An Abstract Logical Approach to Characterizing Strong Equivalence in Logic-based Knowledge Representation Formalisms

Full text

An Abstract Logical Approach to Characterizing Strong Equivalence in
Logic-based Knowledge Representation Formalisms
Ringo Baumann and Hannes Strass
Computer Science Institute
Leipzig University, Germany
Abstract
We consider knowledge representation (KR) formalisms as
collections of finite knowledge bases with a model-theoretic
semantics. In this setting, we show that for every KR formal-
ism there is a formalism that characterizes strong equivalence
in the original formalism, that is unique up to isomorphism
and that has a model theory similar to classical logic.
Introduction
Two knowledge bases T1and T2are (ordinarily) equivalent
if and only if they have the same models. Two knowledge
bases T1and T2are strongly equivalent if and only if for any
arbitrary third knowledge base U, the expansion of T1with
Uis ordinarily equivalent to the expansion of T2with U. In
classical propositional logic, the two notions coincide, but
there are many useful knowledge representation languages
where they are different. Studying strong equivalence of
concrete formalisms is important to gain an insight into the
underlying structure and semantics of the formalism. One
main aim of studying strong equivalence in a concrete form-
alism is to find a so-called characterizing formalism, that
is, another language whose ordinary equivalence coincides
with strong equivalence in the characterized formalism. For
example, strong equivalence of normal logic programs under
the stable model semantics can be characterized by the logic
of here-and-there (Lifschitz, Pearce, and Valverde 2001).
However, such results about the existence of character-
izing formalisms also raise a fundamental question: Does
every formalism have one? In this paper, we answer this
question with a qualified “yes”. More precisely, while not
every formalism has one, we show that the important case
of considering only finite knowledge bases (but still possibly
infinite languages) guarantees the existence of a characteriz-
ing formalism, and that in a very general setting. Existing
results on characterizing formalisms make use of specifics
of each formalism (Lifschitz, Pearce, and Valverde 2001;
Turner 2001). In this paper, we completely abstract away
from formalism specifics and address the core of the prob-
lem, the nature of strong equivalence itself. In fact, we will
not only show the existence of just any characterizing form-
alism, but of characterizing formalisms whose model theory
is uniquely determined (up to isomorphism), and structurally
resembles that of classical logics. At this point, we appeal
to the reader’s intuition on what makes logics classical; we
will later define what we mean by “classical logic” in a pre-
cise mathematical way. Still, we consider this main result
of our paper a surprising and important insight, as it tells
us that for the overwhelming majority of knowledge repres-
entation formalisms, strong equivalence can be approached
using established techniques from classical logic.
While our work is in its essence derived from first prin-
ciples, building mostly upon classical logic and lattice the-
ory, there have been important inspirations. Foremost,
Truszczy´
nski (2006) presented a general, algebraic account
of strong equivalence within approximation fixpoint the-
ory. His setting is indeed quite general, but most of this
generality derives from algebraic commonalities in the se-
mantics of logic programs and default logic. It is not imme-
diately clear, for example, if and how it captures Dung’s ab-
stract argumentation frameworks (AFs, 1995), another im-
portant AI formalism whose strong equivalence has been
studied in the recent past (Oikarinen and Woltran 2011;
Baumann 2016). More precisely, while AFs with all their
semantics can be captured by approximation fixpoint the-
ory (Strass 2013), Truszczy´
nski’s notion of expanding an
operator does not coincide with the corresponding notion of
expanding AFs and his results are not directly applicable.
The paper proceeds as follows. In the next section, we
introduce the general setting in which we derive our results
and present our conception of the term “classical logic”. Af-
terwards, in the main part of the paper, we define charac-
terization logics and show two classes of formalisms that
always possess them.
An Abstract View on Model Theory
What is a classical logic?
