Approximate Reasoning in First-Order Logic Theories

Johan Wittocx and Maarten Mari en and Marc Denecker

Department of Computer Science, K.U. Leuven, Belgium
johan.wittocx,maarten.marien,[email protected]
Abstract
Many computational settings are concerned with nding (all)
models of a rst-order logic theory for a xed, nite do-
main. In this paper, we present a method to compute from
a given theory and nite domain an approximate structure: a
structure that approximates all models. We show conuence
of this method and investigate its complexity. We discuss
some applications, including 3-valued query answering in in-
tegrated and partially incomplete databases, and improved
grounding in the context of model expansion for rst-order
logic.
Introduction
In this paper, a technique is developed to approximate all
models of a given rst-order logic (FO) theory that share
a xed nite domain. The method is an any-time algo-
rithm that takes as input a FO theory and a nite three-
valued structure specifying the domain and (partial) knowl-
edge about the predicate and function symbols, and that pro-
duces a rened three-valued structure that approximates all
models of the theory that are compatible with the input struc-
ture. Such a technique is useful for many reasoning tasks in
the context of incomplete knowledge.
As an example, consider a database application allowing
university students to make up their teaching program by
selecting certain didactic modules and courses. Below, we
illustrate some integrity constraints that might be imposed
on the selections:
(Selected(c
1
) Selected(c
2
)) (1)
m Module(m) Selected(m) (2)
m (Module(m) Selected(m)
c (Course(c) In(m, c) Selected(c))) (3)
The rst constraint states that courses c
1
and c
2
are mutually
exclusive, the second one expresses that at least one module
should be taken and the third one ensures that all courses
of a selected module are selected. Assume that, at some
point in the application, an incomplete database is obtained,
specifying that some student has selected some modules or
courses, that the student has rejected some others, and that
Copyright c 2008, Association for the Advancement of Articial
Intelligence (www.aaai.org). All rights reserved.
he or she is still undecided about the rest. It is possible now
to use the integrity constraints to derive more complete in-
formation about what will be the nal selection of the stu-
dent. E.g., if course c
1
has been selected, we can derive that
c
2
is not going to be in the selection. If c
2
is rejected from
the selection and belongs to module m
2
, then m
2
cannot
be selected. If m
1
is the only module that still can be se-
lected, m
1
will be in the students selection. Etc. Clearly, this
form of inference could be useful for integrity checking and
for query answering in incomplete databases under integrity
constraints (Greco and Molinaro 2007; Cort es-Calabuig et
al. 2006).
In general, a method to perform the above sort of deriva-
tions is by computing the set o of all models of a the-
ory of integrity constraints that extend the given incomplete
database. In the above example, if a course or module is
selected in all models in o, it will occur in the ultimate se-
lection of the student. Vice versa, if it is not selected in
any model in o, it cannot be selected by the student with-
out causing an integrity violation. The problem with the
sketched method is that, in general, the size of o is exponen-
tial in the size of the database, making this method impracti-
cal in realistic applications. Therefore, it is useful to develop
techniques that compute an approximation of the desired re-
sults without having to compute o. The method presented in
this paper accomplishes this by starting from a partial struc-
ture and gradually rening it by propagating the individual
integrity constraints, thus producing a converging sequence
of gradually more precise partial structures, each of which
approximates all elements of o. We show that if the FO
formulas are in a specic form, called Equivalence Normal
Form, the algorithm is tractable, and that each FO theory
can be brought in this form in linear time. To further reduce
the complexity of the method, we investigate a variant of
the algorithm that uses symbolic representations of the par-
tial structures that approximate o. We also briey study the
accuracy of the computed approximation. Finally, we dis-
cuss applications, not only in the context of partially incom-
plete or integrated databases (Grahne and Mendelzon 1999;
Cort es-Calabuig et al. 2007), but also in the context of
grounding for FO model expansion problems (Mitchell and
Ternovska 2005).
Preliminaries
First-Order Logic (FO)
We assume the reader is familiar with FO. We introduce the
conventions and notations used throughout this paper.
A vocabulary consists of , , variables, function and
predicate symbols . Variables are denoted by lowercase let-
ters, predicate and function symbols by uppercase letters.
Tuples of variables are denoted by x, y, etc. Abusing nota-
tion, we also use x to denote the set of variables occurring in
the tuple x. For a formula , we often write [x], to indicate
that x are precisely the free variables of .
A formula is in term normal form (TNF) when all
its atomic subformulas are of the form P(x
1
, . . . , x
n
),
F(x
1
, . . . , x
n
) = x
n+1
or x
1
= x
2
, where x
1
, . . . , x
n+1
are n + 1 different variables. Each formula is equivalent to
a formula in TNF.
The truth values true and false are denoted by t and f. A
-interpretation I consists of a domain Dand an assignment
of appropriate values to each of its symbols, i.e.:
t to and f to ;
an element x
I
D to every variable x ;
a relation P
I
D
n
to every n-ary predicate symbol P
;
a function F
I
: D
n
D to every n-ary function symbol
F .
An interpretation for only the predicate and function sym-
bols of is called a -structure. The restriction of a -
interpretation I to a subvocabulary is denoted by I[

