \([O]^{+}_{G}\psi \in \lnot \mathrm {PT}(\varphi )\) with \(j \in G\), then \([O]_{j}\psi \in \lnot {\mathcal {T}}\) and \([O]_{j}[O]^{+}_{G}\psi \in \lnot {\mathcal {T}}\), where \(O \in \{\mathrm {B},\mathrm {I}\}\).

A \([\mathrm {B}]\)-expanded tableau for a formula \(\varphi \) is a minimal \([\mathrm {B}]\)-expanded tableau \({\mathcal {T}}\) such that \(\varphi \in {\mathcal {T}}\). Given a formula \(\varphi \) we will use \(\mathrm {BT}(\varphi )\) to denote the union of all \([\mathrm {B}]\)-expanded tableaux for \(\varphi \). Notice that any \([\mathrm {B}]\)-expanded tableau is a maximal consistent subset of \(\lnot \mathrm {BT}(\varphi )\). Nodes of a pre-tableau constructed for an input formula \(\varphi \) are labelled with subsets of \(\lnot \mathrm {Cl}(\varphi )\). States are nodes with labels being \([\mathrm {B}]\)-expanded tableaux. This is required because of axioms 5, BG5 and BI5, associated with operators \([\mathrm {B}]_{j}\), and is needed to prevent construction of too long paths in the pre-tableaux.Footnote 8

Algorithm 1 consists of two stages: the stage of pre-tableau construction and the stage of marking nodes. In the first stage the algorithm attempts to construct a pre-tableau based on the input formula. This stage consists of two general steps: the step of state construction and the step of state successors creation. In the step of state construction labels of internal nodes are properly extended with new formulas, resulting in new successor nodes, until a node which is a state is obtained. This step is described in Procedure 2.

figure b

The step of state successors creation is described in three Procedures 3, 4 and 5, creating state successors associated with operators \([\mathrm {B}]_{j}\), \([\mathrm {G}]_{j}\) and \([\mathrm {I}]_{j}\), respectively. We will call these successors \(O_j\)-successors, for \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\). A node which is an \(O_j\)-successor for some \(j \in {\mathcal {A}}\) will be called an O-successor. Additionally, a successor created for a formula \(\xi \) will be called a \(\xi \)-successor. Similar notions can be defined for subsequent states on paths of the pre-tableau. State t is called a \(O_j\)-Successor of state s if t is a descendant of s in the pre-tableau, there are no states between s and t and t is an \(O_j\)-successor. The relations of O-successor and \(\xi \)-Successor are defined analogously. In the presentation we will also refer to similar notions of -predecessors and -Predecessors. The following set of formulas are used in the procedures to define labels of the newly created successors of a state (\(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\), \(j \in {\mathcal {A}}\) and \(\Phi \) is a set of formulas)

$$\begin{aligned} \Phi ^{\lnot [O]_{j}}(\psi )&= \{\sim \!\psi \} \cup \Phi ^{[O]_{j}}. \end{aligned}$$

The definition of set \(\Phi ^{[O]_{j}}\) depends on the axioms associated with \([O]\):

$$\begin{aligned} \Phi ^{[\mathrm {I}]_{j}} = \Phi /[\mathrm {I}]_{j},\quad \Phi ^{[\mathrm {G}]_{j}} = (\Phi /[\mathrm {G}]_{j}) \cup \Phi ^{[\mathrm {I}]_{j}},\quad \Phi ^{[\mathrm {B}]_{j}}&= (\Phi /[\mathrm {B}]_{j}) \cup (\Phi \sqcap j) \end{aligned}$$

and

$$\begin{aligned} \Phi /[O]_{j}&= \{\psi : [O]_{j}\psi \in \Phi \},\\ \Phi \sqcap [O]_{j}&= \{[O]_{j}\psi : [O]_{j}\psi \in \Phi \},\\ \Phi \sqcap \lnot [O]_{j}&= \{\lnot [O]_{j}\psi : \lnot [O]_{j}\psi \in \Phi \},\\ \Phi \sqcap j&= (\Phi \sqcap [\mathrm {B}]_{j}) \cup (\Phi \sqcap [\mathrm {G}]_{j}) \cup (\Phi \sqcap [\mathrm {I}]_{j}) \cup \\&\ \ \ \ (\Phi \sqcap \lnot [\mathrm {B}]_{j}) \cup (\Phi \sqcap \lnot [\mathrm {G}]_{j}) \cup (\Phi \sqcap \lnot [\mathrm {I}]_{j}). \end{aligned}$$

Given a state s and its label L(s), we will write \(L^{[O]_{j}}(s)\) to denote \((L(s))^{[O]_{j}}\) and \(L^{\lnot [O]_{j}}(s,\psi )\) to denote \((L(s))^{\lnot [O]_{j}}(\psi )\).

figure c
figure d
figure e

Procedures 3, 4 and 5 are based on the algorithms described in Halpern and Moses (1992) which need to be considerably extended to address axioms of \({\textsc {Team}}{\textsc {Log}}\). Firstly, they are affected by mixed axioms (see Dziubiński et al. 2007 for the discussion of these aspects). Secondly they need to deal with existence of fixpoint modalities, to avoid construction of too long or infinite paths. When creating \(\lnot [O]_{j}[O]^{+}_{G}\xi \)-successor (with \(O \in \{\mathrm {B},\mathrm {I}\}\) and \(j \in G\)) the sets \(L^{[O]_{j}}(\cdot )\) for all \([O]^{+}_{G}\psi \)-Ancestors are checked. A state t is called a \([O]^{+}_{G}\xi \)-Ancestor of state s if t is an ancestor of s and for every state u on the path from t to s such that \(u \ne s\), there exists \(j \in G\) such that u is a \(\lnot [O]_{j}[O]^{+}_{G}\xi \)-Successor. If the label of a potential \(O_j\)-successor, with \(O \in \{\mathrm {B},\mathrm {I}\}\), of a state s is equal to the label of a successor node n of some \([O]^{+}_{G}\psi \)-Ancestor of s which is on the path from t to s, then construction of the successor of s is blocked by n. This is illustrated in Fig. 2.

Fig. 2
figure 2

Creation of \(\lnot [O]_{j}[O]^{+}_{G}\xi \)-successor of s is blocked by its ancestor n which is a \(\lnot [O]_{j}[O]^{+}_{\xi }\)-successor of a \([O]^{+}_{G}\xi \)-Ancestor t of s. Dotted lines depict sequences of internal nodes (these sequences can be empty, in which case the starting node coincides with the ending state)

To decide whether a node containing a formula of the form \(\lnot [O]^{+}_{G}\xi \) is satisfiable, it has to be checked whether an appropriate sequence of states can be constructed, that would indicate that this formula is satisfiable. Since creation of successors for formulas of the form \(\lnot [O]_{j}[O]^{+}_{G}\xi \), with \(j \in G\), may be blocked by some ancestor node, the decision whether such an appropriate sequence of states can be constructed may have to be suspended until the satisfiability of the ancestors is checked. Therefore for each node n there is a set B(n) associated with it and containing weak ancestorsFootnote 9 that block creation successors of states in the n-subtree of the pre-tableau (c.f. Fig. 2).Footnote 10 Whenever a new node n is created by the algorithm, the associated set of nodes B(n) is set to \(\varnothing \).

