Bisimulation and Expressivity For Conditional Beli PDF
Bisimulation and Expressivity For Conditional Beli PDF
Bisimulation and Expressivity For Conditional Beli PDF
net/publication/279309678
CITATIONS READS
0 54
4 authors, including:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Thomas Bolander on 04 July 2015.
Abstract
Plausibility models are Kripke models that agents use to reason about knowledge
and belief, both of themselves and of each other. Such models are used to interpret the
notions of conditional belief, degrees of belief, and safe belief. The logic of conditional
belief contains that modality and also the knowledge modality, and similarly for the logic
of degrees of belief and the logic of safe belief. With respect to these logics, plausibility
models may contain too much information. A proper notion of bisimulation is required
that characterises them. We define that notion of bisimulation and prove the required
characterisations: on the class of image-finite and preimage-finite models (with respect to
the plausibility relation), two pointed Kripke models are modally equivalent in either of
the three logics, if and only if they are bisimilar. As a result, the information content of
such a model can be similarly expressed in the logic of conditional belief, or the logic of
degrees of belief, or that of safe belief. This, we found a surprising result. Still, that does
not mean that the logics are equally expressive: the logics of conditional and degrees of
belief are incomparable, the logics of degrees of belief and safe belief are incomparable,
while the logic of safe belief is more expressive than the logic of conditional belief. In
view of the result on bisimulation characterisation, this is an equally surprising result.
We hope our insights may contribute to the growing community of formal epistemology
and on the relation between qualitative and quantitative modelling.
1 Introduction
A typical approach in belief revision involves preferential orders to express degrees of belief and
knowledge [19, 25]. This goes back to the ‘systems of spheres’ in [23, 15]. Dynamic doxastic
logic was proposed and investigated in [26] in order to provide a link between the (non-modal
logical) belief revision and modal logics with explicit knowledge and belief operators. A
similar approach was pursued in belief revision in dynamic epistemic logic [5, 34, 30, 8, 37],
that continues to develop strongly [11, 31]. We focus on the proper notion of structural
equivalence on models encoding knowledge and belief simultaneously. A prior investigation
into that is [12], which we relate our results to at the end of the paper. Our motivation is to
find suitable structural notions to reduce the complexity of planning problems. Such plans
are sequences of actions, such as iterated belief revision. It is the dynamics of knowledge and
belief that, after all, motivates our research.
The semantics of belief depends on the structural properties of models. To relate the
structural properties of models to a logical language we need a notion of structural similarity,
known as bisimulation. A bisimulation relation relates a modal operator to an accessibility
∗
DTU Compute, Technical University of Denmark, {mibi,tobo,mhje}@dtu.dk
†
LORIA, CNRS / Université de Lorraine, hans.van-ditmarsch@loria.fr
1
ML MC MR
p p p p p p p p
w1 a w2 a w3 v1 a v2 u1 a u2 a u3
b a b b b a b
w4 w5 v3 u4 u5
q q q q q
relation. Plausibility models do not have an accessibility relation as such but a plausibility
relation. This induces a set of accessibility relations: the most plausible worlds are the
accessible worlds for the modal belief operator; and the plausible worlds are the accessible
worlds for the modal knowledge operator. But it contains much more information: to each
modal operator of conditional belief (or of degree of belief) one can associate a possibly distinct
accessibility relation.This raises the question of how to represent the bisimulation conditions
succinctly. Can this be done by reference to the plausibility relation directly, instead of by
reference to these, possibly many, induced accessibility relations? It is now rather interesting
to observe that relative to the modalities of knowledge and belief, the plausibility relation is
already in some way too rich.
The plausibility model ML on the left in Figure 1 consists of five worlds. The proposition
p is true in the top ones and false in the bottom ones. The reverse holds for q: true at the
bottom and false at the top. The a relations in the model correspond to the plausibility order
w3 >a w2 >a w1 , interpreted such that the smaller of two elements in the order is the most
plausible of the two. Further, everything that is comparable with the plausibility order is
considered epistemically possible. We can then view the model as a standard multi-agent S5
plus an ordering on the epistemic possibilities. As w1 is the most plausible world for a, she
believes p and that b believes ¬p ∧ q. This works differently from the usual doxastic modal
logic, where belief corresponds to the accessibility relation. In the logics of belief that we
study, belief is what holds in the most plausible world(s) in an epistemic equivalence class.
For a the most plausible world in all three p-worlds is w1 , so a believes the same formulas in
all of them.
In w2 agent b knows p. If a is given the information that b does not consider q possible
(that is, the information that neither w1 nor w3 is the actual world), then a believes that b
knows p – or conditional on Kb ¬q, a believes Kb p. Such a statement is an example of the logic
of conditional belief LC defined in Section 3. In LC we write this statement as BaKb ¬q Kb p.
Now examine w3 . We will show that w1 and w3 are modally equivalent for LC : they
agree on all formulas of that language—no information expressible in LC distinguishes the
two worlds. This leads to the observation that no matter where we move w3 in the plausibility
ordering for a, modal equivalence is preserved. Similarly, we can move w2 anywhere we like
except making it more plausible than w1 . If we did, then a would believe Kb p unconditionally,
and the the formulas true in the model would have been changed.
It turns out that moving worlds about in the plausibility order can be done for all models,
as long as we obey one (conceptually) simple rule: Grouping worlds into “modal equivalence
classes” of worlds modally equivalent to each other, we are only required to preserve the
ordering between the most plausible worlds in each modal equivalence class. Only the most
2
plausible world in each class matters.
Another crucial observation is that standard bisimulation in terms of ≥a does not give
correspondence between bisimulation and modal equivalence. For instance, while w1 and w3
are modally equivalent, they are not “standardly” bisimilar with respect to ≥a : w3 has a
≥a -edge to a Kb p world (w2 ), whereas w1 does not. Thus, the straightforward, standard
definition of bisimulation does not work, because no modality in the language corresponds
to the plausibility relation. Instead we have an infinite set of modalities corresponding to
relations derived from the plausibility relation. One of the major contributions of this paper
is a solution to exactly this problem.
Making w3 as plausible as w1 and appropriately renaming worlds gets us MR of Figure
1. Here the modally equivalent worlds u1 and u3 are equally plausible, modally equivalent
and standardly bisimilar. This third observation gives a sense of how we solve the problem
generally. Rather than using ≥a directly, our definition of bisimulation checks accessibility
with respect to a relation ≥R a derived from ≥a and the bisimulation relation R itself. Post-
poning details for later we just note that in the present example the derived relation for ML is
exactly the plausibility relation for MR . This indicates what we later prove: This new derived
relation reestablishes the correspondence between bisimilarity and modal equivalence.
The model MC of Figure 1 is the bisimulation contraction of the right model using standard
bisimilarity. It is the bisimulation contraction of both models with the bisimulation notion
informally defined in the previous paragraph. In previous work on planning with single-agent
plausibility models [2], finding contractions of plausibility models is needed for decidability
and complexity results. In this paper we do this for the first time for multi-agent plausibility
models, opening new vistas in applications of modal logic to automated planning.
Overview of content In Section 2 we introduce plausibility models and the proper and
novel notion of bisimulation on these models, and prove various properties of bisimulation. In
Section 3 we define the three logics of conditional belief, degrees of belief, and safe belief, and
provide some further historical background on these logics. In Section 4 we demonstrate that
bisimilarity corresponds to logical equivalence (on image-finite and preimage-finite models) for
all three core logics, so that, somewhat remarkably, one could say that the content of a given
model can equally well be described in any of these logics. Then, in Section 5 we determine
the relative expressivity of the three logics, including more expressive combinations of their
primitive modalities. The main result here is that the logics of conditional and degrees of belief
are incomparable, and that the logics of degrees of belief and safe belief are incomparable, but
that the logic of safe belief is (strictly) more expressive than the logic of conditional belief. In
Section 6, we put our result in the perspective of other recent investigations, mainly the study
by Lorenz Demey [12], and in the perspective of possible applications: decidable planning.
3
we write [x]R for {x′ ∈ X | (x, x′ ) ∈ R}. A binary relation R on X is image-finite if and only
if for every x ∈ X, {x′ ∈ X | (x, x′ ) ∈ R} is finite. A relation is preimage-finite if and only if
for every x ∈ X, {x′ ∈ X | (x′ , x) ∈ R} is finite. We say R is (pre)image-finite if it is both
image-finite and preimage-finite.
Definition 1 (Plausibility model) A plausibility model for a countably infinite set of propo-
sitional symbols P and a finite set of agents A is a tuple M = (W, ≥, V ), where
• V : W → 2P is a valuation.
For ≥(a) we write ≥a . If w ≥a v then v is at least as plausible as w (for agent a), and the
≥a -minimal elements are the most plausible worlds. For the symmetric closure of ≥a we write
∼a : this is an equivalence relation on W called the epistemic relation (for agent a). If w ≥a v
but v 6≥a w we write w >a v (v is more plausible than w), and for w ≥a v and v ≥a w we
write w ≃a v (w and v are equiplausible). Instead of w ≥a v (w >a v) we may write v ≤a w
(v <a w).
Note that we have required each relation ≥ (a) to be (pre)image-finite. This amounts
to requiring that all equivalence classes of ∼a are finite, while still allowing infinite domains.
This requirement is not part of the definition of plausibility models provided in [8]. We require
it here, since it leads to simplifications without any significant reduction in generality:
1. We will show full correspondence between bisimilarity and modal equivalence for three
different logics over plausibility models. As is the case in standard modal logic, this
correspondence can only be achieved for (pre)image-finite models (the direction from
modal equivalence to bisimilarity only hold for such models, see e.g. [9]). Simply as-
suming (pre)image-finiteness from the outset simplifies the presentation, as we do not
have to repeat this restriction in a large number of places.
2. Some of our later results are going to rely on the existence of a largest autobisimulations
(see below). Usually it is quite trivial to show the existence of a largest bisimulation,
since the union of any set of bisimulations is a bisimulation. However, we need a non-
standard notion of bisimulation for our purposes, and for such bisimulations closure
under union is far from a trivial result.2 Given our first correspondence result between
bisimilarity and modal equivalence, we are however going to get this result for free (see
Section 4.1).
4
Definition 2 (Autobisimulation) Let M = (W, ≥, V ) be a plausibility model. An auto-
bisimulation on M is a non-empty relation R ⊆ W × W such that for all (w, w′ ) ∈ R and for
all a ∈ A:
[forth≥ ] If v ∈ W and w ≥R ′ ′ R ′ ′
a v, there is a v ∈ W such that w ≥a v and (v, v ) ∈ R;
[back≥ ] If v ′ ∈ W and w′ ≥R ′ R ′
a v , there is a v ∈ W such that w ≥a v and (v, v ) ∈ R;
[forth≤ ] If v ∈ W and w ≤R ′ ′ R ′ ′
a v, there is a v ∈ W such that w ≤a v and (v, v ) ∈ R;
[back≤ ] If v ′ ∈ W and w′ ≤R ′ R ′
a v , there is a v ∈ W such that w ≤a v and (v, v ) ∈ R.
Our bisimulation relation is non-standard in the [back] and [forth] clauses. A standard
[forth] condition based on an accessibility relation ≥a would be
Here, R only appears in the part ‘(v, v ′ ) ∈ R’. But in the definition of autobisimulation
for plausibility models, in [forth≥ ], the relation R also features in the condition for applying
[forth≥ ] and in its consequent, namely as the upper index in w ≥R ′ R ′
a v and w ≥a v . This
′
means that R also determines which v are accessible from w, and which v are accessible from
w′ . This explains why we define an autobisimulation on a single model before a bisimulation
between distinct models: We need the bisimulation relation R to determine the plausibility
relation ≥R a from the plausibility relation ≥a on any given model first, before structurally
comparing distinct models.
