Doxastic logic: a new approach

ABSTRACT In this paper, I develop a new set of doxastic logical systems and I show how they can be used to solve several well-known problems in doxastic logic, for example the so-called problem of logical omniscience. According to this puzzle, the notions of knowledge and belief that are used in ordinary epistemic and doxastic symbolic systems are too idealised. Hence, those systems cannot be used to model ordinary human or human-like agents' beliefs. At best, they can describe idealised individuals. The systems in this paper can be used to symbolise not only the doxastic states of perfectly rational individuals, but also the beliefs of finite humans (and human-like agents). Proof-theoretically, I will use a tableau technique. Every system is combined with predicate logic with necessary identity and ‘possibilist’ quantifiers and modal logic with two kinds of modal operators for relative and absolute necessity. The semantics is a possible world semantics. Finally, I prove that every tableau system in the paper is sound and complete with respect to its semantics.


Introduction
In this paper, I develop a new set of doxastic logical systems that include two doxastic operators B and C, two kinds of modal operators for relative and absolute necessity, 'possibilist' quantifiers and an identity sign for necessary identity. B and C are sentential operators that take individual terms and sentences as arguments and give sentences as values. The formal sentence B d A reads 'individual d believes that it is the case that A' and the formal sentence C d A reads 'it is doxastically conceivable for d that A', or 'it is doxastically imaginable to d that A' or 'it is doxastically possible for d that A'. Proof-theoretically, I will develop a set of indexed tableau systems. The meaning of the various symbolic expressions in our systems is described by a kind of possible world semantics. Finally, I prove that every tableau system in the paper is sound and complete with respect to its semantics.
Epistemic and doxastic logic has often been developed as a kind of modal logic (Fagin, Halpern, Moses, & Vardi, 1995;Hintikka, 1962;. There are, however, several well-known problems with various standard epistemic and doxastic systems, for instance, the so-called problem of logical omniscience. It is, therefore, plausible to explore some alternatives to this approach. The systems in this paper avoid many of the problems with the standard systems without totally abandoning the basic modal paradigm. 1 There are many good philosophical reasons to be attracted to the systems developed in this paper. I cannot discuss every possible argument, but I will briefly go through three of the most obvious ones. First, they can be used to solve the problem of logical omniscience. Second, they can be used to explain the validity of many intuitively valid arguments that cannot be proved in standard doxastic logic. Third, they solve the puzzles in a rather parsimonious way: we do not have to introduce any new entities, such as impossible worlds.
Reason 1 suggests that standard doxastic systems are too strong, and Reason 2 suggests that they are too weak. This indicates that the standard systems are in deep trouble. One can avoid problem 1 by making the standard systems weaker. However, then they might become too weak. One can avoid problem 2 by making the standard systems stronger, but then they might become too strong. The solution in this paper seems more promising.
Reason 1: The problem of logical omniscience. The problem of logical omniscience is a well-known puzzle in epistemic and doxastic logic (see Meyer & van der Hoek, 1995, pp. 71-89, for an introduction). According to this problem, the notions of knowledge and belief that are used in ordinary epistemic and doxastic symbolic systems are too idealised. All the following rules of inference hold, for example, in most standard systems (B i A reads 'individual i believes that A'): If A ↔ C is a theorem, then B i A ↔ B i C is a theorem (for every i and A and C) (Belief of equivalent formulas).
If A → C is a theorem, then B i A → B i C is a theorem (for every i and A and C) (Closure under valid implication).
If A is a theorem, then B i A is a theorem (for every i and A) (Belief of valid formulas). If doxastic logic is supposed to be modelling human or human-like agents (nonhuman animals, computers, robots, etc.), then all these inference patterns are unreasonable. It is unlikely that an organism such as a human could ever believe every valid sentence no matter how complex it is. It is also implausible to assume that a human-like agent's beliefs are closed under valid implication; that is, it is unreasonable to believe that she believes every logical consequence of what she believes no matter how complicated it is. Furthermore, it does not seem to be the case that for every A and B that are logically equivalent, if the agent believes A, she also believes B, and vice versa, no matter how complicated these assertions are. That the rules of inference above hold in standard doxastic systems is a consequence of the fact that doxastic (and epistemic) logic traditionally has been developed as a form of normal modal logic. The first rule (Belief of equivalent formulas) is a problem also for many non-normal systems. Several theorems that can be proved in (most) standard systems might also be problematic. Here are some examples: In our systems, none of the rules of inference and none of the formulas above hold. So, we can solve the problem of logical omniscience. From the fact that someone believes something, almost nothing of interest follows. Nevertheless, the inference rules and the formulas above might seem reasonable if we consider a perfectly rational (reasonable or wise) individual. We will, for example, see that a sentence of the following kind Rc → ¬(B c A ∧ B c ¬A) (if c is perfectly rational, then it is not the case that c believes that A and c believes that not-A) holds in certain systems. In other words, in those systems we can prove that no perfectly rational (wise, reasonable) individual has inconsistent beliefs. It might perhaps also be the case that no one, in fact, has any inconsistent beliefs at some time. But even if this were true, it seems unreasonable to assume that it is a logical truth that no one has inconsistent beliefs. Something can be contingently true without being logically true, and it does not seem to be a task for a doxastic logician to decide whether there are any contingent truths of this kind. In any case, the fact that we can use the systems in this paper to solve the problem of logical omniscience is a good reason to study them.
There are many other possible solutions to the problem of logical omniscience. One can use classical modal logic and not normal modal logic as a model for doxastic logic, one can treat the belief-operator as a possibility-operator and not as a necessity-operator, one can make a distinction between implicit and explicit beliefs or between awareness and explicit beliefs, one can introduce the notions of local reasoning and opaque knowledge and beliefs, and principles of implicit beliefs, one can use fusion models or impossible world semantics and one can use various non-modal approaches to doxastic logic to try to solve this problem. 2 It is beyond the scope of this paper to discuss all solutions that have been suggested in the literature in detail. However, every solution that I am aware of seems to me to be either intuitively too strong or intuitively too weak or simultaneously both too strong and too weak. A system is too strong if we can prove too much in it, that is, if we can prove things in this system that are counterintuitive; and it is too weak, if we can prove too little in it, that is, if we cannot prove everything that we want to be able to prove. For example, classical systems seem to be too strong, since Belief of equivalent formulas (see above) still holds in such systems, and systems that use impossible worlds often appear to be too weak. It seems to be intuitively plausible that a perfectly rational individual will not believe any contradiction. However, in systems based on some kind of impossible world semantics we cannot show this. Of course, our intuitions are not infallible, but if we can construct a system that is consistent with our intuitions such a system appears to be preferable to systems that have counterintuitive consequences. Some of the solutions also postulate various new kinds of entities that might be ontologically problematic, for example impossible objects and impossible worlds. The systems in this paper are both partly weaker and partly stronger than many standard systems. Several intuitively problematic sentences and arguments that are valid in standard systems are not valid in our systems, and several intuitively plausible sentences and arguments that are not valid in standard systems are valid in our systems. Consequently, we can avoid many problems with classical doxastic logic and with many other solutions to the problem of logical omniscience. This is a good reason to be interested in the results in this paper.

