Survey

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Survey

Document related concepts

Pragmaticism wikipedia , lookup

Gettier problem wikipedia , lookup

Philosophical skepticism wikipedia , lookup

Epistemology wikipedia , lookup

Rationalism wikipedia , lookup

Analytic–synthetic distinction wikipedia , lookup

List of unsolved problems in philosophy wikipedia , lookup

Transcript

Not Every Truth Can Be Known (at least, not all at once) Greg Restall∗ Philosophy Department The University of Melbourne [email protected] version 1 April 10, 2007 Abstract: According to the “knowability thesis,” every truth is knowable. Fitch’s paradox refutes the knowability thesis by showing that if we are not omniscient, then not only are some truths not known, but there are some truths that are not knowable. In this paper, I propose a weakening of the knowability thesis (which I call the “conjunctive knowability thesis”) to the e:ect that for every truth p there is a collection of truths such that (i) each of them is knowable and (ii) their conjunction is equivalent to p. I show that the conjunctive knowability thesis avoids triviality arguments against it, and that it fares very di:erently depending on another thesis connecting knowledge and possibility. If there are two propositions, inconsistent with one another, but both knowable, then the conjunctive knowability thesis is trivially true. On the other hand, if knowability entails truth, the conjunctive knowability thesis is coherent, but only if the logic of possibility is weak. §1 There are many things that we don’t know to be true. Ignorance is a fact of life. However, it is tempting to think that of the things that are true but not known to be true, each of them could be known. If the significance of a proposition is to be explained in terms of its verification conditions, for example, then if it is true, there must be some verification conditions, and it is tempting to say that we could (at least potentially, in theory) have access to them. So, it is tempting to endorse the claim Every truth is knowable. (♦) which has come to be known as the knowability thesis, and its formalisation (∀p)(p ⊃ ♦Kp) (1) ∗ See http://consequently.org/writing/notevery/ for the latest version of the paper, to post comments and to read comments left by others. ¶ Thanks to Conrad Asmus, Allen Hazen, Lloyd Humberstone, Nick Smith and Timothy Williamson, to audiences at Monash University, the University of Melbourne, and Oxford University, and to commentators at http://consequently. org/writing/notevery/ for helpful discussions. Feedback from anonymous reviewers for this volume was useful in cleaning up the presentation. 1 2 http://consequently.org/writing/notevery/ for any truth, it is possible (♦) that it be known (K). This position is tempting to many, but Fitch has shown that the temptation comes at a very high cost. Using only inference principles that are very tough to reject, we can show that, given the knowability thesis, every truth is, indeed, known.1 1·2 Here is Fitch’s proof that the knowability thesis fails if we are not omniscient. Suppose, for a reductio, that we are ignorant of some truth, so suppose that p is true but not known to be true. Then p ∧ ¬Kp is true. So, by the knowability thesis, this is possibly known: ♦K(p ∧ ¬Kp). Now, this is very hard to take. How could we know that p ∧ ¬Kp? If knowing a conjunction entails knowing the conjuncts, then K(p ∧ ¬Kp) entails Kp and K¬Kp.2 Now knowledge entails truth, so K¬Kp entails ¬Kp, a contradiction. So, by a reductio, it is not possible that K(p ∧ ¬Kp), and we have (using the knowability thesis), refuted the hypothesis of ignorance. If the knowability thesis holds, a much stronger thesis holds too: every truth is not merely knowable, but known. 1·3 This phenomenon has come to be called Fitch’s Paradox, after F. B. Fitch, who first formulated it [3]. This paradox has generated a vast literature, including on the one hand “search and rescue” missions designed to find the true principle underlying knowability thesis, and to save them from similar paradoxical fate, and “seek and destroy” campaigns aimed at hammering more nails into the co?n of any so-called principle of knowability that shows signs of life.3 This paper contains elements of both kinds of discussion. I shall present and motivate yet another revision of the knowability thesis, and then show that this revision is consistent and not subject to Fitch-paradoxical refutation — that is the search and rescue part of the story — and then I will show that this revision is not only consistent but either it is also almost trivially true and therefore, it is not likely to do the work that a verificationist or anti-realist might require of a principle of knowability, or it’s an interesting, controversial thesis about knowledge, which is coherent under certain conditions. 