.
For a variable x and an element d D, I[x/d] is the inter-
pretation that assigns d to x and corresponds to I on all other
symbols. This notation is extended to tuples of variables and
domain elements of the same length. The truth value I() of
a formula in I, and the satisfaction relation [= are dened
as usual (see, e.g., (Enderton 1972)). Abusing notation, we
write I([d]) and I [= [d] instead of I[x/d]([x]), respec-
tively I[x/d] [= [x].
Approximate Structures
In this section, the notion of an approximate structure for
a vocabulary is introduced. To ease the presentation, we
assume from now on that a vocabulary contains no function
symbols. This assumption can be made without loss of gen-
erality, because there exists a standard transformation
1
from
any theory T, to a theory T

containing no function symbols

and such that there is a one-to-one correspondence between
the models of T and T

. A relation R can be approximated

by providing a lower and upper bound for R. The lower
bound expresses which tuples certainly belong to R. Vice
versa, the upper bound states which tuples possibly belong
to R.
1
First transformT to an equivalent theory in TNF. Then replace
each occurrence of an atom F(x) = y in the resulting theory by
G
F
(x, y), where G
F
is a new predicate symbol, and add the sen-
tences x y G
F
(x, y) and x, y
1
, y
2
(G
F
(x, y
1
) G
F
(x, y
2
)
y
1
= y
2
).
Denition 1. An n-ary approximate relation

R is a tuple

R
l
,

R
u
) of two n-ary relations.

R approximates an n-ary
relation R if

R
l
R

R
u
.
An approximate relation

R is called consistent if

R
l

R
u
, i.e., it approximates at least one relation. It is called
exact if

R
l
=

R
u
, i.e., it approximates precisely one rela-
tion. If

R is exact, we identify

R with the unique relation
R it approximates and write R instead of R, R). An ap-
proximate relation

R
1
is less precise than an approximate
relation

R
2
, denoted

R
1

p

R
2
, if

R
l
1


R
l
2
and

R
u
1


R
u
2
.
If

R
1

p

R
2
, then

R
1
approximates all relations approxi-
mated by

R
2
.
Approximate structures are dened similarly to approxi-
mate relations.
Denition 2. An approximate -structure

I with domain D
is a tuple

I
l
,

I
u
) of two -structures with domain D. The
approximate relation P

I
assigned by

I to a predicate symbol
P is the approximate relation P

I
l
, P

I
u
).

I approxi-
mates a -structure I with domain D if for each predicate
symbol P , P
I
is approximated by P

I
.
The notions of consistency, exactness and precision point-
wise extend to approximate structures. I.e.,

I is consis-
tent (exact) if for all predicate symbols P, P

I
is consis-
tent (exact) and

I
1

p

I
2
if for all predicate symbols P,
P

I
1

p
P

I
2
. Note that

I approximates a structure I iff

I
is less precise than the exact approximate structure I, I).
The least precise approximate structure with domain D is
denoted
D

p
. Vice versa, the most precise one is denoted

p
.
D

p
approximates all structures with domain D,
D

p
is inconsistent. The size [

I[ of

I is dened as the size of its
domain D.
The truth value

I() of a formula in an approximate
structure

I with domain D is inductively dened as follows:

I() =

I
l
() if is atomic;

I() = (

I
u
,

I
l
)());

I( ) =

I()

I();

I(x [x, y]) = t iff for some d D,

I([d, y]) = t.
Otherwise,

I(x [x, y]) = f.
Observe that if a structure I is approximated by

I, then

I()
underestimates I(). I.e., if

I() = t, then I() = t.
As such, we say that is certainly true according to

I if

I() = t. Vice versa,

I
u
,

I
l
)() is an overestimation of
I(). I.e., if I() = t, then

I
u
,

I
l
)() = t and we say that
is possibly true in

I. Similarly, is certainly false in

I if

I() = t, or, equivalently,

I
u
,

I
l
)() = f. Finally, is
possibly false in

I if

I() = f.
Denition 3. An approximate structure

I with domain D is
an approximation of a theory T, if it approximates all mod-
els of T with domain D. An approximate structure

J with
domain D is an approximation of T above

I if

I
p

J and
all models of T approximated by

I are approximated by

J.
Approximate Reasoning
As mentioned in the introduction, we are interested in ap-
proximations for a given theory T (e.g., the integrity con-
straints of a database) above a given approximate struc-
ture

I (e.g., the tables of an incomplete database) with -
nite domain D. The most precise approximation that can
be obtained is given by glb

p
(M, M) [ M, M)
p

I and M [= T). We call this the optimal approximation

and denote it by O
T
(

I). If T has no models approximated

by

I, then O
T
(

I) =
D

p
. In particular, this is the case when

I is inconsistent.
Example 1. Consider the theory T
1
= (1); (2); (3) of
the sentences of the introduction and let