Reason 2:
The problem of intuitively valid arguments. There are many arguments that are intuitively valid which cannot be proved in standard doxastic logic. Here is one example: The unmarried teacher argument Every student believes that if the teacher is a bachelor then the teacher is an unmarried man.
Every student believes that the teacher is a bachelor. Susan is a student. Hence, If Susan is perfectly rational (wise), she believes that the teacher is an unmarried man.
This argument is intuitively valid. It seems to be impossible that the premises are true and the conclusion false. Informally, we can reason as follows. From the first and the third premise, it follows that Susan believes that if the teacher is a bachelor, then the teacher is an unmarried man. From the second and the third premise, it follows that Susan believes that the teacher is a bachelor. Hence, Susan believes both that if the teacher is a bachelor then the teacher is an unmarried man and that the teacher is a bachelor. If the contents of Susan's beliefs are true, it follows that the teacher is an unmarried man. Hence, if she is perfectly rational, she believes that the teacher is an unmarried man. Yet, in standard doxastic systems, we cannot prove that the unmarried teacher argument is valid. In standard doxastic logic, we cannot even quantify over believers in any natural way. 3 In general, we cannot adequately symbolise the expressions 'everyone who is such and such believes that' and 'someone who is such and such believes that' in such systems. However, in every system in this paper, we can symbolise these phrases. In Section 7, I will prove that the conclusion follows from the premises in the unmarried teacher argument. Since there are countless other arguments of this kind, this is a good reason to be attracted to the systems in this paper.
Let us now consider an argument that is intuitively invalid. The conscientious student argument (1) Every student believes that if she studies hard she deserves a good grade.
(2) Every student believes that she studies hard. Hence, Every student believes that she deserves a good grade.
If it is true that it is true that if x studies hard x deserves a good grade and it is true that x studies hard, then it is true that x deserves a good grade (for every x). However, if someone is not perfectly rational, she might not have thought about the matter and believe that if she studies hard then she deserves a good grade and also believe that she studies hard even though she does not believe that she deserves a good grade. Someone might be filled with 'irrational' self-doubt. So, the conscientious student argument is intuitively invalid. It seems to be possible that the premises are true and the conclusion false. In Section 7, I will show that the argument is invalid in the class of all models and I will verify this claim by constructing a countermodel. This example will illustrate how we can use the semantic tableau method to generate countermodels and to prove that an argument (or sentence) is invalid.
Reason 3: The problem of economy. Finally, the systems in this paper solve the puzzles above in a rather parsimonious and conservative way. We do not have to introduce any new entities such as impossible worlds, and we do not have to abandon the modal paradigm completely. The results in this paper are conservative in the sense that every system is an extension of classical propositional logic: there are no truth-value gaps and there are no truth-value gluts; every (closed) sentence is either true or false (in a world) and there are no true contradictions. Moreover, the tableau rules for the possibilist quantifiers and the modal operators are classical. The primitive quantifiers are, in effect, used to quantify over absolutely everything, including merely possible objects (if there are any). So, in every possible world they vary over all possible objects, not only over all the things that happen to exist at this world. However, we also define a pair of actualist quantifiers that have existential import (see Definition 2.1). Furthermore, when it comes to the doxastic part of the systems, we do not have to abandon the modal paradigm completely. When the denotation of 'a' is not perfectly rational in a world, 'B a A' and 'C a A' behave as if they were ordinary predicates in this world; but when the denotation of 'a' is perfectly rational in a world, 'B a ' and 'C a ' behave as if they were ordinary modal operators in this world at this time. Consequently, if a is not perfectly rational, almost nothing of interest follows from the proposition that a believes something. However, if we assume that a is perfectly rational, we can derive all sorts of interesting consequences from this proposition. Exactly what it means to be perfectly rational and exactly what follows from the claim that a perfectly rational individual believes something seems to be a matter of choice, and in different systems we can derive different consequences. 4 All in all, I conclude that we have very good reasons to be interested in the systems presented in this paper.
The paper is divided into seven main sections. Section 2 deals with the syntax and Section 3 with the semantics of our systems. In Section 4, I describe the proof theory of our logics and Section 5 includes some examples of theorems. Section 6 contains soundness and completeness proofs for every system and Section 7 includes two examples of derivations in doxastic logic.

Predicates
(iii) For every natural number n > 0, n-place predicate symbols P 1 n , P 2 n , P 3 n . . .. (iv) The monadic existence predicate E and the monadic rationality predicate R.
(viii) The doxastic operators B and C.
Parentheses (x) The brackets ), (. I will use x, y and z . . . for arbitrary variables, a, b, c . . . for arbitrary constants, and s and t for arbitrary terms (with or without primes or subscripts). F n , G n , H n . . . stand for arbitrary n-place predicates. The subscript will be omitted if it can be read off from the context.

Language
I will use the following language, L, in this paper: (i) Any constant or variable is a term. (ii) If t 1 , . . . , t n are any terms and P is any n-place predicate, Pt 1 . . . t n is an atomic formula. (iii) If t is a term, Et ('t exists') is an atomic formula and Rt ('t is perfectly rational [reasonable, wise]') is an atomic formula. (iv) If s and t are terms, then s = t ('s is identical with t') is an atomic formula.   A, B, C, D . . . stand for arbitrary formulas, and , . . . for finite sets of closed formulas. The concepts of bound and free variables, and open and closed formulas, are defined in the usual way. 5 (A)[t/x] is the formula obtained by substituting t for every free occurrence of x in A. The definition is standard. Note that substitutions are performed also within the scope of the doxastic operators. Brackets around formulas are usually dropped if the result is not ambiguous.

