Professional Documents
Culture Documents
Olkhovikov
Inference as Doxastic
Heinrich Wansing Agency. Part I: The Basics
of Justification Stit Logic
Abstract. In this paper we consider logical inference as an activity that results in proofs
and hence produces knowledge. We suggest to merge the semantical analysis of delib-
eratively seeing-to-it-that from stit theory (Belnap et al. in Facing the future: agents and
choices in our indeterminist world, Oxford University Press, New York, 2001) and the
semantics of the epistemic logic with justification from (Artemov and Nogina in Journal
of Logic and Computation 15:1059–1073, 2005). The general idea is to understand proving
that A as seeing to it that a proof of A is (publicly) available. We introduce a semantics
of various notions of proving as an activity and present a number of valid principles that
relate the various notions of proving to each other and to notions of justified knowledge,
implicit knowledge, and possibility. We also point out and comment upon certain principles
our semantics fails to validate.
Keywords: Proofs as acts, Doxastic agency, Epistemic logic, Justification logic, dstit logic.
1. Introduction
Studia Logica
c
https://doi.org/10.1007/s11225-017-9779-z Springer Science+Business Media B.V., part of Springer Nature 2018
G. K. Olkhovikov, H. Wansing
1
If proving is a source of knowledge, and knowledge, as is generally assumed, implies
truth, then an agent proves that A only if A is true. Note that in natural language ‘to
know’ is factive not only in the sense of entailing the truth of its complement but also
in the sense of presupposing the truth of its complement, as the verb ‘to know’ preserves
truth entailment under negation, see, for example, [7]. This may be seen as a pragmatic
property of ‘to know’; in any case it is not captured by the familiar Kripke semantics for
epistemic logic.
Inference as Doxastic Agency...
velop a semantics of ‘to prove’ as an action verb that merges (i) the seman-
tical analysis of deliberatively seeing-to-it-that from stit theory and (ii) the
semantics of Sergei Artemov and Elena Nogina’s [1] epistemic logic with
justification.2 The general idea is to understand proving that A as seeing
to it that a proof of A is (publicly) available. The paper focuses on a con-
ceptual analysis and on a number of conceptual distinctions the suggested
semantical framework allows one to draw. In Section 2 we will recapture the
required prerequisites, namely dstit logic (Section 2.1) and epistemic logic
with justification (Section 2.2). In Section 3, we shall introduce a semantics
of various notions of proving as an activity. In particular, we present a num-
ber of valid principles that relate the various notions of proving with each
other and with notions of justified knowledge, implicit knowledge, and possi-
bility. We also point out and comment upon certain principles our semantics
fails to validate. Finally, in Section 4 we will come back to the distinction
between proofs as abstract objects versus proofs as acts, and close with a
brief summary and an outlook.
2. Prerequisites
In order to keep this paper self-contained, we here present all the prerequis-
ites needed for our discussion of proving as a form of doxastic agency in
Section 3.
2
An interpretation of imagination ascriptions that merges the semantics of deliberat-
ively seeing-to-it-that with the neighbourhood semantics of classical modal logic is pre-
sented in [20, 35]. Other combinations of deliberatively seeing-to-it-that with various modal
operators can be found in bdi-stit logic [26–28], see also [33].
G. K. Olkhovikov, H. Wansing
and, independently, John Horty [11]. In the more recent past, several other
kinds of stit-operators have been suggested and investigated, among them
Jan Broersen’s xstit-operators [4] that express seeing-to-it that something is
the case in the next moment of time, assuming branching time structures
for discrete time.
Syntax The dstit-operators are definable in terms of the so-called cstit-
operators (Chellas-stit-operators), an operator for historical necessity (also
known as the settled-true-operator), and negation. The language of dstit
logic extends the language of classical propositional logic, for instance in
the connectives ¬ (negation) and ∧ (conjunction). Disjunction, ∨, impli-
cation, →, the biconditional, ↔, and the zero-place connectives and ⊥
are then introduced as usual. If j is any index from a finite set of agent
indices and p is any propositional variable from a countably infinite set V ar
of propositional variables, then the formulas of dstit logic are defined as
follows: A := p | ¬A | A ∧ A | [c]j A | A.