I be the approx-
imate structure with domain D = m
1
, m
2
, c
1
, c
2
, c
3
, c
4

given by Module

I
= m
1
, m
2
, Course

I
=
c
1
, c
2
, c
3
, c
4
, In

I
= (m
1
, c
1
); (m
1
, c
3
), (m
2
, c
2
)
and Selected

I
= c
1
, D). Then O
T
1
(

I) assigns
m
1
, c
1
, c
3
, m
1
, c
1
, c
3
, c
4
) to Selected.
The problem of constructing O
T
(

I) is at least as hard as
deciding whether T has a model approximated by

I, which
is intractable. Even for a xed T and varying

I, this decision
problem is NP-complete (Fagin 1974).
The usual complexity measure for algorithms in the con-
text of databases with integrity constraints is called data
complexity, and is dened as the complexity in the size [

I[
of

I. I.e., for this measure, the theory T is considered to
be xed. All complexity results in this paper concern data
complexity. In this section we develop a method to compute
an approximation for T above

I in polynomial time in [

I[.
The method we present consists of computing and com-
bining the optimal approximations for individual sentences
of T, instead of computing the optimal approximation for
T as a whole. The resulting approximate structure

J is not
guaranteed to be as precise as O
T
(

I). Since an individual

sentence can be as complex as a theory, this computation can
be as hard as that of O
T
(

I). We therefore proceed by intro-

ducing a normal form for which this computation is guar-
anteed to be polynomial, and provide a linear reduction of
arbitrary theories to this normal form.
Renement Sequences
For the rest of this paper, let T be a theory and

I an approx-
imate structure with nite domain D.
Similarly as for theories, we associate to each sentence
the operator O

which maps an approximate structure

I
to the optimal approximation of above

I, i.e., O

I) =
glb

p
(M, M) [ M, M)
p

I and M [= ). Observe
that for a given T, M, M)
p

I [ M [=
M, M)
p

I [ M [= T, hence O
T
(

I) is more precise
than O

I). Also note that for any sentence , O

is a
p
-
monotone operator on the lattice of approximate structures
with domain D.
Lemma 4. If is a sentence of T, then O

I) is more pre-
cise than

I and approximates T above

I.
Lemma 4 shows the soundness of the following procedure
for computing approximations

J of T above

I. Start by set-
ting

J =

I. Then choose a sentence of T and replace

J
by O

(

J). Repeat.
Interesting properties of this procedure are termination and
conuence. To state these formally, we introduce the con-
cept of a renement sequence.
Denition 5. A renement sequence for T above

I is a se-
quence

J
i
)
0in
of approximate structures such that

J
0
=

I,

J
i
<
p

J
i+1
and

J
i+1
= O

J
i
) for some sentence
of T. A renement sequence is called terminal if it cannot
be extended anymore. For a terminal renement sequence

J
i
)
0in
, the structure

J
n
is called its limit.
Theorem 6. Every renement sequence for T above

I is
nite and every terminal renement sequence has the same
limit.
We denote the unique limit of a terminal renement se-
quence for T above

I by O
T
(

I).
Example 1 (continued). Let

I
i
)
0i4
be the renement
sequence for T
1
above

I given by

I
1
= O
(1)
(

I
0
),

I
2
=
O
(3)
(

I
1
),

I
3
= O
(2)
(

I
2
) and

I
4
= O
(3)
(

I
3
). One can ver-
ify that c
2
, Selected

I
u
1
, i.e., in

I
1
it is already derived
that c
2
cannot be selected. Also, m
2
, Selected

I
u
2
, m
1

Selected

I
l
3
and c
3
Selected

I
l
4
. Hence,

I
4
= O
T
1
(

I), the
renement sequence is terminal and O
T
1
(

I) = O
T
1
(

I).
It is not necessarily the case that O
T
(

I) is equal to O
T
(

I).
E.g., if T = P Q; P Q and

I =
D

p
, then
O
T
(

I) =
D

p
. However, because T has no model, the
optimal approximation O
T
(

I) is equal to
D

p
.
Also, for two logically equivalent theories T and T

,
O
T
(

I) and O
T

I) are not necessarily the same. E.g., this

is not the case for

I =
D

p
, T = P Q; P Q and
T

= (P Q) (P Q).
Equivalence Normal Form
The length of a terminal renement sequence for T above

I is polynomial
2
in [

I[. Consequently, if for every sentence

of T, O

is computable in polynomial time, a terminal

renement sequence for a xed T can be computed in poly-
nomial time.
We nowdene a normal formfor which O

is computable
in polynomial time.
Denition 7. A TNF sentence is in equivalence normal
form (ENF) if is of the form x (P(x) [y]), such
that P is a predicate symbol or the symbol , y x and
[y] is of the form Q(y), Q(y), Q(y) R(y), Q(y)
R(y), v Q(y, v), v Q(y, v), y
1
= y
2
, where Q and R are
2
Note that the length of a renement sequence can be exponen-
tial in the arity of the symbols occurring in T.
predicate symbols, different from P. A theory is in ENF if
all its sentences are.
Note that for ENF sentences of the form x (P(x)
Q(y) R(y)) and x (P(x) Q(y) R(y)), the variables
occurring in Q and R are exactly the same.
The following transformation (akin to the Tseitin trans-
formation for propositional logic (Tseitin 1968)) reduces ar-
bitrary theories T to ENF theories T