1·4 The knowability thesis, cast as the statement (1), is dead. Fitch’s paradox is a conclusive refutation, and even though many interesting moves are possible with the 1 I indicated before that Kp should be read as “it is known that p.” But known by whom? Kp can be read either as “α knows that p” for some fixed agent α, or “someone knows that p” without too much strain in what follows. The existential reading, which requires that p merely be known by someone, ensures that K is nothing like a normal modal operator. We do not have Kp, Kq ` K(p∧q), since it may well be that someone knows that p and someone (else) knows that q without anyone knowing that p∧q. In what follows, however, we soon move from reading Kp as the straightforward “p is known” to the idealisation “p is a logical consequence of what is known,” and this does satisfy the principle that distributes knowledge over conjunctions: if p and q are consequences of what is known (by someone or other) then so is their conjunction. For any of these readings, the knowability thesis has some bite. It seems like a substantial claim that any truth is knowable by someone or other. It seems like a more substantial claim that any truth is knowable by you. 2 Maybe we could know a conjunction without knowing the conjuncts. No problem: Just interpret Kp as “p is a logical consequence of what I know.” If the knowability thesis works for knowledge, it works for this K too. So, from now on, K will allow for deductive closure: if Kp and p ` q then Kq. 3 Brogaard and Salerno’s “Fitch’s Paradox of Knowability” [2] is a fine guide to this literature. Greg Restall, [email protected] version 1 April 10, 2007 3 http://consequently.org/writing/notevery/ logic in which these principles are couched, defeating the inference from (1) to omniscience [1, 5], these answers do not address the question I take to be asked by Fitch’s paradox. I say this because upon reflection, the principles motivating a knowability thesis in fact undercut its application in a case such as p ∧ ¬Kp. Consider any truth p, of which we are ignorant. Given the knowability thesis we can, indeed, imagine coming to know that p. This is all well and good, but any way we can go about knowing that p makes it no longer the case that ¬Kp. But, ¬Kp was true, and so, maybe it too could be known. If we do not inquire as to the status of p (so we don’t come to know that p is true) but rather take ourselves to consider whether or not Kp, it seems plausible to suppose that we could confirm that ¬Kp. In other words, it’s quite coherent to suppose that there is nothing that we can see that makes it impossible for us to know that ¬Kp. But the conjunction p ∧ ¬Kp is true, and none of the ways we have considered, of coming to know p, or coming to know ¬Kp will provide a way to know both p and ¬Kp. The conjunction p ∧ ¬Kp cannot be known “all at once.” Fitch’s proof, it seems, is not a trick to be avoided or to be explained away but a result to be understood. §2 This reasoning points the way to a possible answer: the Fitch-paradoxical conjunction p ∧ ¬Kp cannot be known “all at once” but it can be known “in pieces.” In particular, the first conjunct can be known (or rather, there seems to be nothing preventing us knowing it), but it cannot be known if the second conjunct is known. Similarly, the second conjunct can be known, but it cannot be known if the first conjunct is known. They cannot be known together. (1) is refuted, but it begs for a reformulation. Instead of saying that any truth could be known, let’s attempt to maintain instead that every truth can be known “in pieces.” That is, for any truth p, there is some collection of truths, each of which could be known, and when taken together, entail the original truth p. In other words, p can be factored into components, each of which is knowable. 2·2 If we could defend the knowability thesis in this weaker form, according to which unknowable truths could be factored into knowable pieces, then we may be able to provide some comfort to the anti-realist who takes meaningfulness to be a matter of knowablility. For the fact that p ∧ ¬Kp is unknowable is no counterexample to its meaningfulness any more than the unknowability of p ∧ ¬p renders it meaningless. No, p ∧ ¬p is meaningful when p is meaningful, because we can understand p and its negation and its conjunction, even if to understand this is to come to see that it can never be known for it can never be true. The same kind of process can be seen in p ∧ ¬Kp, though now we have a conjunction which we can see that we will never know even though it may be true. It is meaningful because it is a conjunction of meaningful claims. 2·3 In fact, one could say that in the original naïve formulation (♦) didn’t mean what is expressed by (1), at least in its application to the statement p ∧ ¬Kp. For the p ∧ ¬Kp is not, in itself, one truth that is knowable, but two. There are two knowable truths here, not one. (This is altogether too tendentious a reading to Greg Restall, [email protected] version 1 April 10, 2007 4 http://consequently.org/writing/notevery/ take seriously, however. Nothing in this paper hangs on the idea that conjunctive knowability is what we really wanted in the first place.) 