All agent indices are assumed to stand for corresponding pairwise distinct
agents from a set Ag of agents.3 A formula A is read as “A is settled true”,
and [c]j A is read as “agent j cstit-realizes A”. The expression [d]j A, which
is read as “agent j dstit-realizes A” or “agent j deliberatively sees to it that
A”, is defined as [c]j A ∧ ¬A, and ♦A (“it is possible that A”) is defined
as ¬¬A.
Since dstit-logic extends classical logic, an axiomatic system for it may
make use of any complete axiomatization of classical logic, for example the
one from [12].
Definition 2.1. The axiom system L for dstit logic consists of the following
axioms and rules:
A0 The axioms of classical propositional logic from [12]
A1 The standard S5-axioms for
A2 The standard S5-axioms for [c]j and every agent index j
A3 A → [c]j A for every agent index j
A4 (♦[c]j1 A1 ∧ . . . ∧ ♦[c]jn An ) → ♦([c]j1 A1 ∧ . . . ∧ [c]jn An ),
provided that all the j1 , . . . , jn are pairwise distinct
R1 A together with A → B implies B (modus ponens)
R2 A implies A (necessitation).
3
For every j ∈ Ag we could write j for the agent index corresponding to j, but omitting
the bar will not give rise to any confusion.
Inference as Doxastic Agency...
4
For convenience, we often use ‘∀’, ‘∃’, ‘&’, ‘⇒’, and ‘⇔’ in the metalanguage to stand
for universal and particular quantification, and Boolean conjunction, implication, and bi-
implication, respectively.
G. K. Olkhovikov, H. Wansing
Let M = T ree, ≤, Ag, Choice, V be a dstit model, and let C be any class
of such models. A formula A is said to be true in M (written as M |= A) iff
M, (m, h) |= A for every (m, h) ∈ M H(M), and A is said to be valid with
respect to C (written as C |= A) iff M |= A for every M ∈ C. Let D be the
class of all dstit models.
Theorem 2.3. (Xu) For any formula A, L A iff D |= A.
Thus the semantics of stit logic extends the traditional modal structure of
states (here named moments) related to one another by a binary accessibility
relation (in this case ≤) by a set of histories and uses the fan Hm of histories
passing through a given moment to represent, for any given sentence A,
a degree in which A’s truth (or falsity) at the given moment m depends
on the future events. Under this reading, a statement like ‘tomorrow there
is going to be a sea-battle’ is true at a given m only for those histories
h ∈ Hm which verify this prediction. In cases such as this one, the truth
value of a sentence, even though evaluated at a given moment-history pair,
depends rather on a history than on a moment. On the opposite end of this
dependence-on-a-future spectrum are statements that deal with ‘hard’ or
accomplished facts like ‘this lawn is green’. Statements of this kind, when
true at a given moment m, will remain true at m whatever happens later.
It is easy to see that statements of the form [d]j A occupy a middle position
between statements of hard facts and statements about future contingencies:
when they are true at a moment, they are true through all of the histories in
a given choice cell of j but not necessarily through other choice cells. Thus
statements like [d]j A behave in a way not unlike statements of accomplished
facts but only within a restricted sub-fan of Hm . It is this idea of agency as
restricted factuality, as something lying in between a hard fact and a mere
contingency, that constitutes a central idea of the stit approach to the logic
of actions.5
5
The idea to distinguish between states of affairs and entities by the degree of their
dependence on future events can also be given an ontological reading. An event is only
dependent on the future when it has not fully revealed itself yet, when it has not yet made it
into being. The lawn revealed itself as green, so future events can not undo this. Tomorrow’s
Inference as Doxastic Agency...
