Erkenn

DOI 10.1007/s10670-014-9647-4

ORIGINAL ARTICLE

Closure of A Priori Knowability Under A Priori

Knowable Material Implication

Jan Heylen

Received: 17 June 2013 / Accepted: 23 May 2014

 Springer Science+Business Media Dordrecht 2014

Abstract The topic of this article is the closure of a priori knowability under a
priori knowable material implication: if a material conditional is a priori knowable
and if the antecedent is a priori knowable, then the consequent is a priori knowable
as well. This principle is arguably correct under certain conditions, but there is at
least one counterexample when completely unrestricted. To deal with this, Ander-
son proposes to restrict the closure principle to necessary truths and Horsten sug-
gests to restrict it to formulas that belong to less expressive languages. In this article
it is argued that Horstens restriction strategy fails, because one can deduce that
knowable ignorance entails necessary ignorance from the closure principle and
some modest background assumptions, even if the expressive resources do not go
beyond those needed to formulate the closure principle itself. It is also argued that it
is hard to find a justification for Andersons restricted closure principle, because one
cannot deduce it even if one assumes very strong modal and epistemic background
principles. In addition, there is an independently plausible alternative closure
principle that avoids all the problems without the need for restriction.

1 Introduction

Ever since the work of Dretske (1970) and Nozick (1981), there has been an
intensive debate in epistemology about the principle of closure of knowledge under
known material implication, which can be formalized as follows:

K / ! w ! K/ ! Kw: 1:1

Jan Heylen is a Postdoctoral Fellow of the Research Foundation Flanders.

J. Heylen (&)
Centre for Logic and Analytic Philosophy, Institute of Philosophy, KU Leuven, Andreas
Vesaliusstraat 2, Box 3220, 3000 Louvain, Belgium
e-mail: [email protected]

123
 J. Heylen

The discussion is often framed in terms of general, so-called subjunctivist theories

about knowledge. Recently, Holliday (2014) formalized the theories of knowledge
proposed by Dretske (1970), Nozick (1981), Heller (1989, 1999) and Sosa (1999) in
a single framework and he showed that (1.1) fails on all of them.1
What is perhaps less known is that Dretske (2005) has also implicitly challenged
the principle of closure of knowability, i.e. possible knowledge, under known
material implication, which can be formalized as follows:

}K / ! w ! }K/ ! }Kw: 1:2

Dretske (2005) claims that we face a choice between: first, accepting that we have
lots of easily obtainable knowledge of light-weight propositions (e.g., there are
cookies in the jar); second, accepting that it is very hard or perhaps impossible to
obtain knowledge about certain heavy-weight propositions (e.g., there is no evil
demon that deceives us); third, accepting that one often knows that light-weight
propositions have heavy-weight implications; fourth, accepting closure of knowl-
edge under known material implication. For discussion of this see the exchange
between Dretske (2005) and Hawthorne (2005). Since Dretske thinks that some of
the heavy-weight propositions are not simply unknown but also unknowable, one
can replace the fourth option, namely closure of knowledge under known material
implication, with closure of knowability under material implication as well.
If one focuses not so much on knowledge in general but on a priori knowledge in
particular, then it seems initially plausible that a priori knowability is closed under
known material implication, which can be expressed as follows:

}Ka / ! w ! }Ka / ! }Ka w: 1:3

Assessing the material soundness of (1.3) is the main goal of this paper. Fritz (2013)
has formalized so-called epistemic two-dimensionalism and closure of a priori
knowability under modus ponens is a theorem in his framework. Yet it is not
entirely uncontroversial. Assume the principle that all tautologies > are a priori
knowable, which can be formalized as follows:

}Ka >: 1:4

It follows from (1.3) and (1.4) that a priori knowability is closed under conjunction
introduction,2 which can be formalized as follows:

}Ka / ^ }Ka w ! }Ka / ^ w: 1:5

Anderson (1993) and Horsten (2000) each offer a counterexample to (1.5). They
each also suggest a way of dealing with these counterexamples. This is the subject
of Sect. 2. I will argue against Horstens proposal in Sect. 3. The justification of

1
If one takes into account the basis for a belief, then the outcome is different (Holliday 2012, appendix
2.D).
2
Interestingly, Holliday (2014) shows that this closure principle is a consequence of the various
subjunctivist theories of knowledge.

123
Closure of A Priori Knowability

Andersons proposal is the subject of Sect. 4. In Sect. 5 I will present an alternative

principle that is more resilient to the problems encountered in this article and that
has good, independent justification.

2 Anderson and Horsten on Closure

Anderson (1993, pp. 89) presents the following counterexample. Assume that (a
priori) knowledge is factive, which can formalised as follows:

Ka / ! /: 2:1
Let / be of the form p $ @p, with @ the actuality operator. Let p mean that I dont
have a priori knowledge of any conjunction.3 Furthermore, suppose that p is actually
true. Next, assume that I know a priori that p $ @p and, therefore, it is a priori
knowable for me. Given (1.4), it is a priori knowable that p _ :p. However, it
cannot be the case that I know a priori that p $ @p and p _ :p, since in that case it
would have to be true that p $ @p, which can only the case if p is true, because @p
is true. But then in that case it is true that I dont know a priori any conjunction and,
consequently, I dont know a priori that p $ @p and p _ :p.
Let us consider how strong the counterexample really is. In the example it is
assumed that it is actually the case that I dont have a priori knowledge of
conjunctions, which is manifestly not the case. Anderson also mentions the
possibility that p means that I have made no inferences. On that reading of p, it
again is actually false. A third and final option Anderson considers is that p means
that everything I know is of a bounded complexity. This is more promising; it could
actually very well be the case. Let us delve a bit deeper.
Anderson does not say which complexity measure he has in mind, but a natural
choice here is the following: the complexity of a formula is the number of steps
needed to construct the formula on the basis of the recursive definition of well-
formed formulas. E.g., the complexity of p $ @p ^ p _ :p is higher than the
complexity of p $ @p. Strictly speaking, the complexity of p is then of the lowest
complexity. However, this may be deemed somewhat misleading. After all, the
complexity of formulas is defined in a metalanguage, so p expresses something that
is properly formulated in a metalanguage. This might be the reason why Horsten
(2000, p. 65, fn. 9) says that [Anderson] gives a counterexample based on self-
reference. In order to make sure that there is nothing paradoxical going on, it may
be wiser to make to use arithmetization to make explicit in the object language what
p means in the metalanguage. However, this raises the question how it is possible to
say that I dont know anything above a certain complexity without exceeding that
complexity. An arithmetized expression of the claim that nothing above a certain
complexity, denoted by the natural number n, is known to be true can be expressed
as follows:
3
As will become clear, this can only be the case if material equivalence is not analysed in terms of the
conjunction of two material implications. Alternatively, let p mean that I dont have a priori knowledge of
any conjunction of the form h $ @h ^ h _ :h.