ML MR
p p p p p p
w1 a w2 a w3 u1 a u2 a u3
b a b b a b
w4 w5 u4 u5
q q q q
Figure 2: The left and right models of Figure 1, with the dotted lines showing the largest
autobisimulations (modulo reflexivity).
Example 1 The models ML and MR of Figure 1 are reproduced in Figure 2. Consider the
relation R = Rid ∪ {(w1 , w3 ), (w3 , w1 ), (w4 , w5 ), (w5 , w4 )}, where Rid is the identity relation
on W . With this R, we get that w1 and w3 are equiplausible for ≥R a:
w1 ≃R
a w3 iff
Min a ([w1 ]R= ∩ [w1 ]a ) ≃a Min a ([w3 ]R= ∩ [w3 ]a ) iff
Min a {w1 , w3 } ≃a Min a {w1 , w3 } iff
w1 ≃a w1
5
We also get that w2 ≥R
a w3 :
w2 ≥R
a w3 iff
Min a ([w2 ]R= ∩ [w2 ]a ) ≥a Min a ([w3 ]R= ∩ [w3 ]a ) iff
Min a {w2 } ≥a Min a {w1 , w3 } iff
w2 ≥a w1
This gives ≥R R
a = {(w1 , w3 ), (w3 , w1 ), (w2 , w3 ), (w2 , w1 )} ∪ Rid . For b, we get ≥b = ≥b . The
autobisimulation R on ML is shown in Figure 2. It should be easy to check that R is indeed
an autobisimulation. To help, we will justify why (w4 , w5 ) is in R: For ≥R b , we have that, as
(w1 , w3 ) ∈ R and w1 ≥R b w 4 , there must be a world v such that w 3 ≥ R v and (w , v) ∈ R.
b 4
This v is w5 .
Note that R is the largest autobisimulation. Based on [atoms] there are only two possible
candidate pairs that could potentially be added to R (modulo symmetry), namely (w1 , w2 )
and (w2 , w3 ). But w2 does not have a b-edge to a q world, whereas both w1 and w3 do. There
is therefore nothing more to add.
The largest autobisimulation for MR is completely analogous, as shown in Figure 2. ⊣
Proof From (w, w′ ) ∈ R= and w ∼a w′ we get [w]R= = [w′ ]R= and [w]a = [w′ ]a and
hence [w]R= ∩ [w]a = [w′ ]R= ∩ [w′ ]a . Thus also Min a ([w]R= ∩ [w]a ) = Min a ([w′ ]R= ∩ [w′ ]a ),
immediately implying w ≃R a v.
we get {w′ } ≥a Min a ([w′ ]R= ∩ [w′ ]a ) and then Min a ([w′′ ]R= ∩ [w′′ ]a ) <a {w′ }. In other words,
for all w′ ∈ W ′ there is a u ∈ W , namely any u ∈ Min a ([w′′ ]R= ∩ [w′′ ]a ), such that u <a w′ .
This contradicts ≥a being a well-preorder on W ′′ . We have now shown 1), 2) and 3). Finally
we show 4): Assume w ≥R a v, that is, Min a ([w]R= ∩ [w]a ) ≥a Min a ([v]R= ∩ [v]a ). This implies
the existence of an x ∈ Min a ([w]R= ∩ [w]a ) and a y ∈ Min a ([w]R= ∩ [v]a ) with x ≥a y. By
choice of x and y we have x ∼a w and y ∼a v. From x ≥a y we get x ∼a y. Hence we have
w ∼a x ∼a y ∼a v, as required.
6
bisim. contr. of MR MC MR
p
p p p p p p
a v1 a v2 u1 a u2 a u3
{u1 , u3 } {u2 }
b b a b
b
v3 u4 u5
q q q
{u4 , u5 }
Figure 3: See Figure 1. The dotted edges show the largest bisimulation between MC and
MR . Model MC is isomorphic to the bisimulation contraction of MR (on the left) and to the
bisimulation contraction of ML (not depicted).
We postpone the proof of this result to Section 4.1, since it is going to follow from the
correspondence between bisimilarity and modal equivalence for our language of conditional
belief (Theorem 1). Let us already now reassure the reader that we are not risking any circular
reasoning here: None of the results that lead to Theorem 1 and hence to Proposition 1 rely
on largest autobisimulations.
7
, V ). For ≥′a and ≥′b take the reflexive closures.
This proposition is not hard to prove, so we do not provide a full proof, but only sketch the
overall idea. First, to prove that the bisimulation contraction (W ′ , ≥′ , V ′ ) of a plausibility
model (W, ≥, V ) is a plausibility model, we simply have to prove that the relations ≥′a are
well-preorders on each ∼′a equivalence class. That is shown by first proving reflexivity of
≥′a , then transitivity of ≥′a , and finally by proving that any non-empty subset has minimal
elements with respect to ≥′a . To show that (W, ≥, V ) is bisimilar to (W ′ , ≥′ , V ′ ), we define
the (functional) relation S : W → W ′ as S = {(w, [w]R ) | w ∈ W } and show that this is a
bisimulation relation (that it satisfied [atoms] and the [back] and [forth] conditions).
Example 4 Consider again the models ML and MR of Figure 2. From the largest bisimula-
tion on ML (shown by dotted edges), we can conclude that MR is the normalisation of ML
(modulo a renaming of the worlds wi to ui , for i = 1, . . . , 5). ⊣
Definition 6 (Logical language) For any countably infinite set of propositional symbols
P and finite set of agents A we define language LCDS
P A by:
where p ∈ P , a ∈ A, and n ∈ N. ⊣
8
The formula Ka ϕ stands for ‘agent a knows (formula) ϕ’, Baψ ϕ stands for ‘agent a believes ϕ
on condition ψ’, Ban ϕ stands for ‘agent a believes ϕ to degree n’, and a ϕ stands for ‘agent
a safely believes ϕ’. (The semantics of these constructs is defined below.) The duals of Ka ,
Baϕ and a are denoted K ba, Bbaϕ and ♦a . We use the usual abbreviations for the boolean
connectives as well as for ⊤ and ⊥, and the abbreviation Ba for Ba⊤ . In order to refer to
the type of modalities in the text, we call Ka a knowledge modality, Baψ a conditional belief
modality, Ban a degree of belief modality, and a a safe belief modality.
In LCDS
P A , if A is clear from the context, we may omit that and write LP
CDS , and if P is
clear from the context, we may omit that as well, so that we get LCDS . The letter C stands
for ‘conditional’, D for ‘degree’, and S for ‘safe’. Let X be any subsequence of CDS, then LX
is the language with, in the inductive definition, only the modalities X (and with knowledge
Ka ) for all agents. In our work we focus on the logic of conditional belief with language LC ,
the logic of degrees of belief with language LD , and the logic of safe belief with language LS .
M, w |= p iff p ∈ V (w)
M, w |= ¬ϕ iff M, w 6|= ϕ
M, w |= ϕ ∧ ψ iff M, w |= ϕ and M, w |= ψ
M, w |= Ka ϕ iff M, v |= ϕ for all v ∈ [w]a
M, w |= Baψ ϕ iff M, v |= ϕ for all v ∈ Min a (JψKM ∩ [w]a )
M, w |= Ban ϕ iff M, v |= ϕ for all v ∈ Min na [w]a
M, w |= a ϕ iff M, v |= ϕ for all v with w a v
where
Min 0a [w]a =
Min a [w]a
[w]a if Min na [w]a = [w]a
Min n+1 [w]a =
a Min na [w]a ∪ Min a ([w]a \ Min na [w]a ) otherwise
Example 5 Consider Figure 1. In the the plausibility model MC we have, for instance:
MC |= Ka p → (Ba Bb q ∧ ¬Ka Bb q): If a knows p (true in v1 and v2 ), a believes, but does
not know, that b believes q. Another example is MC |= Ba¬Bb q Kb ¬q: Conditional on b not
believing q, a believes that b knows ¬q. Only in v2 does ¬Bb q hold; there Kb ¬q holds. A final
b
example is MC |= Ka p → BaKb q Bb q: From v1 and v2 (where Ka p holds), formula K b b q only
9
P P′
p p q p q
a a y a a a
w z x w′ z′ x′
holds in v1 , and conditional to that, the one and only most plausible world v1 satisfies Bb q.
We can repeat this exercise in ML and MR , as all three models are bisimilar and therefore,
as will be proved in the next section, logically equivalent. ⊣
In [5, 34, 20] different layers can contain bisimilar worlds. In our approach they cannot,
because we define belief degrees on the normal plausibility relation. Unlike [27] our semantics
does not allow empty layers in between non-empty layers. If Ean [w]a 6= ∅ and Ean+2 [w]a 6= ∅,
then Ean+1 [w]a 6= ∅. Layers above the maximum degree will be empty, i.e. if there is a
maximum degree n for a at w, as there will always be in our (pre)image-finite models, then
for all degrees k > n, we have Eak [w]a = ∅.
Example 6 In Figure 1, we have that MC |= Ba0 Bb0 q but not MC |= Ba1 Bb0 q. The maximum
degree of belief for a in MC is at either v1 and v2 , where it is 1, so MC |= Ka ϕ ↔ Ba1 ϕ. This
is also true in the other two models. Consider now the models P and P ′ in Figure 4 and an
10
alternative definition of Ban not using a but ≥a (as in [5, 34, 20, 8]). In the ≥a -semantics we
have P |= Ba2 ¬q, as q is false in {y, z, w}. Only when we reach the third degree of belief does
q become uncertain: P 6|= Ba3 ¬q. With a -semantics, 2 is the maximum degree so P 6|= Ba2 ¬q.
This can be seen in the bisimilar model P ′ , where P ′ 6|= Ba2 ¬q. ⊣
The semantics we propose for degrees of belief and safe belief are non-standard. Still,
as we show in the following, these non-standard semantics and the standard semantics for
conditional belief are all bisimulation invariant. This makes the results in Section 5 showing
a non-trivial expressivity hierarchy between these logics even more remarkable.
11
w′ ∼a v ′ . We need to prove M, v ′ |= ψ. From Lemma 2 we have that ≥R a is a well-preorder on
each ∼a -equivalence class. Since w ∼a v we hence get that either w ≥R
′ ′ ′ ′ ′
a v or v ≥a w . We
R ′
can assume w ≥a v , the other case being symmetric. Then since (w, w ) ∈ R and w′ ≥R
′ R ′ ′ ′
a v ,
[back≥ ] gives us a v s.t. (v, v ′ ) ∈ R and w ≥R a v. Lemma 2 now implies w ∼a v, and hence
M, v |= ψ. Since (v, v ′ ) ∈ R, the induction hypothesis gives us M, v ′ |= ψ, and we are done.
Now consider the case ϕ = Baγ ψ. This case is more involved. Assume M, w |= Baγ ψ, that
is, M, v |= ψ for all v ∈ Min a (JγKM ∩ [w]a ). Letting v ′ ∈ Min a (JγKM ∩ [w′ ]a ), we need to show
M, v ′ |= ψ (if Min a (JγKM ∩ [w′ ]a ) is empty there is nothing to show). The standard approach
is to use one of the back conditions to find a v with (v, v ′ ) ∈ R and then show that M, v |= ψ.