Footnote 5 continued
sea battle has not yet revealed itself at all, therefore, only future can tell us about its
status. And the agentive events described by statements like ‘Smith opens the door’ are
revealing themselves even though they have not quite revealed themselves yet. These are
events in the making, entities in statu nascendi. Unlike future sea battles, they are present
already in the world, but unlike green grass and wooden chairs, they are not a fact yet so
that one can only see their results (like the closed door) or notice their influence on the
world once they are finished. This philosophy is also very manifest in the independence
of agents restriction given above for dstit models: actions, even though present already at
a given moment m, can only lead to something tangible (e.g., influence choices of other
agents) once they have fully revealed themselves but not when they themselves are still in
the making.
G. K. Olkhovikov, H. Wansing
The distinction between agentive and factual sentences in stit logic re-
flects the use of modalities available in the stit language for formaliz-
ing action descriptions as opposed to descriptions of more static matters
of fact. It is clear that one would prefer to formalize propositions like
‘Smith opens the door’ in such a way that for any dstit model M one gets
that
M |= Smith opens the door → [d]Smith Smith opens the door,
Let P V ar be the set of all proof variables and let P ol be the set of all proof
polynomials. The language of the Logic of Proofs can again be introduced as
extending the language of classical propositional logic in the connectives ¬
and ∧, the other connectives being defined as usual. If t ∈ P ol and p ∈ V ar,
6
It might seem that
M |= Smith opens the door → [d]Smith the door is open
is a better formalization of the agentive character of a door opening by Smith, whereas
our example seems to be somewhat pleonastic as to Smith’s presence in this whole situ-
ation. However, note that ‘[d]Smith the door is open’ is not quite precise as a presentation
of Smith’s actions; in fact, the latter expression might actually describe a situation where
Smith orders his butler Jones to open the door instead of opening the door himself. At the
same time ‘[d]Smith Smith opens the door’ can only be understood as describing an opening
of the the door by Smith and allows for no alternative readings.
Inference as Doxastic Agency...
All axioms of LP
E1 K(A → B) → (KA → KB)
E2 KA → KKA
E3 KA → A
R1 A together with A → B implies B (modus ponens)
R4 A implies KA
R5 c:A, if A is an axiom from LP0–LP4 or E1–E3,
and c is a proof constant
C1 t:A → KA.
Under the epistemic reading of t:A, the interaction axiom C1 is reminis-
cent of the idea that explicit knowledge implies implicit knowledge.7 Another
interaction axiom considered in [1] is C2: ¬t:A → K¬t:A. With the explicit
negative introspection axiom C2 one can prove the “decidability of evi-
dence”: Kt:A ∨ K¬t:A. The addition of axiom C2 to S4LP results in
a proof system referred to as S4LPN. If CS = {c1 :A1 , c2 :A2 , . . .}, where
each Ai is an axiom of S4LP (S4LPN) and each ci is a proof constant, then
CS is said to be a constant specification, and S4LPCS (S4LPNCS ) is the sub-
system obtained from S4LP (S4LPN) by restricting the rule R5 to formulas
from CS. We consider the axiom C2 as an additional principle that may or
may not be added to the basic systems S4LPCS
Semantics. A model of epistemic logic with justification, or justification
model, is a structure M = W, R, Re , E, V , where W is a non-empty set of
states (possible worlds), R and Re are reflexive and transitive binary rela-
tions on W , and R ⊆ Re . The relation R is used to interpret the operator
K of implicit knowledge, Re is a binary “evidence accessibility relation” on
W , and E is a function that maps pairs consisting of a world and a proof
term t to sets of formulas. Intuitively E maps a pair (u, t) to the formulas for
which t is admissible evidence at world u. The valuation V is an arbitrary
mapping from propositional variables to subsets of W .
7
The logic of general awareness introduced by Fagin and Halpern in [8] is guided by
the idea that explicit belief is implicit belief plus the agent’s awareness of the proposition
in question, and their semantics of explicit belief imposes a syntactic awareness filter on
implicit belief. Although the notion of awareness is, in general, clearly to be distinguished
from the concept of justification, Fagin and Halpern [8, p. 41] emphasize that “[t]he notion
of awareness . . . in this approach is open to a number of interpretations”, including an
interpretation as the agent’s capability to compute the truth value of a formula within a
certain time or space bound. Artemov and Nogina [1] consider formulas t:A as expressing
both justified and explicit knowledge that A, and their semantics for formulas of the form
t:A, too, imposes a syntactic filter on implicit knowledge, cf. Definition 2.7.