, such that the models

of T

restricted to Ts vocabulary are the models of T. Start

by transforming T to TNF. Then replace every sentence of
the resulting theory by . Finally, as long as the theory
contains sentences x (P(x) [y]) that are not in ENF,
replace a subformula [z] of by Q(z), where Q is a new
predicate symbol, and add the sentence z (Q(z) [z])
to the theory. Note that at most one new predicate symbol is
introduced for each node in the parse tree of . Hence, the
size of T

is linear in the size of T.

Example 2. Applying the transformation outlined above
on sentence (2) from the introduction yields the two sen-
tences m Q(m) and m (Q(m) Module(m)
Selected(m)), where Q is a new predicate symbol.
We now show that for any ENF sentence , O

is com-
putable in polynomial time. The method we present is based
on the simple observation that if M is a model of the ENF
sentence y, z (P(y, z) [y]) and d
y
, d
z
are tuples of do-
main elements, then M(P(d
y
, d
z
)) = M([d
y
]). This has
the following consequences:
1. If

I([d
y
]) = t, i.e., [d] is certainly true in

I, then
for any model M of approximated by

I, it holds that
M(P(d
y
, d
z
)) = M([d
y
]) = t. Therefore,

I([d
y
]) =
t implies that O

I)(P(d
y
, d
z
)) = t. Similarly, if [d
y
]
is certainly false in

I, then so is P(d
y
, d
z
) in O

I).
2. Vice versa, if

I(P(d
y
, d
z
)) = t, then M([d
y
]) = t
in any model M of approximated by

I. Therefore, if

I(P(d
y
, d
z
)) = t and there is only one way to ensure that

J([d
y
]) = t in an approximate structure

J
p

I, this is
reected in O

I). We illustrate this on two examples:

If [y] is the formula Q(y) R(y), then the only
way to ensure that O

I)([d
y
]) = t, is by making
both Q(d
y
) and R(d
y
) certainly true in O

I). As
such,

I(P(d
y
, d
z
)) = t, then O

I)(Q(d
y
)) = t and
O

I)(R(d
y
)) = t.
If [y] is the formula Q(y) R(y),

I(P(d
y
, d
z
)) = t
and

I(R(d
y
)) = t, then O

I)(Q(d
y
)) = t.
A similar observation can be made if P(d
y
, d
z
) is cer-
tainly false in

I.
3. For tuples d
y
, d
z
and d

z
of domain elements and
any model M of , M(P(d
y
, d
z
)) = M([d
y
]) =
M(P(d
y
, d

z
)). Therefore, if there exists a tuple d
z
such that

I(P(d
y
, d
z
)) = t, then for all tuples d

z
,
O

I)(P(d
y
, d

z
)) = t. Similarly if

I(P(d
y
, d
z
)) = f.
We can now show how to compute O

. Lemma 8, based
on items 1 and 3 above, shows how to compute P
O

I)
;
Lemma 9, based on item 2 above, does the same for the
predicates that occur in .
Lemma 8. Let be the ENF sentence y, z (P(y, z)
[y]). Denote by [y] the formula z P(y, z) and
by

[y] the formula z P(y, z). If O

I) is
consistent, then for any tuple of domain elements d
y
and d
z
, O

I)
l
(P(d
y
, d
z
)) =

I([d
y
] [d
y
]) and
O

I)
u
(P(d
y
, d
z
)) =

I([d
y
]

[d
y
]). If O

I) is
inconsistent, then either

I is inconsistent or there exist a tu-
ple d
y
such that

I([d
y
] [d
y
]) =

I([d
y
]

[d
y
]).
Lemma 9. Let , [y] and

[y] be as in the previous

lemma and let Q be a predicate that occurs in . If O

I)
is consistent, then O

I)
l
(Q(d)) =

I(Q(d)
ct
[d]) and
O

I)
u
(Q(d)) =

I(Q(d)
cf
[d]), where
ct
and
cf
are the formulas that are dened as follows, depending on
.

ct

cf
Q(y)

Q(y) R(y)

R(y)
v Q(y, v)

(v ,= v

Q(y, v

))

ct

cf
Q(y)

Q(y) R(y) R(y)

v Q(y, v) v

(v ,= v

Q(y, v

))

Observe that equality atoms in an ENF theory occur only

in sentences of the form P(y
1
, y
2
, z) y
1
= y
2
. Since
the interpretation of = is given and xed (by the identity
relation), such sentences can only be used to rene the inter-
pretation of P, which is described by Lemma 8.
Because the evaluation of a formula in an approximate
structure