During the stage of marking nodes, nodes of the pre-tableau are marked either \(\texttt {sat}\), \(\texttt {unsat}\), or \(\texttt {undec}\). A node n being marked \(\texttt {undec}\) indicates that satisfiability of \(\bigwedge L(n)\) could not be decided due to existence of a formula of the form \(\lnot [O]^{+}_{G}\psi \) in its label for which an appropriate sequence of states was not constructed yet. We call such formulas unresolved in a given node, as defined below. In the definition we refer to a notion of \(\xi \)-descendant. A node is a \(\xi \)-descendant if it is a \(\xi \)-successor or a descendant of a \(\xi \)-successor such that there is no state between it and the \(\xi \)-successor.

Definition 16

(Unresolved formula) Let n be a node in a pre-tableau and let \(\lnot [O]^{+}_{G}\psi \in L(n)\) with \(O \in \{\mathrm {B},\mathrm {I}\}\). A formula \(\lnot [O]^{+}_{G}\psi \) is unresolved in n if one of the following holds:

  • n is an internal node and a \(\lnot [O]_{j}[O]^{+}_{G}\psi \)-descendant with \(j \in G\), none of its successors is marked \(\texttt {sat}\), there exists a successor of n marked \(\texttt {undec}\) and \(B(n) \ne \varnothing \),

  • n is a state and a \(\lnot [O]_{j}[O]^{+}_{G}\psi \)-Successor with \(j \in G\), \(B(n) \ne \varnothing \), none of \(\lnot [O]_{k}[O]^{+}_{G}\psi \)-successors of n, with \(k \in G\), is marked \(\texttt {sat}\) and \([O]_{k}\psi \in L(n)\), for all \(k \in G\).

Notice that if \(B(n) = \varnothing \), then a node cannot be marked \(\texttt {undec}\). The stage of marking nodes is described in Procedure 6.

We show first that for any input formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) the algorithm for checking satisfiability terminates. The usual method of showing that is by showing that the heightFootnote 11 of every node in the constructed pre-tableau is bounded. In the case of standard modal systems (e.g. \(\text {K}_{n}\), \(\text {KD}_{n}\), \(\text {KD45}_{n}\)) this is shown by showing that modal depth of formulas in subsequent states on paths of the pre-tableaux is falling with distance from the root. In the case of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) this approach would not work and we need to find a different parameter of the labels that changes with distance from the root. The main problem here are the formulas of the form \([O]^{+}_{G}\psi \) or \(\lnot [O]^{+}_{G}\psi \). This is because if t is an \(O_j\)-Successor of s in a pre-tableau constructed by Algorithm 1, then any formula \([O]^{+}_{G}\psi \in L(s)\) with \(j \in G\) is in L(t) as well. Similarly with a formula of the form \(\lnot [O]^{+}_{G}\psi \in L(s)\), if t is a \(\lnot [O]_{j}[O]^{+}_{G}\psi \)-Successor of s. Moreover, if additionally u is a \(O_k\)-Successor of t and \(k \in G\), then for any formula \(\xi \in \lnot \mathrm {BT}(\psi )\), \(\xi \in \lnot L(t)\) (as it is added during \([\mathrm {B}]\)-expanded tableau formation) and \(\xi \in \lnot L(u)\), as \([O]^{+}_{G}\psi \in L(u)\). Thus formulas of the form \([O]^{+}_{G}\psi \) may carry over to the label of the \(O_j\)-Successor formulas from \(\lnot \mathrm {BT}(\psi )\). Similarly, they may carry over the formulas \([O]_{j}[O]^{+}_{G}\psi \) and \([O]_{j}\psi \), that are added to the label during the closed propositional tableau formation.

To analyse the length of sequences of O-Successors in a pre-tableau constructed by Algorithm 1, we need to separate the formulas in labels of states which are carried by some other formulas from those which are not carried by any other formula. We will say that a formula \([O]^{+}_{G}\psi \) carries a formula \(\xi \) if \(\xi \in \lnot \mathrm {BT}(\psi )\) or \(\xi \in \widetilde{\mathrm {Cl}}([O]^{+}_{G}\psi )\), where \(\widetilde{\mathrm {Cl}}([O]^{+}_{G}\psi ) = \{[O]_{j}\psi : j \in G\} \cup \{[O]_{j}[O]^{+}_{G}\psi : j \in G\}\). Similarly, a formula \(\lnot [O]^{+}_{G}\psi \) carries a formula \(\xi \) if \(\xi \in \lnot \mathrm {BT}(\psi )\) or \(\xi \in \widetilde{\mathrm {Cl}}(\lnot [O]^{+}_{G}\psi )\), where \(\widetilde{\mathrm {Cl}}(\lnot [O]^{+}_{G}\psi ) = \lnot \widetilde{\mathrm {Cl}}([O]^{+}_{G}\psi )\). Given a set of formulas \(\Phi \) and a formula \(\xi \) we will say that \(\xi \) is carried by \(\Phi \) if there is a formula in \(\Phi \) which carries it.

First we will consider the carried formulas of the form \([O]^{+}_{H}\zeta \) or \(\lnot [O]^{+}_{H}\zeta \). Notice that in this case such a formula is carried by some formula \([O]^{+}_{G}\psi \) or \(\lnot [O]^{+}_{G}\psi \) if and only if it is in \(\lnot \mathrm {PT}(\psi )\). Given a set of formulas \(\Phi \), letFootnote 12

$$\begin{aligned} \mathrm {Gr}(\Phi ) = \bigcup _{O \in \{\mathrm {B},\mathrm {I}\}} \left( \left( \Phi \sqcap [O]^{+}\right) \cup \left( \Phi \sqcap \lnot [O]^{+}\right) \right) , \end{aligned}$$

where

$$\begin{aligned} \Phi / [O]^{+}&= \{\psi : [O]^{+}_{G}\psi \in \Phi , \text { for some }G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\},\\ \Phi \sqcap [O]^{+}&= \{[O]^{+}_{G}\psi : [O]^{+}_{G}\psi \in \Phi , \text { for some }G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\},\\ \Phi \sqcap \lnot [O]^{+}&= \{\lnot [O]^{+}_{G}\psi : \lnot [O]^{+}_{G}\psi \in \Phi , \text { for some }G \in \mathrm {P}({\mathcal {A}}) \setminus \{\varnothing \}\}. \end{aligned}$$

Let \({\mathcal {F}}_{\Phi }: \mathrm {P}\left( {{\mathcal {L}}}^{\mathrm {T}}\right) \rightarrow \mathrm {P}\left( {{\mathcal {L}}}^{\mathrm {T}}\right) \) be defined as follows, for \(\Psi \subseteq {{\mathcal {L}}}^{\mathrm {T}}\),

$$\begin{aligned} {\mathcal {F}}_{\Phi }(\Psi )&= \mathrm {Gr}(\Phi ) \setminus \lnot \mathrm {PT}\left( \bigcup _{O \in \{\mathrm {B},\mathrm {I}\}}\lnot \Psi / [O]^{+}\right) , \end{aligned}$$