We will spend this section introducing an abstract no-
tion of logics with model-theoretic semantics and explaining
when we call some of them classical. Formally, we consider
logical languages L, that is, non-empty sets of language
elements. We make no assumption on the internal struc-
ture of pieces of knowledge F∈ L. These pieces of know-
ledge could be formulas of classical propositional logic, nor-
mal logic program rules, or attacks between arguments. A
model-theoretic semantics for a language Luses a set Iof
interpretations and a model function σ: 2L→2Iwith the
intuition that σassigns each language subset T⊆ L, a the-

ory, the set σ(T)of its models. We make no assumptions
on the internal structure of interpretations – there need not
be an underlying vocabulary of atoms or the like (although
in the concrete cases we consider there often will be) that
are the same among syntax and semantics. This is the main
abstraction in our setting. It goes beyond what is known
from classical logic in that meaning is not assigned to lan-
guage elements (formulas), but only to theories, that is, sets
of language elements. This is a necessary requirement for
being able to model a number of established knowledge rep-
resentation formalisms: for example, in normal logic pro-
grams, meaning is not assigned to single rules, but only to
sets thereof. We illustrate our definitions so far by showing
more precisely how existing formalisms can be embedded
into our setting.
Example 1. Consider a set Aof propositional atoms.
Classical propositional logic: The underlying language
LPL is the set of all classical propositional formulas
over Aand can be defined as usual by induction.
The set of interpretations is then given by the set
IPL ={v:A → {t,f}} of all two-valued interpreta-
tions of A. Lastly, σmod (T)is the set of all models of
the theory T⊆ LPL, that is, the set of all interpretations
satisfying all formulas in T.
Normal logic programs: The underlying language
LLP is the set of all normal logic program rules
a0←a1, . . . , am,∼am+1,...,∼anwith 0≤m≤n
and a0, a1, . . . , an∈ A. The set ILP of interpretations
is then the set ILP = 2Aof all possible stable model
candidates. Accordingly, σstb(T)returns the set of
stable models of the theory (normal logic program)
T⊆ LLP (Gelfond and Lifschitz 1988).
Abstract argumentation frameworks (Dung 1995): The un-
derlying language LAF contains the fundamental build-
ing blocks of AFs, that is, arguments and at-
tacks: LAF ={({a},∅),({a, b},{(a, b)})|a, b ∈ A}.
Extension-based semantics can be incorporated by set-
ting IF= 2Aand, depending on the argumentation
semantics ρwe use, we set σρ(T) = ρ(FT), where
FT= (S(A,R)∈TA, S(A,R)∈TR)is the AF associated to
T⊆ LAF .♦
A consequence function for a language Lis a function
Cn : 2L→2Lthat assigns a given set Tof language ele-
ments another set Cn(T)of language elements. Intuitively,
Cn(T)is understood to be the set of logical consequences
of all formulas in T. Given a language, we can define the
consequence function in terms of the semantics. In words,
the set of consequences of a given theory Tis the union of
all theories Ssuch that any model of Tis a model of S.
Definition 1. Let Lbe a language and σ: 2L→2Ibe a
model function. Define the consequence function
Cσ: 2L→2L, T 7→ [
S⊆L,
σ(T)⊆σ(S)
S♦
For classical logic LPL, this definition coincides with the
standard notion of logical consequence. It will be of great in-
terest in this paper that certain algebraic properties of the se-
mantics induce certain useful properties of the consequence
relation. We now introduce the most important properties.
Definition 2. Let Lbe a language.
•A model function σ: 2L→2Iis antimonotone iff for all
T1, T2∈2L:T1⊆T2=⇒σ(T2)⊆σ(T1).
•A consequence operator Cn : 2L→2Lis monotone iff
for all T1, T2∈2L:T1⊆T2=⇒Cn(T1)⊆Cn(T2).
•A consequence operator Cn : 2L→2Lis increasing iff
for all T∈2L, we find T⊆Cn(T).♦
It is easy to show that for each antimonotone semantics σ,
Definition 1 induces a monotone consequence function. To
save some space in what follows, we define a logic as a tuple
(L,I, σ)consisting of a language L, an interpretation set I,
and a model function σ: 2L→2I.
Standard and strong equivalence
This paper is chiefly about characterizing strong equivalence
in one logic via standard equivalence in another logic. We
will now formally introduce these concepts.
Definition 3. Let (L,I, σ)be a logic and T1, T2⊆ L theor-
ies. We say that T1and T2are
•ordinarily equivalent iff σ(T1) = σ(T2);
•strongly equivalent iff ∀U⊆ L :σ(T1∪U) = σ(T2∪U).
Model function σhas the replacement property if and only
if ordinary equivalence implies strong equivalence. ♦
What properties must a logic possess in order for standard
and strong equivalence to coincide? Maybe it suffices that
the logic has a monotone consequence function?