From this M, v ′ |= ψ will follow, using the induction hypothesis. It is tempting to directly
apply [back≥ ] to w′ ≥R ′ ′ R ′ ′
a v (or [back≤ ] to w ≤a v ) to produce such a v with (v, v ) ∈ R. But
unfortunately, we will not be able to conclude that such a v is in Min a (JγKM ∩[w]a ), and hence
not that M, v |= ψ. More work is needed. Instead we will first find a y in Min a (JγKM ∩ [w]a ),
then find a y ′ with (y, y ′ ) ∈ R, and only then apply [back≥ ] to y ′ ≥R ′
a v to produce the
required v. The point is here that our initial choice of y in Min a (JγKM ∩ [w]a ) will ensure
that v is in Min a (JγKM ∩ [w]a ).
As mentioned, we want to start out choosing a y in Min a (JγKM ∩ [w]a ), so we need to
ensure that this set is non-empty. By choice of v ′ we have v ′ ∈ JγKM and v ′ ∼a w′ . From
v ′ ∼a w′ we get that w′ ≥R ′ ′ R ′ ′
a v or w ≤a v , using Lemma 2. Since also (w, w ) ∈ R, we can
apply [back≥ ] or [back≤ ] to get a u such that (u, v ′ ) ∈ R and either w ≥R R
a u or w ≤a u. From
′ ′
(u, v ) ∈ R and v ∈ JγKM , we get u ∈ JγKM , using the induction hypothesis. From the fact
that either w ≥R R
a u or w ≤a u we get w ∼a u, using Lemma 2. Hence we have u ∈ JγKM ∩[w]a .
This shows the set JγKM ∩ [w]a to be non-empty. Hence also Min a (JγKM ∩ [w]a ) is non-empty,
and we are free to choose a y in that set. Since y ∼a w, Lemma 2 gives us that either y ≥R a w
R ′ ′
or w ≥a y, so we can apply [forth≤ ] or [forth≥ ] to find a y with (y, y ) ∈ R and either
y ′ ≥R ′ ′
a w or w ≥a y .
R ′
Claim 1. y ′ ≥R ′
a v .
Proof of claim 1. We need to prove Min a ([y ′ ]R= ∩ [y ′ ]a ) ≥a Min a ([v ′ ]R= ∩ [v ′ ]a ). We first
prove that [y ′ ]R= ∩ [y ′ ]a ⊆ JγKM ∩ [w′ ]a :
Since v ′ is chosen minimal in JγKM ∩[w′ ]a and [y ′ ]R= ∩[y ′ ]a ⊆ JγKM ∩[w′ ]a we get Min a ([y ′ ]R= ∩
[y ′ ]a ) ≥a {v ′ } ≥a Min a ([v ′ ]R= ∩ [v ′ ]a ), as required. This concludes the proof of the claim.
By choice of y ′ we have (y, y ′ ) ∈ R, and by Claim 1 we have y ′ ≥R ′
a v . We can now finally, as
promised, apply [back≥ ] to these premises to get a v s.t. (v, v ) ∈ R and y ≥R
′
a v.
12
• x ∈ [w]M : By choice of x we have x ∼a v. Since y ≥R a v, Lemma 2 implies v ∼a y. By
choice of y we have y ∼a w, so in total we get x ∼a v ∼a y ∼a w, as required.
We proceed now to show the converse, that modal equivalence with regard to LC implies
bisimulation. The proof has the same structure as the Hennessy-Millner approach, though
appropriately modified for our purposes. Given a pair of image-finite models M and M ′ ,
the standard approach is to construct a relation R ⊆ D(M ) × D(M ′ ) s.t. (w, w′ ) ∈ R if
M, w ≡ M ′ , w′ . Using ♦-formulas, it is then shown that R fulfils the requirements for being a
bisimulation, as such formulas denote what is true at worlds accessible by whatever accessi-
bility relation is used in the model. This means that modally equivalent worlds have modally
equivalent successors, which is then used to show that R fulfils the required conditions. For
our purposes this will not do, as we only have K b a -formulas (i.e. for ∼a ). Instead, our equiv-
ψ
ba ϕ, each such formula corresponding to accessibility to
alent to ♦-formulas are of the form B
the most plausible ψ-worlds from all worlds in an equivalence class. What we want are for-
mulas corresponding to specific links between worlds, so we first establish that such formulas
exists. We thus have formulas with the same function as ♦-formulas serve in the standard
approach.
Proposition 5 Modal equivalence with respect to LC implies bisimilarity. ⊣
Proof Assume (M1 , w) ≡C (M2 , w′ ). We wish to show that (M1 , w) ↔ (M2 , w′ ). Let M =
M1 ⊔ M2 be the disjoint union of M1 and M2 . We then need to show that Q = {(v, v ′ ) ∈
D(M ) × D(M ) | M, v ≡C M, v ′ } is an autobisimulation on M . Note that as ≡C is an
equivalence relation, so is Q. We first show that ♦-like formulas talking about the ≥Q a-
relations between specific worlds in M exist.
Claim 1. Let w and w′ be worlds of the model M = (W, ≥, V ) where w ≥Q ′
a w . Further
C ′ C
let ϕ ∈ L be any formula true in w . There then exists a formula ψ ∈ L such that
baψ ϕ.
([w]Q ∪ [w′ ]Q ) ∩ [w]a = JψKM ∩ [w]a and M, w |= B
Proof of Claim 1. If two worlds s and s′ are not modally equivalent, there exists some
distinguishing formula Ψs,s′ with M, s |= Ψs,s′ and M, s′ 6|= Ψs,s′ . As ∼a is image-finite (since
both ≥a and its converse are) the following formula is finite:
^
Ψt = {Ψt,t′ | t ∼a t′ ∧ (t, t′ ) 6∈ Q}
The formula Ψt distinguishes t from all the worlds in [t]a that it is not modally equivalent to.
If there are no such worlds, Ψt is the empty conjunction equivalent to ⊤.
We now return to our two original worlds w and w′ . With the assumption that M, w′ |= ϕ,
we show that ψ = Ψw ∨ Ψw′ is a formula of the kind whose existence we claim. First note that
13
JΨw KM ∩ [w]a contains only those worlds in [w]a that are modally equivalent to w, exactly
as [w]Q ∩ [w]a does. As JΨw KM ∪ JΨw′ KM = JΨw ∨ Ψw′ KM we have ([w]Q ∪ [w′ ]Q ) ∩ [w]a =
JΨw ∨Ψw′ KM ∩[w]a . To get M, w |= Bbaψ ϕ we need to show that ∃v ∈ Min a (JΨw ∨ Ψw′ KM ∩[w]a )
s.t. M, v |= ϕ. Pick an arbitrary v ∈ Min a ([w′ ]Q ∩ [w′ ]a ). We will now show that this has the
required properties.
Let T = JΨw ∨Ψw′ KM ∩[w]a . Since T = ([w]Q ∪[w′ ]Q )∩[w]a , Lemma 1 gives u ≃Q a w or u ≃a
Q
We now proceed to show that Q fulfils the conditions for being an autobisimulation on M
(Definition 2). [atoms] is trivial. Next we show [forth≥ ]. Let (w, w′ ) ∈ Q (i.e. (M, w) ≡C
(M, w′ )) and w ≥Q ′ ′
a v. We then have that [forth≥ ] is fulfilled if ∃v ∈ W , s.t. w ≥a v and
Q ′
(v, v ′ ) ∈ Q (i.e. (M, v) ≡C (M, v ′ )). To this end, we show that assuming for all v ′ ∈ W ,
w ′ ≥Q ′ C ′
a v implies (M, v) 6≡ (M, v ) leads to a contradiction. This is analogous to how Q is
shown to be a bisimulation in standard Hennesey-Millner proofs.
We first show that ≥Q a is image-finite. First recall that by assumption on plausibility
models, ≥a is (pre)image finite, that is, both ≥a and ≤a are image-finite. It follows that
∼a = ≥a ∪ ≤a is image-finite as well. If a relation is image-finite, then so is any subset of the
relation. Therefore, as ≥Q Q Q
a ⊆ ∼a , ≥a must be image-finite. Hence the set of ≥a -successors
of w′ , S = {v ′ | w′ ≥Q ′ ′ ′
a v } = {v1 , . . . , vn } is also finite. Having assumed that v and none of
′
the vi s are modally equivalent, we have that there exists a number of distinguishing formulae
′ ′ ′ ′ ′
ϕvi , one for each vi′ , such that M, v |= ϕvi and M, vi′ 6|= ϕvi . Therefore, M, v |= ϕv1 ∧ · · · ∧ ϕvn .
′ ′
For notational ease, let ϕ = ϕv1 ∧ · · · ∧ ϕvn .
With M, v |= ϕ, Claim 1 gives the existence of a formula ψ, such that ([w]Q ∪[v]Q )∩[w]a =
JψKM ∩ [w]a and M, w |= B baψ ϕ. Due to modal equivalence of w and w′ , we must have
M, w′ |= B baψ ϕ. This we have iff ∃u′ ∈ Min a (JψKM ∩ [w′ ]a ), s.t. M, u′ |= ϕ. By construction
of ϕ, no world vi′ exists such that w′ ≥Q ′ ′ ′ Q
a vi and M, vi |= ϕ, so we must have u >a w . As
′
14
us w′ ≃Q ′′ ′ Q ′′ ′ Q ′ ′ Q ′
a w . From this and v ≥a w we get v ≥a w and hence w ≤a v , as required. This
concludes proof of [forth≤ ]. As for [back≥ ] getting to [back≤ ] is easy and left out.
Proof (of Proposition 1) First note that neither the semantics of LC nor the proofs of
Proposition 4 and 5 rely on the existence of largest autobisimulations. Hence we can use these
in proving the proposition. Given a plausibility model M = (W, ≥, V ) we define a relation
R by R = {(w, v) ∈ W 2 | M, w ≡C M, v}. Since modal equivalence implies bisimilarity
(Theorem 5), R is a bisimulation relation (indeed, R is exactly the relation shown to be an
autobisimulation in the proof of Proposition 5). Now we have to show that R is the largest
autobisimulation. If it was not, there would exist an autobisimulation R′ with R′ − R 6= ∅. By
definition of R, R′ would then contain at least one pair (w, v) with M, w 6≡C M, v. However,
since bisimilarity implies modal equivalence (Proposition 4), this contradicts R′ being an
autobisimulation. Hence R must be the largest autobisimulation. It only remains to prove
that R is an equivalence relation. However, this is trivial given its definition in terms of modal
equivalence.
(ii) If v ∈ M inn+1
a [w]a then w a v . ⊣
Proof The truth of (i) easily comes from the definition of Min na . For (ii), we consider two
exhaustive cases for v. Either v ∈ M inn+1 n
a [w]a \ M ina [w]a in which case w a v follows from
a -minimality, since by assumption w ∈ [w]a \ M ina [w]a . Otherwise v ∈ M inna [w]a , and so
n
from w 6∈ M inna [w]a and (i) it follows that w ≻a v and hence also w a v.
15
Now getting to the meat of this section, showing bisimulation correspondence for LD , we
first show that bisimilar worlds belong to spheres of all the same degrees.