Inference as Doxastic Agency...
The first four conditions in this series, in fact, just explicate what we have
informally stated in this section thus far, where the fourth condition states
the principle that everything getting onto the whiteboard stays on the white-
board. The other conditions might seem less obvious, so we briefly comment
on them.
No new proofs guaranteed The fact that a proof is in h∈Hm (Act(m, h))
means that a proof is presented at all histories through a given moment.
In other words, the presence of that proof to the community is already an
accomplished fact. This can only happen if this proof was presented at an
earlier moment where its existence was dependent on the actions by some
agents in the community. Otherwise the only option would be that a new
proof which was not presented earlier springs into existence as a hard fact.
This possibility is ruled out by our restriction.
Presenting a new proof makes histories divide This restriction is just an
application of the principle behind the ‘no choice between undivided his-
tories’ restriction of dstit models. Since presenting new proofs is typically
due to activity by agents, the presentation of different proofs already con-
stitutes a choice by the presenting agents and thus enforces the branching
of histories.
Future always matters Those moments which are related to m by the
causal order ≤ are all the variants of the future for m. It is quite natural
that the possible future of a given state m is always relevant to what is
compatible with what is implicitly known in m and to weighing the evidence
in m, and this is exactly what this restriction says in addition to R ⊆ Re .
Inference as Doxastic Agency...
Indeed, one half of this is just the no new proofs guaranteed constraint, and
the other half can be shown as follows. Assume that m < m, h ∈ Hm and
that t ∈ P ol is such that t ∈ Act(m , h). Let h ∈ Hm be arbitrary. It is clear
that h is undivided from h at m , therefore, by presenting a new proof makes
histories
divide constraint, we get that Act(m , h) = Act(m , h ). Hence t ∈
h ∈Hm (Act(m , h )). It follows then by the expansion of presented proofs
that also t ∈ h ∈Hm (Act(m, h )).
On the other hand, note that within the above system of semantic con-
straints, one can also equivalently replace the expansion of presented proofs
and no new proofs guaranteed conditions with (1). Indeed, the no new proofs
guaranteed constraint is just one half of (1). As for the expansion of pre-
sented proofs constraint, assume that m < m, h ∈ Hm and that t ∈ P ol is
such that t ∈ Act(m , h). Then:
t∈ (Act(m1 , g)) = (Act(m, g)) ⊆ Act(m, h).
m1 <m,g∈Hm g∈Hm
We now proceed to define the language of our logic. It will extend both
the language of dstit logic and the language of S4LP. The semantics of
stit formulas carries over from basic stit logic since justification-stit models
extend dstit models. The semantics for S4LP formulas carries over with the
following minor modifications:
M, (m, h) |= KA ⇔ (∀m ∈ T ree)((mRm & h ∈ Hm )
(2)
⇒ M, (m , h ) |= A)
M, (m, h) |= t:A ⇔ (A ∈ E(m, t) &
(3)
(∀m ∈ T ree)((mRe m & h ∈ Hm ) ⇒ M, (m , h ) |= A)),
where M is a justification-stit model and (m, h) ∈ M H(M).
The new components in the justification-stit language are certain modal-
ities representing the proving activity of agents. W.r.t. these modalities both
G. K. Olkhovikov, H. Wansing
Agentive Factual
Explicit j proves A by t A has been proven by t
P rove(j, t, A) P roven(t, A)
Implicit j proves A A has been proven
P rove(j, A) P roven(A)
is a negative condition for the truth of the assertion that someone proves
A.8 There is also a positive condition: proving A must ensure that a proof of
A is presented. One way of writing these conditions would be, for instance:
M, (m, h) |= P rove(j, A) ⇔ (∀h ∈ Choicem
j (h))(∃t ∈ Act(m, h ))
(M, (m, h ) |= t:A)
& (∀s ∈ P ol)(∃h ∈ Hm )(M, (m, h ) |= s:A ⇒ s ∈
/ Act(m, h )).