Definition 3.1 (MODEL):
A model M is a relational structure D, W, R, D, v , where D is a non-empty set of individuals (the domain), W is a non-empty set of possible worlds, R is a binary alethic accessibility relation (R is a subset of W × W), D is a ternary doxastic accessibility relation (D is a subset of D × W × W), and v is an interpretation function.
R is used to define the truth conditions for sentences that begin with the alethic operators and ♦, and D is used to define the truth conditions for sentences that begin with the doxastic operators B and C. Informally, Rωω says that the possible world ω is alethically (relatively) accessible from the possible world ω, and Dδωω that the possible world ω is doxastically accessible to the individual δ from the possible world ω, or that δ can see ω from ω.
Every constant in our language is a kind of rigid designator, it refers to the same individual in every possible world. In other words, the valuation function v assigns every constant c an element v(c) of D. The extension of a predicate, however, may change from world to world and it may be empty in a world, that is, v assigns every possible world ω in W and n-place predicate P a subset v ω (P) (the extension of P in ω) of D n . We shall say that v ω (P) is the set of n-tuples that satisfy P in the world ω.
The predicate R has a special meaning in our systems. 'Rc' says that c is perfectly rational, perfectly reasonable or perfectly wise. Exactly what these expressions mean can, of course, be debated. By imposing various conditions on the doxastic accessibility relation D (Section 3.3), we obtain several different interpretations of the predicate R. Consistency is one prima facie plausible condition, but we might also want to include some very strong properties in our concept of perfect wisdom (rationality), for example infallibility or doxastic omniscience. If v (c) is in the extension of R at the possible world ω, this means that v(c) is perfectly rational, reasonable or wise in ω. R functions as an ordinary predicate. Hence, an individual δ may be in R's extension in one possible world even though δ is not in R's extension in every possible world. Consequently, the fact that an individual δ is perfectly rational, reasonable or wise in a possible world does not entail that δ is perfectly rational, reasonable or wise in every possible world. We can, if we want, add the extra assumption that every perfectly rational individual is necessarily perfectly rational to any system in this paper (see the semantic condition C−UR (Table 4) in Section 3.3.4). R plays an important role in our systems, as should be obvious from the definitions of the truth conditions for sentences of the forms B a A and C a A (see Section 3.2 below). In Section 4.2, we will see that R also plays an important role in our various tableau rules. 6 The valuation function assigns extensions to so-called matrices. Given any closed doxastic formula of the form B t A or C t A, we shall construct its matrix as follows. Let m be the least number greater than every n such that x n occurs bound in A. From left to right, replace every occurrence of an individual constant with x m , x m+1 , etc. The result is the formula's matrix.
. . , a n /x 1 , . . . , x n ] is the result of replacing x 1 by a 1 , and ··· , and x n by a n in A. A[a 1 , . . . , a n /x 1 , . . . , x n ] will be abbreviated as A[a 1 , . . . , a n / → x ]. If M is any matrix of the form B t A or C t A with free variables x 1 , . . . , x n , then v ω (M) ⊆ D n . Note that M always includes at least one free variable. 7 Let M be a matrix where x m is the first free variable in M and a m is the constant in M[a 1 , . . . , a n / → x ] that replaces x m . Then the truth conditions for closed doxastic formulas of the form M[a 1 , . . . , a n / → x ], when v ω (Ra m ) = 0, are defined in terms of the extension of M in ω (see condition (ii) in Section 3.2 below). v ω (=) = { d, d : d ∈ D}, i.e. the extension of the identity predicate is the same in every possible world (in a model). This means that all identities (and non-identities) are both absolutely and relatively necessary. The existence predicate E functions as an ordinary predicate. The extension of this predicate may vary from one world to another. Ec is true in a possible world intuitively means that v (c) exists in this world.
Let M be a model. Then the language of M, L(M), is obtained by adding a constant k d such that v(k d ) = d to the language for every member d ∈ D. Thus, every object in the domain of a model has at least one name in our language while several different constants may refer to one and the same object.

Truth conditions
Let us consider the truth conditions for some sentences in our language. (1 = True and 0 = False; '∀ω ∈ W' is read as 'for all possible worlds ω in W'; and '∃ω ∈ W' is read as 'for some possible world ω in W'. The truth conditions for the omitted sentences are standard.) We extend the interpretation function so that every closed formula A is assigned exactly one truth-value v ω (A) in each world ω.
Let M be a matrix where x m is the first free variable in M and a m is the constant in M[a 1 , . . . , a n / → x ] that replaces x m . Then the truth conditions for closed doxastic formulas of the form M[a 1 , . . . , a n / → x ], when v ω (Ra m ) = 0, are given in (ii) below.
(ii) v ω (M[a 1 , . . . , a n / is not an element in v ω (R), then C a A is assigned a truth-value in ω in a way that does not depend on the value of A (see condition (ii) above).
and are substitutional quantifiers. Nevertheless, we can also call them 'possibilist' because they, in effect, vary over every object in the domain and the domain is the same in every possible world. Intuitively, conditions (xi) and (xii) mean the following. If v(a) is not perfectly rational in a possible world, B a A and C a A behave as if they are ordinary predicates in this world; and if v(a) is perfectly rational in a possible world, B a and C a behave as ordinary modal operators in this world.
Here is an example to help explain condition (ii) above. Consider the closed doxastic formula B c Pcd. The matrix of this formula is B x 1 Px 2 x 3 (note that the first occurrence of c is replaced by x 1 and the second by The interpretation function v assigns extensions to matrices in possible worlds. So, v(c), v (c) We are now in a position to define some important semantic concepts.

Conditions on models
In this section, I will introduce some conditions that can be used to classify different kinds of models. The conditions concern the formal properties of the accessibility relations, the relationships between the various accessibility relations and the relationships between the accessibility relations and the valuation function. Table 1 includes information about the alethic accessibility relation. The conditions in this table are mentioned in most introductions to modal logic (see, for example, Blackburn, de Rijke, & Venema, 2001;Chellas, 1980;Fitting & Mendelsohn, 1998;Garson, 2006;Priest, 2008). Given almost any concept of necessity, it is reasonable to assume that the alethic accessibility relation is an equivalence relation and, hence, that it satisfies all conditions in Table 1. Some clauses (C−dD, C−dT and C−d4) in Table 2 have been mentioned in the literature (Fitting, Thalmann, & Voronkov, 2001). All other conditions are new. By combining the clauses in this section in various ways, we can generate many different doxastic systems. Exactly which conditions we should accept will depend on what we mean or should mean by 'perfectly rational' (reasonable, wise). There might be good reasons to accept all (or almost all) conditions in this section. Having said that, it might also be interesting to see what follows if we accept some smaller class. The more conditions we accept, the more content we pack into the concept of rationality (wisdom). Perhaps we should distinguish between different concepts of rationality and talk about 'rationality 1 ', 'rationality 2 ', etc., and not just about 'rationality'. Then, we can use different systems to explicate these different notions. Even though the conditions in this section should be more or less self-explanatory, I have added a few comments about some of the new clauses. There are many interesting relationships between the various conditions that I do not have space to discuss in this paper. I will, however, mention a few connections. Some combinations may be more philosophically interesting than others. Some combinations might be philosophically implausible. Table 1. Conditions on the relation R.

Condition
Formalisation of condition Table 2. Conditions on the relation D.

Condition
Formalisation of condition Table 2 are similar to the conditions in Table 1, and to some well-known clauses that are often used in epistemic and doxastic logic. Nevertheless, there are also some important differences. R is a binary relation, while D is a ternary relation. 'C' in 'C−dD' stands for 'condition' and 'd', for 'doxastic'. C−dD is called 'C−dD' because it is similar to the well-known condition D (as in 'Deontic') in ordinary alethic (modal) logic. Similar remarks apply to the other clauses in this section. If it is clear from the context that we are talking about a semantic condition, I will often omit the initial C. The (semantic) condition of (doxastic) omniscience. According to condition C−dO, it holds that for every (individual) δ and for all (possible worlds) ω and ω : if ω is doxastically accessible from ω to δ, then ω is identical to ω . In other words, in every world there is at most one possible world that is doxastically accessible from this world (to an individual), namely this world itself. This condition corresponds to the tableau rule T−dO in Table 10. In any model that satisfies this condition, the principle of (doxastic) omniscience, (Table 17), which says that every perfectly rational (wise) individual believes everything that is true, is valid. It is not plausible to assume that actual human beings have access to every fact whatsoever, but it might be reasonable to assume that perfectly wise individuals are different. 'dO' in 'C−dO' stands for 'doxastic omniscience' since it is reasonable to call an individual who believes everything that is true doxastically omniscient.