Example 2. Consider the language L={a, b}with inter-
pretation set I={1,2}and model function σgiven by
σ(∅) = σ({a}) = {1,2};σ({b}) = {2};σ({a, b}) = ∅
It is easy to verify that the semantics σis antimonotone and
(thus) its consequence function Cσis monotone:
Cσ(∅) = Cσ({a}) = {a};Cσ({b}) = Cσ({a, b}) = {a, b}
However, while ∅and {a}are ordinarily equivalent, they
are not strongly equivalent, which can be seen by extending
both with the theory {b}:σ(∅∪{b}) = σ({b}) = {2}and
σ({a}∪{b}) = σ({a, b}) = ∅, with ∅ 6={2}.♦
So having a monotone consequence function is, by itself,
insufficient to guarantee the replacement property. We can
however identify a property that is strong enough to guar-
antee replacement on its own. We call it the intersection
property, because it basically says that the semantics of a
theory can be obtained by only considering the semantics of
the singleton sets constituting the theory.
Definition 4. Let (L,I, σ)be a logic. Its model function
σ: 2L→2Ihas the intersection property iff for all T⊆ L:
σ(T) = \
F∈T
σ({F})♦

It follows from the definition that for any two theor-
ies T1, T2⊆ L, we have that σ(T1∪T2) = σ(T1)∩σ(T2).
The intersection property is a certain locality, independence,
or compositionality criterion. In particular, the intersection
property entails that σ(∅) = I. Towards an explanation of
Example 2 we can now remark that its model function σ
does not have the intersection property:
σ({a, b}) = ∅ 6={2}={1,2}∩{2}=σ({a})∩σ({b})
Indeed, this is necessarily so: as we will show next (and
as is easy to show), satisfying the intersection property is
sufficient for satisfying the replacement property.
Proposition 1. Let (L,I, σ)be a logic. If σsatisfies the
intersection property, then standard equivalence coincides
with strong equivalence.
Notably, monotonicity properties were not even needed in
the above result. So why is it that all formalisms we know
of that have the replacement property also happen to have
monotone consequence functions? It holds because σhav-
ing the intersection property implies that σis antimonotone
(and this in turn implies that Cσis monotone).
Proposition 2. Let (L,I, σ)be a logic where σhas the in-
tersection property. Then σis antimonotone.
It is easy to see that classical propositional logic LPL has
the intersection property simply by definition: the standard
model semantics is typically firstly defined for single formu-
las ϕ∈ LPL and then generalized to theories Tby setting
σmod(T) = Tϕ∈Tσmod ({ϕ}).
Characterization Logics
From now on we omit Ifrom the presentation of
logics and thus write (L, σ), since concrete interpret-
ations are immaterial for strong equivalence. Fur-
thermore, we consider subsets of 2Las domain of σ,
namely the cases dom(σ)=2L(called full logics) and
dom(σ) = 2Lfin ={T∈2L|Tis finite}(finite-theory
logics), the restriction of Lto finite knowledge bases.
Definition 5. Let (L, σ)be a logic. Define the binary
relation strong equivalence ≡σ
s⊆dom(σ)×dom(σ)by
T1≡σ
sT2⇐⇒ ∀U∈dom(σ) : σ(S∪U) = σ(T∪U).♦
It is straightforward to show that ≡σ
sis an equivalence
relation; we denote the equivalence class of a theory
T∈dom(σ)⊆2Lby [T]σ
s. We recall that for all theories
T1, T2⊆ L, we have T1∈[T2]σ
siff [T1]σ
s= [T2]σ
s.
Given an arbitrary logic (L, σ), we want to find a char-
acterizing classical logic, that is, a semantics σ0that has the
intersection property and whose ordinary equivalence coin-
cides with strong σ-equivalence. Such logics get a name.
Definition 6. Let (L, σ)be a (full) logic. The logic (L, σ 0)
is a (full) characterization logic for (L, σ)if and only if:
1. ∀T1, T2⊆ L :σ0(T1) = σ0(T2)⇐⇒ [T1]σ
s= [T2]σ
s;
2. ∀T ⊆ 2L:σ0(ST∈T T) = TT∈T σ0(T).♦
Property (2) is the intersection property; we refer to (1) as
the characterization property. We will start our analysis of
characterization logics with showing that they are unique up
to isomorphism. More precisely, for any model function σ,
the algebras corresponding to the model theories of any two
characterizing model functions σ0and σ00 are isomorphic.