(6)
The first of the conjuncts in the above clause corresponds to the positive
condition, the second to the negative one. Note the order of quantifiers in
the positive condition: it says that when proving A, an agent ensures that
some proof of A is presented; but this proof may look different depending
on how the future unfolds. In our opinion, this clause captures the creative
character of proving things which manifests itself in the situation that one
cannot typically predict what the final proof will look like before the proving
process is actually finished. The negative clause, on the other hand, says that
none of A’s proofs have yet become an accomplished fact. In this formulation
the matter seems to be simple enough; however, once we try to take a closer
look at the details, the complications begin. Note, for example, that our
clause ensures that no proof of A is established as an accomplished fact;
however it does not exclude a scenario where the process of proving A is an
accomplished fact. Indeed, consider the example in Figure 1.
8
One might feel that our condition of novelty is too restrictive, given the actual usage
of contexts like ‘j proves A’ in natural language. Indeed, it is common to say things like
‘Smith proves Gödel completeness theorem in his Logic 101 course’. However, this type
of situations is already excluded by our setting. Our community is engaged in finding out
new proofs rather then teaching them, and given that every proof, once presented to the
community, remains presented forever, any attempts at exposing known proofs would have
void results, since the state of Act, the community’s whiteboard, would remain unchanged.
Of course, one could still claim that even if the repetition of known proofs makes no sense
in our setting, the presentation of new proofs for already known propositions might well
have it. This latter point is legitimate; however, we also think that in cases of new proofs for
old truths the logical emphasis is naturally on the proof itself rather than the proposition
proved; that is to say, it is far more natural to describe situations of this type in explicit
mode. One does not normally say that Jerzy L os [16] proved the compactness theorem
for first-order logic since by the time L os presented his proof, this theorem was known
for some 20 years. The most common description would be something like ‘Los proved
the compactness theorem for first-order logic using the ultrafilter construction’. Thus the
discussions of novel proofs for known truths are better accounted for when dealing with the
semantics of an explicit agentive proving modality, and in fact, our clause for P rove(j, t, A)
takes these considerations into an account.
G. K. Olkhovikov, H. Wansing
h h
t t
m
Act(m, h) Act(m, h )
Examples like this may (or may not) motivate us to consider strengthen-
ing the negative condition for P rove(j, A), say in the following form:
M, (m, h) |= P rove (j, A) ⇔ (∀h ∈ Choicem
j (h))(∃t ∈ Act(m, h ))
(M, (m, h ) |= t:A) & (∃h ∈ Hm ) (7)
(∀s ∈ Act(m, h ))(M, (m, h ) |= s:A).
Here one requires that were it not for the actions of j, the proving agent, A
might not end up being proved. As this alternative and stricter version of an
implicit agentive proving modality also seems to us a legitimate reading of
it, we would like to keep it in our logic. We achieve this by using a somewhat
different notation so as to make it distinguishable from P rove(j, A), which
we interpret according to (6).
We now turn to the explicit proving modalities, beginning with
P roven(t, A). Again, the intuitions behind this factual version seem to be
simple enough: the presentation of t to the community must take place as
an accomplished fact, and further, t must constitute an acceptable proof of
A. These intuitions are then summed up in the following clause:
M, (m, h) |= P roven(t, A) ⇔ (∀h ∈ Hm )
(t ∈ Act(m, h ) & M, (m, h) |= t:A). (8)
With the agentive version of an explicit proving modality, matters are
less clear. When one asserts that j proves A by t, one would at least expect
that j presents t to the community and that t in fact proves A. This is a
positive condition. Then a negative condition would involve at least that t
Inference as Doxastic Agency...
was not presented earlier, for otherwise, given the perfect memory of the
community, j’s actions would have zero influence on the state of discussions
within the community. If we do not add any further conditions, we are left
then with the following definition:
M, (m, h) |= P rove(j, t, A) ⇔ (∀h ∈ Choicem
j (h))(t ∈ Act(m, h )
(9)
& M, (m, h ) |= t:A) & (∃h ∈ Hm )(t ∈ Act(m, h )).