Conditions on the relation D Some of the conditions in
The (semantic) condition of infallibility. C−dT says that the doxastic accessibility relation is 'reflexive': every possible world is doxastically accessible from itself (for every individual). This condition corresponds to the tableau rule T−dT in Table 10. In any model that satisfies this condition, the principle of infallibility, (Table 17), is valid. This principle says that everything a perfectly wise individual believes is true. We call C−dT the 'infallibility condition' since it is reasonable to say that an individual who does not believe anything that is false is infallible. Again, this condition does not entail that no one has any false beliefs. We can only conclude that no one who is perfectly rational has any false beliefs.
Together C−dO and C−dT say that every agent has doxastic access to the world where she is and to no other world. If we accept both of these conditions, we can show that every perfectly rational individual is both omniscient and infallible, that is, that the following schema is valid x(Rx → (B x A ↔ A)). This formula says that every perfectly rational being believes that A if and only if A. Note that the following sentences are not valid even if we assume both C−dO and C−dT: are usually not accepted in doxastic logic. Nonetheless, it might be interesting to note that something like x(Rx → (B x A → A)) has a long tradition in philosophy. The Stoics seem to have accepted this proposition, or something very similar. Thus, according to Diogenes Laertius, the Stoics thought that ··· the wise man will never form mere opinion, that is to say, he will never give assent to anything false··· . the wise are infallible, not being liable to error. (See Diogenes Laertius (1925), Book VII,Zeno,p. 227) The concept of doxastic omniscience is a neglected topic in the literature on epistemic and doxastic logic. It is an interesting concept that deserves further investigation. Some might think it is obvious that we should include the notion in the concept of perfect wisdom (rationality); otherwise, how could we call this wisdom perfect? Others might think that it is obvious that perfect rationality does not entail doxastic omniscience since rationality only has to do with internal consistency among beliefs, or because they believe that this condition (together with some other plausible thesis) has unreasonable consequences. 'Facts' do not entail anything about beliefs (not even the beliefs about a perfectly rational or wise individual), and beliefs do not entail anything about 'facts'. If this is true, we should neither accept the condition of doxastic omniscience nor the condition of infallibility. On the other hand, is it not reasonable to call a being that is infallible and omniscient wiser than a being who is not? If it is, how can the latter be perfectly wise (rational)? In the next section, I will consider a 'weaker' condition of omniscience, the so-called semantic condition of necessityomniscience (Section 3.3.3). I will also consider a weaker form of 'infallibility', so-called possibility-infallibility. 8 Table 3 concern the relationship between the doxastic accessibility relation and the alethic accessibility relation. They correspond to the tableau rules in Table 13.

Conditions concerning the relation between R and D The conditions in
because it is similar to the wellknown alethic (modal) condition C−4, and similarly for C−ad5. 'CMP' in 'C − CMP' is an abbreviation of 'Conceivability Must Permutation', and 'BMP' in 'C − BMP' is an abbreviation of 'Belief Must Permutation'.
The (semantic) condition of (doxastic) necessity-omniscience. C − MB says that for every (individual) δ, for every (possible world) ω and for every (possible world) ω , ω is doxastically accessible from ω (to δ) only if ω is alethically accessible from ω. In other words, if C − MB holds, then it is not the case that δ can see ω from ω if ω is not alethically accessible from ω. C − MB corresponds to the tableau rule T − MB in Table 13. In every class of models that satisfies this condition, the principle of (doxastic) necessityomniscience, x(Rx → ( A → B x A)), which says that every perfectly wise individual believes every necessary truth, is valid. If we assume C − MB, we can also establish the validity of the principle of consequence-consistency, (Table 20), which says that every perfectly wise person believes every necessary implication of the things she believes. So, this condition is philosophically Table 3. Conditions concerning the relation between R and D.

Condition
Formalisation of condition quite interesting. The concept of necessity-omniscience is 'weaker' than the concept of omniscience. If someone believes everything that is true, she believes everything that is necessarily true, but the converse does not hold. It is possible to believe everything that is necessarily true without believing everything that is (in fact) true (given that there are things that are true but not necessarily true). The (semantic) condition of possibility-infallibility. According to C − BP, for every (individual) δ, there is for every (possible world) ω a (possible world) ω such that δ can see ω from ω and ω is alethically accessible from ω. In other words, in every possible world, there is at least one possible world that is alethically and doxastically accessible (to δ). C − BP corresponds to the tableau rule T − BP in Table 13. This condition is similar to condition C−dD ( Table 2) In every class of models that satisfy this condition, the principle of possibility-infallibility is valid: In other words, according to this condition, a perfectly wise individual does not believe anything impossible. This is an intuitively interesting principle. If c believes something that is impossible, c's belief will inevitably be false. C − BP is also similar to the semantic condition of infallibility. If everything a person believes is true, then everything she believes is possible (given that everything that is true is possible). However, the converse does not hold. From the fact that everything a person believes is possible, it does not follow that everything she believes is (in fact) true (given that there are things that are possible but that are not true). So, the concept of possibility-infallibility is 'weaker' than the concept of infallibility.
The (semantic) condition of the necessity of beliefs. C−ad4 says that for every (individual) δ, for every (possible worlds) ω, ω and ω , if ω is alethically accessible from ω and ω is doxastically accessible from ω to δ, then ω is doxastically accessible from ω to δ. C−ad4 corresponds to the tableau rule T−ad4 in Table 13. In the class of models that satisfy this condition (and C−UR or C−FTR in Table 4), the principle of the necessity of beliefs holds, x(Rx → (B x A → B x A)), which says that if a perfectly wise individual believes something, it is (relatively) necessary that she believes it. Note that the following formula is not valid, even if we assume C−ad4, x(Rx → (B x A → UB x A)). It is still possible that a perfectly rational being believes something without it being the case that it is absolutely necessary that she believes it.
The (semantic) condition of the necessity of non-beliefs. According to C−ad5, for every (individual) δ, for every (possible worlds) ω, ω and ω , if ω is alethically accessible from ω and ω is doxastically accessible from ω to δ, then ω is doxastically accessible from ω to δ. C−ad5 corresponds to the tableau rule T−ad5 in Table 13. In the class of models that satisfy this condition (and C−UR or C−FTR in Table 4), the principle of the necessity of non-beliefs holds, x(Rx → (¬B x A → ¬B x A)), which says that if a perfectly wise individual does not believe something, it is (relatively) necessary that she Table 4. Conditions on the valuation function v.