I can be done in polynomial time in the size of

I, lemmas 8 and 9 provide a method to compute O

I) in
polynomial time for any ENF sentence . Indeed, rst as-
sume that O

I) is consistent and compute it according to

the characterization provided by the lemmas. Then check
whether the result is consistent. If so, then it is indeed
O

I). Else, by lemma 8, O

I) =
D

p
.
Theorem 10. Let T be a theory in ENF and

I an approxi-
mate structure with nite domain D. Then a terminal rene-
ment sequence for T above

I is computable in time polyno-
mial in [

I[.
Hence, the following algorithm can be used to compute an
approximation for an arbitrary theory T over a vocabulary
above

I.
1. Transform T to a theory T

over

in ENF and extend

I
to

by assigning (, D
n
) to every n-ary predicate sym-
bol P

.
2. Construct a (terminal) renement sequence

I
i
)
0in
for
T

above

I and return

I
n
[

.
Note that this is an any-time algorithm: the renement se-
quence constructed in step 2 can be terminal, but this is not
necessary. The following examples illustrate what the algo-
rithm can achieve.
Example 1 (continued). If T
1
is transformed to an ENF
theory T

as described below denition 7, then a terminal

renement sequence for T

above

I is equal to O
T
(

I). I.e.,
the algorithm above on T
1
and

I computes the optimal re-
sult.
Example 3. Consider the theory T
2
, taken from
some planning domain, consisting of the sentence
a
0
, a
p
, t
0
Prec(a
p
, a
0
) Do(a
0
, t
0
) t
p
(t
p
<
t
0
Do(a
p
, t
p
)). This sentence describes that some
action a
0
with precondition a
p
can only be performed at
timepoint t
0
if a
p
is performed at an earlier timepoint t
p
.
Let

I be an approximate structure. Then the algorithm
above derives that if t
0
is smaller than the ith timepoint
and there exists a chain of actions a
1
, . . . , a
i
such that

I(Prec(a
1
, a
0
) . . . Prec(a
i
, a
i1
)) holds, then a
0
can
certainly not be performed at timepoint t
0
.
In an actual implementation of the algorithm, it is impor-
tant to construct a good renement sequence. I.e., a short se-
quence with a sufciently precise last element. To nd such
a sequence, a good heuristic is needed to decide at each step,
which operator O

to apply to extend the sequence.

A simple proposal consists of keeping track of a set C of
predicates whose approximation recently changed. To de-
cide which O

to apply, one could search for the predicate

P C with the greatest difference between its current and
previous approximation. Then, apply O

for each con-

taining an occurrence of P. Finally, if Ps approximation
did not change by applying these operators, remove P from
C.
Also, to efciently compute the next approximate struc-
ture in a renement sequence, one should try to avoid recom-
puting the information that is already present in the previous
approximate structures. This can be done in a similar way as
the semi-naive evaluation technique (Ullman 1988) avoids
to recompute already known answers to a datalog query.
Optimality
As mentioned above, it is not necessarily the case that
O
T
(

I) = O
T
(

I). This raises the question how well O

T
(

I)
approximates O
T
(

I). Generally spoken, the less axioms

T contains, the closer O
T
(

I) is to O
T
(

I), with O
T
(

I) =
O
T
(

I) if T is a singleton. But since translation to ENF tends

to split up a theory in many small sentences, the question of
the accuracy of the computed approximation is certainly an
important one.
Answering this question is far from trivial. Below, we
present two simple lemmas that can be helpful in proving
O
T
(

I) = O
T
(

I) for a given T and

I. Finding more elabo-
rate results is part of future work. Other preliminary results,
in the more restricted setting of locally closed databases, are
presented in (Cort es-Calabuig et al. 2007).
Lemma 11. Let

I be an approximate structure and let
T be the theory ,
1
, . . . ,
n
and T

the theory
, ,
1
, . . . ,
n
. If one of the following conditions is sat-
ised, then O
T
(

I) = O
T

I):
there is no predicate symbol that occurs in both and ;

I() = t or

I() = t.
Lemma 12. Let T and T

be as in lemma 11 and let

I be an
approximate structure. If

I
i
)
0in
is a renement sequence
for T

above

I and O
T
(

I
n
) = O
T

I
n
), then O
T
(

I) =
O
T

I).
The following example illustrates the use of lemmas 11
and 12 to prove that for a certain T and

I, the transformation
to ENF does not result in a less optimal approximation of T
above

I.
Example 4. Consider the theory T
3
over consisting
of the sentence x UGCourse(x) (GCourse(x)
Difficult(x)), which states that a course is either for un-
dergraduates, or for graduates and difcult. An ENF trans-
formation T
4
of T
3
consists of, e.g., the three sentences
x S
1
(x) (4)
S
1
(x) UGCourse(x) S
2
(x) (5)
S
2
(x) GCourse(x) Difficult(x) (6)
We show that O
T
4
(