The operator \({\mathcal {F}}_{\Phi }\), when applied to a set of formulas \(\Psi \), removes from \(\Phi \) all the formulas from \(\mathrm {Gr}(\Phi )\) which are carried by \(\Psi \).

figure f

Given a set of formulas \(\Phi \) and a formula \(\psi \) we will say that \(\psi \) is uncarried in \(\Phi \) if \(\psi \in \Phi \) and \(\psi \) is not carried by \(\Phi \). We will be interested in sets of formulas which are carry-free, that is \(\Phi \) such that all the formulas in \(\Phi \) are uncarried in it. Given \(i \in {\mathbb {N}}\), let

$$\begin{aligned} F^{(i)}_{\Phi } = \left\{ \begin{array}{ll} \varnothing , \hfill &{} \text {if }i = 0, \hfill \\ {\mathcal {F}}_{\Phi }\left( F_{\Phi }^{(i-1)}\right) , \hfill &{} \text {if }i > 0 \hfill \end{array} \right. \end{aligned}$$

and let \(F_{\Phi }^{(\infty )} = \lim _{i \rightarrow \infty } F_{\Phi }^{(i)}\). As we show below, for any \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}\) the limit \(F_{\Phi }^{(\infty )}\) exists.

Lemma 1

For any \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}\), \(F_{\Phi }^{(\infty )}\) exists.

Let \(\widehat{\mathrm {Gr}}(\Phi ) = F_{\Phi }^{(\infty )}\), that is

$$\begin{aligned} \widehat{\mathrm {Gr}}(\Phi ) = \mathrm {Gr}(\Phi ) \setminus \lnot \mathrm {PT}\left( \bigcup _{O \in \{\mathrm {B},\mathrm {I}\}}\lnot \widehat{\mathrm {Gr}}(\Phi ) / [O]^{+}\right) . \end{aligned}$$

The set \(\widehat{\mathrm {Gr}}(\Phi )\) it the maximal carry-free subset of \(\mathrm {Gr}(\Phi )\) containing all the formulas which are uncarried in \(\mathrm {Gr}(\Phi )\).

Now we turn to the uncarried formulas which are not in \(\mathrm {Gr}(\Phi )\) but can ‘carry’ other formulas to successor labels. These are those formulas in \(\Phi \) which are (possibly negated) formulas of the form \([O]_{j}\xi \). More precisely, we will be interested in those of such formulas which are not carried by \(\widehat{\mathrm {Gr}}(\Phi )\) nor are elements of \(\widetilde{\mathrm {Cl}}(\Phi )\), where \(\widetilde{\mathrm {Cl}}(\Phi ) = \bigcup _{\psi \in \mathrm {Gr}(\Phi )} \widetilde{\mathrm {Cl}}(\psi )\). The set of such formulas isFootnote 13

$$\begin{aligned} \mathrm {Ind}(\Phi ) = \left( \bigcup _{j \in {\mathcal {A}}} \Phi \sqcap j\right) \setminus \left( \widetilde{\mathrm {Cl}}(\Phi ) \cup \lnot \mathrm {BT}\left( \bigcup _{O \in \{\mathrm {B},\mathrm {I}\}}\lnot \widehat{\mathrm {Gr}}(\Phi ) / [O]^{+}\right) \right) . \end{aligned}$$

Although the modal depth of \(\widehat{\mathrm {Gr}}(\cdot )\) of labels may not change between O-Successors in the sequence of states in a pre-tableau constructed by Algorithm 1, there is one more parameter of these sets that will change. Given a formula \(\psi \) and \(O \in \{\mathrm {B},\mathrm {I}\}\), we define a setFootnote 14

$$\begin{aligned} \mathrm {ag}(\psi ,[O]^{+}) = \left\{ \begin{array}{ll} G, \hfill & \text {if }\psi \text { is of the form }[O]^{+}_{G}\xi \text { or }\lnot [O]^{+}_{G}\xi , \hfill \\ {\mathcal {A}}\ \cup \{\omega \}, \hfill & \text {otherwise}, \hfill \end{array} \right. \end{aligned}$$

where \(\omega \notin {\mathcal {A}}\). Given a set of formulas \(\Phi \ne \varnothing \) and \(O \in \{\mathrm {B},\mathrm {I}\}\) we define

$$\begin{aligned} \mathrm {ag}(\Phi ,[O]^{+}) = \left\{ \begin{array}{ll} \bigcap _{\psi \in \Phi } \mathrm {ag}(\psi ,[O]^{+}), \hfill & \text {if }\Phi \ne \varnothing , \hfill \\ {\mathcal {A}}\ \cup \{\omega \}, \hfill & \text {otherwise}. \hfill \end{array} \right. \end{aligned}$$

Notice that \(\omega \in \mathrm {ag}(\Phi ,[O]^{+})\) implies that there are no formulas of the form \([O]^{+}_{G}\xi \) nor \(\lnot [O]^{+}_{G}\xi \) in \(\Phi \). Also, when formulas are removed from \(\Phi \), then \(\mathrm {ag}(\Phi ,[O]^{+})\) either remains unchanged or increases.

When analysing how labels of subsequent states change, we will divide them into subsets (levels) of different modal depth of formulas and then we will look at the sets \(\mathrm {ag}\left( \cdot ,[O]^{+}\right) \) at different levels. Given a set of formulas \(\Phi \), let \(\Phi _d =\{\psi \in \Phi : \mathrm {dep}(\psi ) = d\}\). Also, let

$$\begin{aligned} \mathrm {ag}\left( \Phi ,[O]^{+},d\right) = \mathrm {ag}\left( \Phi _d,[O]^{+}\right) . \end{aligned}$$

Notice that \(\mathrm {ag}\left( \Phi ,[O]^{+},d\right) \) is well defined even for \(d > \mathrm {dep}(\Phi )\). Is it simply \({\mathcal {A}}\cup \{\omega \}\) then. Similarly for levels \(d \le \mathrm {dep}(\Phi )\) at which there are no formulas of the form \([O]^{+}_{G}\xi \) nor \(\lnot [O]^{+}_{G}\xi \). Notice also, that \(\mathrm {ag}\left( \Phi ,[O]^{+},0\right) = {\mathcal {A}}\cup \{\omega \}\).

The main difficulty in showing the boundaries on the lengths of paths in pre-tableau constructed by Algorithm 1 is in showing that sequences of \(\mathrm {B}\)-Successors and \(\mathrm {I}\)-Successors are properly bounded. The key lemmas to this result are Lemma 2 and Lemma 3 proven below. The lemmas give properties of \(\widehat{\mathrm {Gr}}(\cdot )\) and \(\mathrm {Ind}(\cdot )\) that follow from modal context restrictions \({\mathbf{R}}_1\) and \({\mathbf{R}}_2\). In the case of \(\mathrm {I}\)-successors restriction \({\mathbf{R}}_1\) is assumed only. This restriction guarantees that if a formula of the form \([O]^{+}_{H}\xi \) is carried by a formula \([O]^{+}_{G}\psi \), then it must be that \(G \cap H = \varnothing \). Similarly, if a formula of the form \([O]_{j}\xi \) is carried by a formula \([O]^{+}_{G}\psi \), then it must be that \(j \notin G\). Thus if \(j \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {I}]^{+}, d\right) \) and there are no formulas of the form \([\mathrm {B}]^{+}_{G}\psi \in \lnot \Phi _d\) at levels \(d > D\), then any formula of the form \([\mathrm {I}]^{+}_{H}\xi \in \lnot \Phi \) with modal depth \(\ge D\) must be uncarried in \(\Phi \). Also, if a formula \([\mathrm {I}]_{j}\xi \) is in \(\lnot \widetilde{\mathrm {Cl}}(\Phi )\), then it must be in \(\lnot \widetilde{\mathrm {Cl}}(\widehat{\mathrm {Gr}}(\Phi ))\).