123
 J. Heylen

:9xKa T x ^ x [ n; 2:2

with Tx a truth predicate and with n the Peano numeral of the number n. The
Godel code of the above formula is higher than n, since even the Godel code of n is
considerably greater than n. The Godel code of
:9xKa T x ^ x [ n $ @:9xKa T x ^ x [ n 2:3

is higher still. But in that case (2.3) is not actually known, because nothing above
complexity n is actually known. Moreover, it is not knowable, since (2.2) is false in any
world in which (2.3) is known, while @ (2.2) is true in that world, thereby making (2.3)
false. It is not my intention to claim that the above considerations are a definite
refutation of Anderson (1993)-style counterexamples. However, I think that the bur-
den of proof has shifted to those who want to provide such a counterexample.4
Horsten (2000, pp. 5051) also targets (1.5). Consider the following two sentences:

}Ka 8xKa x 2 fxj9y 2 N x 2  yg $ x 2 _ x 4; 2:4

}Ka 8xKa x 2 fxj9y 2 N x 2  yg $ x 2 _ x 4 _ x 6: 2:5

The first one says that is is a priori knowable that every number of which it is a
priori known that it is an even number is identical to two or to four. The second one
says that it is a priori knowable that every number of which it is a priori known that
it is an even number is identical to two or to four or to six. Combining the above
with (1.5) and using elementary first-order, arithmetical and modal reasoning, one
can derive a contradiction via

}Ka 6 2 fxj9y 2 N x 2  yg ^ :Ka 6 2 fxj9y 2 N x 2  yg:

Deeming the assumptions plausible, Horsten (2000) rejects (1.5).
Both Anderson (1993) and Horsten (2000) reject (1.5), which is entailed by (1.3)
and (1.4). They both think that on that basis (1.3) is to be rejected. Notwithstanding
their rejection of (1.3), they both think that a restricted version of (1.3) can be
upheld, although they each favour a different restriction.
Anderson (1993, p. 7) accepts the principle when the a priori knowable
propositions are also necessary, which can be formalized as follows:

}Ka / ! w ^ h/ ! w ! }Ka / ^ h/ ! }Ka w: 2:6

The counterexample he used against (1.3) involved a contingent proposition, viz.
p $ @p and, hence, it does not work against (2.6). Furthermore, Horstens (2000)
counterexample also involved contingent truths, viz. truths of the form Ka / and
:Ka /. Moreover, (2.6) still has a large class of applications. Analytic truths (e.g. all

4
Hawthorne (2005, p. 40, fn. 4) mentions a similar example: Suppose I know P and know necessarily P
[!] I will never perform a deductive inference again. Without specifying what P is and without
explaining how one could get a priori knowledge of this, the example is underdeveloped.

123
Closure of A Priori Knowability

vixens are foxes) and mathematical truths (e.g. 2 3 5) are all thought to be
necessary. So far, the negative justification for (2.6). The positive justification
offered by Anderson (1993, p. 9) is, however, not relevant. He says the following:
Our knower has the potential to have both of the necessities p and q in mind
and to put them together into a conjunction. If both of them are known a priori,
and he believes the conjunction as a result of his inference, then he surely has
a priori knowledge of that conjunction.
[] We know by proof the necessary truths theorem A and theorem B. Can it
be that if we infer their conjunction and believe it because of the inference we
have made, we still may not necessarily have a priori knowledge of that
conjunction? Surely this is not a possibility.
The problem with these considerations is that that they do not directly support
(1.5) or rather5

}Ka / ^ }Ka w ^ h/ ^ hw ! }Ka / ^ w : 2:7

Instead, they support something along the following lines:

Ka / ^ Ka w ^ h/ ^ hw ! }Ka / ^ w: 2:8
I will later return to the logical difference between (2.7) on the one hand and (2.8)
on the other hand.
Horsten (2000, pp. 57, 60) adopts a closure principle that, unlike Andersons
(1993) (2.7), has non-trivial applications to contingent formulas. His closure
principle says that, if two conjuncts are each a priori knowable, then it is a priori
knowable that each conjunct is possibly true, which can formalized as follows:

}Ka / ^ }Ka w ! }Ka }/ ^ }w: 2:9

It can easily be seen that both Andersons (1993) counterexample and Horstens
(2000) counterexample fail if (1.5) is replaced by (2.9). Unfortunately, the only
reason given to accept (2.9) is that, if }Ka / ^ }Ka w, then it is true that }/ ^ }w
(Horsten 2000, p. 56). This follows in modal system K from the assumption and
(2.1). It is acknowledged that the argument does not establish the soundness of this
aggregation principle beyond all possible doubt (Horsten 2000, p. 66, fn. 15). Note
that there is to some extent agreement with Anderson in the case of necessary
formulas. For suppose that / and w are necessary. Then it is true that / ^ w.
Analogous reasoning then leads one to the conclusion that }Ka / ^ w.
Horsten (2000, p. 51) is also prepared to go beyond (2.9). He thinks that the
problem is due to the fact that his counterexample is formulated in a very expressive
language. He notes that the systems of intensional mathematics found in Shapiro
(1985) and Horsten (1994) each contain a version of the relevant closure principle, but
the languages in which these closure principles are formulated is less expressive than
the language that is used to give the counterexample to the closure principle. In
5
From this and closure of a priori knowability under tautological equivalence one can deduce (2.6).

123
 J. Heylen

particular, the language used by Shapiro (1985) is the extension of the language of
Peano Arithmetic (PA), LPA , extended with a single operator expressing complex a
priori knowability, which will be expressed here with hKa i, whereas the language used
by Horsten (1994, 1998) is the extension of LPA with both a possibility operator, },
and an a priori knowledge operator, Ka . Call the first language LEA and the second
language LMEA . Neither language contains the set membership relation, 2, or the set-
abstraction notation, fxj. . .g. So, neither (2.4) nor (2.5) can be formulated in LEA or
LMEA . Similarly, the actuality operator does not belong to either LEA or LMEA , so
Andersons (1993) counterexample is also inexpressible in those languages.
The positive proposals are related to one another. To see this, it is useful to define
a translation from LEA to LMEA that systematically replaces hKa i with }Ka .
Definition 2.1 Let F be a translation from LEA to LMEA as follows (Heylen
2013, p. 93):
1. if / is an atomic formula, then /F /;
2. if / is :w, then /F :wF ;


3. if / is w ! h, then /F wF ! hF ;
4. if / 9xw, then /F 9xwF ;
5. if / is Kw, then /F }Ka wF .
For instance, 9y9z y zzF is just 9y9z y zz but hKa i9y9z y zzF
is }Ka 9y9z y zz. Shapiros closure principle can then be expressed in LMEA
as follows:



}Ka /F ! wF ! }Ka /F ! }Ka wF : 2:10

If one assumes PA and quantified modal system S5, then one can prove that all
formulas in the range of the F-translation are true if possibly true, which can be
formalized as follows:

}/F ! /F : 2:11

The details can be found in Horsten (1994, p. 287) and Heylen (2013, p. 95). If one
combines this result with (2.1), then one can prove that (2.10) is deducible from
Andersons (1993) (2.6). Ergo, Shapiros (1985) version of the closure principle
satisfies both Andersons (1993) restriction to necessary truths and Horstens (2000)
restriction to less expressive languages. The same does not hold for Horstens
(1994, 1998) version of the closure principle, which allows instantiation with
contingent truths, e.g. Ka / and :Ka /.
The main goal of this paper is to further assess the material soundness of the
principle of closure of a priori knowability under a priori knowable material
implication. First, Horstens restriction strategy will be examined. It will be shown
that the relevant closure principle in combination with a few very plausible other
principles leads to a highly implausible consequence. No expressive resources other
than the ones needed to formalize the relevant closure principle are used. It is
however crucial for obtaining this result that one can instantiate the closure

123
Closure of A Priori Knowability

principle with certain contingent formulas. This takes Horstens (1994, 1998)
version off the table but it still leaves Andersons (1993) and Shapiros (1985)
versions on the table. This is the subject of Sect. 3.
Second, Andersons restriction strategy will be examined. In particular, I will
look at the closure principle restricted to one kind of necessary truths, namely
arithmetical ones. It will be shown that if the arithmetical complexity of the
consequent is low, the closure principle is provable, provided some fairly weak
background assumptions are made. However, it is not provable when also
arithmetically more complex sentences are allowed in the consequent, even if one
throws in a substantial number of modal, epistemic and modal-epistemic principles.
This does not amount to a refutation of the closure principle restricted to necessary
truths, but it does raise the question of how exactly to justify it.
Third, I will present an alternative principle, which epistemic logicians and
mainstream epistemologists have good reason to accept. Furthermore, it manages to
avoid both the counterexamples offered by Anderson (1993) and Horsten (2000) and
it is immune to the Socratic objection. Moreover, it is not restricted to necessary
truths. This is the subject of Sect. 5.