2·4 Now consider what it is for a sentence to be a conjunction of knowable sentences. (In what follows, I will call these ‘knowables’ for short). From the perspective of pure logic it matters not whether the original sentence is complex or atomic. For whatever may be expressed by a complex sentence may be expressed by an atomic sentence too. In whatever model theory we like, if we have a model in which a complex sentence is interpreted in some way, then as far as logic is concerned, any simple sentence may be interpreted in just that way too.4 But suppose that our original sentence was a complex sentence like p ∧ ¬Kp. This sentence is unknowable. If this is because it expresses an unknowable sentence, then if we interpret the atomic sentence q as “meaning the same thing” as p ∧ ¬Kp, then it seems that q will (relative to this interpretation, of course) be unknowable as well. But q has no conjuncts at all: it is a simple sentence. Have we sunk the “factoring” analysis before it could set sail? 2·5 This factoring analysis may survive if we are prepared to agree that while the sentence q from our example has no explicit conjuncts, it may have conjuncts implicitly. The sentence q is equivalent (relative to this model, again) to the conjunction p ∧ ¬Kp. As far as logic is concerned this will su?ce for a factorisation. We will say that q is conjunctively knowable (relative to a model) if it is equivalent (relative to that model) to a conjunction, each of whose conjuncts are knowable (relative to that model). 2·6 This is my proposed revision of the knowability principle: (♥) Every truth is conjunctively knowable. In the rest of this paper, we will examine the fate of this principle.5 §3 For the proposal to be formally evaluated, it must be stated more sharply. This thesis assumes a number of exotic elements of logical vocabulary, such as propositional equivalence, propositional quantification, epistemic and modal operators. To properly state this thesis will require a great deal of machinery. The syntax of the claim is straightforward enough. We may formalise one version of this claim as follows: (∀p) p ⊃ (∃q, r) (p = q ∧ r) ∧ ♦Kq ∧ ♦Kr (2) 4 This phenomenon underlies one of the substitutional properties of formal logics. If φ is a tautology containing the atomic sentence p, then φ 0 , found by replacing p everywhere by another formula B is also a tautology. 5 After writing a draft of this paper, Joe Salerno brought to my attention Risto Hilpinen’s paper “On a Pragmatic Theory of Meaning and Knowledge,” [4] in which he argues that a Peircean pragmatism motivates a conjunctive knowability principle just like this. I must leave it to the reader to determine whether or not the results of the investigation below are congenial to the pragmatist project. Greg Restall, [email protected] version 1 April 10, 2007 5 http://consequently.org/writing/notevery/ This version posits a very strong version of conjunctive knowability: every proposition may be factored into a conjunction of two propositions, each of which is knowable. 6 3·2 The work comes in when we are required to characterise the logical properties of propositional quantification, propositional identity and the modal and epistemic operators. Thankfully, for our purposes, we need not attempt to pin down the right principles governing ♦, K, (∀p) and =. Qua logician my job is to investigate the consistency of (♥) and its formalisation (2). Fitch showed that (1) is inconsistent with deeply plausible modal and epistemic principles. I will show that (2) does not su:er that fate. (2) is consistent, and compatible with the strong principles of modal and epistemic reasoning. To do this, we need not find the right principles of such reasoning. In doing this, it is acceptable to overshoot and require too much. I will provide a class of models that show that the revised knowability thesis (♥) and its formalisation (2) can be absolutely unrestrictedly true at no cost to ignorance or to many other epistemic or modal principles. (There will, however, be an important caveat to be discussed in Section 5.) 3·3 Our logic will be the incredibly strong modal epistemic logic in which ♦ and K are both governed by the principles of the logic s5. This is unrealistic in the extreme, for it commits us to wild epistemic principles such as the claim that if p is true then we must know that we don’t know that ¬p (if p then K¬K¬p) and even that if we don’t know something we know that we don’t know it (if ¬Kp then K¬Kp). Neither of these principles is particularly plausible (even if we take Kp to mean that p is a consequence of what we know) but we will use such a strong logic nevertheless, since nearly every epistemic or modal principle endorsed by someone or other is valid in this logic: s5♦ ⊕ s5K . 