Lemma 2

Let \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) and let \(D \ge 0\) and \(j \in {\mathcal {A}}\) be such that \(j \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {I}]^{+}, d\right) \) and \(\omega \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {B}]^{+}, d\right) \), for all \(d \ge D + 1\). Then the following hold:

  1. (i)

    if \([\mathrm {I}]^{+}_{G}\psi \in \lnot \Phi \) with \(j \in G\) and \(\mathrm {dep}([\mathrm {I}]^{+}_{G}\psi ) \ge D\), then \([\mathrm {I}]^{+}_{G}\psi \in \lnot \widehat{\mathrm {Gr}}(\Phi )\),

  2. (ii)

    if \(\mathrm {dep}(\mathrm {Ind}(\Phi )) \le D\) and \([\mathrm {I}]_{j}\psi \in \lnot \Phi \) with \(\mathrm {dep}([\mathrm {I}]_{j}\psi ) \ge D+1\), then \([\mathrm {I}]_{j}\psi \in \lnot \widetilde{\mathrm {Cl}}\!\left( \widehat{\mathrm {Gr}}(\Phi )\right) \).

The analogous lemma for \(\mathrm {B}\)-Successors differs from the case of \(\mathrm {I}\)-Successors in two aspects. If t is a \(\mathrm {B}_j\)-Successor of state s, then, by construction of the algorithm, \(L(s) \sqcap j \subseteq L(t)\). For this reason the set \(\mathrm {Ind}(\Phi ) \sqcap j\) rather than the set \(\mathrm {Ind}(\Phi )\) is used. Secondly the lemma has an additional point that requires restriction \({\mathbf{R}}_2\). The point is crucial for having bounds on the lengths of sequences of \(\mathrm {B}\)-Successors in the pre-tableau.

Lemma 3

Let \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) and let \(D \ge 0\) and \(j \in {\mathcal {A}}\) be such that \(j \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {B}]^{+}, d\right) \) and \(\omega \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {I}]^{+}, d\right) \), for all \(d \ge D + 1\). Then the following hold:

  1. (i)

    if \([\mathrm {B}]^{+}_{G}\psi \in \lnot \Phi \) with \(j \in G\) and \(\mathrm {dep}([\mathrm {B}]^{+}_{G}\psi ) \ge D\), then \([\mathrm {B}]^{+}_{G}\psi \in \lnot \widehat{\mathrm {Gr}}(\Phi )\),

  2. (ii)

    if \(\mathrm {dep}(\mathrm {Ind}(\Phi ) \sqcap j) \le D\) and \([\mathrm {B}]_{j}\psi \in \lnot \Phi \) with \(\mathrm {dep}([\mathrm {B}]_{j}\psi ) \ge D+1\), then \([\mathrm {B}]_{j}\psi \in \lnot \widetilde{\mathrm {Cl}}\!\left( \widehat{\mathrm {Gr}}(\Phi )\right) \),

  3. (iii)

    if \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\), \(\mathrm {dep}(\mathrm {Ind}(\Phi ) \sqcap j) \le D\), \(\omega \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(\Phi ), [\mathrm {I}]^{+}, d\right) \), for all \(d \ge D\) and \([O]_{j}\psi \in \lnot \Phi \) with \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\) and \(\mathrm {dep}([O]_{j}\psi ) \ge D+1\), then \([O]_{j}\psi \in \lnot \widetilde{\mathrm {Cl}}\!\left( \widehat{\mathrm {Gr}}(\Phi )\right) \) and \(O = \mathrm {B}\).

In what follows we will concentrate on sequences of \(\mathrm {B}\)-Successors. The general approach in the case of \(\mathrm {I}\)-Successors is similar and easier. In proofs, given in the Appendix, we indicate where the differences in proofs for these to cases lie. For the detailed proofs we refer the reader to Dziubiński (2011). Lemma 3 allows us to analyse the origins of formulas in labels of \(\mathrm {B}_j\)-Successors. The corollary below points out the sources of formulas in the successor state with modal depth not smaller than \(\mathrm {dep}(\mathrm {Ind}(\cdot )) \sqcap j\) of the predecessor state. Roughly speaking all such formulas are either added when the label of the successor state is being closed or the formulas are carried by the uncarried formulas from the label of the predecessor state, or are uncarried formulas inherited from the label of the predecessor state. The analogous corollary for \(\mathrm {I}\)-successors would concern set \(\mathrm {Ind}(\cdot )\) instead of \(\mathrm {dep}(\mathrm {Ind}(\cdot )) \sqcap j\).

Corollary 1

Let t be an \(\mathrm {B}_j\)-Successor of state s in the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\), with \(D \ge 0\) such that \(\mathrm {dep}(\mathrm {Ind}(L(s)) \sqcap j) \le D\), \(j \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(L(s)), [\mathrm {B}]^{+}, d\right) \) and \(\omega \in \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(L(s)), [\mathrm {I}]^{+}, d\right) \), for all \(d \ge D + 1\). Then for all \(\psi \in L(t)\) with \(\mathrm {dep}(\psi ) \ge D\) one of the following holds

  1. (i)

    \(\psi \in L(s) \sqcap j\) or

  2. (ii)

    \(\psi \in \widetilde{\mathrm {Cl}}(L(t))\) or

  3. (iii)

    there exists \([\mathrm {B}]^{+}_{G}\xi \in \lnot \widehat{\mathrm {Gr}}(L(s))\) with \(j \in G\) such that \(\psi \in \lnot \mathrm {BT}(\xi )\) or

  4. (iv)

    \(\psi \) is of the form \([\mathrm {B}]^{+}_{G}\eta \) with \(j \in G\) and \(\psi \in \lnot \widehat{\mathrm {Gr}}(L(s))\) or

  5. (v)

    \(\psi \) is of the form \(\lnot [\mathrm {B}]^{+}_{G}\eta \) with \(j \in G\), \(\psi \in \widehat{\mathrm {Gr}}(L(s))\), \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\eta \in L(s)\) and t is a \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\eta \)-Successor of s.

We will also need the following auxiliary lemma which will be useful in analysing the lengths of sequences involving \(\mathrm {B}\)-successors. The lemma extends a similar one used in the analysis of the complexity of \({\textsc {Team}}{\textsc {Log}}\) without fixpoint modalities in Dziubiński et al. (2007).

Lemma 4

Let t be a \(\mathrm {B}_j\)-Successor of state s in the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\). Then the following hold for \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\):

  1. 1.

    \(\lnot [O]_{j}\xi \in L(s)\) iff \(\lnot [O]_{j}\xi \in L(t)\).