I)[

= O
T
3
(

I)[

for any approximate

structure

I for which UGCourse

I
is exact, i.e.,

I contains
complete information about the undergraduate courses.
Denote by T
5
the theory (4); (5) (6) and by T
6
the the-
ory (4) (5) (6). Because clearly, O
T
6
(

I) = O
T
6
(

I)
and O
T
6
(

I)[

= O
T
3
(

I)[

, it is sufcient to prove that

O
T
6
(

I) = O
T
4
(

I).
First observe that O
(4)
(

I) assigns the approximate re-

lation D, D) to S
1
, i.e. O
(4)
(

I)(S
1
(d)) = t for each
d D. Hence, O
(4)
(

I)((4)) = t and therefore, by

lemma 11, O
T
5
(O
(4)
(

I)) = O
T
6
(O
(4)
(

I)). By lemma 12,

also O
T
5
(

I) = O
T
6
(

I).
Because UGCourse

I
is exact, also O
(5)
(O
(4)
(

I))((5)) =
t. As before, we conclude by lemma 11 and 12 that
O
T
4
(

I) = O
T
5
(

I).
Symbolic Approximate Reasoning
The presented algorithm to compute an approximation for T
above some approximate structure

I with domain D will of-
ten be too expensive for practical purposes. If the vocabulary
of T contains a predicate P with arity n, then the approxi-
mate relation assigned to P can contain up to 2 [D[
n
tuples.
For a large D, storing and manipulating such an approxima-
tive relation can become infeasible in practice. Note that n is
not necessarily small. Even if the original theory T contains
no predicates with a large arity, transforming T to an ENF
theory creates predicates with arity equal to the number of
free variables of the subformula in T it represents.
To obtain a more practical algorithm, the computed rene-
ment sequence can be represented in a compact, symbolic
way, independent of D. In this section, we show how such a
symbolic representation can be obtained.
Symbolic Approximate Structures
Let T be an ENF theory over a vocabulary and let be a
vocabulary, not necessarily related to .
Denition 13. A symbolic approximate -structure

over
is an assignment of a tuple

l
P
[x],

u
P
[x]) to each n-ary
predicate P , where

l
P
[x] and

u
P
[x] are two formulas
over with n free variables.
Given a xed -structure I, to each symbolic approxi-
mate -structure

over , an approximate -structure is
associated, by evaluating it in I. As such,

can be seen
as a symbolic representation of this associated approximate
-structure.
Denition 14. The evaluation of a symbolic approximate -
structure

over in a -structure I is the approximate -
structure I(

), dened by P
I(

)
= d [ I [=

l
P
[d], d [
I [=

u
P
[d]).
Example 5. Let be the vocabulary containing two n-
ary relation symbols P
l
and P
u
for every n-ary relation
symbol P . Let

I be an approximate -structure,

the symbolic approximate structure over assigning

P
l
(x), P
u
(x)) to every P and I the -structure as-
signing P

I
l
to P
l
and P

I
u
to P
u
for every P . Then the
evaluation of

in I is precisely

I. Hence, given I,

can be
seen as a symbolic representation of

I.
For a formula over , its evaluation in a symbolic ap-
proximative structure

is the sentence

(), obtained by,
for every predicate symbol P , substituting each posi-
tive occurrence of an atomP(x) by
l
P
[x] and each negative
occurrence by
u
P
[x]. First evaluating in

and then eval-
uating the result in a -structure I, corresponds to evaluating
in I(

):
Lemma 15. Let be a formula over ,

a symbolic ap-
proximative structure over and I a -structure. Then
I(

()) = I(

)().
Let be an ENF sentence. Based on lemmas 8 and 9, we
dene the operator

O

on the set of symbolic approximate

structures. This operator is the symbolic variant of O

, in
the sense that for any symbolic approximate structure

, it
holds that I(

O

)) = O

(I(

)).
Denition 16. Let be the ENF sentence y, z (P(y, z)
[y]).

O

is the operator on the set of symbolic approximate

structures dened by
(

O

))
l
P
=

([x] z P(x)) and (

O

))
u
P
=
(

([x] z P(x)));
for an atom Q(y) that occurs in , (

O

))
l
Q
=

(
ct

Q(y)) and (

O

))
u
Q
= (

(
cf
Q(y))), where
ct
and
cf
are dened as in lemma 9.

) corresponds to

on all predicates that do not oc-
cur in .
Because of lemmas 8, 9, and 15, we have the following
result, as desired.
Proposition 17. For every ENF sentence , symbolic ap-
proximate structure

over and -structure I, it holds
that I(

O

)) = O

(I(

)) if O

(I(

)) is consistent. If
O

(I(

)) is inconsistent, then so is I(

O

)).
Symbolic Renement Sequences
Let

be a symbolic approximate structure. Similarly to a
renement sequence, a symbolic renement sequence for T
above

is a sequence

i
)
0in
of symbolic approximate
structures such that

0
=

and for each i < n, there is a
sentence in T such that

i+1
=

O

i
).
Based on proposition 17, an approximate structure for T
above

I can be obtained by rst dening an approximate
structure

over a vocabulary and a -structure such that
I(

) =

I. Example 5 shows a possible denition of these.
Then construct a symbolic renement sequence

i
)
0in
for T above

. Finally, return I(

n
).
However, using symbolic renement sequences has its
costs. Note that for any symbolic approximate structure

and ENF sentence , the formulas assigned by

O

) are
strictly larger than the ones assigned by

. Also, a symbolic
renement sequence is not guaranteed to be nite. These
problems can sometimes be avoided by replacing the formu-
las assigned by