3 Closure and the Socratic objection

Horsten (2000, pp. 6162) shows that (1.3) is not a theorem of his system. In fact,
one needs only a subtheory of his system. The argument makes use of three
principles that have not been introduced before. The first is a principle of weak
positive introspection and the second is a principle of weak negative introspection,
which can be formalized as follows:

Ka / ! }Ka Ka /; 3:1

:Ka / ! }Ka :Ka /: 3:2

The third is what Horsten calls a density principle:

}:Ka /: 3:3

The above is claimed to be the opposite of logical omniscience (Horsten

2000, p. 51), but this claim is too strong, since even logically omniscient agents
may be ignorant about truths that are not theorems of the system, e.g. proposition
letters. Given these principles, one can reconstruct his arguments as a mutual
inconsistency result.6 One advantage is that the focus shifts from whether or not
(1.3) is a theorem of the system in Horsten (2000) to whether or not (1.3) is
6
In Horstens proof the density principle is invoked to infer }Ka :Ka h, where that principle by itself only
allows the inference of }:Ka h. However, one can give a charitable interpretation of the proof in which
the desired conclusion follows from (3.3) and (3.2), which is a trivial consequence of the axiom schemes
of his theory.

123
 J. Heylen

compatible with a number of other principles that may belong to other theories as
well. An additional benefit is that one no longer has to consider the plausibility of
certain factual assumptions such as (2.4) and (2.5), which were used in Horstens
(2000) counterexample to (1.3).
Theorem 3.1 If K4  T; T (1.3), T (1.4) T (2.1, T (3.1), T (3.2), T (3.3),
then T is inconsistent.
Proof Let h be p _ :p. By (1.4), it follows that }Ka h. It is a K4-consequence of
the latter and (3.1) that }Ka Ka h. By (3.3), it is also the case that }:Ka h. It is a K4-
consequence of the latter and (3.2) that }Ka :Ka h. By (1.5), it follows that

}Ka Ka h ^ :Ka h:
It is a K-consequence of the latter and (2.1) that }Ka h ^ :Ka h, which contradicts
a theorem of extensions of K. h
A philosopher wishing to uphold (1.3), could choose to drop modal system K4
(1.4), (2.1), (3.1), (3.2) or (3.3). I will first show that dropping (3.2) and (3.3) is not
solution.
Lemma 3.1 If K  T, T (1.3), T (1.4) T (2.1), then

T :}Ka / ^ }Ka :/:

Proof Suppose that }Ka / ^ }Ka :/. It follows that }Ka / ^ :/. Given modal
system K and (2.1), it follows that }/ ^ :/. This contradicts a theorem of
extensions of modal system K. h
Theorem 3.2 If K4  T; T (1.3), T (1.4), T (2.1), then:
1. if T (3.1), then T }Ka :Ka / ! h:Ka /;
2. if T (3.2), then T }Ka Ka / ! hKa /.

Proof By Lemma 3.1, it is the case that:

1. }Ka :Ka / ! :}Ka Ka /;
2. }Ka Ka / ! :}Ka :Ka /.
Next, note that the following are K4-consequences of (3.1) and (3.2) respectively:
1. :}Ka Ka / ! :}Ka /;
2. :}Ka :Ka / ! :}:Ka /.

Finally, the following are tautological consequences of the above:

1. }Ka :Ka / ! :}Ka /;
2. }Ka Ka / ! :}:Ka /.

123
Closure of A Priori Knowability

Corollary 3.1 If K4  T, T (1.3), T (1.4), T (2.1), T (3.1), T (3.2) , then:

1. T Ka / ! hKa /;
2. T :Ka / ! h:Ka /.
Dropping (3.3) is not a solution, since one could then still prove a highly
undesirable result (Corollary 3.1), namely that knowledge and ignorance are
necessary. Dropping (3.3) and (3.2) is also not a solution, since one could then still
prove that a priori knowable ignorance entails necessary ignorance (Theorem 3.2).7
Suppose that one has a list of proofs that one has obtained. Next, one wants to check
whether one has proved some conjecture. After having gone through the entire list
and not finding the conjecture on it, one knows that one does not have a proof of it.
It would be absurd to conclude that one consequently cannot have a proof of the
conjecture. Similarly, the result that known ignorance entails necessary ignorance is
too strong. Knowing about ones ignorance does not preclude learning. Since
Socrates famously stated that knowledge of ones ignorance is the beginning of
wisdom, I call this the Socratic objection to (1.3).
Theorem 3.2 is bad news for Horsten (1994, 1998), because his system of modal-
epistemic arithmetic (MEA) contains all of (1.3), (1.4), (2.1) and (3.1). Still,
someone who would want to uphold (1.3) could choose to drop (3.1). This is to no
avail either, if one accepts that (a priori) knowledge is closed under conjunction
elimination, which can be formalized as follows:

Ka / ^ w ! Ka / ^ Ka w: 3:4

With only (2.1), (1.4) and (3.4) and modal system K in the background, one can still
level the Socratic objection to (1.3).
Lemma 3.2 If K  T, T (1.3), T (1.4), T (2.1), T (3.4) , then

T :}Ka / ^ }Ka :Ka /:

Proof Suppose that }Ka / ^ }Ka :Ka /. It follows that }Ka / ^ :Ka /. Given
(3.4), it is K-consequence of the latter that }Ka / ^ Ka :Ka /. Assuming (2.1), it is
a K-consequence of the latter that }Ka / ^ :Ka /, which contradicts a theorem of
extensions of K.8 h

7
The principle of strong negative introspection for (a priori) knowledge says that ignorance entails
known ignorance, which can be formalized as :Ka ! Ka :Ka /. This is not equal to the conclusion that
knowable ignorance entails necessary ignorance. Moreover, Theorem 3.2 does not assume weak negative
introspection, let alone strong negative introspection. It is perhaps also worth pointing out that, if strong
negative introspection of (a priori) knowledge is combined with the principle of factivity of (a priori)
knowledge, then one can deduce that falsity entails known ignorance (Hendricks 2006, p. 87). This is a
particularly strong consequence.
8
The reader familiar with the paradox of Fitch (1963) will recognize this sentence, albeit that it contains
an a priori knowledge operator rather than a (general) knowledge operator. However, the argument does
not start from the assumption of weak verificationism, namely that all truths are knowable. Indeed, it
would hardly be credible that all truths are a priori knowable. Rather, my argument starts from (1.5),
which is derivable from (1.3) and (1.4). For discussion of the paradox of Fitch (see Williamson 2000,
ch. 12).