Lemma 5 If (M1 , w1 ) ↔ (M2 , w2 ) then for all n ∈ N, w1 ∈ Min na [w1 ]a iff w2 ∈ Min na [w2 ]a .
w ∈ Ean [w]a and w′ 6∈ Ean [w′ ]a . Because w′ ∈ [w′ ] \ Min na [w′ ]a , we know that n is not the
maximum degree, so there must be some world v ′ ∈ Ean [w′ ]a which by definition means
that v ′ 6∈ Min an−1 [w′ ]a . With v ′ ∈ Ean [w′ ]a ⊆ Min na [w′ ]a and and w′ 6∈ Min na [w′ ]a , Lemma
4 gives w′ ≻a v ′ , i.e. w′ a v ′ and v ′ 6a w′ . By [back≥ ] there is a v s.t. w a v and
(v, v ′ ) ∈ Rmax . Because v ′ 6∈ M inan−1 [w′ ]a we cannot have v ∈ Min an−1 [w]a , as we could
then again have chosen a smaller n making either i) or ii) true. Thus v ∈ [w]a \ Min an−1 [w]a .
As w ∈ Min na [w]a , Lemma 4 gives v a w, so by [forth≥ ] there is a u′ s.t. v ′ a u′ and
(w, u′ ) ∈ Rmax .
With (w, w′ ) ∈ Rmax and (w, u′ ) ∈ Rmax , we have (w′ , u′ ) ∈ Rmax . As w′ ∼a u′ (we have
w a v ′ and v ′ a u′ ), Lemma 1 gives w′ ≃R
′
a
max
u′ , i.e. w′ a u′ and w′ a u′ . As w′ 6∈
n ′ n
Min a [w ]a , we then have u′ 6∈ Min a [w′ ]a . As u′ 6∈ Min na [w′ ]a while v ′ ∈ Ean [w′ ]a ⊆ Min na [w′ ]a ,
Lemma 4 then gives u′ ≻a v ′ . But this contradicts v ′ a u′ , concluding the proof.
Proof Assume (M1 , w1 ) ↔ (M2 , w2 ). Let M = (W, ≥, V ) denote the disjoint union of M1 and
M2 . Then there exists an autobisimulation R on M with (w1 , w2 ) ∈ R. Using Proposition 1,
let Rmax denote the largest autobisimulation on M . Then R ⊆ Rmax . We need to prove that
(M, w1 ) ≡D (M, w2 ).
We will show that for all (w, w′ ) ∈ Rmax , for all ϕ ∈ LD , M, w |= ϕ iff M, w′ |= ϕ (which
then also means that it holds for all (w, w′ ) ∈ R). We proceed by induction on the syntactic
complexity of ϕ. The propositional and knowledge cases are already covered by Proposition4,
so we only go for ϕ = Ban ψ.
Assume M, w |= Ban ψ. We need to prove that M, w′ |= Ban ψ, that is M, v ′ |= ψ for all
v ′ ∈ Min na [w′ ]a . Picking an arbitrary v ′ ∈ Min na [w′ ]a , we have [w′ ]a = [v ′ ]a from Lemma 2,
and either w′ a v ′ or w′ a v ′ (so we also have v ′ ∈ Min na [v ′ ]a ). Using [back≥ ] or [back≤ ] as
appropriate, we get that there is a v such that w a v or w a v, and (v, v ′ ) ∈ Rmax . From
this, v ′ ∈ Min na [v ′ ]a , and Lemma 5 we get v ∈ Min na [v]a , allowing us to conclude v ∈ Min na [w]a
from [w]a = [v]a . With the original assumption of M, w |= Ban ψ we get M, v |= ψ. As
(v, v ′ ) ∈ Rmax , the induction hypothesis gives M, v ′ |= ψ. As v ′ was chosen arbitrarily in
Min na [w′ ]a this gives M, w′ |= Ban ψ. Showing that M, w′ |= Ban ψ implies M, w |= Ban ψ is
completely symmetrical and therefore left out.
16
We now get to showing that modal equivalence for the language of degrees of belief im-
plies bisimilarity. Trouble is, that the Ban modality uses the largest autobisimulation for
deriving the relation a . This makes it difficult to go the Hennesey-Millner way of showing
by contradiction that the modal equivalence relation Q is an autobisimulation.
Instead, we establish that modal equivalence for LD implies modal equivalence for LC .
We go about this by way of a model and world dependent translation of LC formulas into LD
formulas (Definition 10). This translation has two properties. First, the translated formula
is true at M, w iff the untranslated formula is (Lemma 7)—a quite uncontroversial property.
More precisely, letting M = (W, ≥, R) be a plausibility model, then for any w ∈ W , γ ∈ LC
where σM,w (γ) is the translation at M, w: M, w |= γ ⇔ M, w |= σM,w (γ). Assume further
that we have some M ′ , w′ such that (M, w) ≡D (M ′ , w′ ). As σM,w (γ) is a formula of LD we
can conclude M ′ , w′ |= σM,w (γ). So in all we get that
The second property is that the translation of γ is the same for worlds modally equivalent
for LD (Lemma 8): If (M, w) ≡D (M ′ , w′ ) then σM,w (γ) = σM ′ ,w′ (γ). This then gives
Combining (*) and (**) gives that if (M, w) ≡D (M ′ , w′ ) then M, w |= γ iff M ′ , w′ |= γ for any
γ ∈ LC , i.e. that (M, w) ≡C (M ′ , w′ ). As shown in the previous section, modal equivalence
for LC implies bisimilarity (Proposition 5), and we can therefore finally conclude that modal
equivalence for LD implies bisimilarity (Proposition7).
Lemma 6 For a plausibility model M , a world w ∈ D(M ), agent a ∈ A and a formula ψ of
LC , if JψKM ∩ [w]a 6= ∅, there is a unique natural number k for which Min a (JψKM ∩ [w]a ) ⊆
Eak [w]a (= Min a ([w]a \ Min ak−1 [w]a )).
Proof Let S = JψKM ∩ [w]a . We first show that all worlds in Min a S are equiplausible with
respect to a .
Take any two worlds v1 , v2 ∈ Min a S. We wish to show v1 ≃R a v2 , i.e. Min a ([v1 ]R ∩
[v1 ]a ) ≃a Min a ([v2 ]R ∩[v2 ]a ), where R is the largest autobisimulation on M . With Proposition4
(bisimilarity implies modal equivalence for LC ) and for i = 1, 2, we have that [vi ]R ⊆ JψKM .
Hence [vi ]R ∩[vi ]a = [vi ]R ∩[w]a ⊆ JψKM ∩[w]a = S. With vi ∈ Min a S and vi ∈ [vi ]R ∩[vi ]a ⊆ S,
we have vi ∈ Min a ([vi ]R ∩ [vi ]a ) (if an element of a set A is minimal in a set B ⊇ A, then
it is also minimal in A). From this we can conclude that Min a ([vi ]R ∩ [vi ]a ) ≃a {vi }. Since
v1 ≃a v2 we get Min a ([v1 ]R ∩ [v1 ]a ) ≃a {v1 } ≃a {v2 } ≃a Min a ([v2 ]R ∩ [v2 ]a ), concluding the
proof of v1 ≃R a v2 .
Due to (pre)image-finiteness of M , [w]a is finite. This means that for any v ∈ [w]a there is a
unique natural number k for which v ∈ Eak [w]a . As all worlds in Min a S are a -equiplausible,
we have that Min a S ⊆ Eak [w]a for some unique k.
Having established that if JψKM ∩ [w]a 6= ∅ then there does indeed exist a unique k st.
Min a (JψKM ∩ [w]a ) ⊆ Eak [w]a , we have that the following translation is well-defined.
Definition 10 (Translation σM,w ) Let M = (W, ≥, V ) be a plausibility model and γ ∈ LC
be given. We write σM,w (γ) for the translation of γ at M, w into a formula of LD defined as
follows:
σM,w (p) = p
σM,w (¬ϕ) = ¬σM,w (ϕ)
σM,w (ϕ1 ∧ ϕ2 ) = σM,w (ϕ1 ) ∧ σM,w (ϕ2 )
17
( W
b k W{σM,v (ψ) | v ∈ [w]a } if JψKM ∩ [w]a 6= ∅
Bak {σM,v (ψ → ϕ) | v ∈ [w]a } ∧ B
ψ a
σM,w (Ba ϕ) = W
Ka {σM,v (¬ψ) | v ∈ [w]a } if JψKM ∩ [w]a = ∅
where k is the natural number such that Min a (JψKM ∩ [w]a ) ⊆ Eak [w]a . As Ka ϕ is definable
in LC as Ba¬ϕ ⊥, we need no Ka ϕ-case in the translation. ⊣
18
ψ
show W M, w |= Ba ϕ.′ First take the case where W JψKM ∩ [w′ ]a = ∅. Then σM,w′ (Baψ ϕ) =
Ka {σM,v ′ ] }, i.e. M, w |= K ′ ′
W (¬ψ) | v ∈ [w a {σM,v′ (¬ψ) | v ∈ [w ]a }. This means that
′ a
M, v |= {σM,v′ (¬ψ) | v ∈ [w ]a } for all v ∈ [w]a , i.e. for any v ∈ [w]a there is a v ′ ∈ [w′ ]a
′ ′
such that M, v |= σM,v′ (¬ψ). Applying the induction hypothesis, we get M, v |= ¬ψ for all
v ∈ [w]a . Thus JψKM ∩ [w]a = ∅ and we trivially have M, w |= Baψ ϕ.
Now take the case JψKM ∩ [w ′ ] 6= ∅. Letting S ′ = Min (JψK ∩ [w ′ ] ) and k ′ be s.t. S ′ ⊆
a a M W a
′ ′ W
k ′
Ea [w ]a , we have W k
M, w |= Ba {(σM,v′ (ψ → ϕ) | v ′ ∈ [w′ ]a } ∧ W Bbak′ {σM,v′ (ψ) | v ′ ∈ [w′ ]a }.
′
From M, w |= Bak {σM,v′ (ψ → ϕ) | v ′ ∈ [w′ ]a } we have M, v |= {σM,v′ (ψ → ϕ) | v ′ ∈ [w′ ]a }
′
for all v ∈ Min ka [w]a , i.e. for any v ∈ [w]a there is a v ′ ∈ [w′ ]a such that M, v |= σM,v′ (ψ → ϕ).
′
Applying theWinduction hypothesis, we get M, v |= ψ → W ϕ for all v ∈ Min ka [w]a . From
M, w |= B b k′ {σM,v′ (ψ) | v ′ ∈ [w′ ]a } we have M, v |= {σM,v′ (ψ) | v ′ ∈ [w′ ]a } for some
a
′
v ∈ Min ka [w]a , i.e. there is a v ∈ [w]a and a v ′ ∈ [w′ ]a such that M, v |= σM,v′ (ψ). Applying
′
the induction hypothesis gets us M, v |= ψ. Thus we have M, w |= Bak (ψ → ϕ) ∧ B bak′ ψ (where
ψ, ϕ ∈ LC ).
From M, w |= B bak′ ψ we have that JψKM ∩ [w]a 6= ∅, so Lemma 6 gives the existence of
a k, s.t. Min a (JψKM ∩ [w]a ) ⊆ Min ka [w]a . We also have from M, w |= B bak′ ψ that k ≤ k′ ,
′ ′
so Min a (JψKM ∩ [w]a ) ⊆ Min ka [w]a . With M, w |= Bak (ψ → ϕ) we get M, v |= ψ → ϕ
for all v ∈ Min a (JψKM ∩ [w]a ), then M, v |= ϕ for all v ∈ Min a (JψKM ∩ [w]a ), and finally
M, w |= Baψ ϕ.
We have now gotten to the second of the two promised properties; that the translation is
the same for worlds modally equivalent for LD .