  2. 2.

    \([O]_{j}\xi \in L(s)\) iff \([O]_{j}\xi \in L(t)\).

  3. 3.

    \(L^{[O]_{j}}(s) = L^{[O]_{j}}(t)\).

  4. 4.

    \(L^{\lnot [O]_{j}}(s, \xi ) = L^{\lnot [O]_{j}}(t, \xi )\).

We are now ready to state the lemma about the bounds on the length of a sequence of \(\mathrm {B}\)-Successors with unchanged modal depth of labels in the pre-tableau constructed by Algorithm 1 for some input \(\varphi \). To assess lengths of sequences of \(\mathrm {B}\)-Successors with unchanged modal depth of labels, we will show that the sets \(\mathrm {ag}\left( \widehat{\mathrm {Gr}}(\cdot ), [\mathrm {B}]^{+},d\right) \) and \(\mathrm {ag}\left( \widehat{\mathrm {Gr}}(\cdot ),[\mathrm {I}]^{+},d\right) \) must gradually increase proceeding top down, from \(d = \mathrm {dep}(\Phi )\) to \(d = 1\). For this reason we will need to assess for how long these sets may remain unchanged at different levels. This is expressed by the following properties of states, for \(O \in \{\mathrm {B},\mathrm {I}\}\):

  • P O1 \(\mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(L(s)), [O]^{+}, d\right) = \mathrm {ag}\!\left( \widehat{\mathrm {Gr}}(L(t)), [O]^{+}, d\right) \),

We say that the sequence of states \(s_0,\ldots ,s_m\) satisfies P O1 if for all \(0 < k \le m\), states \(s_{k-1}\) and \(s_k\) satisfy P O1.

Additional factor that needs to be taken into account is the set of formulas of the form \(\lnot [O]^{+}_{G}\xi \) at different levels of the set \(\widehat{\mathrm {Gr}}(\cdot )\). The following property states that this set remains unchanged between two states:

  • P O2 \(\left( \widehat{\mathrm {Gr}}(L(s)) \sqcap \lnot [O]^{+}\right) _d = \left( \widehat{\mathrm {Gr}}(L(t)) \sqcap \lnot [O]^{+}\right) _d\).

We say that the sequence of states \(s_0,\ldots ,s_m\) satisfies P O2, if for all \(0 < k \le m\), states \(s_{k-1}\) and \(s_k\) satisfy P O2.

Lemma 5

The maximal length of sequence of \(\mathrm {B}\)-Successors with unchanged modal depth of labels in the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2|{\mathcal {A}}|}\right) \).

Analogous lemma for sequences of \(\mathrm {I}\)-successors can be shown with restriction \({\mathbf{R}}_1\) only.

Lemma 6

The maximal length of a sequence of \(\mathrm {I}\)-Successors with unchanged modal depth of labels in the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2|{\mathcal {A}}|}\right) \).

Now we are ready to prove the lemma on bounds on the height and state heightFootnote 15 of the pre-tableau constructed by Algorithm 1 for an input formula satisfying modal context restriction \({\mathbf{R}}_2\). The height is bounded by a polynomial depending on \(|\varphi |\), while the state height is bounded by a polynomial depending on \(\mathrm {dep}(\varphi )\).

Lemma 7

The state height of the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is \(\le {\mathcal {O}}(\mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1})\) and its height is \(\le {\mathcal {O}}(|\varphi | \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1})\).

Proof

For any node n in a pre-tableau constructed by the algorithm \(|L(n)| \le (2 |{\mathcal {A}}| + 1) |\varphi |\), as \(L(n) \subseteq \lnot \mathrm {Cl}(\varphi )\). Thus the path between any subsequent states s and t can contain at most \((2 |{\mathcal {A}}| + 1) |\varphi | - 1\) internal nodes. Moreover, for any states s and t such that t is a descendant of s it must be that \(\mathrm {dep}(L(t)) \le \mathrm {dep}(L(s))\).

If s and t are states, such that t is an \(\mathrm {G}\)-Successor of s, then \(\mathrm {dep}(L(t)) < \mathrm {dep}(L(s))\). Thus any sequence of states in the pre-tableau can contain at most \(\mathrm {dep}(\varphi )\) \(\mathrm {G}\)-Successors. Also, if s, t and u are states such that t is an \(\mathrm {B}_j\)-Successor of s and u is an \(\mathrm {I}_k\)-Successor of t with \(j \ne k\), then it holds that \(\mathrm {dep}(L(u)) < \mathrm {dep}(L(s))\). By Lemma 4 and construction of the algorithm, if s, t and u are states such that t is an \(\mathrm {B}_j\)-Successor of s and u is an \(O_k\)-Successor of t, where \(O \in \{\mathrm {B},\mathrm {G},\mathrm {I}\}\), then it must hold that \(j \ne k\). By Lemma 6, the maximal length of a sequence of \(\mathrm {I}\)-Successors with the same modal depth of labels in a pre-tableau constructed by the algorithm is \(\le {\mathcal {O}}(\mathrm {dep}(\varphi )^{2 |{\mathcal {A}}|})\). Similarly, by Lemma 5, the maximal length of a sequence of \(\mathrm {B}\)-Successors with the same modal depth of labels in a pre-tableau constructed by the algorithm is \(\le {\mathcal {O}}(\mathrm {dep}(\varphi )^{2 |{\mathcal {A}}|})\). Hence any sequence of nodes in the pre-tableau must be of length \(\le {\mathcal {O}}(|\varphi | \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1})\) and contains \(\le {\mathcal {O}}(\mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1})\) states. \(\square \)

Since the height of the pre-tableau constructed by the algorithm for an input formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is bounded and the number of successor of any state is bounded as well so we have the following proposition as a corollary of Lemma 7.

Proposition 3

For any input formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) Algorithm 1 terminates.

For soundness and validity, we show that Algorithm 1 is sound and valid for any formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\) for which it terminates. Thus as long as we can show that the algorithm terminates for a fragment of \({{\mathcal {L}}}^{\mathrm {T}}\), we have a sound and valid method of checking \({\textsc {Team}}{\textsc {Log}}\) satisfiability.

Proposition 4

Suppose that Algorithm 1 terminates on a formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}\). Then \(\varphi \) is satisfiable iff Algorithm 1 returns \(\texttt {sat}\) on the input \(\varphi \).

By Proposition 3, Algorithm 1 terminates, so we have the following corollary from Proposition 4, stating soundness and validity of the algorithm.

Proposition 5

A formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is satisfiable iff Algorithm 1 returns \(\texttt {sat}\) on the input \(\varphi \).

By Lemma 7, Algorithm 1 can be run on a Turing machine using space bounded from above by a polynomial depending on \(|\varphi |\). Thus the \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem for \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is PSPACE solvable. It is also PSPACE hard as it is PSPACE hard for \({{\mathcal {L}}}^{\mathrm {T}}\) without fixpoint modalities (Dziubiński et al. 2007). Thus we have the following theorem.

Theorem 2

The \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is PSPACE complete.

As Lemma 7 and proof of Proposition 4 suggest, bounding modal depth of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) makes the \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem NPTIME complete.