123
 J. Heylen

Theorem 3.3 If K  T, T (1.3), T (1.4), T (2.1), T (3.4) , then

T }Ka :Ka / ! :}Ka /.
So, philosophers who want to uphold (1.3) and avoid the Socratic objection have
to reject modal system K, (1.4), (2.1) or (3.4). Horsten (2000, pp. 54, 60) is willing
to accept even S5 for the diamond operator. Anderson (1993, p. 2) is more
cautious,9 but he is also willing to go beyond K, since he accepts T. Tautologies are
paradigm examples of a priori knowable truths. It is the standard view in
epistemology that knowledge in general is factive (Williamson 2000, pp. 3341).
That knowledge distributes over conjunction is a very reasonable assumption
(Williamson 2000, pp. 275283). Given the high plausibility of (1.4), (2.1), (3.4)
and the weakness of modal system K, the only reasonable option seems to be to
reject (1.3). Nothing hinges on the expressive power of the language: it suffices that
the language contains the modal possibility operator }, the a priori knowledge
operator Ka and the usual logical connectives. Therefore, Horstens (2000)
suggestion that (1.3) is safe when formulated in a less expressive language, viz.
LMEA , is false.
The next question is whether the Socratic objection can be raised against
Andersons (1993) (2.6) and Horstens (2000) (2.9) as well, if modal system K4
(1.4), (2.1), (3.1) and (3.4) are in the background. The answer is no.
Theorem 3.4 There is an awareness model of L [ f}; Ka g such that

M  K4; 1:4; 2:1; 3:1; 3:4

and

M  2:6; 2:9;

but M 6 }Ka :Ka p ! :}Ka p.

Proof Let M be an awareness model, with the the following characteristics:
1. W fw1 ; w2 g;
2. RM W  W;
3. RE fhw1 ; w1 i; hw2 ; w2 ig;
4. V is a function from proposition letters q and possible worlds w 2 W such that:
(a) V p; w1 V p; w2 1;
(b) V q; w1 V q; w2 , for every other proposition letter;
5. Aw1 L [ f}; Ka g and Aw2 is the set of formulas of L [ f}; Ka g
recursively defined as follows:
(a) :Ka p is in Aw2 ;
(b) if / 2 Aw2 , then Ka / 2 Aw2 ;
(c) nothing else is in Aw2 .
9
He refers to Salmon (1989), who argues that modal axiom scheme 4 is not correct for the logic of
metaphysical or counterfactual possibility. Williamson (2007) argues that even S5 is correct for
metaphysical or counterfactual possibility.

123
Closure of A Priori Knowability

The key clauses in the definition of M; w  / are the one for / }w and the one
for / Ka w:
M; w  }w if and only if M; w0  w for all w0 such that wRM w0 ;
M; w  Ka w if and only if w 2 Aw and M; w0  w for all w0 such that wRE w0 .

Given the definition of RE , the last conjunct of the last clause can be simplified to:
M; w  w. Moreover, given the definition of Aw1 , the case of w1 can be simplified
to: M; w1  Ka w if and only if M; w1  w.
Since RM is an equivalence relation, M  S5 and, hence, M  K4. Since RE is a
reflexive relation, M  (2.1). Clearly, M  (1.4), because for every w 2 W, there is
a w0 2 W, namely w1 , such that wRM w0 and Aw0 L [ f}; Ka g. It follows that
/ 2 Aw1 . For any tautology /, it is the case that M  /, i.e. M; w0  / for every
w0 2 W and, a fortiori, M; w1  /. Therefore, it follows that M; w1  Ka /.
Consequently, M  }Ka /, because w0 RM w1 for every w0 2 W. Hence, M  (1.4).
It is also the case that M  (3.1). If M; w1  Ka /, then M; w1  Ka Ka / and, given
the reflexivity of RM , it is also the case that M; w1  }Ka Ka /. The case of w2 can
be proved on the basis of the construction of Aw2 and the reflexivity of RM .
Finally, M  (3.4). The case of w1 is trivial as before. The case of w2 follows from
the fact that no formula of the form / ^ w belongs to Aw2 and, therefore,
M; w2 6 Ka / ^ w, which makes (3.4) trivially true.
Next, M  (2.6). The reason is that, if M; w0  }Ka / ! w ^ h/ ! w and
M; w0  }Ka / ^ h/, then M; w0  hw. Since RM is an equivalence relation, the
latter entails that M; w1  w. Hence, M; w1  Ka w and, given the properties of RM ,
it is also the case that M; w1  }Ka w and M; w2  }Ka w. Furthermore, M 
(2.9). Suppose that M; w0  }Ka / ^ }Ka w. Since the model makes (2.1) and S5
true, it follows that M; w1  }/ ^ }w. Therefore, M; w1  Ka }/ ^ }w.
Finally, let us check that the model is a counterexample to }Ka :Ka p ! :}Ka p.
First, note that M; w2  Ka :Ka p, since :Ka p 2 Aw2 and M; w2  :Ka p, because
Ka p 62 Aw2 . Second, M; w2  }Ka p, since M; w1  Ka p (because M; w1  p)
and w2 RM w1 . h
In this section it was shown that (1.3) entails that a priori knowable ignorance
entails necessary ignorance, which was called the Socratic objection to (1.3). The
consequence was deduced with the help of principles that Horsten (1994, 1998)
accepts (Theorem 3.2) or with the help of principles that are highly plausible on
independent grounds (Theorem 3.3). Apart from the usual logical connectives,
neither result depends on expressive resources beyond those needed to express (1.3).
Therefore, Horstens (2000) restriction strategy to salvage (1.3) from counterex-
amples, which was discussed in Sect. 2, fails. A hopeful conclusion of the
investigations in this section is that both Andersons (1993) (2.6) and Horstens
(2000) (2.9) are secure against the Socratic objection (Theorem 3.4). In the next
section it is considered whether these principles can also be positively justified.
9
He refers to Salmon (1989), who argues that modal axiom scheme 4 is not correct for the logic of
metaphysical or counterfactual possibility. Williamson (2007) argues that even S5 is correct for
metaphysical or counterfactual possibility.

123
 J. Heylen

4 Closure and the Justification Challenge

The good news about Andersons (1993) (2.6) and Horsten (2000) (2.9) is that they
both block Andersons and Horstens counterexamples (Sect. 2) and that they are
immune to the Socratic objection (Sect. 3). Unfortunately, neither principle was
adequately justified by Anderson and Horsten (Sect. 2). In this section I will
investigate the prospects of giving such a justification. In Sect. 4.1 it will be argued
that there is partial justification for (2.6). However, it will also be argued that it is
very challenging to find an adequate justification for (2.6) or (2.9) in their full
generalitysee Sect. 4.2.