O

) by logically equivalent, but smaller

ones. Such smaller, equivalent formulas can be computed
by simplication algorithms for FO (Goubault 1995).
Applications
In this section, we discuss some applications of computing
an approximation of an FO theory.
Model Expansion
Searching approximations for a theory T above an approx-
imate structure

I is strongly related to the setting of model
expansion for FO (MX(FO)). This setting has been proposed
as a framework to solve NP problems declaratively (Mitchell
and Ternovska 2005).
Denition 18. Given a theory T over a vocabulary , a vo-
cabulary and a nite -structure I

, the model ex-

pansion (MX(FO)) search problem for input , T, , I

) is
the problem of nding models M of T that expand I

, i.e.,
M[

= I

. The MX(FO) decision problem for the same in-

put is the problem of deciding whether such a model exists.
The techniques developed in this paper can be used to
give approximate answers for MX(FO) problems with input
, T, , I

). Let Dbe the domain of I

and let

I

be the ap-
proximate structure with domain D that assigns P
I

, P
I

)
to every predicate symbol P and , D
n
) to all other
n-ary predicate symbols P . Then an approximation

J
of T above

I

is an approximation to the answers for the

MX(FO) search problem. If

J is inconsistent, then certainly
the answer to the MX decision problem is negative.
The existing solvers for MX(FO) (Mari en, Wittocx, and
Denecker 2006; Mitchell et al. 2006) work by rst ground-
ing T, i.e., creating a propositional theory T
G
equivalent to
T, and then applying a SAT solver. To obtain a compact
grounding T
G
, subformulas in T
G
that are certainly true or
false according to

I

are replaced by respectively and .

In (Wittocx, Mari en, and Denecker 2008), it was observed
that an even more compact grounding can be obtained by
rst computing an approximation

J for T above

I

and then
substituting subformulas that are certainly true or false ac-
cording to

J. Experiments in that paper showed that com-
puting and using

J almost never incurs overhead, while the
size of T
G
decreases. Moreover, often the time to create the
grounding decreases drastically.
Incomplete Databases
A recent trend in databases is the development of approx-
imate methods to reason about databases with incomplete
knowledge. The incompleteness of the database may stem
from the use of null values, or of a restricted form of closed
world assumption as in (Cort es-Calabuig et al. 2007), or it
arises from integrating a collection of local databases each
based on its own local schema into one virtual database over
a global schema (Grahne and Mendelzon 1999). In all these
cases, the data complexity of certain and possible query an-
swering is computationally hard (co-NP, respectively NP).
For this reason fast (and often very precise) polynomial ap-
proximate query answering methods are developed, which
compute an underestimation of the certain, and an overesti-
mation of the possible answers.
The tables of an incomplete database are naturally rep-
resented as an approximate structure

I. The integrity con-
straints, local closed world assumption or mediator scheme
corresponds to a logic theory T. Approximate answering a
query [x] boils down to computing two relations R
l
and
R
u
consisting of tuples d which are true in all models, re-
spectively some model, of T above

I. Hence, a possible
technique to compute R
l
and R
u
consists of computing a
renement sequence

I
i
)
0in
for T above

I and then com-
puting

I
n
([x]), respectively the complement of

I
n
([x]).
It turns out that this method of approximate query an-
swering generalizes the algorithm of (Cort es-Calabuig et
al. 2006). Computing a symbolic renement sequence

i
)
0in
and then computing

n
([x]), respectively

n
([x]), generalizes the query rewriting technique pre-
sented in (Cort es-Calabuig et al. 2007).
Related Work
The notion of approximation in our work is quite differ-
ent from the one in most other works about approximate
reasoning in logic theories. E.g., the approach started by
Levesque (Levesque 1988) and further explored by, a.o.,
Schaerf and Cadoli (Schaerf and Cadoli 1995) performs ap-
proximate query answering for one query, while our method
computes an approximate structure approximating all mod-
els of a theory. Their methods could hardly be used for, e.g.,
grounding. The same holds for the knowledge compilation
approach, started by Selman and Kautz (Selman and Kautz
1991), which is often applied in description logics.
In the literature, we nd mathematical constructs similar
to the ones used in this paper in two other research areas:
rough set theory (Pawlak 1992) and four-valued logic (Ack-
ermann 1967). In rough set theory, attributes of elements of
a universe U are used to classify each x U as certainly in,
possibly in, or certainly not in a rough set X U of inter-
est. In our work, the computed sets P
O
T
(