Lemma 8 Given plausibility models M and M ′ , for any w ∈ D(M ) and w′ ∈ D(M ′ ), if
(M, w) ≡D (M ′ , w′ ) then for any formula γ ∈ LC , σM,w (γ) = σM ′ ,w′ (γ).
Proof We show this by another induction on the modal depth of γ. For the base case of
modal depth 0 we trivially have σM,w (γ) = σM ′ ,w′ (γ).
For the induction step we, as before, only deal with γ = Baψ ϕ. Note first that every
world in [w]a is modally equivalent to at least one world in [w′ ]a . If that wasn’t the case,
there would be some formula LD -formula ϕ true somewhere in [w]a and nowhere in [w′ ]a .
Then M, w |= K b a ϕ while M ′ , w′ 6|= K
b a ϕ, contradicting (M, w) ≡D (M ′ , w′ ). A completely
analogous argument gives that every world in [w′ ]a is modally equivalent to at least one world
in [w]a . Thus JψKM ∩ [w]a = ∅ iff JψK′M ∩ [w′ ]a = ∅. We thus have two cases, either both
σM,w (Baψ ϕ) and σM ′ ,w′ (Baψ ϕ) are Ka -formulas, or both are Bak -formulas.
We deal first W with the case where both translations are KaW -formulas. Here we have
σM,w (Baψ ϕ) = Ka {σM,v (¬ϕ) | v ∈ [w]a } and σM ′ ,w′ (Baψ ϕ) = Ka {σM ′ ,v′ (¬ϕ) | v ′ ∈ [w′ ]a }.
As already shown, for all v ∈ [w]a there is a v ′ ∈ [w′ ]a such that (M, w) ≡D (M ′ , v ′ ), and vice
versa. The induction hypothesis gives σM,v (¬ϕ) = σM ′ ,v′ (¬ϕ) for all these vs and v ′ s. Then
W W
{σM,v (¬ϕ) | v ∈ [w]a } = {σM ′ ,v′ (¬ϕ) | v ′ ∈ [w′ ]a } and thus σM,w (Baψ ϕ) = σM ′ ,w′ (Baψ ϕ).
TakeW now the case where both translations
W are Bak -formulas. A similar argument
W as above
gives {σ ′ ′
WM,v (ψ → ϕ) | v ∈ [w]a } = {σM ,v (ψ → ϕ) | v ∈ [w ]a } and {σM,v (ψ) | v ∈
′ ′
[w]a } = {σM ′ ,v′ (ψ) | v ′ ∈ [w′ ]a }. Letting k and k′ be the indices chosen in the translation
of σM,w (Baψ ϕ) and σM ′ ,w′ (Baψ ϕ) respectively, we have σM,w (Baψ ϕ) = σM ′ ,w′ (Baψ ϕ) if k = k′ .
Assume towards a contradiction that k > k′ . Lemma 6 now gives Min a (JψKM ∩ [w]a ) ∩
′ ′
Min ka [w]a = ∅, so M, v |= ¬ψ for all v ∈ M inka [w]a . With Lemma 7 we have M, v |= σM,v (¬ψ)
k ′ k ′ W
for all v ∈ M ina [w]a and thus also that M, w |= Ba {σM,v (¬ψ) | v ∈ [w]a }. From Lemma
′ ′
6 we also have Min a (JψKM ∩ [w′ ]a ) ⊆ M inka [w′ ]a , so M ′ , v ′ 6|= ¬ψ for some v ′ ∈ M inka [w′ ]a .
′
From here we use Lemma 7 to conclude M ′ , v ′ 6|= σM ′ ,v′ (¬ψ) for some v ′ ∈ M inka [w′ ]a and
k ′ W
thus M , w 6|= Ba {σM ′ ,v′ (¬ψ) | v ′ ∈ [w′ ]a }. By the work done so far, this also means
′ ′
19
′ W
M ′ , w′ |6 = Bak {σM,v (¬ψ) | v ∈ [w]a } which contradicts (M, w) ≡D (M ′ , w′ ). The case when
k′ > k is completely symmetrical, and the proof if thus concluded.
Putting this together with Theorem 5 (modal equivalence for LC implies bisimilarity), we
have that two worlds which are modally equivalent in LD are also modally equivalent in LC
and therefore bisimilar.
Proof of claim. To prove (i), let Si denote the largest autobisimulation on Mi . If we can
show Si ⊆ Ri we are done. Since Si is an autobisimulation on Mi , it must also be an
autobisimulation on M1 ⊔ M2 . Thus, clearly, Si ⊆ R, since R is the largest autobisimulation
on M1 ⊔ M2 . Hence, since Si ⊆ D(Mi ) × D(Mi ), we get Si = Si ∩ (D(Mi ) × D(Mi )) ⊆
R ∩ (D(Mi ) × D(Mi )) = Ri . This shows Si ⊆ Ri , as required.
We now prove (ii). Since w ∈ D(Mi ) we get [w]a ⊆ D(Mi ). Since Ri = R ∩ (D(Mi ) ×
D(Mi )) this implies [w]R ∩ [w]a = [w]Ri ∩ [w]a . Now note that since R is the largest auto-
bisimulation on M1 ⊔ M2 and Ri is the largest autobisimulation on Mi , we have R = R= and
Ri = Ri= , by Proposition 1 (the largest autobisimulation is an equivalence relation). Hence
20
from [w]R ∩ [w]a = [w]Ri ∩ [w]a we can conclude [w]R= ∩ [w]a = [w]R=
i
∩ [w]a , and then finally
Min a ([w]R= ∩ [w]a ) = Min a ([w]R=
i
∩ [w]a ).
We now prove (iii). Note that if w ≥R Ri
a v or w ≥a then w ∼a v (by Lemma 2). So in
proving w ≥R Ri
a v ⇔ w ≥a v for w ∈ D(Mi ), we can assume that also v ∈ D(Mi ). We then
get:
w ≥R
a v ⇔ Min a ([w]R= ∩ [w]a ) ≥a Min a ([v]R= ∩ [v]a )
⇔ Min a ([w]R=
i
∩ [w]a ) ≥a Min a ([v]R=
i
∩ [v]a ) by (ii), since w, v ∈ D(Mi )
R
⇔ w ≥a v.
i
R R
can from w2 ≥a 2 v2 conclude w2 ≥a v2 . Since R is an autobisimulation, we can now apply
[back≥ ] to (w1 , w2 ) ∈ R and w2 ≥R R
a v2 to get a v1 with w1 ≥a v1 and (v1 , v2 ) ∈ R. Using (iii)
R R
again we can conclude from w1 ≥a v1 to w1 ≥a 1 v1 , since w1 ∈ D(M1 ). By (i), R1 is the
largest autobisimulation on M1 , so w1 ≥R a v1 is by definition the same as w1 a v1 . Since
1
As for the previous logics, the converse also holds, that is, modal equivalence with regard
to LS implies bisimulation. This is going to be proved as follows. First we prove that any
conditional belief formula ϕC can be translated into a logically equivalent safe belief formula
ϕS . This implies that if two pointed models (M, w) and (M ′ , w′ ) are modally equivalent in LS ,
they must also be modally equivalent in LC : Any formula ϕC ∈ LC is true in (M, w) iff its
translation ϕS ∈ LS is true in (M, w) iff ϕS is true in (M ′ , w′ ) iff ϕC is true in (M ′ , w′ ). Now
we can reason as follows: If two pointed models (M, w) and (M ′ , w′ ) are modally equivalent
in LS then they are modally equivalent in LC and hence, by Theorem 5, bisimilar. This is the
result we were after. We postpone the full proof until Section 5.1, which is where we provide
the translation of conditional belief formulas into safe belief formulas (as part of a systematic
investigation of the relations between the different languages and their relative expressivity).
Here we only state the result:
As for the two previous languages, LC and LD , we now get the following bisimulation char-
acterisation result.
21
4.4 Combining characterisation results
By combining Theorems 4, 6 and 8 from the previous subsections we immediately have the
following result.
For example, we also have that (M, w) ↔ (M ′ , w′ ) iff (M, w) ≡CDS (M ′ , w′ ), or that (M, w) ↔ (M ′ , w′ )
iff (M, w) ≡DS (M ′ , w′ ). Also, to be explicit, we now have that
In other words, the information content of a pointed plausibility model is equally well de-
scribed in any of these logics. This seems to suggest that it does not matter which logic you
use to describe the information content of such a model, apart from the usual considerations
of succinctness. Still, this is not the case: our logics are not equally expressive. This will now
be addressed in the next section.
5 Expressivity
In this section we will determine the expressivity hierarchy of the logics under consideration.
Abstractly speaking, expressivity is a yardstick for measuring whether two logics are able to
capture the same properties of a class of models. More concretely in our case, we will for
instance be interested in determining whether the conditional belief modality can be expressed
using the degrees of belief modality (observe that the translation in Section 4.2 depends on
a particular model). With such results at hand we can for instance justify the inclusion or
exclusion of a modality, and it also sheds light upon the strengths and weaknesses of our
doxastic notions. To start things off we now formally introduce the notion of expressivity
found in [38].
Definition 11 Let L and L′ be two logical languages interpreted on the same class of models.
• For ϕ ∈ L and ϕ′ ∈ L′ , we say that ϕ and ϕ′ are equivalent (ϕ ≡ ϕ′ ) iff they are true in
the same pointed models of said class.3
22
Below we will show several cases where L 6≦ L′ ; i.e. that L′ is not at least as expressive
as L. Our primary modus operandi (obtained by logically negating L ≦ L′ ) will be to show
that there is a ϕ ∈ L, where for any ϕ′ ∈ L′ we can find two pointed models (M, w), (M ′ , w′ )
such that
In other words, for some ϕ ∈ L, no matter the choice of ϕ′ ∈ L′ , there will be models which
ϕ distinguishes but ϕ′ does not, meaning that ϕ 6≡ ϕ′ .
Our investigation will be concerned with the 7 distinct languages that are obtained by
considering each LX such that X is a non-empty subsequence of CDS. In Section 5.1 our
focus is on safe belief, and in Section 5.2 on degrees of belief. Using these results, we provide
in Section 5.3 a full picture of the relative expressivity of each of these logics, for instance
showing that we can formulate 5 distinct languages up to equal expressivity. We find this
particularly remarkable in light of the fact that our notion of bisimulation is the right fit for
all our logics.
Proof We let M = (W, ≥, V ) be any plausibility model with w ∈ W , and further let a
denote the normal plausibility relation for an agent a in M . We will show that M, w |=
Baψ ϕ ↔ (K baψ → K b a (ψ ∧ a (ψ → ϕ))). To this end we let X = Min a (JψKM ∩ [w]a ).
Immediately we have that if X = ∅ then no world in [w]a satisfies ψ, thus trivially yielding
both M, w |= Baψ ϕ and M, w |= K baψ → K b a (ψ ∧ a (ψ → ϕ)). For the remainder we therefore
assume X is non-empty. We now work under the assumption that M, w |= Baψ ϕ and show
that this implies M, w |= Kbaψ → K b a (ψ ∧ a (ψ → ϕ)).
23
Consequently this also means that M, w |= Kbaψ → Kb a (ψ ∧ a (ψ → ϕ)), thus completing the
proof of this direction.
For the converse assume now that M, w |= Kbaψ → K b a (ψ ∧ a (ψ → ϕ)). As X 6= ∅ there
is a world u ∈ W s.t. w ∼a u and M, u |= ψ ∧ a(ψ → ϕ). Therefore we have M, u′ |= ψ → ϕ
for all u a u′ .