3·4 This logic has models of the usual kind for modal logics. Here a model is a quadruple hW, R♦ , RK , [[·]]i where W is a non-empty set of worlds, R♦ and RK are accessibility relations on W and [[·]] is a function assigning to each atomic sentence (for example, p) an interpretation—a set of worlds (in this case, [[p]]). In this modal logic we place no restrictions on which sets can be used to interpret sentences. All sets may be propositions in our model. The set [[p]] is the set of worlds in which p is true. In the usual way, the interpretation function is extended to assign sets It could be that something could be factored into three knowable conjuncts but not two. As far as I can see, there is no natural upper limit to the number of conjuncts one could require in a formalisation of (♥), so a perfectly general formulation would perforce be quite complex indeed, as it would have to quantify over collections of propositions (there are some propositions which factorise p). This seems to be the right condition: 6 ^ (∀p) p ⊃ (∃P)(( P = p) ∧ (∀q)(Pq ⊃ ♦Kq)) (2 0 ) where P is the second order propositional variable used in the second-order propositional existential quantifier ∃P ranging over classes of propositions (which is “really” a third-order V quantifier over objects, since propositions are “really” zero-place properties) and the “connective” sends a class of propositions to its conjunction. Greg Restall, [email protected] version 1 April 10, 2007 6 http://consequently.org/writing/notevery/ to arbitrary sentences in the language. For conjunction, disjunction, the material conditional and negation we use the standard boolean operations. A conjunction p ∧ q is true at the worlds where both p and q are true: [[p ∧ q]] = [[p]] ∩ [[q]], and similarly for the other Boolean connectives. The relations R♦ and RK are used to model the operators ♦ and K respectively. In our case R♦ and RK are both equivalence relations. R♦ is the equivalence relation governing ♦: » ♦φ is true at w i: φ is true at a world in w’s R♦ equivalence class. We will call the worlds in w’s R♦ equivalence class the modal alternatives of w.7 Another way to understand the interpretation is as a function from propositions to propositions. [[♦φ]] is the union of all modal equivalence classes overlapping [[φ]]. Think of approximating the proposition [[φ]] by equivalence classes, counting in our approximation any equivalence class that at least partly overlaps the original proposition: [[φ]] is approximated by its closure. Similarly, an equivalence relation RK governs K: » Kφ is true at w i: φ is true at all worlds in w’s RK equivalence class. We will call the worlds in w’s RK equivalence class the epistemic alternatives of w. [[Kφ]] is the set containing all epistemic equivalence classes totally included in [[φ]]. Here, [[φ]] is approximated by its epistemic interior. The modal alternatives of w need not be the same worlds as the epistemic alternatives: a modal alternative need not be an epistemic alternative (we can know things that are not necessary) and an epistemic alternative need not be a modal alternative (we can be ignorant of some necessary truths). Now for the propositional quantifier: » (∃p)φ is true at a world w if and only if for some set X of worlds, φ is true at w when we take the formula p occurring unbound in φ to be true at exactly the worlds in X. » For identity, we will say that φ = ψ is true at a world just when [[φ]] = [[ψ]]. This su?ces to ensure that we may, for example, infer from φ = ψ that θ(φ) = θ(ψ). Any formula containing φ (for example, ♦Kφ) is true in the same worlds as the formula found by replacing those φs by ψs (in this case, ♦Kψ). An argument is s5♦ ⊕ s5∃p,= valid if for every model, if the premises are true in K a world in that model, the conclusion is true in that world too. A sentence is an s5♦ ⊕ s5∃p,= tautology if and only if it is true in every world in every model. K 3·5 Let me reiterate: This model theory is not to be endorsed as giving us the “true picture” of knowledge, possibility, propositional quantification and propositional identity. It is intended as a grab-bag sizeable enough to catch all principles thought 7 This leads to a slight infelicity: w counts as a modal alternative of itself. Greg Restall, [email protected] version 1 April 10, 2007 7 http://consequently.org/writing/notevery/ to govern epistemic modal logic with propositional quantifiers. If we can find models that both validate the conjunctive knowability principle (♥) and allow for ignorance, then they show that no principle true in these models collapses conjunctive knowability into omniscience. 3·6 Here is a simple model in which (♥) holds. There are four worlds {a1 , a2 , b1 , b2 }. a1 a2 b1 b2 Figure 1: A simple s5♦ ⊕ s5∃p,= frame K The modal accessibility relation R♦ relates a1 to b1 and a2 to b2 ; the epistemic accessibility relation RK is orthogonal to the modal relation: it relates a1 to a2 and b1 to b2 . So, a world’s modal alternatives are those worlds sharing a number, and its epistemic alternatives are those sharing a letter. In Figure 1 (and in all other diagrams) solid lines join epistemic alternatives, and dashed lines join modal alternatives. 