To do that, we first show that the model theory of any char-
acterization logic is a complete lattice, that is, a partially
ordered set where each subset of the carrier set has both a
greatest lower bound (glb) and a least upper bound (lub).
Proposition 3. Let (L, σ)be a full logic with characteriz-
ation logic (L, σ0). The pair σ02L,⊆is a complete
lattice where glb Vand lub Ware given such that for all
K ⊆ σ02L,
^
K∈K
K=\
K∈K
Kand _
K∈K
K=^
L∈Ku
L
with Ku=L∈σ02L∀K∈ K :K⊆L.
After these necessary preliminaries, we now present the
result on uniqueness of characterization logics.
Theorem 4. Let (L, σ)be a full logic with characteriza-
tion logics (L, σ0)and (L, σ00 ). Then the complete lattices
σ02L,⊆and σ002L,⊆are isomorphic.
Thus if a classical characterization logic exists, it is (up
to isomorphism on its model theory) uniquely determined.
However, as we show next, in some cases there simply is no
characterization logic.
Example 3. Let L=Nbe the natural numbers and I 6=∅
arbitrary. We define the semantics σ: 2L→2Isuch that
σ(T) = ®∅if Tis finite,
Iotherwise.
There are two strong equivalence classes: [∅]σ
s, the set of all
finite subsets of N, and [N]σ
s, the set of all infinite subsets of
N. Assume that (L, σ0)is a characterization logic for (L, σ).
By the model intersection property, we get
σ0(N) = σ0 [
n∈N
{n}!=\
n∈N
σ0({n}) = σ0(∅)
in contradiction to the characterization property. ♦
We proceed with some useful properties needed to show that
there is a sub-class of full logics guaranteeing the existence
of a characterization logic in contrast to the unrestricted case
as shown in Example 3. Most importantly, strong equival-
ence classes have an expansion property: It is not completely
obvious, but it follows easily from the definition of strong
equivalence that two strongly equivalent theories can both
be expanded (via set union) with the same theory and are
again strongly equivalent; the converse holds as well. Fur-
thermore, the union of two strongly equivalent theories is
again strongly equivalent to the two theories.
Lemma 5. Let (L, σ)be a full logic and T , T1, T2⊆ L.
1. Strong equivalence is invariant to expansion:
[T1]σ
s= [T2]σ
s⇐⇒ ∀U⊆ L : [T1∪U]σ
s= [T2∪U]σ
s
2. Each strong equivalence class is a join-semilattice:
T1, T2∈[T]σ
s=⇒T1∪T2∈[T]σ
s

It follows in particular that in the case of logics (L, σ)
with Lfinite, each strong equivalence class [T]σ
shas a ⊆-
greatest element that equals the union of all elements. How-
ever, for logics with infinite L, this need not be the case:
in the logic of Example 3, the class [∅]σ
shas no maximal
elements, in particular no greatest element; the class [N]σ
s
has no minimal elements, in particular no least element. We
will see that having a ⊆-greatest element in each equival-
ence class is sufficient for the existence of a characterization
logic. We therefore decided to name this class of logics.
Definition 7. Let (L, σ)be a logic.
1. For T⊆ L define ‘
[T]σ
s=SS∈[T]σ
s
S.
2. (L, σ)is covered if and only if ∀T⊆ L :‘
[T]σ
s∈[T]σ
s.♦
Roughly, the existence of greatest elements in equivalence
classes guarantees that these classes are closed under arbit-
rary set union. Clearly any finite logic is covered. Further-
more, two familiar representatives of covered logics are clas-
sical logic and abstract argumentation theory. In the former
case, it is clear that arbitrary unions of families of equivalent
theories are again theories that are equivalent to each of its
members. In the latter case it is not immediately clear but
can be shown with reasonable effort. We conclude this sec-
tion with its main theorem showing that any full logic being
covered possesses a characterization logic.
Theorem 6. Let (L, σ)be a logic. If (L, σ )is covered then
a characterization logic for (L, σ)is given by (L, σ0)with
σ0: 2L→22L, T 7→ [
S∈2L,
T⊆d
[S]σ
s
[S]σ
s
Finite-Theory Characteriziation Logics
In the field of knowledge representation it is a common as-
sumption that knowledge bases are finite. This is indeed not
overly limiting, as finite knowledge bases will be most rel-
evant for practical purposes. The following definition trans-
lates this assumption into our setting: the finite-theory ver-
sion of a given logic (or simply, a finite-theory logic) con-
siders only the finite knowledge bases of a language.