However, note that this reading of P rove(j, t, A), again, allows for a sit-
uation when (i) A is being proved (possibly by different proofs) throughout
the elements of Hm . Moreover, it even allows for a situation when (ii) A was
actually proved earlier by another proof so that j just announces another
proof of a known fact to the community. While in situations like (i) and (ii)
one can still say, in some sense, that j proves A with t, in other cases this
modality seems to be used in a narrower meaning. Again, one may want to
imply either (a) that A was not proved to the community earlier or (b) were
it not for the efforts of j to present t to the community, A might not have
been proven. Now, these stronger versions of the negative condition were
already discussed w.r.t. implicit agentive proving modalities so we can just
borrow the respective conjuncts from their definitions. This gives us two
further modalities:
M, (m, h) |= P rove (j, t, A) ⇔
(∀h ∈ Choicem
j (h))(t ∈ Act(m, h ) & M, (m, h ) |= t:A)
& (∀s ∈ P ol)(∃h ∈ Hm )(M, (m, h ) |= s:A ⇒ s ∈
/ Act(m, h )),
(10)
M, (m, h) |= P rove (j, t, A) ⇔
(∀h ∈ Choicem
j (h))(t ∈ Act(m, h ) & M, (m, h ) |= t:A) &
(∃h ∈ Hm )(∀s ∈ Act(m, h )(M, (m, h ) |= s:A). (11)
Let M = T ree, ≤, Ag, Choice, R, Re , E, Act, V be any justification-stit
model, and let C be any class of such models. A formula A is said to be true
in M (written as M |= A) iff M, (m, h) |= A for every (m, h) ∈ M H(M),
and A is said to be valid with respect to C (written as C |= A) iff M |= A
for every M ∈ C.
A formula is said to be universally valid (or just valid) iff it is valid with
respect to the class of all justification-stit models.
We now proceed to discussing logical properties of various modalities
defined in our justification stit environment. First, we check that both
modalities for acts of proving and justification stit realizations of S4LP
G. K. Olkhovikov, H. Wansing
modalities display the intended properties w.r.t. the distinction between fac-
tual and agentive propositions. The matter is summed up in the following
theorem:
Thus, one may observe that all of the above defined factual proving
modalities, as well as the justification stit realizations of modalities bor-
rowed from S4LP turn out to be factual, just as expected. As for agentive
proving modalities, these are seen to be mostly strictly agentive w.r.t. the
proving agent, with the notable exception of P rove(j, A), which is agentive
but not strictly agentive. The reason for this is the weakness of the negative
condition for P rove(j, A) which, as we have already shown, does not preclude
P rove(j, A) from being true throughout a given history fan Hm . Note how-
ever that this weakness does not destroy the property of strict agentiveness
in case of explicit modalities P rove(j, t, A) and P rove (j, t, A) even though
the latter uses the same negative condition as P rove(j, A), and the negative
condition for the former is even weaker. Such are the beneficial effects of
explicitly mentioning the proof polynomial involved in the situation: though
P rove(j, A) can still obtain throughout Hm together with P rove(j, t, A) or
P rove (j, t, A) being true at a given pair (m, h), even the weak negative con-
ditions associated with these explicit modalities are already strong enough
to enforce that even in this case, A will be proved by proofs different from
t in other histories running through m.
Inference as Doxastic Agency...
Note that part one of this theorem is actually an easy corollary of the
above Theorem 3.1. Indeed, assume that a formula A → [d]j A is universally
valid for some j ∈ Ag. This means that A → ¬A is universally valid
by the definition of the dstit-operator. Assume then, that t:A is satisfied
in some model M at some (m, h) ∈ M H(M). We then get M, (m, h) |=
t:A, therefore, M, (m, h) |= A by Theorem 3. On the other hand, we
get M, (m, h) |= A by LP4 and further, M, (m, h) |= ¬A by validity of
A → ¬A. This gives us a contradiction, therefore t:A cannot be satisfied.