3·7 (♥) says that any proposition true at the world of evaluation is a conjunction of two propositions which, at the world of evaluation, are knowable. What are the propositions in our model that are knowable at any world? Any proposition true at both a1 and a2 is knowable at all worlds, since it is known at a1 and a2 (and hence, it is possibly known there), and at b1 , the world a1 is possible, and at b2 , a2 is possible, so at b1 or at b2 , this proposition is also possibly known. So, if {a1 , a2 } ⊆ [[φ]], then φ is knowable at any point in the model. Similarly, any proposition true at both b1 and b2 is knowable at every world. And these propositions are the only propositions knowable at any world. The propositions which can not be known are ∅, each singleton proposition {a1 }, {b1 }, etc., and the two diagonal propositions {a1 , b2 } and {a2 , b1 } and the modal alternative propositions {a1 , b1 } and {a2 , b2 }. All other propositions are knowable, from the point of view of every world. 3·8 It will be helpful to consider why for any interpretation of p, the proposition denoted by p ∧ ¬Kp is not knowable at any world. If [[p]] = X ⊆ W, then [[Kp]] consists of the interior epistemic approximation of X, and [[¬Kp]], then, is the union of all equivalence classes not totally inside X. So, its intersection with X (the set [[p ∧ ¬Kp]] consists of the union of all X-overlapping parts of epistemic equivalence classes that overlap X but do not fall completely inside X.8 In the case where [[p]] = {a1 , a2 , b1 }, [[Kp]] = {a1 , a2 } and so [[¬Kp]] = {b1 , b2 }, and 8 In topological terms it is the part of the (epistemic) boundary of X that is also inside X. Greg Restall, [email protected] version 1 April 10, 2007 8 http://consequently.org/writing/notevery/ [[p ∧ ¬Kp]] = {b1 }. This proposition is not knowable, because it contains no epistemic equivalence classes as a subset. 3·9 In this model, every true proposition is a conjunction of two knowable propositions. The singleton {a1 } is the conjunction of {a1 , a2 } and {a1 , b1 , b2 }. The same goes for each other singleton. The pair {a1 , b1 } is the conjunction of {a1 , a2 , b1 } and {a1 , b1 , b2 }. The same goes for each other pair. It follows that (2) is true in our model. a1 a2 b1 b2 Figure 2: {a1 } as a conjunction of knowables 3 · 10 So, we have shown that (2) survives consistently and coherently. In this model there is much ignorance (a proposition true at a1 alone is true there but not known to be true), yet every proposition is a conjunction of two knowable propositions. The principle (♥) of conjunctive knowability is secure. Truth and knowability can be intimately connected, even if not every truth is knowable. What the friend of knowability cannot have whole she is allowed to have if she will accept it in two pieces. 3 · 11 At this point, the story takes a di:erent turn. Conjunctive knowability is secure, but it either is almost certainly not what the verificationist wants, or it is too high a price for the verificationist to pay. In the rest of this paper I shall show that if knowability does not entail truth (so some falsehoods are knowable while not being known), then conjunctive knowability, in the form of (2) is not only true, but it’s very hard to refute in an epistemic modal logic. It puts precious few constraints on knowledge or necessity, and so, is not useful as criterion for favouring one theory over another. If principles acceptable to the realist lead them to accept (2) while maintaining their realism, then (2) will do no good as a principle designed to favour the anti-realist. On the other hand, if knowability entails truth, then any non-trivial account of conjunctive knowability is inconsistent with plausible modal principles. (In particular, with the modal principle of transitivity: ♦♦p ` ♦p.) §4 So, consider what we have done so far. We have a model of s5♦ ⊕ s5∃p,= in which K conjunctive knowability is satisfied. It turns out that this is not a one-o: a:air. In an epistemic modal logic like s5♦ ⊕ s5∃p,= , and its much weaker cousins in K which the modal and epistemic accessibility relations satisfy fewer constriants, (2) turns out to be very easy to validate. Not only are there many models in which Greg Restall, [email protected] version 1 April 10, 2007 9 http://consequently.org/writing/notevery/ (2) is true, it turns out that (2) is a consequence of other, unproblematic modal and epistemic principles. In particular, (2) follows from the following thesis about possible knowledge, satisfied in the models we have seen: (∃q)(♦Kq ∧ ♦K¬q) (3) This is relatively uncontroversial, given one understanding of how possibility and knowledge (or the consequences of what is known) are connected. Provided that, for some q, both q and ¬q are possibly true (and this is not too