Definition 8. Given a full logic (L, σ), the finite-theory ver-
sion (L, σ
fin)of (L, σ)is defined by the semantics
σ
fin :2Lfin →σ2Lwith σ
fin(T) = σ(T)
where 2Lfin ={T∈2L|Tis finite}.♦
For finite-theory restrictions of logics, we adequately re-
lax our requirements on characterization logics.
Definition 9. Let (L, σ)be a full logic and (L, σ
fin)its finite-
theory version. We say that (L, σ 0
fin)is a finite-theory char-
acterization logic for (L, σ)if and only if:
1. ∀T1, T2∈2Lfin :σ0
fin(T1) = σ0
fin(T2)iff [T1]σ
fin
s= [T2]σ
fin
s;
2. ∀T1, T2∈2Lfin :σ0
fin(T1∪T2) = σ0
fin(T1)∩σ0
fin(T2).♦
The second item requires binary intersection only; this is due
to the fact that arbitrary unions of theories are not necessar-
ily finite, and thus their semantics might not be well-defined.
As we did in the general case before, we first analyze the
algebraic structure of the resulting model theories. We show
that the model theory of any finite-theory characterization
logic forms a lattice, that is, a partially ordered set where
each non-empty finite subset has both a greatest lower bound
and a least upper bound. (This is in contrast to complete
lattices in the general case.) The proof is, although similar
in procedure, slightly more involved than in the general case.
Proposition 7. Let (L, σ
fin)be a finite-theory lo-
gic with characterization logic (L, σ0
fin). Denoting
K=σ0
fin(T)T∈2Lfin, the pair (K,⊆)is a lattice
where glb and lub are given such that for all K1, K2∈ K:
K1∧K2=K1∩K2and K1∨K2=^{K1, K2}u
where {K1, K2}u={K∈ K | K1⊆K, K2⊆K}.
As before, we can show (with reasonable effort) that finite-
theory characterization logics are unique up to isomorphism.
Theorem 8. Let (L, σ)be a finite-theory logic hav-
ing two finite-theory characterization logics (L, σ0
fin)and
(L, σ00
fin). Denoting the sets K0=σ0
fin(T)T∈2Lfin
and K00 =σ00
fin(T)T∈2Lfin, the lattices (K0,⊆)and
(K00,⊆)are isomorphic.
The following theorem shows that any logic possesses
a finite-theory characterization logic. This means that the
most important case for knowledge representation behaves
well in the sense that characterization logics always exist.
Theorem 9. Let (L, σ)be a full logic. Then a finite-theory
characterization logic for (L, σ)is given by (L, σ0
fin)with
σ0
fin :2Lfin →22L, T 7→ [
S∈(2L)
fin,
T⊆S
[S]σ
fin
s
Intuitively, in this canonical construction of a characteriza-
tion semantics σ0
fin (akin to Herbrand interpretations in first-
order logic), the model set of a theory Tis the set of all the-
ories that are strongly equivalent to some supertheory of T.
References
Baumann, R. 2016. Characterizing equivalence notions for labelling-
based semantics. In KR (this volume).
Dung, P. M. 1995. On the acceptability of arguments and its fun-
damental role in nonmonotonic reasoning, logic programming and n-
person games. Artificial Intelligence 321–357.
Gelfond, M., and Lifschitz, V. 1988. The stable model semantics for
logic programming. In ICLP, 1070–1080.
Lifschitz, V.; Pearce, D.; and Valverde, A. 2001. Strongly equivalent
logic programs. ACM Transactions on Computational Logic 2(4):526–
541.
Oikarinen, E., and Woltran, S. 2011. Characterizing strong equivalence
for argumentation frameworks. Artificial Intelligence 175(14–15):1985–
2009.
Strass, H. 2013. Approximating operators and semantics for abstract
dialectical frameworks. Artificial Intelligence 205:39–70.
Truszczy´
nski, M. 2006. Strong and uniform equivalence of nonmono-
tonic theories – an algebraic approach. Annals of Mathematics and Ar-
tificial Intelligence 48(3–4):245–265.
Turner, H. 2001. Strong equivalence for logic programs and default
theories (made easy). In LPNMR, 81–92.