4.1 Partial Justification

Let us start with the partial justification for (2.6). One could reason as follows. If a
material implication and its antecedent are each a priori knowable, then the material
implication and its antecedent are each possibly true. Suppose that they each belong
to a category of formulas that are such that, if they are possibly true, then they are
true. This category is, of course, a subset of the category of strict implications (i.e.
necessary material implications) with necessary antecedents. In that case the
material implication and its antecedent are true and, therefore, its consequent is true
as well. Suppose furthermore that the consequent belongs to a category of formulas
that are such that, if they are true, then they are a priori knowable. In that case the
consequent is a priori knowable as well. The natural question to ask is whether there
are examples of material implications, antecedents and consequents with the
mentioned properties. The answer is yes.
Let us consider material implications with arithmetical antecedents and
consequents. First one needs a definition of a particular class of arithmetical
formulas. It is a consequence of (2.11) that these are true if possibly true. Moreover,
let us consider not just any arithmetical consequent, but so-called 9-rudimentary
formulas. These are defined as follows.
Definition 4.1 A formula / 2 LPA is a rudimentary formula if and only if:
1. if / is an atomic formula of LPA , then / is a rudimentary formula;
2. if w is a rudimentary formula, then so is :w;
3. if w and h are rudimentary formulas, then so is w ^ h;
4. if w is a rudimentary formula and if t is a term of LPA , then 8x x\t ! w and
9x x\t ^ w are rudimentary formulas.
An 9-rudimentary formula is a formula of the form 9xw, with w a rudimentary
formula. A 8-rudimentary formula is a formula of the form 8xw, with w a
rudimentary formula (Boolos et al. 2003, p. 204)
It is a an important theorem that an 9-rudimentary sentence / is true in the
standard interpretation of arithmetic if and only if Q w (Boolos et al. 2003, pp.
199, 208), with Q the system of minimal arithmetic.

123
Closure of A Priori Knowability

Let MEAy be the extension of PA with modal system S5 (2.1) and the closure of
a priori knowability under provability in Q, which can be formalized as follows:

Q / ) Q }Ka /: 4:1
Theorems of Q constitute another class of paradigm examples of a priori knowable
truths.
Theorem 4.1 For any formula / 2 LPA , for any 9-rudimentary sentence w , it is
the case that MEAy }Ka / ! w ! }Ka / ! }Ka w.
Proof Note that MEAy }/ ! / (Heylen 2013, p. 95). It follows that

MEAy }Ka / ! /

(Horsten 1994, p. 287). Suppose that }Ka / ! w. Therefore, / ! w. Suppose

that }Ka /. Hence, /. By modus ponens, it follows that w. Suppose that w is an 9-
rudimentary formula. Either w is true in the standard interpretation of arithmetic or
not. If it is true, then Q w. If it is false, then Q :w. By (4.1), Q }Ka w or
Q }Ka :w. The second option is incompatible with the hypothetical fact that w,
given that }Ka :w ! :w. Hence, }Ka w. h
Corollary 4.1 For any formula / 2 LPA , for any 9 -rudimentary sentence w , it
is the case that MEAy }Ka / ! w ^ h/ ! w ! }Ka / ^ h/ ! }Ka w.
So, it is possible to find justification for a restricted version of Andersons (1993)
(2.6), namely the version in which the antecedent is an arithmetical formula and the
consequent is an arithmetical formula of a certain low complexity. This raises the
question what happens when the consequent is an arithmetical formula of a higher
complexity. For these formulas it is not provable that they are a priori knowable if
true, given (4.1).

4.2 No More Justification

It will be argued that that one cannot similarly justify (2.6) or (2.9) for all
arithmetical cases, even if one makes very strong modal and epistemic assumptions.
The argumentation strategy is as follows. I will introduce a theory that contains very
strong modal and epistemic principles. Then I am going to show that neither
Andersons (1993) (2.6) nor Horstens (2000) (2.9) are theorems in this theory. This
means that even with some very strong modal, epistemic and modal-epistemic
principles there is no deductive justification for (2.6) or (2.9). The challenge is then
to find some other (perhaps non-deductive or informal) justification for (2.6) or
(2.9). The thought is that, if it is very challenging to find justification for (2.6) and
(2.9) when restricted to arithmetic, it is also very challenging to find justification for
those principles when they are unrestricted.
Let MEAyy be the extension of PA with modal system S5 (2.1) and with the
following other epistemic principles. First, add the principle that all theorems of
MEAyy are a priori known, which can be formalized as follows:

123
 J. Heylen

MEAyy / ) MEAyy Ka /: 4:2

Second, add the principle that a priori knowledge is closed under a priori known
material implication, which can be formalized as follows:

Ka / ! w ! Ka / ! Ka w: 4:3
Third, add the principle that all a priori knowledge is a priori known, which can be
formalized as follows:

Ka / ! Ka Ka /: 4:4
All three principles are surely too strong, but this only makes the negative result
stronger. I will show that neither (2.6) nor (2.9) is a theorem of MEAyy .
The proof will make use of a couple of facts about about undecidable arithmetical
sentences (Boolos et al. 2003, ch. 17, 18). Godels first incompleteness says that
there is no consistent, complete, axiomatizable extension of Q. If PA is consistent,
then it is not complete. In particular, there is a sentence, GPA , i.e. the sentence such
that PA GPA $ :9yProvPA pGPA q; y, and 0PA GPA . Moreover, if PA is x-
consistent, then 0PA GPA . It follows that both PA [ fGPA g and PA [ f:GPA g are
consistent. Let us refer to the first theory as PA and to the second theory as PA .
What will be needed is that neither theory proves GPA . It follows from the first
incompleteness theorem that 0PA GPA . It is true that, if PA /, then PA /. What
we need, is that this can be formalized in PA, i.e. PA 9yProvPA p/q; y
! 9yProvPA p/q; y. If this is the case, then PA :9yProvPA p/q; y !
:9yProvPA p/q; y. In particular, PA GPA ! GPA . But then 0PA GPA . So, the
question is whether one can prove that PA 9yProvPA p/q; y
! 9yProvPA p/q; y. The details depend on the details of arithmetization (Boolos
et al. 2003, ch. 15) and they will be omitted here, but the definition of
ProvPA p/q; y is identical to the definition of ProvPA p/q; y, except that the
clause axiome is defined as

logicalaxiome _ PAaxiome

in the second case and as

logicalaxiome _ PAaxiome _ G
odelsentencee
in the first case. Given these definitions, the theorem is essentially a case of dis-
junction introduction. Finally, note that both GPA and GPA are true in the standard
model of arithmetic and that neither GPA nor GPA are 9-rudimentary formulas.
Theorem 4.2 There is an interpretation M of LMEA such that M  MEAyy
but M 6 }Ka GPA ! GPA ^ hGPA ! GPA ! }Ka GPA ^ hGPA !
}Ka GPA and M 6 }Ka GPA ^ }Ka GPA ! GPA ! }Ka }GPA ^ }
GPA ! GPA :