Condition
Formalisation of condition does not believe it. We can also show that the principle of the necessity of conceivability is valid in this class of models, x(Rx → (C x A → C x A)), which says that if something is (doxastically) conceivable to a perfectly wise individual, then it is (relatively) necessary that it is conceivable to her. Note that the following formula is not valid, even if we assume C−ad5, x(Rx → (¬B x A → U¬B x A)). It is still possible that a perfectly rational being does not believe something without it being the case that it is absolutely necessary that she does not believe it. Nor is the following formula a theorem Other (semantic) conditions. C − CMP corresponds to the tableau rule T − CMP in Table 13 and C − BMP to the tableau rule T − BMP in the same table. In the class of models that satisfy C − CMP (and C−UR or C−FTR in Table 4), we can show that the following sentence is valid x(Rx → (C x A → C x A)); and in the class of models that satisfy C − BMP (and C−UR or C−FTR in Table 4) we can show that the following formula is valid x (Rx → (B x A → B x A)). These sentences are kinds of permutation principles.
According to the condition of omniscience as well as the condition of necessityomniscience (in combination with infallibility), wisdom includes more than just inner consistency or perfect reasoning skills; it also involves correct 'factual' beliefs or at least correct beliefs about 'facts' that are (relatively) necessary. For example, you cannot be truly wise without having true beliefs about necessary conditions for various important ends. If we want to pack such 'factual' beliefs (knowledge) into the concept of wisdom, it might be plausible to accept those conditions. If we want a weaker notion of wisdom or rationality, we can omit them and concentrate on conditions that have to do with the consistency of an individual's beliefs; we can, for example, accept condition C−dD. As already pointed out, it might be the case that we want to use different analyses of the concept of wisdom (rationality) for different purposes.

Conditions on the valuation function v in a model
The semantic conditions C−FTR and C−UR (Table 4) correspond to the tableau rules T−FTR and T−UR (Table 12), respectively. It follows from C−UR that every perfectly rational individual is necessarily perfectly rational (See Section 4.2.8, Table 12, for more on this). 9 Finally, to show how one can justify the claims in this section, I will verify the proposition that the principle of (doxastic) necessity-omniscience, x(Rx → ( A → B x A)), is valid in the class of all models that satisfy C − MB. Suppose that x(Rx → ( A → B x A)) is not valid in the class of all models that satisfy C − MB. Then x(Rx → ( A → B x A)) is false in some possible world ω in some model M in this class. Hence, for some there is a possible world ω in M that is doxastically accessible from ω to c in M in which A is false. It follows that ω is alethically accessible from ω in M, for ω is doxastically accessible from ω to c in M and M satisfies C − MB. Therefore, A is true in ω in M, since A is true in ω in M and ω is alethically accessible from ω in M. But this is absurd. Hence, our assumption cannot be true. It follows that x(Rx → ( A → B x A)) is valid in the class of all models that satisfy C − MB. Q.E.D.

Model classes and the logic of a class of models
We can use the conditions mentioned in Section 3.3 to obtain a classification of the set of all models into various kinds. Let M(C1, . . . , Cn) be the class of (all) models that satisfy the conditions C1, . . . , Cn. Then, M(C − dT, C − d4, C − d5) is the class of (all) models that satisfy the conditions C−dT, C−d4 and C−d5, etc.
We can say that the set of all sentences in a language that are valid in a class of models M is the (logical) system of M, or the logic of M, in symbols S(M). For example, ) is the class of sentences (in our language) that are valid in the class of (all) models that satisfy the conditions C−dT, C−d4 and C−d5.
We can define a large set of systems by using this classification of model classes. In the next section, I will introduce a set of semantic tableau systems that exactly correspond to these semantically defined logics.

Semantic tableaux
In Section 4, I will introduce a set of tableau rules. Then, I will show how these rules can be used to construct a large set of non-equivalent tableau systems. All systems are extensions of ordinary propositional logic. All systems also include rules for a pair of 'possibilist' quantifiers, a pair of 'absolute' necessity operators and a pair of 'relative' necessity operators. The propositional part of the systems is similar to systems introduced by Jeffrey (1967) and Smullyan (1968) , and the modal part is similar to systems discussed by, among others, Priest (2008). For more information about the tableau method and various kinds of tableau systems, see, for example, D'Agostino, Gabbay, Hähnle, and Posegga (1999) and Fitting and Mendelsohn (1998). 10

Tableau rules
In this section, I will introduce a set of tableau rules that can be used to construct a large set of tableau systems (Section 4.3). They should be more or less self-explanatory. Nevertheless, I will briefly discuss some of the new rules. Table 5 4

.2.2. Basic alethic rules (ba-rules)
see Table 6 4.2.3. Basic doxastic rules (bd-rules) see Table 7 Intuitively, 'Rc, i' in the doxastic rules says that the individual denoted by 'c' is perfectly rational, reasonable or wise in the possible world denoted by 'i', and 'iDcj' says that the possible world denoted by 'j' is doxastically accessible to the individual  Table 6. Basic alethic rules. Table 7. Basic doxastic rules.
denoted by 'c' in the possible world denoted by 'i'. The basic doxastic rules hold for every constant c (i.e. c can be replaced by any constant in these rules). Table 8 Note that a and c in the quantifier rules are rigid constants-we never instantiate with variables; a is any constant on the branch and c is a constant new to the branch.

Alethic accessibility rules (a-rules)
The alethic accessibility rules in Table 9 correspond to the semantic conditions in Table 1.

Doxastic accessibility rules (d-rules)
The doxastic accessibility rules in Table 10 correspond to the semantic conditions in Table 2. Table 11 We could use a more restricted CUT rule, CUTR, where 'A' in CUT is replaced by 'Rc' where c is a constant (that occurs as an index to some doxastic operator) on the branch. In fact, in the completeness proofs we will use CUTR and not CUT. Yet, CUT is often more useful in proving many theorems and derived rules.

The CUT-rule (CUT), (CUTR) and the (world) identity rules see
There

are two identity rules: T−IdI and T−IdII (both abbreviated Id). α(i) is a line in a tableau that includes 'i', and α(j) is like α(i) except that 'i' is replaced by
is A, i we only apply the rule when A is atomic or of the form B t A or C t A given that ¬Rt, i is on the branch. Table 9. Alethic accessibility rules. Table 10. Doxastic accessibility rules. Table 11. CUT and (world) identity rules.

CUT T−IdI T−IdII
CUTR (or CUT), T−IdI and T−IdII are included in every system in this paper. Still, in every system that does not include T−dO, the identity rules are redundant and can, in principle, be omitted.

Transfer rules, etc.
'FT' in 'T−FTR' is an abbreviation of 'Forward Transfer', and 'R' in 'T−FTR' and 'T−UR' of 'Rationality'. The tableau rules in Table 12 correspond to the semantic conditions in Table 4.
In every system that includes T−FTR and T − MB (Table 13), we can prove that the following sentence is a theorem x(Rx → B x Rx), which says that everyone who is perfectly rational (wise) believes that she is perfectly rational (wise).
If a system includes T−FTR, T − MB and T−dD (Table 10), we can prove that x(Rx → C x Rx) is a theorem in this system. This formula says that everyone who is perfectly rational is such that it is conceivable to her that she is perfectly rational.
In every system that includes T−UR, we can prove the following sentence x(Rx → URx), which says that every perfectly rational (wise) individual is necessarily perfectly rational (wise).
The transfer rules (the rules in Table 12) are not included in every system. Whether they should be added seems to be a matter of choice. In standard doxastic systems, that every believer is perfectly rational is usually built into the logic. The systems in the present paper are, therefore, more flexible.
where l is new is new where A is of a certain form (see below, Section 4.2.10) Table 14 In the identity rules R stands for 'reflexive', S for 'substitution (of identities)', N for 'necessary identity' and D for '(doxastic) accessibility'. The star in (T − R =) indicates that (T − R =) is a rule without premises; we may add t = t, i to any open branch in a tree. (T − S =) is applied only 'within worlds', and we usually only apply the rule when A is atomic. However, we shall also allow applications of the following kind.