di?cult to imagine) then it is not much more di?cult to conclude that for some q, both q and ¬q are possibly known. Of course, a circumstance in which one knows q is, perforce, one in which ¬q is not known, and vice versa, for there to be a q such that ♦Kq and ♦K¬q, there must be at least two distinct modal alternatives, one at which q is known, and the other at which ¬q is known. All that requires is that we have two modal alternatives whose epistemic closures do not intersect, like so. Figure 3: Two inconsistent knowables Given a q such that both it and its negation are knowable, we can prove that for any true p, there are knowable p1 and p2 where p = p1 ∧ p2 and both p1 and p2 are knowable. If q is such that (♦Kq ∧ ♦K¬q), then we may choose p1 to be p ∨ q and p2 to be p ∨ ¬q. Then by simple boolean reasoning, p1 ∧ p2 = (p∨q)∧(p∨¬q) = p. However, since ♦Kq, we have ♦K(p∨q).9 Similarly, since ♦K¬q, we have ♦K(p ∨ ¬q). Both p ∨ q and p ∨ ¬q are knowables, regardless of how unknowable p might be! (2) is a trivial consequence of the trivial truth (∃q)(♦Kq ∧ ♦K¬q). It looks as if (2) tells us little about the connection between truth and knowability. 4·2 Where can the fan of conjunctive knowability resist this analysis? It might be thought that a friend of relevance would quail at the identification of p with (p∨q)∧(p∨¬q), as well they should. The inference from p to (p∨q)∧(p∨¬q) is valid in almost every logic you care to mention, as it is found by composing the inferences from p to p ∨ q, and from p to p ∨ ¬q and from these to their conjunction. All are simple lattice moves. The problem with relevance is in the other direction. To get from (p ∨ q) ∧ (p ∨ ¬q) we need q ∧ ¬q ` p, and this is By distribution of both K and ♦ over logical consequence: since q ` p ∨ q, then Kq ` K(p ∨ q) (remember, we read K(p ∨ q) as “p ∨ q is a consequence of what is known”) and so, ♦Kq ` ♦K(p ∨ q): all are reasonable principles. 9 Greg Restall, [email protected] version 1 April 10, 2007 10 http://consequently.org/writing/notevery/ relevantly invalid. Nonetheless, rejecting this identity10 is not going to stop this argument from getting o: the ground. The crucial premise in the argument was that q and its negation were both knowable, and could be used in the factorisation of p. There is no requirement that q and its negation be used for this purpose. Provided that we are given two incompatible propositions (say, q1 and q2 ) that are knowable—so q1 , q2 ` ⊥ for the trivial proposition ⊥, and ♦Kq1 and ♦Kq2 — then even in relevant logics, the sentences p and (p ∧ q1 ) ∨ (p ∧ q2 ) are true in exactly the same situations. Blocking the inference from (p ∧ q1 ) ∨ (p ∧ q2 ) (with the rule q1 , q2 ` ⊥) to p requires blocking the distribution of conjunction over disjunction, not any odd behaviour about negation or relevance. So, pleading relevance or paraconsistency will not give the fan of conjunctive knowability (or its enemy, for that matter) a straightforward way out of the problem.11 4·3 Denying that one can infer K(p ∨ q) from Kp is not going to help, either, for as we have seen, we can replace talk of what is known by talk of what is a consequence of what is known (at least in decidable logics), and clearly, if p is a consequence of what is known, so is p ∨ q. Furthermore, possibly being a logical consequence of what is known is not very far removed from being possibly known, so reading Kp throughout as “p is a consequence of what is known” does little violence to the principles in question, and it validates the inferences used in our deduction. So requiring high standards for knowledge, so high that logical consequence can lead you from what is known to what is not, is also not a way out of the problem. 4·4 So, conjunctive knowability is not only consistent, but it is trivially so, if possibility and knowledge are connected as given by (3), that is, if possible knowledge can sometimes outrun truth, just as Fitch’s paradox has shown us that truth can sometimes outrun possible knowledge. §5 Nonetheless, (3) is by no means uncontroversial.12 What if we reject (3) and hold, instead, that only truths may be possibly known? So, let us embrace (4): (4) ♦Kp ` p 5·2 Before proceeding, I wish to do away with a bad argument for (4). No-one should argue as follows: “What is possibly known must be true, because of necessity, what is known is true. It is, therefore, impossible for what is known to be false. It follows that if something is possibly known, it is true.” 10 Which, it must be said, is not the same identity as that between p and (p ∧ q) ∨ (p ∧ ¬q), a factorisation seen again and again in di:erent kinds of reasoning. 11 A not-quite-straightforward way out of the problem is to deny that there are any incompatible pairs of propositions. To be sure, in many relevant logics, there is no way to construct formulas φ and ψ such that φ, ψ ` ⊥. Nonetheless, in most models for such logics there are ways to interpret φ and ψ such that their conjunction is absolutely inconsistent. Merely take [[φ]] and [[ψ]] to have empty intersection, so their conjunction is true nowhere. 