123
Closure of A Priori Knowability

Proof Let MEA be identical to MEAyy , extended with / ! Ka /, for any

/ 2 LMEA . Let M be an awareness model, with the following characteristics:10
1. W fw1 ; w2 g;
2. D N;
3. RM W 2 ;
4. RE fhw1 ; w1 i; hw2 ; w2 ig;
5. V is N, the standard interpretation of the language of arithmetic;
6. S the set of variable assignment functions;
7. A is a function from W and S to LMEA , defined as follows:
(a) for any variable assignment a, Aw1 ; a f/ j C1 a MEA /g, with
C1 a fGPA g [ ft t0 j denM;a t denM;a t0 g;
(b) for any variable assignment a, Aw2 ; a f/ j C2 a MEA /g, with
C2 a fGPA ! GPA g [ ft t0 j denM;a t denM;a t0 g.
The key clause in the recursive definition of M  / is the clause for / Ka w:
M; w; a  Ka w if and only if w 2 Aw; a and M; w0 ; a;  w for all w0 such that
wRE w0 . Simplifying, this means that M; wi ; a  Ka w if and only if Ci a MEA w
and M; wi ; a;  w.
We need to check that, if MEAyy /, then M  /. The proof is by induction on
the complexity of proof in MEAyy .
Since RM is an equivalence relation, M  S5. Since RE is a reflexive relation,
M  (2.1). Now consider the case of (4.2). Suppose that MEAyy }Ka /, inferred
from MEAyy /. It follows from the latter by construction of Aw; a that / 2 Aw,
for any w 2 W. By the induction hypothesis, M  /. Therefore, M; w; a  Ka /
and, by the reflexivity of RM , it is also the case that M; w; a  }Ka /. Next,
consider the case of (4.3). Suppose that M; w; a  Ka / ! w ^ Ka /. It follows by
reflexivity of RE and the clauses for conjunction and material implication that
M; w; a  w. It is also the case that / ! w; / 2 Aw; a. Since Aw; a is by its
construction closed under modus ponens, it is also the case that w 2 Aw; a.
Therefore, M; w; a  Ka w. Finally, consider the case of (4.4). Suppose that
M; w; a  Ka / for some w, a. Then / 2 Aw; a and, by construction,
Ka / 2 Aw; a. Since fw0 j wRE w0 g fwg, it follows that M; w; a  Ka Ka /. As
for the first-order and arithmetical part, I will leave the proof to the reader, except
for noting that the ft t0 j denM;a t denM;a t0 g part of Aw; a is important for
proving a Principle of Replacement and a Principle of Agreement, which are used in
the proofs of universal instantiation and universal generalisation.
The model provides us with a counterexample to (1.3). First, note that GPA and
GPA are true in the standard model of arithmetic and, therefore, M  GPA and
M  GPA ! GPA . Second, note that, by definition, GPA 2 Aw1 ; a and

10
The model that will be constructed is of the type defined in (Heylen 2013, Sect. 4, especially
Definitions 4.10, 4.11). Definition 4.10 (Heylen 2013, p. 103) hard codes the Principle of Replacement,
but it ought to hard code the Principle of Agreement as well. The first is needed for proving the soundness
of Universal Instantiation. The second is needed to prove the soundness of Universal Generalisation. In
the model constructed here the Principle of Agreement and the Principle of Replacement come out true by
construction of the awareness sets.

123
 J. Heylen

GPA ! GPA 2 Aw2 ; a. Consequently, M; w1 ; a  Ka GPA and M; w2 ; a 

Ka GPA ! GPA . Therefore, M; w1 ; a  }Ka GPA and M; w1 ; a 
}Ka GPA ! GPA . Since M; w; a  / for all / that are true in the standard
model of arithmetic and for all w 2 W, it is also the case that M; w1 ; a  hGPA and
M; w1 ; a  hGPA ! GPA . The question then is whether there is a w 2 W such
that M; w; a  Ka GPA . It is sufficient for a negative answer that GPA 62 Aw1 ; a
and GPA 62 Aw2 ; a. For this it needs to be shown that C1 a 0MEA GPA and
C2 a 0MEA GPA .
I will demonstrate that the two cases under consideration can be reduced to
GPA 0PA GPA and GPA ! GPA 0PA GPA respectively. First, for every /, if
Ci a MEA /, then Ci aE PA /E , with E the eraser translation from LMEA to
LPA that erases all occurrences of } and Ka . Note that the eraser translation of any
identity sentence belonging to Ci a belongs to LPA and that GPA E GPA ,
GPA ! GPA E GPA ! GPA and GPA E GPA . Therefore, the case of
C1 a MEA GPA reduces to the case of C1 a PA GPA and the case of
C2 a MEA GPA reduces to the case of C2 a PA GPA . Next, if t1 and t2 are
closed terms and if t1 t2 2 Ci a, then PA t1 t2 , since all true rudimentary
sentences are theorems of Q. Therefore, one can take them out of Ci a without
affecting the deducibility. Call the reduced set Ci a0 . If Ci a0 PA GPA , then there
is a finite subset Ci a00 of Ci a0 such that Ci a00 PA GPA . Let ti tj , ..., tk tl be
the identity statements belonging to C1 a00 . If C1 a00 PA GPA , then
GPA ; ti tj ^ . . . ^ tk tl PA GPA . Let x1 ; . . .; xn be the free variables in the
assumption. By the instantiation rule for the existential quantifier, if

GPA ; ti tj ^ . . . ^ tk tl PA GPA , then GPA ; 9x1 . . .9xn ti tj ^ . . . ^ tk tl
PA GPA . Indeed, GPA and GPA are sentences, so they do not contain any free
variables. Note that ti tj ^ . . . ^ tk tl is satisfied in the standard model relative


to a and, therefore, 9x1 . . .9xn ti tj ^ . . . ^ tk tl is true in the standard model as
well. The latter sentence is a 9-rudimentary sentence and, consequently, it is a
theorem of PA. Therefore, one can take it out without affecting the deducibility.
Analogous reasoning can be used in the second case. Hence, one can further reduce
the two cases to GPA PA GPA and GPA ! GPA PA GPA respectively.
The essential point is that neither PA nor PA [ fGPA ! GPA g entail GPA . The
first point was made in the introduction to the proof. The second point follows from
the fact that 0PA GPA or :GPA 0PA GPA , which was also noted in the
introduction to the proof.
The second question is whether there is a w 2 W such that
M; w; a  Ka GPA ^ GPA ! GPA ;

which is answered in the negative if

GPA ^ GPA ! GPA 62 Aw1 ; a

and
GPA ^ GPA ! GPA 62 Aw2 ; a:

123
Closure of A Priori Knowability

The latter is the case if

Ci a 0MEA GPA ^ GPA ! GPA :

This can again be reduced to

GPA 0PA GPA ^ GPA ! GPA

and

GPA ! GPA 0PA GPA ^ GPA ! GPA ;

which follow from the two facts noted in the introduction to the proof. h
Let us recapitulate. It was proved that neither Andersons (1993) (2.6) nor
Horstens (2000) (2.9) are provable even when they are restricted to arithmetic and
very strong modal and epistemic assumptions are made (Theorem 4.2). This is in a
way disappointing, since it was also proved that, if the antecedent in (2.6) is an
arithmetical formula and if the consequent is an arithmetical formula of a certain
low complexity, then (2.9) follows from modest modal and epistemic assumptions
(Theorem 4.1). Of course, it is open to defenders of (2.6) or (2.9) to suggest
additional principles, e.g. (3.2), which do the trick and which are independently
plausible. However, I think that it is fair to say that at this stage of the debate it is up
to them to make a move.
The foregoing highlights the difficulty of finding a positive justification for (2.6)
or (2.9). It was argued in Sect. 2 that the attempts by Anderson (1993) and Horsten
(2000) to justify those principles fail. In this section it has been argued that (2.6) is
justified on two conditions. First, the material implication and its antecedent have to
be true if possibly true. Second, the consequent has to be a priori knowable if true.
When the material implication is arithmetical and the consequent is an arithmetical
formula of a certain low complexity, then those two conditions are satisfied. The
challenge is to justify (2.6) when the second assumption is not made or is not
provable given some background assumptions. This should not be mistaken for the
claim that the second assumption is false. The case when the material implication is
arithmetical but the consequent is of a higher arithmetical complexity, is a nice case
in point. Again, it is not assumed that the formula in question, viz. GPA , is really a
priori unknowable.
In fact, Myhill (1960, p. 463) has made an interesting case for the a priori
knowability of the arithmetical statement that says that PA is consistent. The
consistency statement for PA is equivalent with the Godel sentence for PA (Boolos
et al. 2003, pp. 233234). Godel (1995, p. 290 ff.) claimed that the following
disjunction is true: either the mind is not a Turing Machine or there exist absolutely
unknowable diophantine equations, which are 8-rudimentary formulas (Godel
(1995, p. 156 ff.). Ever since Godel advanced his disjunction, attempts have been
made to argue for one of the disjuncts. In particular, Leitgeb (2009) and Horsten
(2009) have investigated the limits of knowledge about arithmetic.11 But to repeat,
11
For a critical reply to Horsten (2009), see Heylen (2010).