As for Theorem 3.3(2), it is less obvious and its proof is instructive enough
to at least sketch it here, for this proof provides a nice illustration of the
meaning and use of almost all semantical restrictions introduced above. We
have two cases to consider. In Case 1, we assume that m ∈ T ree is the
last moment w.r.t. ≤. This means that there is only one history h passing
through m. But then P rove(j, A) cannot be true at (m, h), and therefore,
by LP4, t:P rove(j, A) cannot hold at (m, h) as well. Indeed, if P rove(j, A)
holds at (m, h) then by the positive condition of P rove(j, A), there must
be an s ∈ P ol such that s ∈ Act(m, h) and s proves A. On the other
hand, by the negative condition for P rove(j, A), since s proves A, there
must be a history h passing through m such that s ∈ / Act(m, h). Since we
already assumed that s ∈ Act(m, h) and h is the only history passing through
m, this is a contradiction, showing that P rove(j, A) is false at (m, h). For
Case 2, we assume that m is not the last moment, that is to say, that
there is a moment m such that m < m . Choose an arbitrary history h
passing through m and assume that t:P rove(j, A) is true at (m, h). Since
Inference as Doxastic Agency...
histories are defined as maximal chains of moments, we may assume that all
histories passing through m are in Choicem j (h) — otherwise we can choose
another m . Since ≤ is subsumed by the evidence relation Re , we know that
P rove(j, A) must be true at (m , h ) for every h passing through m . Now,
by m < m we know that all histories passing through m are undivided at
m, therefore, since presenting new proofs makes histories divide, we know
that the sets Act(m, h ) coincide for all histories h passing through m . By
LP4 we know that P rove(j, A) is true at (m, h) so that in every Act(m, g)
for g in Choicem j (h) we must have a proof s of A present in Act(m, g). Now,
it is the case that all histories passing through m are in Choicem j (h) and
their sets of presented proofs coincide, therefore, we may assume that there
is a proof s of A which is present in Act(m, h ) for all h ∈ Hm . By expansion
of presented proofs, this means that s must be also present in Act(m , h )
for all h ∈ Hm . But, since s is a proof of A, this clearly denies the negative
condition for P rove(j, A) at (m , h ), therefore P rove(j, A) cannot hold at
any such moment-history pair, which gives us a contradiction.
Having shown that proving modalities display some very natural and
expected properties w.r.t. some philosophically meaningful distinctions, we
now turn to a brief characterization of the logical relations among these
modalities. The following theorem sums up entailment relations between
them:
9
But note that the inverse implications are in fact valid.
Inference as Doxastic Agency...
4. Conclusion
We started with the remark that often a distinction is drawn between proofs
as abstract objects and proofs as actions. The semantics of justification stit
logic does justice to both conceptions of proofs. There is first of all the set P ol
of proof polynomials. This set of proof terms demarcates a set of objectively
given poofs, and for every moment-history pair (m, h) from a justification-
stit model, the function Act provides a subset of P ol representing the set of
proofs that are available to Ag, the assumed scientific community. Moreover,
the semantics allows one to interpret various modalities for acts of proving
and to draw a distinction between agentive and factive ascriptions of proving
as an activity. In particular, we can express that an agent j proves A or
that j proves A by proof t, and we can semantically evaluate such agentive
sentences.
The suggested analysis of proving as doxastic activity may be seen as
affecting our understanding of what proof theory is about. Dag Prawitz [22,
p. 66] famously characterized general proof theory as “a study of proofs in
their own right where one is interested in general questions about the nature
and structure of proofs.” If, in addition, to proofs as objects we are interested
in proving as an activity, general proof theory also encompasses the study
of proofs not only as programs, but also as concrete doxastic actions. The
combination of stit theory and epistemic logic with justification opens a road
to such a study.