Claim 2. Let x ∈ X be arbitrarily chosen, then M, x |= ϕ.
Proof of claim 2. Since x ∈ X we have by definition that M, x |= ψ. It is sufficient to
prove that u a x because this implies M, x |= ψ → ϕ and hence M, x |= ϕ as required.
To show u a x we assume towards a contradiction that u 6a x. Now let R denote the
largest bisimulation on M and consider any x′ ∈ Min a ([x]R= ∩ [x]a ). As R= and ∼a are both
reflexive, we have x ≥a x′ . From u 6a x we therefore have a u′ ∈ Min a ([u]R= ∩ [u]a ) s.t.
u′ 6≥a x′ , u ∼a u′ and (u, u′ ) ∈ R (thus also x′ >a u′ ). Since u′ ∼a u and u ∼a w we have
also u′ ∼a w, and additionally from x ≥a x′ and x′ >a u′ we can conclude that x ≥a u′ and
u′ 6≥a x. Using M, u |= ψ and (u, u′ ) ∈ R we apply Corollary 1 which implies M, u′ |= ψ. As
x ∈ X, u′ ∼a w and x ≥a u′ it must be the case that u′ ∈ X. From u′ 6≥a x we also have
that x 6∈ X, but this this contradicts our initial assumption that x ∈ X. We therefore have
u a x and hence that M, x |= ϕ which completes the proof of the claim.
Recalling that M, w |= Baψ ϕ iff M, x |= ϕ for all x ∈ X, Claim 2 readily shows this
direction, and thereby completes the proof.
From Corollary 3 we have that any expressivity result for LS also holds for LCS , and
similarly for LDS and LCDS . In other words, the conditional belief modality is superfluous in
terms of expressivity when the safe belief modality is at our disposal. What is more, we can
now finally give a full proof of Proposition 9.
Proof (of Proposition 9) Let (M, w) and (M ′ , w′ ) be plausibility models which are modally
equivalent in LS . For any ϕC ∈ LC it follows from Corollary 2 that there is a ϕS ∈ LS s.t.
ϕC ≡ ϕS . Therefore
≡S
M, w |= ϕC ⇔ M, w |= ϕS ⇐=⇒ M ′ , w′ |= ϕS ⇔ M ′ , w′ |= ϕC
24
We proceed by induction on ϕ and let i ∈ {1, 2, 3}. When ϕ is a propositional symbol r in P ′ ,
we have that r 6= q and so r ∈ V (wi ) iff r ∈ V ′ (wi′ ), thus completing the base case. Negation
and conjunction are readily shown using the induction hypothesis.
For ϕ = Ka ψ we have that M, wi |= Ka ψ iff M, v |= ψ for all v ∈ {w1 , w2 , w3 }, since
[wi ]a = {w1 , w2 , w3 }. Applying the induction hypothesis to each element this is equivalent
to M ′ , v ′ |= ψ for all v ′ ∈ {w1′ , w2′ , w3′ } iff M ′ , wi′ |= Ka ψ (as [wi′ ]a = {w1′ , w2′ , w3′ }), which
completes this case. Continuing to consider ϕ = Baγ ψ we can simplify notation slightly,
namely Min a (JγKM ∩ [wi ]a ) = Min a JγKM since [wi ]a = W . The same holds for each world wi′
of M ′ .
Claim 1. For M and M ′ we have that wi ∈ Min a JγKM iff wi′ ∈ Min a JγKM ′ .
Proof of Claim 1. For M we have that w3 >a w2 and w2 >a w1 , and similarly w3′ >′a w2′ and
w2′ >′a w1′ for M ′ . Thus the claim follows from the argument below.
wi ∈ Min a JγKM ⇔
(IH)
M, wi |= γ and there is no j <a i s.t. M, wj |= γ ⇐=⇒
M ′ , wi′ |= γ and there is no j <a i s.t. M ′ , wj′ |= γ ⇔
wi′ ∈ Min a JγKM ′
We now have that M, wi |= Baγ ψ iff M, v |= ψ for all v ∈ Min a JγKM . Applying both the
induction hypothesis and Claim 1, we have that this is equivalent to M, v ′ |= ψ for all
v ′ ∈ Min a JγKM ′ iff M ′ , wi′ |= Baγ ψ.
Finally we consider the case of ϕ = Ban ψ. To this end we note that the union of {(w1′ , w3′ )}
and the identity relation on W is the largest bisimulation on M ′ (this relation cannot be
extended and still satify [atoms]). As w1′ and w3′ are bisimilar, it follows from Corollary 1
that M ′ , w1′ |= ψ iff M ′ , w3′ |= ψ (∗).
Claim 2. For n ∈ N we have that M, w |= ψ for all w ∈ Min na [wi ] iff M ′ , w′ |= ψ for all
w′ ∈ Min na [wi′ ].
Proof of Claim 2. We treat three exhaustive cases for n.
(IH) (∗)
• n = 0: M, w |= ψ for all w ∈ Min 0a [wi ] ⇔ M, w1 |= ψ ⇐=⇒ M ′ , w1′ |= ψ ⇐ ⇒ M ′ , w3′ |= ψ.
0
Therefore M, w |= ψ for all w ∈ Min a [wi ] is equivalent to M , w |= ψ for all w′ ∈
′ ′
{w1′ , w3′ }, and as Min 0a [wi′ ] = {w1′ , w3′ } this concludes this case.
(IH)
• n = 1: Since Min 1a [wi ] = {w1 , w2 } we have that M, w |= ψ for all w ∈ {w1 , w2 } ⇐=⇒
M ′ , w′ |= ψ for all w′ ∈ {w1′ , w2′ }. Using (∗) this is equivalent to M ′ , w′ |= ψ for all
w′ ∈ {w1′ , w2′ , w3′ }. By this argument and the fact that Min 1a [wi′ ] = {w1′ , w2′ , w3′ }, we
can conclude M, w |= ψ for all w ∈ Min 1a [wi ] ⇔ M ′ , w′ |= ψ for all w′ ∈ Min 1a [wi′ ] as
required.
We have that M, wi |= Ban ψ iff M, w |= ψ for all w ∈ Min na [wi ]. Applying Claim 2 this
is equivalent to M ′ , w′ |= ψ for all w′ ∈ Min na [wi′ ] iff M ′ , wi′ |= Ban ψ, thereby completing the
final case of the induction step. It follows that (M, w3 ) ≡CD ′ ′
P ′ (M , w3 ) as required.
Proposition 11 LS 6≦ LCD .
25
a
a a
M w1 w2 w3
a a
p, q q
′a
′a ′a
M′ w1′ w2′ w3′
a a
p
Figure 5: Two single-agent plausibility models and their normal plausibility relations (dashed
arrows). As usual reflexive arrows are omitted.
Proof Consider the formula ♦a p of LS with p ∈ P , and take some arbitrary formula ϕCD ∈
LCD
P . As ϕCD is finite and P is countably infinite, there will be some q 6= p not occurring in
ϕCD . Letting P ′ = P \ {q} this means that ϕCD ∈ LCD P ′ . This choice of p and q can always
be made, and consequently there also exists models M and M ′ as given in Figure 5. The
largest bisimulation on M is the identity as no two worlds have the same valuation. At the
same time {(w1′ , w1′ ), (w1′ , w3′ ), (w2′ , w2′ ), (w3′ , w1′ ), (w3′ , w3′ )} is the largest bisimulation on M ′ .
This gives rise to the normal plausibility relations a (for M ) and ′a (for M ′ ) depicted in
Figure 5 using dashed edges.
Since w3 a w2 and M, w2 |= p it follows that M, w3 |= ♦a p. Furthermore we have that
the image of w3′ under ′a is {w1′ , w3′ }. This means that there is no v ′ ∈ W ′ s.t. w3′ ′a v ′ and
M ′ , v ′ |= p, and consequently M ′ , w3′ 6|= ♦a p. At the same time we have by Lemma 9 that
M, w3 |= ϕCD iff M ′ , w3′ |= ϕCD . Therefore using the formula ♦a p of LS , for any formula of
ϕCD ∈ LCD there are models which ♦a p distinguishes but ϕCD does not, and so ♦a p 6≡ ϕCD .
Consequently we have LS 6≦ LCD as required.
To further elaborate on this result, what is really being put to use here is the ability of the
safe belief modality to (at least in part) talk about propositional symbols that do not occur in
a formula. This is an effect of the derived relation a depending on the largest bisimulation.
26
a
a a
M y x2 x1
a a
p p, q
a
p
M′ a
y′ x′
′a
Figure 6: Two single-agent plausibility models and their normal plausibility relations (dashed
arrows). As usual reflexive arrows are omitted.
Proof We will show the following stronger version of this lemma: For i ∈ {1, 2} : (M, xi ) ≡SP ′
(M ′ , x′ ) and (M, y) ≡SP ′ (M ′ , y ′ ). We proceed by induction on ϕ ∈ LSP ′ , showing that:
for i ∈ {1, 2} : M, xi |= ϕ iff M ′ , x′ |= ϕ and M, y |= ϕ iff M ′ , y ′ |= ϕ (1)
For the base case we have ϕ = r for some r ∈ P \ {q}. Because r 6= q it is clear that r ∈ V (x1 )
iff r ∈ V ′ (x′ ). Since we also have V (x2 ) = V ′ (x′ ) and V ′ (y) = V (y ′ ) this completes the
base case. The cases of negation and conjunction are readily established using the induction
hypothesis, and ϕ = Ka ψ is shown just as we did in the proof of Lemma 9. Before proceeding
we recall that A = {a} and note that for any w ∈ W we have [w]a = {x1 , x2 , y}, as well as
[w′ ]a = {x′ , y ′ } for any w′ ∈ W ′ . Moreover, the largest bisimulation on M and M ′ respectively
is the identity relation, meaning that ≥a =a and ≥′a =′a . For the case of ϕ = a ψ we can
therefore argue as follows.
(IH)
M, x1 |= a ψ ⇔ M, x1 |= ψ ⇐=⇒ M ′ , x′ |= ψ ⇔ M ′ , x′ |= a ψ
(IH)
M, x2 |= a ψ ⇔ (∀i ∈ {1, 2} : M, xi |= ψ) ⇐=⇒ M ′ , x′ |= ψ ⇔ M ′ , x′ |= a ψ
(IH)
M, y |= a ψ ⇔ (∀w ∈ W : M, w |= ψ) ⇐=⇒ (∀w′ ∈ W ′ : M ′ , w′ |= ψ) ⇔ M ′ , y ′ |= a ψ
In fact the last line is essentially the case of Ka ψ, as the image of y under a is W (and W ′
is the image of y ′ under ′a ). This concludes our proof by induction, shows (1) and allows us
to conclude that (M, x1 ) ≡SP ′ (M ′ , x′ ).
Proposition 12 LD 6≦ LS .
Proof Consider the formula Ba1 p ∈ LD with p ∈ P , and additionally take any formula
ϕS ∈ LSP . As ϕS is finite and P is countably infinite, there will be some q 6= p which does
not occur in ϕS . With P ′ = P \ {q} we therefore have ϕS ∈ LSP ′ . As we can always make
such a choice of p and q, this means that there always exists models (M, x1 ), (M ′ , x′ ) of the
form given in Figure 6.
As in the proof of Lemma 10 the largest bisimulation on M and M ′ is the identity and
so Min 1a [x1 ]a = {x1 , x2 } and Min 1a [x′ ]a = {x′ , y ′ }. Consequently M, x1 |= Ba1 p whereas
M ′ , x′ 6|= Ba1 p. Since ϕS ∈ LSP ′ it follows from Lemma 10 that M, x |= ϕS iff M ′ , x′ |= ϕS .