Identity rules see
Intuitively, these formulas are reasonable. It seems to be necessarily true that if Samuel Clemens believes that the evening star is the morning star and Samuel Clemens is Mark Twain, then Mark Twain believes that the evening star is the morning star. In every system that includes (T − N =), we can prove that all identities and non-identities are (absolutely and relatively) necessary, that is, we can prove all of the following theorems: x y(x = y → Ux = y), x y(x = y → x = y), x y(¬x = y → U¬x = y) and x y(¬x = y → ¬x = y). Recall that every constant is treated as a rigid designator in this paper.

Tableau systems and some basic proof-theoretical concepts
A tableau system is a set of tableau rules. A doxastic tableau system (or logic) is a tableau system that includes all propositional rules, all basic alethic rules, all basic doxastic rules, the rules for the possibilist quantifiers, CUTR, T−IdI and T−IdII, and all identity rules. The smallest doxastic logic will be called D. By adding various additional rules, we obtain a large class of stronger doxastic systems.
A tree is a structure that looks, something, like this (for more information on the concept of a tree, see, for example, Smullyan, 1966Smullyan, , 1968: The dots are called nodes and the top node is called the root. Nodes without successors are called tips or leaves. Any path from the root down a series of arrows to a tip is called a branch. A (semantic) tableau is a tree like this where the nodes have the following form: A, i, where A is a formula in L and i ∈ {0, 1, 2, 3, . . .}, or something of the form iRj, iDcj or i = j where i, j ∈ {0, 1, 2, 3, . . .} and c is a constant in L. The arrows in a tree indicate relations among the nodes. Arrows may be omitted if the structure of the tree can be seen without them.
Different tableau systems contain different tableau rules which, intuitively, tell us how to 'extend branches' from given nodes. A

branch in a tableau is closed iff there is a formula A and a number i, such that both A, i and ¬A, i occur on the branch; it is open just in case it is not closed. A tableau itself is closed iff every branch in it is closed; it is open iff it is not closed.
Semantic tableaux can be used to check validity of sentences and logical consequence between sets of sentences and sentences. It can be seen as a systematic search for a model in which the class of every sentence on a branch is satisfiable. If the tableau is closed, there is no such model.
We are now in a position to define some important proof-theoretical concepts. Let S be any system in this paper in the following definitions and let an S-tableau be a tableau generated in accordance with the rules in S.

Definition 4.1:
(i) (PROOF IN A SYSTEM). A proof of A in S is a closed S-tableau for ¬A, 0, that is, a closed S-tableau whose root consists of ¬A, 0. (DERIVATION IN A SYSTEM). A derivation in the system S of B from the set of formulas , is a closed S-tableau whose initial list comprises A, 0 for every A ∈ and ¬B, 0. The sentences in are called the premises of the derivation and B is called the conclusion of the derivation. The initial list of a tableau consists of the first nodes in this tableau whose satisfiability we are testing. (iv) (PROOF-THEORETIC CONSEQUENCE IN A SYSTEM). B is a proof-theoretic consequence of the set of formulas in S or B is derivable from a set of formulas in S ( S B) iff there is a derivation of B in S from , that is, iff there is a closed S-tableau whose initial list comprises A, 0 for every A ∈ and ¬B, 0.

Examples of theorems
In this section, I will mention some sentences that can be proved in various systems. The informal reading of the formulas should be obvious. Every sentence in Table 15 is a theorem in every system in this paper, every sentence in Table 16 is a theorem in every system that includes the tableau rule T−dD, etc. Table 15. Some theorems in every system.
Theorem System Table 16. Some theorems in dD.

Theorem Systems
The principle of doxastic omniscience The principle of doxastic necessity-omniscience The principle of the necessity of non-beliefs The principle of the necessity of conceivability Table 19. Some Barcan-like theorems.
Theorem System Theorem System All of the following sentences (schemas) are theorems in every system in this paper: Tables 17-20

Soundness and completeness theorems
In this section, I will prove that every system in this paper is sound and complete with respect to its semantics. The concepts of soundness and completeness are defined as usual. Proof: The proof is by recursion on formulas; 'the IH' refers to the induction hypothesis.
Suppose that v 1ω (Ra m ) = 0, that M is a matrix where x m is the first free variable in M and that a m is the constant in M[a 1 , . . . , a n / → x ] that replaces x m .
The cases for the other alethic operators are similar. The case for C c B is similar.
The case for the particular quantifier is similar.
The arguments for the other primitive alethic operators are similar.
Either v ω (Rt) = 1 or v ω (Rt) = 0. We have already shown that the result holds if v ω (Rt) = 0. Accordingly, suppose that v ω (Rt) = 1. Since x is the only free variable, t cannot be a variable distinct from x. So, t is either x or a constant. Suppose t is x.
The case for C t is similar.
so the result is trivial. Accordingly, suppose that x and y are distinct. Then, The case for the particular quantifier ( ) is similar.

Soundness theorem
Let M = D, W, R, D, v be any model and B any branch of a tableau. Then B is satisfiable in M iff there is a function f from 0, 1, 2, . . . to W such that If these conditions are fulfilled, we say that f shows that B is satisfiable in M.

Lemma 6.4 (Soundness Lemma): Let B be any branch of a tableau and M be any model. If B is satisfiable in M and a tableau rule is applied to it, then there is a model M and an extension of B, B , such that B is satisfiable in M .
Proof: The proof is by induction on the height of the derivation. Let f be a function that shows that the branch B is satisfiable in M.
Connectives and the modal operators. Straightforward. (B). Suppose that Rc, i, B c C, i, and iDcj are on B, and that we apply the B-rule. Then we get an extension of B that includes C, j. Since B is satisfiable in M, B c C is true in f (i) and Rc is true in f (i). Furthermore, for any i and j such that iDcj is on B, Dv(c)f (i)f (j). Thus by the truth conditions for B c C, C is true in f (j).
(C). Suppose that Rc, i, C c B, i are on B and that we apply the C-rule to get an extension of B that includes nodes of the form iDcj and B, j. Since B is satisfiable in M, C c B is true in f (i) and Rc is true in f (i). Hence, for some ω in W, Dv(c)f (i)ω and B is true in ω [by the truth conditions for C c B and the fact that Rc is true in f (i)]. Let f be the same as f except that f (j) = ω. Since f and f differ only at j, f shows that B is satisfiable in M.