12 Thanks to Nick Smith for pressing me on this point. Greg Restall, [email protected] version 1 April 10, 2007 11 http://consequently.org/writing/notevery/ This contains a modal fallacy. We have attempted to infer from the innocuous “it is impossible for what is known to be false” (¬♦(Kp ∧ ¬p)) to the much stronger “if something is possibly known, it is true” (which as a material implication is ¬♦Kp ∨ p). In the first, the truth of p (or its not being false) is under the scope of the possibility operator, and in the second it is not. That is a bad argument for (4). If you contemplate (4), do not do so for that reason. 5·3 In the case of an epistemic modal logic modelled with an accessibility relation R♦ for possibility and RK for knowledge, (4) is straightforward to guarantee: we need simply that (∀x)(∀y)(xR♦ y ⊃ yRK x) (4 0 ) for then, when we are at x and we have some modally accessible world (call it y) in which every epistemically accessible world has p true, p is true at x, since x is epistemically accessible from y. Conversely, if we have some x and y where xR♦ y but not yRK x, then if p is true at everywhere other than x (or, if you like, everywhere epistemically accessible from y, if everywhere other than x seems like overkill) then at y, Kp is true, and hence at x, ♦Kp is true. However, at x, p is false. So, if we are allowed to assign the extension of a proposition at whim in our models (and it is hard to see why not) then condition (4 0 ) corresponds precisely to the validity of (4). 5·4 Similarly, we can say, precisely, what condition on RK and R♦ corresponds to conjunctive knowability in its weakest possible form. First note that, in a given model, if p is conjuctively knowable when [[p]] is a singleton set (so p is true at one world only) then every proposition is conjunctively knowable. (If φ is true at x, and if p is true at x alone, then consider the propositions, each knowable, which jointly entail p. These jointly entail φ—relative to that model—too, which shows that φ is also conjunctively knowable.) So, what does it take for a proposition true at x alone to be conjunctively knowable? Well, we must find for any world y distinct from x, a proposition which is knowable but not true at y. If that is not found, the conjunction of all knowable propositions will not entail p, since it will also be true at y, where p is not true. So, we require the following condition (∀x)(∀y)(x 6= y ⊃ ∃z(xR♦ z ∧ ¬zRK y)) (5) for if (5) does not hold, then any z modally accessible from x will include y as epistemically accessible, so no proposition false at y will be knowable from x, as it will not be known at any modally accessible worlds. 5·5 It follows that normal epistemic modal models for conjunctive knowability satisfy (5). Alas if (4 0 ) and (5) both hold, then if R♦ is transitive, it is trivial in the sense that xR♦ y only if x = y. Here is why: if (4 0 ) holds, then ¬zRK y means that ¬yR♦ z, which when substituted in (4 0 ) gives (∀x)(∀y)(x 6= y ⊃ ∃z(xR♦ z ∧ ¬yR♦ z)) Greg Restall, [email protected] version 1 April 10, 2007 12 http://consequently.org/writing/notevery/ but if x and y are non-identical and yR♦ x, then whenever xR♦ z by transitivity yR♦ x, which contradicts what we have assumed. 5·6 We have a syntactic proof of this modal collapse as well. We can show that (4), ♦♦p ` ♦p and conjunctive knowability (in the most general form (2 0 )), ensure that ♦p ` p, in the presence of propositional quantification. 5·7 Here is the proof: Suppose ♦p. So there is some possible circumstance in which p is true. Consider one. In this circumstance p is true, so there are propositions r1 , r2 , . . . , which together entail p, and each of which are possibly known. So, we have ♦Kr1 , ♦Kr2 , etc., and r1 , r2 , . . . ` p. Now consider the actual circumstance in which ♦p is true. In this circumstance, each ♦Kri is possible: that is, ♦♦Kri . But possible possibility is (we assume) possibility, so we have ♦Kri for each i. But by (4), ♦Kri ` ri , so each ri is true. But r1 , r2 , . . . ` p, so p is true too. In other words, we have inferred p from ♦p. 5·8 So, we cannot have (4), (5), ♦♦p ` ♦p and the non-triviality of ♦. One, at least, must go. Which one is to go? I am tempted to do away with (4), but we have already seen what can be done without (4): it makes conjunctive knowability all too easy. Making R♦ trivial is unacceptable, for then the only possibilities will be truths, so if every proposition is a conjunction of knowables, it will be a conjunction of knowns, and hence, every truth will be a consequence of what is known, making all ignorance vanish. We avoid Fitch’s paradox and its heirs by denying the premise that we are not omniscient. To do away with (5) is to give up the task of exploring the consequences of conjunctive knowability. The only remaining option here (given the machinery of