Various ramifications and refinements of the basic justification stit logic
of the present paper are possible and will be dealt with in Part II of this
paper [21]. The role of the proof variables in S4LP, S4LPCS , and therefore
also in justification stit logic, remains unclear to some extent. Whilst it is
clear that the proof constants stand for proofs, in S4LP and S4LPCS there
is no variable-binding operation that binds the proof variables, and there is
no assignment function that maps proof variables to proofs from a given,
moment-dependent domain of proofs. In Part II we shall consider what we
call E-notions that allow one to express that a certain proof t or some proof
of a formula A is presented to the assumed community. Moreover, we shall
G. K. Olkhovikov, H. Wansing
References
[1] Artemov, S., and E. Nogina, Introducing Justification into Epistemic Logic, Journal
of Logic and Computation 15:1059–1073, 2005.
[2] Belnap, N.D., Before Refraining: Concepts for Agency, Erkenntnis 34:137–169, 1991.
[3] Belnap, N.D., M. Perloff, and M. Xu, Facing the Future: Agents and Choices in
our Indeterminist World. Oxford UP, New York, 2001.
[4] Broersen, J., A complete stit logic for knowledge and action, and some of its appli-
cations, in M. Baldoni et al. (eds.), Declarative Agent Languages and Technologies VI,
6th International Workshop, DALT 2008, LNCS vol. 5397, Springer, Berlin, 2009, pp.
47–59.
[5] Davidson, D., The Logical Form of Action Sentences, in N. Rescher (ed.), The Logic
of Decision and Action, University of Pittsburgh Press, Pittsburgh, 1967, pp. 81–120.
[6] van Ditmarsch, H., W. van der Hoek, and B. Kooi, Dynamic Epistemic Logic,
Springer, Berlin, 2007.
[7] Egré, P., Question-embedding and factivity, Grazer Philosophische Studien 77:85–
125, 2008.
[8] Fagin, R., and Y. Halpern, Belief, Awareness, and Limited Reasoning, Artificial
Intelligence 34:39–76, 1988.
[9] Goldman, A., A Causal Theory of Knowing, The Journal of Philosophy 64:357–372,
1967.
[10] Harel, D., D. Kozen, and J. Tiuryn, Dynamic Logic, MIT Press, Cambridge, MA,
2000.
[11] Horty, J., An alternative stit operator, unpublished manuscript, Philosophy Depart-
ment, University of Maryland.
[12] Kleene, S.C., Introduction to Metamathematics, North-Holland, Amsterdam, 1952.
[13] von Kutschera, F., Bewirken, Erkenntnis 24:253–281, 1986.
[14] Lehrer, K., and T. Paxson, Knowledge: Undefeated justified true belief, The Journal
of Philosophy 66:1–22, 1969.
Inference as Doxastic Agency...
[32] Thielscher, M., Reasoning Robots, The Art and Science of Programming Robotic
Agents, Springer, Dordrecht, 2005.
[33] Wansing, H., Doxastic Decisions, Epistemic Justification, and the Logic of Agency,
Philosophical Studies 128:201–227, 2006.
[34] Wansing, H., Proofs, disproofs, and their duals, in L. Beklemishev, V. Goranko and
V. Shehtman (eds.), Advances in Modal Logic 2010, College Publications, London,
2010, pp. 483–505.
[35] Wansing, H., Remarks on the logic of imagination. A step towards understanding
doxastic control through imagination, Synthese 194:2843–2861, 2017, published online
October 2015, https://doi.org/10.1007/s11229-015-0945-4.
[36] Wansing, H., Falsification, natural deduction, and bi-intuitionistic logic, Journal of
Logic and Computation 26:425–450, 2016, published online July 2013, https://doi.
org/10.1093/logcom/ext035.
[37] Xu, M., Axioms for deliberative STIT, Journal of Philosophical Logic 27:505–525,
1998.
G. K. Olkhovikov, H. Wansing
Department of Philosophy II
Ruhr University, Bochum
Universitätsstraße 150
44780 Bochum
Germany
grigory.olkhovikov@rub.de;
grigory.olkhovikov@gmail.com
H. Wansing
Heinrich.Wansing@rub.de
G. K. Olkhovikov
Department of Philosophy
Ural Federal University
19 Mira street
Ekaterinburg
Russia 620002