Theorem 3

For any fixed k, if modal depth of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) is bounded by k, then the \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem for them is NPTIME complete.

Proof

By Lemma 7 and the construction of \({\textsc {Team}}{\textsc {Log}}\) tableau based on the pre-tableau constructed by Algorithm 1 presented in Proposition 4, the size of the tableau for a satisfiable formula j is bounded by \({\mathcal {O}}\!\left( ((2|{\mathcal {A}}| + 1)|\varphi |)^{\mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1}}\right) \). Hence, if modal depth of \(\varphi \) is bounded by k, then the size of the tableau is bounded by \({\mathcal {O}}\!\left( ((2|{\mathcal {A}}| + 1)|\varphi |)^{k^{2|{\mathcal {A}}|+1}}\right) \). This means that the satisfiability of \(\varphi \) with bounded modal depth can be checked by the following non-deterministic Algorithm 7.

figure g

Since tableau \({\mathcal {T}}\) constructed by Algorithm 7 is of polynomial size, so checking if it is a tableau for \(\varphi \) can be realized in polynomial time. This shows that satisfiability of \(\varphi \) can be checked in NPTIME. The problem is also NPTIME complete, as the satisfiability problem for propositional logic is NPTIME hard. \(\square \)

figure h

6.2 Complexity of \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\)

The algorithm for checking \({\textsc {Team}}{\textsc {Log}}\) satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) requires a different approach since, as Proposition 1 shows, a model for such formulas may contain an exponentially long path. Therefore we modify Algorithm 1, designed for checking the \({\textsc {Team}}{\textsc {Log}}\) satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) in polynomial space. The difference lies in using a new procedure for \(\mathrm {B}\)-successors creation, specifically for the formulas of the form \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \) with \(j \in G\). Since the satisfying sequence for such a formula may have exponential length with respect to the size of the set of formulas, the algorithm using a polynomial space cannot attempt to construct such a sequence storing it fully in the memory, as it was done in the case of Algorithm 1. For this reason, the new algorithm constructs a pre-tableau just like Algorithm 1, creating \(\mathrm {G}\)- and \(\mathrm {I}\)-successors in the same way, but stopping creation of \(\mathrm {B}\)-successors for formulas of the form \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \) when certain condition is satisfied. In such case Function 9 is used for checking if the label of the \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \)-successor is satisfiable. If it is decided by Function 9 that the label of the \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \)-successor is not satisfiable, then the state is marked \(\texttt {unsat}\). Otherwise, the decision on how the state should be marked depends on the other successors and the same procedure of marking nodes, as the one used in Algorithm 1, is applied. The new algorithm is referred two as Algorithm 2, and its formulation differs from the formulation of Algorithm 1 by the procedure of \(\mathrm {B}\)-successors creation being replaced with a new one, called Procedure 8.

In the algorithm we are referring to the following sets, defined for a given set of formulas \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}\) and \(G \subseteq {\mathcal {A}}\):

$$\begin{aligned} \Phi / [\mathrm {B}]^{+}_{G}&= \{\psi : [\mathrm {B}]^{+}_{H}\psi \in \Phi \text { and } G \subseteq H\},\\ \Phi \sqcap [\mathrm {B}]^{+}_{G}&= \{[\mathrm {B}]^{+}_{H}\psi : [\mathrm {B}]^{+}_{H}\psi \in \Phi \text { and } G \subseteq H\},\\ \Phi ^{[\mathrm {B}]^{+}_{G}}&= (\Phi / [\mathrm {B}]^{+}_{G}) \cup (\Phi \sqcap [\mathrm {B}]^{+}_{G}). \end{aligned}$$

Given a formula \(\psi \), \(G \subseteq {\mathcal {A}}\), \(j\in G\) and a set of formulas \(\Phi \) such that \(\{[\mathrm {B}]_{j}\psi ,\psi \} \subseteq \Phi \) as an input, Function 9 decides whether the set \(\Phi ^{\lnot [\mathrm {B}]_{j}}([\mathrm {B}]^{+}_{G}\psi )\) is satisfiable or not. To describe the idea of this algorithm, let \(\Psi _1\) and \(\Psi _2\) be sets of formulas. Given \(k \in {\mathcal {A}}\), we say that \(\Psi _1\) and \(\Psi _2\) are connected with k if \(\Psi _1 \sqcap k = \Psi _2 \sqcap k\). Moreover, given a set \(H \subseteq {\mathcal {A}}\), we say that \(\Psi _1\) and \(\Psi _2\) are H-connected if they are connected with some \(k \in H\). Let \(\Gamma \) be a set of formulas and let \({\mathcal {S}}(\Gamma )\) be the set of all minimal sets of formulas containing \(\Gamma \) as a subset that are \([\mathrm {B}]\)-expanded tableaux. Given \(H \subseteq {\mathcal {A}}\), let \({\mathcal {G}}_{H}(\Gamma ) = (V,E)\) be an undirected graph such that V consists of all elements \(\Psi \in {\mathcal {S}}(\Gamma )\) such that Algorithm 2 returns \(\texttt {sat}\) on input \(\bigwedge \Psi \) and for all \((\Psi _1,\Psi _2) \in V \times V\), \((\Psi _1,\Psi _2) \in E\) iff they are H-connected. A path in \({\mathcal {G}}_{H}(\Gamma )\) is a sequence \(\Gamma _0,\ldots ,\Gamma _n\) of elements of V such that for all \(1 \le i \le n\), \(\Gamma _{i-1}\) and \(\Gamma _{i}\) are H-connected. The length of path \(\Gamma _0,\ldots ,\Gamma _n\) is n. Given a path \(\Gamma _0,\ldots ,\Gamma _n\) of length \(n \ge 1\) in \({\mathcal {G}}_{H}(\Gamma )\) we call a sequence \(j_1,\ldots ,j_n\) of elements from H such that for each \(1 \le i \le n\), \(\Gamma _{i-1}\) and \(\Gamma _{i}\) are connected with \(j_i\), a sequence associated with path \(\Gamma _0,\ldots ,\Gamma _n\). If \(n = 0\), then the sequence associated with the path is the empty sequence \(\varepsilon \). Given two sets of formulas \(\Psi _0\) and \(\Psi _1\), we say that \(\Psi _1\) is reachable from \(\Psi _0\) in \({\mathcal {G}}_{H}(\Gamma )\) (in n steps) iff there exists a path \(\Gamma _0,\ldots ,\Gamma _n\) in \({\mathcal {G}}_{H}(\Gamma )\) such that \(\Psi _0 = \Gamma _0\) and \(\Psi _1 = \Gamma _n\).