normal epistemic modal logics and their possible worlds models) is to explore the rejection of the transitivity of R♦ . As a result, we will examine what follows if we deny the inference from ♦♦p to ♦p. 5·9 Denying transitivity of R♦ is a severe price to pay to save conjunctive knowability. It turns out that it is enough. In the remaining paragraphs of this section I will show that we may maintain (4), making every knowable a truth, and (5), making every truth conjunctively knowable, without concluding that every truth is known. A model showing this is relatively simple. The worlds are the (positive and negative) integers Z. We have xR♦ y i: y = x or y = x + 1. (Notice that this is not transitive, since 0R♦ 1 and 1R♦ 2, but we don’t have 0R♦ 2. Nonetheless, it is reflexive, so at the very least, p ` ♦p.) We have xRK y i: y = x or y = x − 1. (Notice that this is not transitive either, so we do not have Kp ` KKp, but it is reflexive, so Kp ` p, as one would hope.) x−2 x−1 x x+1 x+2 Figure 4: The Model 5 · 10 That is the model. Let’s see how it manages to satisfy (4 0 ) and (5). We have satisfied Greg Restall, [email protected] version 1 April 10, 2007 13 http://consequently.org/writing/notevery/ (4 0 ) by fiat: if xR♦ y, then y = x or y = x + 1, in which case either x = y or x = y − 1, ensuring that yRK x. So, (4 0 ) is satisfied, ensuring that ♦Kp ` p. Conjunctive knowability, in the form of (5), is satisfied too. If x 6= y, then there is always some z where xR♦ z but not zRK y. If we don’t have xRK y, then choosing x for z will su?ce (since xR♦ x always). If we do have xRK y, then since x 6= y, we have y = x − 1. Then choose x + 1 for z. We have xR♦ z (z is one step up from x) but we don’t have yRK z (y is two steps down from z, which is just too far). 5 · 11 How does this model work? At every point, x, knowledge is a little limited because x is epistemically indistinguishable from x − 1. Only propositions true at both x and x − 1 may be known at x. Nonetheless, the world x + 1 is modally accessible from x, and at this world, x − 1 is not epistemically accessible but x is. This means that any proposition true at x is a conjunction of two knowable propositions. If p is true at the set X (including x) then consider two propositions q1 and q2 , true at X ∪ {x − 1} and X ∪ {x + 1} respectively. q1 , true at X ∪ {x − 1}, is known at x (and so is possibly known at x) and q2 , true at X ∪ {x + 1}, is known at x + 1 (and so is also possibly known at x). In this case, as in our other models, every proposition true at a point is a conjunction of two knowable propositions. Nonetheless, not every proposition is known: at every point there is ignorance. x−2 x−1 x x+1 x+2 Figure 5: {x} as a conjunction of two propositions knowable at x 5 · 12 So, if knowability entails truth then we can maintain the conjunctive knowability thesis in the form of (2), but only at the cost of rejecting the s4 principle for possibility: ♦♦p 6` ♦p. §6 Fitch’s paradox shows us that not every proposition is knowable: at least all at once. Fitch’s paradoxical sentence is an example of a proposition that cannot be known, but which can nonetheless be split into pieces, each conjunct of which can be known. It turns out that this modest fallback position is coherent. We may coherently hold that every proposition can be factored into a conjunction, each of which are knowable. Thinking of this in terms of possible worlds, it comes quite close to one original consideration in motivating of knowability. Propositions divide possible worlds into those that are in and those that are out. Conjunctive knowability tells us that for any world that a proposition takes to be out, we can know something that would rule out that world. Think of the discriminations that a proposition makes as constituted by all of the worlds inconsistent with it. According to conjunctive knowability, no proposition makes a discriminiation essentially beyond our grasp. This is coherent. If two inconsistent propositions are knowable, Greg Restall, [email protected] version 1 April 10, 2007 14 http://consequently.org/writing/notevery/ then conjunctive knowability is coherent but trivial. If knowability entails truth, then conjunctive knowability is both coherent and substantial. §7 references [1] jc beall. “Fitch’s Proof, Verificationism, and the Knower Paradox”. Australasian Journal of Philosophy, 78(2):241–247, 2000. [2] berit brogaard and joe salerno. “Fitch’s Paradox of Knowability”. In Stanford Encyclopedia of Philosophy. 2004. [3] f. b. fitch. “A Logical Analysis of Some Value Concepts”. Journal of Symbolic Logic, 28(2):135–142, 1963. [4] risto hilpinen. “On a Pragmatic Theory of Meaning and Knowledge”. Cognitio: Revista de Filosofia, 5(2):150–167, 2004. [5] heinrich wansing. “Diamonds are a Philosopher’s Best Friends: The Knowability Paradox and Modal Epistemic Relevance Logic”. Journal of Philosophical Logic, 31:591–612, 2002. Greg Restall, [email protected] version 1 April 10, 2007