What this proves is that using the formula Ba1 p of LD , no matter the choice of formula ϕS
of LS there will be models which Ba1 p distinguishes but ϕS does not, hence Ba1 p 6≡ ϕS . From
this follows LD 6≦ LS as required.
27
p0 p1 pk q, r q
w0 a w1 a ... a wk a a y
Mk x
p0 p1 pk q q, r
a a ... a a a
Nk w0′ w1′ wk′ y′ x′
Figure 7: Two single-agent plausibility models. We’ve omitted reflexive arrows and for the
sake of readability also some transitive arrows.
We find that this result is quite surprising. Again it is a consequence of our use of the
largest bisimulation when defining our semantics. The purpose of x1 in model M (which is
otherwise identical to M ′ ) is to inject an additional belief sphere without adding any factual
content from P ′ , as that could allow the safe belief formula ϕS to distinguish x1 from x2 .
At this point it might seem as if all hope was lost for the conditional belief modality,
however our final direct result somewhat rebuilds the reputation of this hard-pressed modality.
To this end we define for any k ∈ N the language LDk , which contains every formula of LD
for which if Ban ϕ occurs then n ≤ k. In other words formulas of LDk talk about belief to at
most degree k, which comes in handy as we investigate the relative expressive power of LD
and LC .
Lemma 11 Let k ∈ N be given, and let (M k , w0 ) and (N k , w0′ ) denote the two plausibil-
ity models presented in Figure 7. Then we have that (M k , w0 ) and (N k , w0′ ) are modally
equivalent in LDk .
Proof We prove a stronger version of this lemma, namely that (M k , wi ) ≡Dk (N k , wi′ ) for
0 ≤ i ≤ k, (M k , x) ≡Dk (N k , x′ ) and (M k , y) ≡Dk (N k , y ′ ).
Key to this proof is the fact that x (resp. y) has the same valuation as x′ (resp. y ′ ), and
that x is more plausible than y whereas y ′ is more plausible than x′ . We proceed by induction
on ϕ ∈ LDk . In the base case ϕ is a propositional symbol, and so as the valuation of each wi
matches that of wi′ (0 ≤ i ≤ k), x matches x′ and y matches y ′ this completes the base case.
The cases of negation and conjunction readily follow using the induction hypothesis, and for
ϕ = Ka ψ the argument is essentially that used in the proof of Lemma 9.
Lastly we consider ϕ = Baj ψ for any 0 ≤ j ≤ k, and recall that this is sufficient as ϕ ∈ LDkP .
As neither model contains two worlds with the same valuation, the largest autobisimulation
on either model is the identity, and so both models are normal. With the epistemic relation
of agent a being total, we have for all w ∈ W that Min ja [w]a = {w0 , . . . , wj } and similarly for
all w′ ∈ W ′ that Min ja [w′ ]a = {w0′ , . . . , wj′ }. We therefore have
(IH)
∀w ∈ W : M k , w |= Baj ψ ⇔ ∀v ∈ {w0 , . . . , wj } : M k , v |= ψ ⇐=⇒
∀v ′ ∈ {w0′ , . . . , wj′ } : N k , v ′ |= ψ ⇔ ∀w′ ∈ W ′ : N k , w′ |= Baj ψ
as required. Observe that we can apply the induction hypothesis since j ≤ k, and that
importantly x, y are not in Min ja [w]a , and x′ , y ′ are not in Min ja [w′ ]a . Thus we have shown
that (M k , w0 ) ≡Dk (N k , w0′ ) thereby completing the proof.
Proposition 13 LC 6≦ LD .
Proof Consider now Baq r belonging to LC and any formula ϕD ∈ LD . Since ϕD is finite
we can choose some k ∈ N such that ϕD ∈ LDk . Because p0 , . . . , pk , q, r is taken from the
countably infinite set P , no matter the choice of k there exists pointed plausibility models
(M k , w0 ) and (N k , w0′ ) as presented in Figure 7.
28
(C)DS
(25) (27)
(20) (22)
(C)S (16) CD
C (10) D
Figure 8: Summary of expressivity results for our logics. An arrow X −−→ X ′ indicates that
′ ′
LX is more expressive than LX . A zig-zag line between X and X ′ means that LX and LX
are incomparable. The abbreviation (C)DS means both CDS and DS, and similarly for
C(S) indicating both CS and S. Labels on arrows and zig-zag lines signify from where the
result is taken in Table 1.
To determine the truth of Baq r in (M k , w0 ) and (N k , w0′ ) respectively we point out that
[[q]]M k = {x, y} and [[q]]N k = {y ′ , x′ }. Therefore we have that Min a ([[q]]M k ∩ [w0 ]a ) = {x}
and Min a ([[q]]N k ∩ [w0′ ]a ) = {y ′ }. Since M k , x |= r and N k , y ′ 6|= r, it follows M k , w0 |= Baq r
whereas N k , w0′ 6|= Baq r. By Lemma 11 we have that M k , w0 |= ϕD iff N k , w0′ |= ϕD . With
this we have shown that taking the formula Baq r of LC , there is for any ϕD ∈ LD pointed
plausibility models which Baq r distinguishes but ϕD does not, thus Baq r 6≡ ϕD . It follows that
LC 6≦ LD as required.
We have now shown that the degrees of belief modality cannot capture the conditional
belief modality. What this really showcases is that for Baψ ϕ, ψ potentially enables us to talk
about worlds of arbitrarily large degree. This sets it apart from the degrees of belief modality,
and causes for instance a difference in expressivity.
Now comes our main result, which shows the relative expressivity between the logic of
conditional belief, the logic of degrees of belief and the logic of safe belief.
Theorem 4 LC < LS , LC ⊲⊳ LD , LD ⊲⊳ LS .
29
# Result Inferred from
(1) LC ≦ LS Corollary 3.
S CD
(2) L 6≦ L Proposition 11.
S C
(3) L 6≦ L LC ≦ LCD from (a), LS 6≦ LCD from (2) and applying (d).
(4) LC < LS LC ≦ LS from (1), LS 6≦ LC from (3).
D S
(5) L 6≦ L Proposition 12.
S D
(6) L 6≦ L LD ≦ LCD from (a), LS 6≦ LCD from (2) and applying (d).
(7) LD ⊲⊳ LS LD 6≦ LS from (5), LS 6≦ LD from (6).
(8) LC 6≦ LD Proposition 13.
D C
(9) L 6≦ L LC ≦ LS from (1), LD 6≦ LS from (5) and applying (d).
(10) LC ⊲⊳ LD LC 6≦ LD from (8), LD 6≦ LC from (9).
(11) LCD 6≦ LC LD ≦ LCD from (a), LD 6≦ LC from (9) and applying (e).
C CD
(12) L <L LC ≦ LCD from (a), LCD 6≦ LD from (13).
(13) LCD 6≦ LD LC ≦ LCD from (a), LC 6≦ LD from (8) and applying (e).
(14) LD < LCD LD ≦ LCD from (a), LCD 6≦ LD from (13).
(15) LCD 6≦ LS LD ≦ LCD from (a), LD 6≦ LS from (5) and applying (e).
(16) LS ⊲⊳ LCD LS 6≦ LCD from (2), LCD 6≦ LS from (15).
(17) LCDS ≦ LDS LCDS ≡ LDS from Corollary 3 and Definition 11.
(18) LC ≦ LDS LC ≦ LCDS from (a), LCDS ≦ LDS from (17) and applying (b).
(19) LDS 6≦ LC LS ≦ LDS from (a), LS 6≦ LC from (3) and applying (e).
C DS
(20) L <L LC ≦ LDS from (18), LDS 6≦ LC from (19).
(21) LDS 6≦ LD LS ≦ LDS from (a), LS 6≦ LD from (6) and applying (e).
(22) LD < LDS LD ≦ LDS from (a), LDS 6≦ LD from (21).
(23) LCD ≦ LDS LCD ≦ LCDS from (a), LCDS ≦ LDS from (17) and applying (b).
(24) LDS 6≦ LS LCD ≦ LDS from (23), LCD 6≦ LS from (15) and applying (e).
(25) LS < LDS LS ≦ LDS from (a), LDS 6≦ LS from (24).
(26) LCD 6≦ LDS LS ≦ LDS from (a), LS 6≦ LCD from (2) and applying (e).
(27) LCD < LDS LCD ≦ LDS from (23), LCD 6≦ LDS from (26).
Table 1: Derivation of the relative expressivity of our logics. Each of the references (a), (b),
(d) and (e) refer to properties stated at the start of Section 5.3. Bold faced numbers are
illustrated in Figure 8.
Beyond showing the above theorem, Table 1 fully accounts for the relative expressivity
between LC , LD , LS , LCD and LDS . Finally, using Corollary 3 and property (c) we have
that any expressivity result for LS holds for LCS and similarly for LDS and LCDS . A more
pleasing presentation of these results is found in Figure 8.
30
In other words, bisimulation corresponds to modal equivalence in all three logics. Our ex-
pressivity results can be summarised as (Theorem 4)
LC < LS
LC ⊲⊳ LD
LD ⊲⊳ LS
The logic of conditional belief is less expressive than the logic of safe belief, the logic of
conditional belief and the logic of degrees of belief are incomparable, as are the logic of
degrees of belief and the logic of safe belief.
Our results on bisimulation characterisation suggest that, in some sense, the three logics
are the same, whereas our results on expressive power suggest that, in another sense, the three
logics are different. It is therefore a good moment to explain how to interpret our results.
The bisimulation characterisation result in Corollary 1 says that the information content
of a given plausibility model is equally well described in the three logics. Now consider an
even more specific case: a finite model; and consider a characteristic formula of that model
(these can be shown to exist for plausibility models along the lines of [29, 36]—where we note
that we take models, not pointed models). For a model M this gives us, respectively, formulas
ϕC D S C D
M , ϕM , and ϕM . Then the bisimulation characterisation results say that ϕM , ϕM , and ϕM
S
are all equivalent. Now a characteristic formula is a very special formula with a unique model
(modulo bisimilarity). For other formulas that do not have a singleton denotation (again,
modulo bisimilarity) in the class of plausibility models, this equivalence cannot be achieved.
That is the expressivity result. For example, given that LC < LS , there is a safe belief formula
that is not equivalent to any conditional belief formula. This formula should then describe
a property that has several non-bisimilar models. It is indeed the case that the formula ♦a p
used in the proof of Proposition 11 demonstrating LC < LS has many models! It is tempting
to allow ourselves a simplication and to say that the expressivity hierarchy breaks down if we
restrict ourselves to formulas with unique models.4
Finally, we must point out that in the publication on single-agent bisimulation [4, p. 285],
we posed the following conjecture:
In an extended version of the paper we are confident that we will prove that the
logics of conditional belief and knowledge, of degrees of belief and knowledge, and
both with the addition of safe belief are all expressively equivalent.
It therefore seems appropriate to note that we have proved our own confident selves resound-
ingly wrong!
Bisimulation Prior to our work Demey discussed the model theory of plausibility models
in great detail in [12]. Our results add to the valuable original results he obtained. Demey
does not consider degrees of belief; he considers knowledge, conditional belief and safe belief.