To decide the satisfiability of \(\bigwedge \Phi ^{\lnot [\mathrm {B}]_{j}}([\mathrm {B}]^{+}_{G}\psi )\), Function 9 checks whether there exist two sets of formulas \(\{\Psi _0,\Psi _1\} \subseteq {\mathcal {S}}((\Phi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\), with \(H = \mathrm {ag}\!\left( \!\left( \Phi \sqcap [\mathrm {B}]^{+}_{\{j\}}\right) \cup \{\lnot [\mathrm {B}]^{+}_{G}\psi \},[\mathrm {B}]^{+}\right) \), such that

  • \((\Phi \sqcap j) \setminus ((\Phi \sqcap [\mathrm {B}]_{j}) \cup (\Phi \sqcap \lnot [\mathrm {B}]_{j})) \subseteq \Psi _0\) and

  • either there exists \(k \in H\) such that Algorithm 2 returns \(\texttt {sat}\) on the input \(\bigwedge \left( \Psi _1^{[\mathrm {B}]_{k}} \cup (\Phi /[\mathrm {B}]^{+}_{H}) \cup \{\sim \!\psi \}\right) \) and \(\Psi _1\) is reachable from \(\Psi _0\) in \({\mathcal {G}}_H((\Phi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\) with path \(\Gamma _0,\ldots ,\Gamma _n\) such that if \(n = 0\), then \(j \ne k\), and if \(n \ge 1\), then there exists \(j_n \in H \setminus \{k\}\) such that \(\Gamma _{n-1}\) and \(\Gamma _{n}\) are connected with \(j_n\).

  • or \(\Psi _1\) is reachable from \(\Psi _0\) in \({\mathcal {G}}_H((\Phi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\) and there exists \(k \in G \setminus H\) such that either Algorithm 2 returns \(\texttt {sat}\) on the input \(\bigwedge \left( \Psi _1^{[\mathrm {B}]_{k}} \cup \Phi ^{[\mathrm {B}]^{+}_{H \cup \{k\}}} \cup \{\sim \!\psi \} \right) \) or Algorithm 2 returns \(\texttt {sat}\) on \(\bigwedge \left( \Psi _1^{[\mathrm {B}]_{k}} \cup \Phi ^{[\mathrm {B}]^{+}_{H \cup \{k\}}} \cup \{\psi ,\lnot [\mathrm {B}]^{+}_{G}\psi ,[\mathrm {B}]_{k}\psi ,\lnot [\mathrm {B}]_{k}[\mathrm {B}]^{+}_{G}\psi \}\right) \)

To check reachability, Function 10 is used. Given sets of formulas \(\Phi _1\), \(\Psi \) and \(\Phi _2\), sets \(H \subseteq {\mathcal {A}}\) and \(F \subseteq H\), \(p \in H\) and \(K \ge 0\), Function 10 checks if there exists a set of formulas \(\Gamma \in {\mathcal {S}}(\Phi _1)\) such that Algorithm 2 returns \(\texttt {sat}\) on input \(\bigwedge \Gamma \) and \(\Phi _2\) is reachable from \(\Gamma \) in \({\mathcal {G}}_{H}(\Phi _1)\) in at most \(2^K - 1\) steps with a path \(\Gamma _0,\ldots ,\Gamma _n\) such that if \(n = 0\), then \(p \notin F\) and if \(n \ge 1\), then there exists \(j_n \in H \setminus F\) such that \(\Gamma _{n-1}\) and \(\Gamma _n\) are connected with \(j_n\). The set F with which the algorithm is called will always be either \(\varnothing \) or a singleton. It is used to forbid, in certain situations, one of the possible connections between the last two sets in the constructed sequence.

figure i

The idea of the algorithm is based on the idea of Savitch’s algorithm for checking reachability in graph that uses quadratic logarithmic space with respect to |V| (c.f. Papadimitriu 1994). Notice that all the sets in \({\mathcal {S}}((\Psi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\) have the same number of elements and if \(\Gamma \in {\mathcal {S}}((\Psi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\), then \(|{\mathcal {S}}((\Psi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})| \le 2^{|\Gamma |}\). Thus to check reachability in \({\mathcal {G}}_H((\Phi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\) it is enough to check whether there is reachability in at most \(2^{|\Gamma |} - 1\) steps, where \(\Gamma \in {\mathcal {S}}((\Psi / [\mathrm {B}]^{+}_{H}) \cup \{\psi \})\).

The procedure of marking nodes remains as in Algorithm 1, however the notion of unresolved formula used by it is different in the case of formulas of the form \(\lnot [\mathrm {B}]^{+}_{G}\psi \). The modification is related to the fact that in the case of any \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \)-Successor state s and any formula \(\lnot [\mathrm {B}]_{k}[\mathrm {B}]^{+}_{G}\psi \in L(s)\) with \(k \ne j\), either the \(\lnot [\mathrm {B}]_{k}[\mathrm {B}]^{+}_{k}\psi \)-successor of s is created or Function 9 is used to check the satisfiability of \(L^{\lnot [\mathrm {B}]_{k}}(s,[\mathrm {B}]^{+}_{G}\psi )\). Hence the only situation in which such a formula can be unresolved is when \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \in L(s)\) and all the other formulas from \(\lnot \widetilde{\mathrm {Cl}}([\mathrm {B}]^{+}_{G}\psi )\) appear positively in L(s). Unresolved formula of the form \(\lnot [\mathrm {B}]^{+}_{G}\psi \) is defined as follows.

Definition 17

(Unresolved formula) Let n be a node in a pre-tableau and let \(\lnot [\mathrm {B}]^{+}_{G}\psi \in L(n)\). A formula \(\lnot [\mathrm {B}]^{+}_{G}\psi \) is unresolved in n if one of the following holds:

  • n is an internal node and a \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \)-descendant with \(j \in G\), none of its successors is marked \(\texttt {sat}\), there exists a successor of n marked \(\texttt {undec}\) and \(B(n) \ne \varnothing \),

  • n is a state and a \(\lnot [\mathrm {B}]_{j}[\mathrm {B}]^{+}_{G}\psi \)-Successor with \(j \in G\), \(B(n) \ne \varnothing \), \([\mathrm {B}]_{k}[\mathrm {B}]^{+}_{G}\psi \in L(n)\), for all \(k \in G \setminus \{j\}\), and \([\mathrm {B}]_{k}\psi \in L(n)\), for all \(k\in G\).

figure j

We show first that for any input \(\Phi \subseteq {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) Algorithm 2 terminates. Notice that Lemma 6 stating the bounds on the length of the sequence of \(\mathrm {I}\)-successors in the pre-tableau holds for Algorithm 2 as well, as it uses the same procedure of \(\mathrm {I}\)-successors creation as Algorithm 2. The procedure of \(\mathrm {B}\)-successors creation is changed in Algorithm 2 and the following lemma, stating the bounds on the length of a sequence of \(\mathrm {B}\)-successors, can be shown.

Lemma 8

The maximal length of sequence of \(\mathrm {B}\)-Successors with unchanged modal depth of labels in the pre-tableau constructed by Algorithm 2 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2|{\mathcal {A}}|}\right) \).

The following lemma states bounds on the state height of a pre-tableau constructed by Algorithm 2 for an input formula satisfying modal context restriction \({\mathbf{R}}_1\).

Lemma 9

State height of the pre-tableau constructed by Algorithm 2 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1}\right) \) and its height is \(\le {\mathcal {O}}\!\left( |\varphi | \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1}\right) \).

The two lemmas above imply that Algorithm 2 terminates for any input satisfying modal context restriction \({\mathbf{R}}_1\), as stated below.

Proposition 6

For any input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) Algorithm 2 terminates.

Algorithm 2 is sound and valid, as stated below.