123
 J. Heylen

the argument developed here does not depend on the assumption that Godels
second disjunct is true or on the more general assumption that there exist a priori
unknowable propositions, although the argument does assume that this has not been
ruled out.
In Sect. 1 it was pointed out that Dretske (2005) has explicitly criticized (1.1) and
that he has implicitly challenged (1.2). In addition, his counterexample also affects
logical competence, since it is easy to know that you have hands and it is easy to
know that, if you have hands, you are not a handless brain-in-a-vat, but it is
impossible to know that you are not a handless brain-in-a-vat. Of course, this does
not concern the notion of a priori knowledge. Interesting from our current
perspective is that there is an analogy between Dretskes argument and the proof of
Theorem 4.2: GPA is an easily knowable sentence of lower arithmetical complexity;
GPA is an arithmetically more complex sentence that is not knowable; GPA ! GPA
is knowable. A disanalogy is that, unlike the statement that connects having hands
with not being a handless brain-in-a-vat, GPA ! GPA is not a logical truth. To turn
the proof in a Dretske-style argument against (1.3) it is necessary but not sufficient
to argue for the existence of a priori unknowable arithmetical sentences of a certain
complexity. This falls outside the scope of this article.12

5 Closure and Logical Competence

If it is already challenging to find a deductive justification for (2.6) and (2.9) when
they are restricted to arithmetic, it is also very challenging to find a deductive
justification for those principles when they are unrestricted. As I will argue next, the
justification challenge is even stronger than one might think, since there is a good
alternative, which in one form or another is upheld by both epistemic logicians and
mainstream epistemologists. Since the work of Kuhn (1962) and Lakatos (1970),
philosophers of science are very much aware of the fact that scientists only leave a
paradigm or research program if there is an alternative paradigm or research
program available.
It is an important task for epistemic logicians to steer between the Scylla of
logical omniscience and the Charybdis of logical incompetence: epistemic agents
should be represented as agents that do not know all logical truths or all logical
consequences of what they already know, but for whom those logical truths and
logical consequences are knowable. Besides (1.1), some examples of logical
omniscience are:

K / ^ w ! K/ ^ Kw;
K/ ^ Kw ! K / ^ w;
K/ ! K / _ w;
K::/ ! K/:

12
The paradox of Fitch (1963) is not of any help here, since even if p is an arithmetical statement, the
formula p ^ :Ka p is not an arithmetical statement.

123
Closure of A Priori Knowability

It can easily be seen that all the examples of logical omniscience follow from the
closure of knowledge under logical theoremhood and (4.3). Together these two
principles are equivalent to the following generalized logical omniscience rule:

Cw ) K C Kw; 5:1

with K C fK/ j / 2 Cg. Some examples of logical competence are:

K / ^ w ! }K/ ^ }Kw;
K/ ^ Kw ! }K / ^ w;
K/ ! }K / _ w;
K::/ ! }K/:

The examples of logical competence can easily be seen to subsume under the
following generalized rule of logical competence:

Cw ) K C }Kw: 5:2

Logical omniscience has been widely criticized, because it presents human knowers
as having divine cognitive capabilities. However, one should also avoid presenting
human knowers as being cognitively handicapped.13 If you know a material con-
ditional and you know its antecedent and if you are logically competent, then you
can expand your knowledge by applying modus ponens, although you may not have
done so yet. Given enough time, attention and other resources, logically competent
reasoners should be able to know the consequent of a conditional if they already
know the antecedent of that conditional and the conditional itself, even if they have
not devoted the time, attention and other resources to it.
A different position between logical omniscience and logical incompetence is the
claim that epistemic agents have the possibility to know all the logical consequences
of what they can know. Examples of this general principle are:

}K / ^ w ! }K/ ^ }Kw;
}K/ ^ }Kw ! }K / ^ w;
}K/ ! }K / _ w;
}K::/ ! }K/:

The above examples all subsume under the following rule of closure of knowability
under logical deducibility: which can be formalized as follows:

Cw ) }K C }Kw: 5:3

Williamson (2000, p. 117) and Hawthorne (2005, p. 2930) defend a closure

principle that is closely related to the logical competence principle (5.2). In the

13
See Heylen (2013, p. 99) for the problem of logical incompetence that presents itself for syntactical
models, awareness models and impossible worlds models, which were introduced to avoid the problem of
logical omniscience.

123
 J. Heylen

words of Hawthorne (2005, p. 29): If one knows some premises and competently
deduces Q from those premises, thereby coming to believe Q, while retaining ones
knowledge of those premises throughout, one comes to know that Q. There seem to
be four differences between what Williamson calls inuitive closure on the one
hand and (5.2) on the other hand. The first difference is that between a temporal
conception of knowability, expressed by coming to know, and a modal conception
of knowability, expressed by being able to know. Based only on this distinction
and taking into account that metaphysical or counterfactual possibility is weaker
than temporal possibility, the modal version of the closure principle is weaker and,
therefore, more easily justified than the temporal version. It is no surprise then that
this is compensated by the other three differences, which are all about restricting the
temporal version of the closure principle. The second difference is that Hawthorne
says that the conclusion of the deduction should come to be believed. The third
difference is that one should retain knowledge of the premises throughout. The
reason he gives for this is that one may get counterevidence to the premises in the
meantime. The fourth difference is that only deductions with at least one premise
are considered. Whether one should also impose these restrictions on the logical
competence principle or not, the crucial point is that the intuitive closure principle
defended by mainstream epistemologists is a version of (5.2) rather than a version of
(5.3), since knowledge of the premises rather than possible knowledge of the pre-
mises is required.
The general rules (5.1)(5.3) are related to the closure principles with a priori
knowledge. Clearly, (4.3) is an instantiation of (5.1) and, moreover, (5.1) is
equivalent to (4.3) and closure of a priori knowledge under logical theoremhood.
Horsten (2000, p. 60) endorses an axiom scheme from which an instantiation of
(5.2) follows:

Ka / ! w ! Ka / ! Ka w: 5:4
In Sect. 2 it was pointed out that, when Anderson (1993, p. 9) tries to give a
justification for his closure principle, he actually ends up supporting (2.8), which
can now be seen to be a special case of (5.2). Finally, it can easily be seen that (5.3)
entails (1.3) and is, in fact, equivalent to (1.3) and closure of a priori knowability
under logical theoremhood. Interestingly, neither (5.1) nor (5.2) justify either (2.6)
or (2.9). It is easily checked that (5.1) and (5.2), when restricted to a priori
knowledge and when formulated in terms of provability-in-MEAyy , are derivable
rules in MEAyy . Hence, one can use Theorem 4.2 to back up this claim.
Let us summarize the situation. There is an alternative closure principle, called a
logical competence principle, for which a reasonable case can be made that is
based on the aims of epistemic logic and based on the work of mainstream
epistemologists. In addition, Anderson (1993) and Horsten (2000) endorse some
instantiations of the principle already. Moreover, the logical competence principle
does not entail (1.3), even if one adds some strong modal and epistemic
assumptions. So, it looks like there is an alternative to (1.3). This leaves us with
checking whether it is a good alternative.