Our plausibility models are what [12] refers to as uniform and locally connected epistemic
4
If we consider infinitary versions of the modalities in our logical languages, in other words, common
knowledge and common belief modalities, we preserve the bisimulation characterisation results (for a more
refined notion of bisimulation) but it is then to be expected that all three logics become equally expressive
(oral communication by Tim French).
31
p p
Mi w1 w2 w3 ··· wi−1 wi
a a a a a
p
M2 v1 v2
a
plausibility models; he also considers models with fewer restrictions on the plausibility func-
tion. But given [12, Theorem 35], these types of models are for all intents and purposes
equivalent to ours. The semantics for conditional belief and knowledge are as ours, but his
semantics for safe belief is different (namely as in [7]). The difference is that in his case an
agent safely believes ϕ if ϕ is true in all worlds as least as plausible as the current world,
whereas in our case it is like that but in the normalised model. This choice of semantics
has several highly significant implications as we will return to shortly.
In line with his interpretation of safe belief as a standard modality, Demey’s notion of
bisimulation for plausibility models is also standard. For example, whereas we require that
[forth≥ ] If v ∈ W and w ≥R ′ ′ R ′ ′
a v, ∃v ∈ W such that w ≥a v and (v, v ) ∈ R,
for example
ϕ4 = ♦a (♦a (♦a (♦a (⊤ ∧ ¬p) ∧ p) ∧ ¬p) ∧ p).
We now have that for any i ≥ 1, Mi , wi |= ϕi ∧ ¬ϕi+1 , which makes this a distinguishing
formula between (Mi , wi ) from (Mi+2 , wi+2 ). In fact, the semantics of a allow us to count
the number of worlds in Mi . In this sense Demey’s logic is immensely expressive.
Again referring to Figure 9, consider M3 , the model with a most plausible ¬p world, a
less plausible p world and an even less plausible ¬p world. In the logic LC of conditional
belief w1 and w3 of M3 are modally equivalent. Hence they also ought to be bisimilar. But
in Demey’s notion of bisimilarity they are not. Hence we have a mismatch between modal
32
equivalence and bisimilarity, which is not supposed to happen: it is possible for two worlds to
be modally equivalent but not bisimilar. Demey also was aware of this, of course. To remedy
the problem one can either strengthen the notion of modal equivalence or weaken the notion
of bisimilarity. Demey chose the former (namely by adding the safe belief modality to the
conditional belief modality), we chose the latter. Thus we regain the correspondence between
bisimilarity and modal equivalence. Baltag and Smets [7] achieve the same via a different
route: they include in the language special propositional symbols, so-called S-propositions.
The denotation of an S-proposition can be any subset of the domain. This therefore also
makes the language much more expressive.
We believe that in particular for application purposes, weakening the notion of bisimu-
lation, as we have done, is preferable over strengthening the logic, as in [7, 12]. This come
at the price of a more complex bisimulation definition (and, although we did not investigate
this, surely a higher complexitity of determining whether two worlds are bisimilar), but, we
venture to observe, also a very elegant bisimulation definition given the ingenious use of the
bisimulation relation itself in the definition of the forth and back conditions of bisimulation.
We consider this one of the highlights of our work.
Expressivity In [7] one finds many original expressivity results. Our results copy those,
but also go beyond. We recall Table 1 for the full picture of our results, and the main results of
those namely LC < LS , LC ⊲⊳ LD , and LD ⊲⊳ LS . The first, LC < LS , is originally found in [7,
page 34, Equation 1.7], and we obtained it using the same embedding translation. However, it
may be worth to point out that in our case this translation still holds for the (in our opinion)
more proper bisimulation preserving notion of safe belief. Baltag and Smets’ S-propositions
are arbitrary subsets of the domain, the (unnecessarily) far more expressive notion of safe
belief. Baltag and Smets also discuss degrees of belief but do not obtain expressivity results
for that, so LC ⊲⊳ LD may be considered novel and interesting. In artificial intelligence, the
degrees of belief notion seems more widely in use than the conditional belief notion, so an
informed reader had better be aware of the incomparability of both logics and may choose
the logic to suit his or her needs. The result that LD ⊲⊳ LS could possibly also be considered
unexpected, and therefore valuable.
33
setting the multi-agent plan existence problem is undecidable [10]. But by placing certain
restrictions on the planning problem it is possible to find decidable fragments even in the
multi-agent case, for example, event models with propositional preconditions [39].
Acknowledgements
Hans van Ditmarsch is also affiliated to IMSc (Institute of Mathematical Sciences), Chennai,
as research associate. He acknowledges support from European Research Council grant EPS
313360. Preliminary versions of the results in this paper can be found in the PhD theses of
Mikkel Birkegaard Andersen [3, Chapter 4] and Martin Holm Jensen [17, Chapter 5].
References
[1] M.B. Andersen, T. Bolander, and M.H. Jensen. Conditional epistemic planning. In Proc.
of 13th JELIA, LNCS 7519, pages 94–106. Springer, 2012.
[2] M.B. Andersen, T. Bolander, and M.H. Jensen. Don’t plan for the unexpected: Planning
based on plausibility models. Logique et Analyse, 2015, To Appear.
[3] Mikkel Birkegaard Andersen. Towards Theory-of-Mind agents using Automated Planning
and Dynamic Epistemic Logic. PhD thesis, Ph. D. Dissertation, Technical University of
Denmark, 2015.
[4] Mikkel Birkegaard Andersen, Thomas Bolander, Hans P. van Ditmarsch, and Mar-
tin Holm Jensen. Bisimulation for single-agent plausibility models. In Stephen Cranefield
and Abhaya Nayak, editors, Australasian Conference on Artificial Intelligence, volume
8272 of Lecture Notes in Computer Science, pages 277–288. Springer, 2013.
[5] G. Aucher. A combined system for update logic and belief revision. In Proc. of 7th
PRIMA, pages 1–17. Springer, 2005. LNAI 3371.
[6] A. Baltag and S. Smets. The logic of conditional doxastic actions. In New Perspectives on
Games and Interaction, Texts in Logic and Games 4, pages 9–31. Amsterdam University
Press, 2008.
[7] A. Baltag and S. Smets. A qualitative theory of dynamic interactive belief revision. In
Proc. of 7th LOFT, Texts in Logic and Games 3, pages 13–60. Amsterdam University
Press, 2008.
[8] Alexandru Baltag and Sonja Smets. A qualitative theory of dynamic interactive belief
revision. In Giacomo Bonanno, Wiebe van der Hoek, and Michael Wooldridge, editors,
Logic and the Foundations of Game and Decision Theory (LOFT7), volume 3 of Texts
in Logic and Games, pages 13–60. Amsterdam University Press, 2008.
[9] P. Blackburn, M. de Rijke, and Y. Venema. Modal Logic, volume 53 of Cambridge Tracts
in Theoretical Computer Science. Cambridge University Press, Cambridge, UK, 2001.
[10] T. Bolander and M.B. Andersen. Epistemic planning for single and multi-agent systems.
Journal of Applied Non-classical Logics, 21(1):9–34, 2011.
[11] K. Britz and I. Varzinczak. Defeasible modalities. In Proc. of the 14th TARK, 2013.
[12] L. Demey. Some remarks on the model theory of epistemic plausibility models. Journal
of Applied Non-Classical Logics, 21(3-4):375–395, 2011.
34
[13] Kit Fine. In so many possible worlds. Notre Dame Journal of Formal Logic, 13(4):516–
520, 1972.
[14] N. Friedman and J.Y. Halpern. A knowledge-based framework for belief change - part i:
Foundations. In Proc. of 5th TARK, pages 44–64. Morgan Kaufmann, 1994.
[15] A. Grove. Two modellings for theory change. Journal of Philosophical Logic, 17:157–170,
1988.
[16] J.Y. Halpern. Reasoning about Uncertainty. MIT Press, Cambridge MA, 2003.
[17] Martin Holm Jensen. Epistemic and Doxastic Planning. PhD thesis, Ph.D. Dissertation,
Technical University of Denmark, 2014.
[18] S. Kraus and D. Lehmann. Knowledge, belief and time. Theoretical Computer Science,
58:155–174, 1988.
[20] N. Laverny. Révision, mises à jour et planification en logique doxastique graduelle. PhD
thesis, Institut de Recherche en Informatique de Toulouse (IRIT), Toulouse, France,
2006.
[21] W. Lenzen. Recent work in epistemic logic. Acta Philosophica Fennica, 30:1–219, 1978.
[22] W. Lenzen. Knowledge, belief, and subjective probability: outlines of a unified system of
epistemic/doxastic logic. In V.F. Hendricks, K.F. Jorgensen, and S.A. Pedersen, editors,
Knowledge Contributors, pages 17–31, Dordrecht, 2003. Kluwer Academic Publishers.
Synthese Library Volume 322.
[23] D.K. Lewis. Counterfactuals. Harvard University Press, Cambridge (MA), 1973.
[24] B. Löwe, E. Pacuit, and A. Witzel. DEL planning and some tractable cases. In Proc. of
LORI 3, pages 179–192. Springer, 2011.
[25] T.A. Meyer, W.A. Labuschagne, and J. Heidema. Refined epistemic entrenchment. Jour-
nal of Logic, Language, and Information, 9:237–259, 2000.
[26] K. Segerberg. Irrevocable belief revision in dynamic doxastic logic. Notre Dame Journal
of Formal Logic, 39(3):287–306, 1998.
[27] W. Spohn. Ordinal conditional functions: a dynamic theory of epistemic states. In W.L.
Harper and B. Skyrms, editors, Causation in Decision, Belief Change, and Statistics,
volume II, pages 105–134, 1988.
[28] R. Stalnaker. Knowledge, belief and counterfactual reasoning in games. Economics and
Philosophy, 12:133–163, 1996.
[29] J. van Benthem. One is a lonely number: on the logic of communication. In Logic
colloquium 2002. Lecture Notes in Logic, Vol. 27, pages 96–129. A.K. Peters, 2006.
[30] J. van Benthem. Dynamic logic of belief revision. Journal of Applied Non-Classical
Logics, 17(2):129–155, 2007.
[31] J. van Benthem. Logical Dynamics of Information and Interaction. Cambridge University
Press, 2011.
35
[32] W. van der Hoek. On the semantics of graded modalities. Journal of Applied Non-
Classical Logics, 2(1), 1992.
[33] W. van der Hoek. Systems for knowledge and beliefs. Journal of Logic and Computation,
3(2):173–195, 1993.
[34] H. van Ditmarsch. Prolegomena to dynamic logic for belief revision. Synthese (Knowl-
edge, Rationality & Action), 147:229–275, 2005.
[35] H. van Ditmarsch. Comments on ‘The logic of conditional doxastic actions’. In New
Perspectives on Games and Interaction, Texts in Logic and Games 4, pages 33–44. Am-
sterdam University Press, 2008.
[36] H. van Ditmarsch, D. Fernández-Duque, and W. van der Hoek. On the definability of
simulation and bisimulation in epistemic logic. Journal of Logic and Computation, 2012.
doi:10.1093/logcom/exs058.
[37] H. van Ditmarsch and W.A. Labuschagne. My beliefs about your beliefs – a case study
in theory of mind and epistemic logic. Synthese, 155:191–209, 2007.
[38] H. van Ditmarsch, W. van der Hoek, and B. Kooi. Dynamic Epistemic Logic, volume
337 of Synthese Library. Springer, 2007.
[39] Quan Yu, Ximing Wen, and Yongmei Liu. Multi-agent epistemic explanatory diagnosis
via reasoning about actions. In Francesca Rossi, editor, IJCAI. IJCAI/AAAI, 2013.
36