Proposition 7

A formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is satisfiable iff Algorithm 2 returns \(\texttt {sat}\) on the input \(\varphi \).

The following theorem states lower and upper bounds on the complexity of the \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\). To show the theorem we have to show that Algorithm 2 can be executed in space polynomial with respect to the size of the input formula.

Theorem 4

The \({\textsc {Team}}{\textsc {Log}}\) satisfiability problem for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) is PSPACE complete.

The tableau constructed in proof of Proposition 7 can have exponential depth with respect to the input formula. For that reason an algorithm similar to that used in proof of Theorem 3 for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\) with modal depth bounded by a constant that works in polynomial time cannot be used in the case \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) with modal depth of formulas bounded by a constant. In fact finding such an algorithm may be very difficult, as the satisfiability problem for formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) with modal depth bounded by 2 is PSPACE hard, as stated below.

Theorem 5

The problem of checking \({\textsc {Team}}{\textsc {Log}}\) satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}}\) with modal depth bounded by 2 is PSPACE complete.

6.3 Restriction \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\)

To check \({\textsc {Team}}{\textsc {Log}}\) satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) Algorithm 1 can be used. We will show that the algorithm terminates on any input from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) and that its state height is bounded by a polynomial depending on modal depth of the input formula. We start with a lemma stating bounds on the length of a sequence of \(\mathrm {B}\)-Successors with unchanged modal depth of labels in the pre-tableau.

Lemma 10

The maximal length of sequence of \(\mathrm {B}\)-Successors with unchanged modal depth of labels in the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2|{\mathcal {A}}|}\right) \).

Notice that the bounds given in Lemma 10 are of the same order as in the case of restriction \({\mathbf{R}}_2\). However, the constant factor in the case of restriction \({\mathbf{R}}_{1(c)}\) is different. This factor depends exponentially on c, as all the maximal consistent subsets of the sets of literals of the form \([O]_{j}\) with \(O \in \{\mathrm {G},\mathrm {I}\}\) and \(j \in G\) in direct scope of operators \([\mathrm {B}]^{+}_{G}\) in the labels of sets may need to be enumerated until the branch expansion stops.

Bounds on the state height of the pre-tableau constructed by Algorithm 1 for an input formula satisfying modal context restriction \({\mathbf{R}}_{1(c)}\) are stated in the lemma below.

Lemma 11

State height of the pre-tableau constructed by Algorithm 1 for an input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is \(\le {\mathcal {O}}\!\left( \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1}\right) \) and its height is \(\le {\mathcal {O}}\!\left( |\varphi | \mathrm {dep}(\varphi )^{2 |{\mathcal {A}}| + 1}\right) \).

Since the size of a pre-tableau constructed by the algorithm for an input formula from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is bounded so Algorithm 1 terminates on any input \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\).

Proposition 8

For any input formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) Algorithm 1 terminates.

Since Algorithm 1 terminates on any input from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) so, by Proposition 4, it is also sound and valid for checking the \({\textsc {Team}}{\textsc {Log}}\) satisfiability of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\), as stated in the proposition below.

Proposition 9

A formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is satisfiable iff Algorithm 1 returns \(\texttt {sat}\) on the input \(\varphi \).

Moreover, since the state height of the pre-tableau constructed by Algorithm 1 for an input formula \(\varphi \in {{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is bounded by a polynomial depending on \(\mathrm {dep}(\varphi )\) and c, so the size of the pre-tableau constructed for \(\varphi \) in proof of Lemma 11 on the basis of this pre-tableau has the size which is bounded by a polynomial depending on \(|\varphi |\) with the degree depending on \(\mathrm {dep}(\varphi )\). Thus we have the following theorem, analogous to Theorem 3 for \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{2 }}}\). Proof of the theorem is analogous to proof of Theorem 3.

Theorem 6

For any fixed k, if modal depth of formulas from \({{\mathcal {L}}}^{\mathrm {T}}_{\mathbf{R _\mathbf{1 }}(c)}\) is bounded by k, then the \({\textsc {Team}}{\textsc {Log}}\) satisfiability for them is NPTIME complete.

7 Conclusions

In this paper we presented a family of language restrictions for modal logics for multiagent systems that can be used to provide NPTIME solvability of the satisfiability problem. The family of restrictions, called modal context restrictions, generalizes modal depth restriction. We applied two restrictions of this kind to one of existing multiagent formalisms called \({\textsc {Team}}{\textsc {Log}}\) and studied the complexity of the satisfiability problem of the restricted language of the formalism.

Like in the case of other modal formalisms for multiagent systems with fixpoint modalities, such as common beliefs or mutual intentions, the satisfiability problem of \({\textsc {Team}}{\textsc {Log}}\) is EXPTIME complete even if the modal depth of formulas is bounded by 2. For this reason a language restriction which would be less forbidding than modal depth restriction would be needed to reduce the complexity of the problem. In the paper we introduced three restrictions called \({\mathbf{R}}_1\), \({\mathbf{R}}_2\) and \({\mathbf{R}}_1(c)\). In the case of the least restrictive one of them, called \({\mathbf{R}}_1\), the problem remains PSPACE hard even if modal depth of formulas is bounded by 2. In the case of the most restrictive one, called \({\mathbf{R}}_2\), combining it with restricting modal depth of formulas by a constant results in NPTIME completeness of the satisfiability problem. Since restriction \({\mathbf{R}}_2\) is too strong, at least in situations when aspects of multiagent systems combining informational and motivational attitudes are specified, like for example collective intentions and collective commitments, we proposed a refinement of restriction \({\mathbf{R}}_1\) called \({\mathbf{R}}_1(c)\). Combining this restriction with restricting modal depth of formula results in NPTIME solvability of the satisfiability problem.

The restrictions of the language studied in this paper do not lead to tractable fragments of the formalisms considered. However, we were able to find NPTIME complete fragments, even in the case of full \({\textsc {Team}}{\textsc {Log}}\), which originally has EXPTIME complete satisfiability problem. Two possible approaches could be undertaken to address this issue: reducing the satisfiability of the NPTIME complete fragments to some other NPTIME complete problems for which well performing, heuristics based algorithms exist, or studying further restrictions of the language that could lead to PTIME solvable satisfiability problem. The first of these approaches was successfully used by Kacprzak, Lomuscio and Penczek in Kacprzak et al. (2004a, 2004b), where model checking of temporal modal logic is studied. The authors reduce this problem to the problem of satisfiability of propositional calculus (SAT) and use existing SAT-solvers for it. Applying a similar approach to the NPTIME complete fragments of \({\textsc {Team}}{\textsc {Log}}\) could be a promising direction for further research. For the second approach, different language restrictions that were already studied in the literature could be considered. Firstly, it would be interesting to investigate the Horn fragment of \({\textsc {Team}}{\textsc {Log}}\). In Nguyen (2000) Linh Nguyen studied Horn fragments of various basic multimodal logics and he found out that when modal depth of formulas is bounded by a constant, then the satisfiability problem is PTIME complete in the case of several of standard multimodal logics.

Another possibility would be to look at restrictions of propositional operators used in formulas, following the approach of Bauland et al. (2006) and Bauland et al. (2009) and to \(\hbox {CTL}^{*}\) and CTL in Meier et al. (2008).