Furthermore, by definition Dv(c)f (i)f (j), and B is true in f (j).
(¬B) and (¬C). Similar. ( ). Suppose that xA, i is on B and that we apply the -rule to get an extension of B that includes a node of the form Accordingly, we can take M to be M.
( ). Suppose that xA, i is on B and that we apply the -rule to get an extension of B that includes a node of the form A [c/x] , v(s) = v(t). Suppose that A is atomic and has the following form: Pa 1 . . . s . . . a n . Then Pa 1 . . . s . . . a n is true in f (i), i.e. v(a 1 ), . . . , v(s), . . . , v(a n ) is an element in P's extension in f (i). Accordingly, v(a 1 ), . . . , v(t), . . . , v(a n ) is an element in P's extension in f (i), i.e. Pa 1 . . . t . . . a n is true in f (i) in M. So, we may take M to be M.
Let M be a matrix where x m is the first free variable in M and a m is the constant in M[a 1 , . . . , a, . . . , a n / → x ] (M[a 1 , . . . , b, . . . , a n / → x ]) that replaces x m . Furthermore, to illustrate, suppose we have a = b, i, M[a 1 , . . . , a, . . . , a n / → x ], i and ¬Ra m , i on the branch (and that a m is not a (b)), and that we apply (T − S =) to obtain an extension of the branch that includes M[a 1 , . . . , b, . . . , a n / It follows that M[a 1 , . . . , b, . . . , a n / → x ] is true in f (i). In conclusion, the extension of B is satisfiable in M. (All other cases are proved similarly.) Theorem 6.5 (Soundness Theorem): Every system S in this paper is sound with respect to its semantics.
Proof: Suppose that B does not follow from in M, where M is the class of models that corresponds to S. Then every premise in is true and the conclusion B false in some world ω in some model M in M. Consider an S-tableau whose initial list consists of A, 0 for every A ∈ and ¬B, 0, where '0' refers to ω. Then the initial list is satisfiable in M. Every time we apply a rule to this list it produces at least one extension that is satisfiable in a model M in M (by the Soundness Lemma). So, we can find a whole branch such that every initial section of this branch is satisfiable in some model M in M. If this branch is closed, then some sentence is both true and false in some possible world in M . Yet, this is impossible. Hence, the tableau is open. Accordingly, B is not derivable from in S. In conclusion, if B is derivable from in S, then B follows from in M.

Completeness theorem
In this section, I will prove that every system in this paper is complete with respect to its semantics. However, first I will introduce some important concepts.
Intuitively, a complete tableau is a tableau where every rule that can be applied has been applied. In this sense, there may be several different (complete) tableaux for the same sentence or set of sentences, some more complex than others, and the tableau rules may be applied in different orders. To produce a complete tableau, we can use the following method (which is usually not the simplest one). (1) For every open branch on the tree, one at a time, start from its root and move towards its tip. Apply any rule that produces something new to the branch (if the application of a rule would result in just repeating lines already on the same branch of the tableau, it should not be applied). For example, is applied at most once to a node of the form xA, i. We do not apply any rules to a branch that is closed. Some rules may have several possible applications, e.g.
and . Then make all applications at once. (2) When we have done this for all open branches on the tree, we repeat the procedure. Some rules introduce new possible worlds, for example T−aD and T − BP. If a rule introduces a new possible world, it is applied once at the tip of every open branch at the end of every cycle when we have moved through all nodes. If a system includes several different rules that introduce new possible worlds (R1, R2 . . .), we alternate between them. The first time we use R1 once; the second time we use R2 once, etc. Before we conclude a cycle and begin to move through all nodes again we split the end of every open branch in the tree and add Rc, i to the left node and ¬Rc, i to the right node, for every constant c (that occurs as an index to some doxastic operator on the tree) and i on the branch. If there is still something to do according to this method, the tableau is incomplete; if not, it is complete.
Proof: The proof is by induction on the complexity of A.
Suppose that M is a matrix where x m is the first free variable and a m is the constant in M[a 1 , . . . , a n / → x ] that replaces x m and that v ω [i] (Ra m ) = 0. Then: Then since the branch is complete, the B-rule has been applied and for every j such that iDcj is on B, D, j is on B. By the induction hypothesis, D is true in every ω [j] such that Dv(c)ω [i] ω [j] . Since Rc, i is on B, v(c) is perfectly rational in ω [i] . It follows that B c D is true in ω [i] , as required.
Suppose that M is a matrix where x m is the first free variable and a m is the constant in M[a 1 , . . . , a n / → x ] that replaces x m and that v ω [i] (Ra m ) = 0. Then: ¬M[a 1 , . . . , a n / Then the ¬B-rule has been applied to ¬B c D, i and we have C c ¬D, i on B. For the branch is complete. Then the C-rule has been applied to C c ¬D, i, since the branch is complete. Hence, for some new j, iDcj and ¬D, j occur on B. By the induction hypothesis, Dv(c)ω [i] ω [j] , and D is false in ω [j] . Since Rc, i is on B, v (c) is perfectly rational in ω [i] . Consequently, B c D is false in ω [i] , as required.
(¬C). Similar as for (¬B). Quantifiers. (¬ ). Suppose that ¬ xD, i is on the branch. Since the tableau is complete (¬ ) has been applied. So, x¬D, i is on the branch. Again, since the tableau is complete ( ) has been applied. Thus, for all c ∈ C, ¬D [c/x], i is on the branch. Consequently, v ω [i] (D [c/x] (¬ ). Straightforward.

Theorem 6.8 (Completeness Theorem):
Every system in this paper is complete with respect to its semantics.
Proof: First we prove that the theorem holds for our weakest system D. Then we extend the theorem to all extensions of this system. Let M be the class of models that corresponds to D. Suppose that B is not derivable from in D: then it is not the case that there is a closed D-tableau whose initial list comprises A, 0 for every A in and ¬B, 0. Let t be a complete D-tableau whose initial list comprises A, 0 for every A in and ¬B, 0.  [k] . Then iDcj and iRk occur on B [by the definition of an induced model]. Since B is complete (T − CMP) has been applied. Hence, for some l, jRl and kDcl are on B. Accordingly, for some ω [l] , Rω [j] ω [l] and Dv(c)ω [k] ω [l] , as required [by the definition of an induced model].