123
Closure of A Priori Knowability

It can be proved that (5.2) is immune to the Socratic objection. This can also be
proved with the help of the model used in the proof of Theorem 5.2, if one replaces
GPA ! GPA in the definition of Aw2 ; a with :GPA , which does not affect the
proof. It is then case that M; w2 ; a  Ka :Ka GPA . First, :Ka GPA 2 Aw2 ; a,
because / $ Ka / 2 Aw2 ; a for every / and :GPA 2 Aw2 ; a. Second,
M; w2 ; a  :Ka GPA , since GPA 62 Aw2 ; a. But M; w1 ; a  Ka GPA . I leave it to
the reader to check that (3.1) and (3.4) are true on the model as well. Furthermore,
(5.2) blocks Horstens (2000) counterexample, since (2.4) and (2.5) are of the form
}Ka /, not of the form Ka /. It was pointed out in Sect. 2 that there are reasons to
doubt the efficacy of Andersons counterexample, so I will ignore that example
here. It should also be emphasized that (5.2) does all this without being restricted,
unlike Andersons (2.6) or Horstens (2000) (2.9).
To sum up, there is a good, if not better, alternative to (1.3). This strengthens the
justification challenge made in Sect. 4.

6 Summary

The subject of this article is the closure of a priori knowability under a priori
knowable material implication. In a restricted version it is arguably correct: if the
material implication and the antecedent are true if possibly true, and if the
consequent is a priori knowable if true, then the principle is correct. These
conditions are fulfilled when the material implication is arithmetical and when the
consequent has a certain low arithmetical complexity (see Sect. 4.1). Philosophers
and logicians have endorsed the closure principle even when not all of the
conditions mentioned earlier are satisfied. For instance, Fritz (2013) accepts the
principle full stop. Anderson (1993) and Horsten (2000) are more cautious. They
each present counterexamples, although it was pointed out that Andersons
counterexample is not unequivocally successful. In reaction to their own counte-
rexamples they each put forward a certain restriction strategy. Horsten suggests that
the closure principle is safe when the language is not highly expressive and
Anderson proposes that the closure principle is secure when restricted to necessary
truths (see Sect. 2).
Against Horstens restriction strategy I presented the so-called Socratic
objection: even with very few expressive resources and with modest background
assumptions, one can deduce from the closure principle that knowable ignorance
implies necessary ignorance, which is a highly implausible consequence (see
Sect. 3). Against Andersons restriction strategy I put the so-called justification
challenge: if even with very strong modal and epistemic background assumptions,
one cannot deduce the closure principle restricted to necessary, arithmetical truths,
the question arises how exactly to justify it (see Sect. 4.2). To raise the stakes, I
introduced a so-called logical competence principle, which says that if all the
premises of a logical deduction are known, then the conclusion is knowable. It is
claimed that an adequate epistemic logic, which steers between the horn of logical
omniscience and the horn of logical incompetence, should adopt the logical
competence principle. Moreover, it was pointed out that there is support in

123
 J. Heylen

mainstream epistemology for versions of the logical competence principle. Finally,

it was shown that the logical competence principle is not hit by the Socratic
objection or Horstens counterexample and that it accomplishes this without any
restrictions. Taking into consideration the existence of a superior alternative, the
question arises why one should try to find a justification for Andersons restricted
closure principle at all.

References

Anderson, C. A. (1993). Toward a logic of a priori knowledge. Philosophical Topics, 21(2), 120.
Boolos, G. S., Burgess, J. P., & Jeffrey, R. C. (2003). Computability and Logic (4th ed.). New York:
Cambridge University Press.
Dretske, F. I. (2005). Is knowledge closed under known entailment? The case against closure. In E. Sosa.,
& M. Steup (Eds.), Contemporary debates in epistemology (pp. 1326). Oxford: Blackwell.
Dretske, F. I. (1970). Epistemic operators. Journal of Philosophy, 67(24), 10071023.
Fitch, F. B. (1963). A logical analysis of some value concepts. Journal of Symbolic Logic, 28(2),
135142.
Fritz, P. (2013). A logic for epistemic two-dimensional semantics. Synthese, 190(10), 17531770.
Godel K, (1995). Collected Works: Volume III Unpublished Essays and Lectures. Oxford University
Press.
Hawthorne, J. (2005). The case for closure. In E. Sosa., & M. Steup (Eds.),Contemporary debates in
epistemology. Oxford: Blackwell.
Heller, M. (1989). Relevant alternatives. Philosophical Studies, 55(1), 2340.
Heller, M. (1999). Relevant alternatives and closure. Australasian Journal of Philosophy, 77(2), 196208.
Hendricks, V. (2006). Mainstream and formal epistemology. Cambridge: Cambridge University Press.
Heylen, J. (2010). Descriptions and unknowability. Analysis, 70(1), 5052.
Heylen, J. (2013). Modal-epistemic arithmetic and the problem of quantifying in. Synthese, 190(1),
89111.
Holliday, W. H. (2012). Knowing what follows: Epistemic closure and epistemic logic. PhD thesis,
Stanford University.
Holliday, W. H. (2014). Epistemic closure and epistemic logic I: Relevant alternatives and subjunctivism.
Journal of Philosophical Logic, 162. doi:10.1007/s10992-013-9306-2.
Horsten, L. (1994). Modal-epistemic variants of Shapiros system of epistemic arithmetic. Notre Dame
Journal of Formal Logic, 35(2), 284291.
Horsten, L. (1998). In defense of epistemic arithmetic. Synthese, 116(1), 125.
Horsten, L. (2000). Models for the logic of possible proofs. Pacific Philosophical Quarterly, 81(1),
4966.
Horsten, L. (2009). An argument concerning the unknowable. Analysis, 69(2), 240242.
Kuhn, T. S. (1962). The structure of scientific revolutions. Chicago: The University of Chicago Press.
Lakatos, I. (1970). Falsification and the methodology of scientific research programmes. In Criticism and
the growth of knowledge (pp. 91195). Cambridge: Cambridge University Press.
Leitgeb, H. (2009). On formal and informal provability. In O. Bueno., & O. Linnebo (Eds.), New Waves
in Philosophy of Mathematics. Palgrave Macmillan.
Myhill, J. (1960). Some remarks on the notion of proof. The Journal of Philosophy, LVII(14), 461471.
Nozick, R. (1981). Philosophical Explanations. Cambridge: Harvard University Press.
Salmon, N. (1989). The logic of what might have been. The Philosophical Review, 98(1), 334.
Shapiro, S. (1985). Epistemic and intuitionistic arithmetic. In S. Shapiro (Ed.), Intensional mathematics.
Amsterdam: North-Holland
Sosa, E. (1999). How must knowledge be modally related to what is known? Philosophical Topics, 26(1/
2), 373384.
Williamson, T. (2000). Knowledge and its limits. Oxford: Oxford University Press.
Williamson, T. (2007). The philosophy of philosophy. Oxford: Blackwell.

123