Examples of derivations in our systems, the unmarried teacher argument and the conscientious student argument
In this section, I will show that the conclusion in the unmarried teacher argument is derivable from the premises in every system in this paper and that the conscientious student argument is invalid in the class of all models (Section 1). It follows that the unmarried teacher argument is valid (in the class of all models) [by the Soundness Theorem]. The argument can be symbolised in the following way. Premises: 'Sx' says that x is a student, 'Bt' says that the teacher is a bachelor, 'Ut' says that the teacher is an unmarried man, and 's' refers to Susan. This argument is intuitively valid, but it seems impossible to prove this in any standard doxastic systems. As we saw in the introduction, this is a good reason to be attracted to the systems developed in the present paper.
To prove that the conclusion is derivable from the premises, we construct a semantic tableau that begins with all premises and the negation of the conclusion. Since this tableau is closed and we have only used the rules of the basic system D, it constitutes a derivation of the conclusion from the premises in this system. Consequently, the conclusion follows from the premises in the class of all models. Here is our proof. ('MP' stands for the derived rule 'Modus Ponens'.) Let us now turn to the conscientious student argument. Let Sx stand for 'x is a student', Hx for 'x studies hard', and Dx for 'x deserves a good grade'. Then the argument can be symbolised in the following way: (1) x (Sx → B x (Hx → Dx)) (Every student believes that if she studies hard she deserves a good grade), (2) x(Sx → B x Hx) (Every student believes that she studies hard), Hence, (3) x(Sx → B x Dx) (Every student believes that she deserves a good grade).
To prove that an argument is not valid we construct an open complete tableau that begins with the premises and the negation of the conclusion. Then we use an open branch in the tree to read off a countermodel. More precisely, if A 1 , . . . , A n are the premises in the argument and B is the conclusion, then we construct a semantic tableau that begins with A 1 , 0, . . . , A n , 0 and ¬B, 0. Here is a tableau for the conscientious student argument. The open branch to the left in this tree closes in a few more steps (it is left to the reader to verify this claim). However, the right branch is open and complete. Accordingly, the whole tableau is open (and complete). Consequently, the conclusion in the conscientious student argument is not derivable from the premises. 11 It follows that the argument is invalid (by the completeness results in Section 6). 12 Let us establish these claims. The branch to the right is open and complete. So, we can use it to read off a countermodel M. The matrix of B c Dc is B x 1 [c] is the only thing that is in the extension of S in ω 0 . R and D are empty. v ω 0 (B x 1 Dx 2 ) is the extension of B x 1 Dx 2 in ω 0 , v ω 0 (B x 1 (Hx 2 → Dx 3 )) is the extension of B x 1 (Hx 2 → Dx 3 ) in ω 0 , and v ω 0 (B x 1 Hx 2 ) is the extension of B x 1 Hx 2 in ω 0 . If ¬Ra m , i occurs on a branch B and M is an n-place matrix with instantiations on the branch (where x m is the first free variable in M and a m is the constant in M[a 1 , . . . , a n /x 1 , . . . , x n ] that replaces x m ), then [a 1 ], . . . , [a n ] is an element of v ω [i] (M) iff M[a 1 , . . . , a n /x 1 , . . . , x n ], i occurs on B. ¬Rc, 0 occurs on our branch above, while B x 1 Dx 2 [c, c/x 1 , x 2 ], 0 (that is, B c Dc, 0) does not occur on this 3. However, see Fitting et al. (2001) and Corsi and Orlandelli (2013). 4. Some might think that the systems that are developed in this paper are too weak.
Almost nothing follows from the proposition that an agent that is not perfectly rational believes something. But is it not reasonable to assume that, for example, the following principles hold also for individuals that are not perfectly rational: B c (A ∧ B) → B c A and B c A ↔ ¬C c ¬A. Two anonymous reviewers independently raised this worry. Should there not be some connections between the truth-values of different beliefs (or propositions about beliefs), for example between B c (A ∧ B) and B c A, even for individuals that are not perfectly rational? As one reviewer pointed out, being not perfectly rational does not mean being ignorant. However, this does not strike me as a particularly serious problem. The point is not that there are no connections between different beliefs. The point is that (for example) B c (A ∧ B) does not entail B c A (for every c). This does not necessarily mean that there are no interesting relations between the truth-values of B c (A ∧ B) and B c A (for example). Many people who believe that A-and-B probably also believe that A; but, according to our systems, it is not logically necessary that someone who believes that A-and-B also believes that A. Obviously, B c (A ∧ B) does not entail that it is not the case that B c A in any system in this paper. All our systems are consistent with the proposition that everyone who believes that A-and-B also, in fact, believes that A. It is, in principle, possible to add x(B x (A ∧ B) → B x A) as an 'axiom' to any system in this paper or to use this formula as a premise in various derivations. Yet, it is obviously conceivable that someone believes that A-and-B without believing that A, and we can think and reason about individuals that are not perfectly rational. In classical doxastic logic this is impossible, we cannot speak about and reason about an agent that, for example, believes that A-and-B without believing that A (at least we cannot do this in any natural way). This is clearly implausible. So even if it were true that everyone who believes that A-and-B also, in fact, believes that A, it would not be a logical truth. I think that one of the reviewers is correct when he or she says that 'being not perfectly rational does not mean being ignorant'. But the systems that I discuss do not entail that an agent is ignorant if she is not perfectly rational. A person that is not perfectly rational can still be rational in many ways. 5. For more on these concepts, see, for example Bostock (1997, p. 79), Church (1996 and Epstein (2006, p. 65). The first occurrence of x in a formula of the form B x A (and C x A) is free. 6. For more on the concept of rationality, see Mele (2004). 7. The idea of using matrices is borrowed from Priest (2005, Ch. 1-2). 8. In the philosophy of religion, the concept of epistemic omniscience has been discussed for a long time (see, for example, Taliaferro, 1998, Chapter 5). We can say that an individual is epistemically omniscient iff she knows everything (i.e. everything that is true). If knowledge implies belief, it follows that everyone who is epistemically omniscient is doxastically omniscient. Theistic philosophers usually think that God is epistemically omniscient. If this is correct (and God exists), then God is doxastically omniscient. It also follows that if an individual is epistemically omniscient and doxastically consistent, she is infallible. And if someone is infallible and epistemically omniscient, she believes something iff she knows it (given that knowledge implies belief ). If knowledge implies truth, it follows that every (epistemically) omniscient individual knows something iff it is true and, hence, believes something iff it is true. For more on epistemology in general and the concept of knowledge in particular, see, for example, Niiniluoto, Sintonen, and Woleński (2004) and Sosa and Kim (2000). I will not say anything more about epistemic omniscience or about the philosophy of religion in this paper. 9. C−UR is a theoretically interesting condition that is philosophically problematic. If all perfectly rational individuals necessarily are perfectly rational, a being that is in fact not perfectly rational cannot be perfectly rational in some other possible world. Yet, no actual human being is, in fact, perfectly rational (or at least so it seems). If this is true, no actual human being could have been perfectly rational. Nevertheless, it seems perfectly possible and interesting to consider what would have been the case if some actual human being would have been perfectly rational. So, if we want to use counterfactuals with such 'antecedents', we should probably reject C−UR. Of course, our systems do not include any symbols for counterfactuals. But they could, in principle, be augmented with operators of this kind. 10. Epistemic and doxastic logic have usually been studied axiomatically. Ajspur, Goranko, and Shkatov (2013) introduces a tableau-based decision procedure for a multiagent epistemic logic. See also Halpern and Moses (1992). 11. We have used the weakest system D to construct our tableau. Hence, when I say that the conclusion is 'not derivable', I mean that it is 'not derivable in the system D'. Yet, it is possible to show that the conclusion is not derivable from the premises in any system in this paper. 12. Since we use D, 'invalid' here means 'invalid in the class of all models'. However, it is also invalid in all other classes of models we consider in this paper. It is left to the reader to verify this.