I)
can be viewed
as non-exact sets similar to rough sets but, in contrast, they
are are not given explicitly but computed from T and

I.
Also, (consistent) approximate relations and structures
can be viewed as (three-valued) four-valued relations and
structures, but in multi-valued logics, relation symbols de-
note three-valued relations, whereas we use standard two-
valued FO-logic and use the approximate relations only to
approximate collections of possible relations.
As pointed out in the previous section, the work pre-
sented here is strongly related to certain approximate meth-
ods in databases with incomplete knowledge, for instance in
the context of integration of distributed databases (Lenz-
erini 2002; Grahne and Mendelzon 1999), or of locally
closed databases (Cort es-Calabuig et al. 2006; 2007). Our
methods could be useful in generalizations of locally closed
databases, e.g., for databases with integrity constraints.
In (Wittocx, Mari en, and Denecker 2008) we presented
in more detail how symbolic approximations can be used
for grounding in the context of MX(FO). The non-symbolic
algorithm, with its results on termination, conuence, com-
plexity and optimality is not described in that paper. Also,
the symbolic method presented there is less precise and less
general than the one in this paper.
Conclusions
We presented a method to compute an approximation of all
models with a given nite domain of given FO theory. For
an important class of theories, the data-complexity of the
method is polynomial. Some preliminary results about pre-
cision were stated. We also presented how to obtain sym-
bolic representations of approximations, which can improve
the efciency of the method. Finally, we discussed some ap-
plications in the context of databases and model expansion
and mentioned related work.
Acknowledgments
Johan Wittocx is research assistant of the Fonds voor Weten-
schappelijk Onderzoek - Vlaanderen (FWO Vlaanderen).
Appendix
Proof of lemma 4 First note that for every model M of T
approximated by

I, it holds that M, M)
p

I and M [=
. Hence, by the denition of O

, O

I) approximates T
above

I.
Denote by / the set M, M)
p

I [ M [= . If
/ = , then O

I) =
D

p
, which is clearly more precise
than

I. On the other hand, if / ,= , then

I is a
p
-lower
bound of /and hence, also in this case glb
leq
p
(/)
p

I.
Proof of theorem 6 Because the domain D of

I is -
nite, there are only a nite number of approximate structures

J
p

I. As such, every renement sequence above

I is -
nite.
To prove conuence, let

J
i
)
0in
and

K
j
)
0jm
be two terminal renement sequences and denote by

i
and
j
the sentences of T such that respectively
O

i
(

J
i
) =

J
i+1
and O

j
(

K
j
) =

K
j+1
. Then,
because for each sentence , O

is a
p
-monotone
operator,

K
m
= O

m1
(O

m2
(. . . (O

0
(

I
n
))))
p
O

m1
(O

m2
(. . . (O

0
(

J
n
)))) =

J
n
. Similarly,

J
n

p

K
m
. Hence,

J
n
=

K
m
.
Proof of lemmas 8 and 9 The proofs of lemmas 8 and 9
merely consist of a simple, but long and tedious case-by-
case analysis. As an example, we prove the case where is
the ENF sentence y, z (P(y, z) v Q(y, v)).
Denote by

J the approximate structure suggested by the
lemmas, i.e., the approximate structure such that for any
d
y
, d
z
, d
v
D:

J(P(d
y
, d
z
)) =

I(z P(d
y
, z) v Q(d
y
, v));

J(P(d
y
, d
z
)) =

I(z P(d
y
, z) v Q(d
y
, v));

J(Q(d
y
, d
v
)) =

I(Q(d
y
, d
v
) z P(d
y
, z));

J(Q(d
y
, d
v
)) =

I(Q(d
y
, d
v
) (z P(d
y
, z)
v

(v

,= d
v
Q(d
y
, v

))));

J corresponds to

I on all symbols that do not occur in .
One can easily check that

J
p

I. From the discussion
above lemma 8, it follows that

J
p
O

I).
Assume O

I) is inconsistent. Then either

I is inconsis-
tent or one of the following is true:
for some d
y
, d
z
, d
v
D,

I(P(d
y
, d
z
)) = t and

I(Q(d
y
, d
v
)) = t;
for some d
y
, d
z
D,

I(P(d
y
, d
z
)) = t and

I(v Q(d
y
, v)) = t.
In all cases, also

J is inconsistent.
Now assume O

I) is consistent. It is sufcient to
prove that for any atom P(d
y
, d
z
), respectively Q(d
y
, d
v
)
such that

J(P(d
y
, d
z
)) =

J(P(d
y
, d
z
)) = f, re-
spectively

J(Q(d
y
, d
v
)) =

J(Q(d
y
, d
v
)) = f, also
O

I)(P(d
y
, d
z
)) = O

I)(P(d
y
, d
z
)) = f, respectively
O

I)(Q(d
y
, d
v
)) = O

I)(Q(d
y
, d
v
)) = f.
Assume that

J(P(d
y
, d
z
)) =

J(P(d
y
, d
z
)) = f.
Let M be a model of approximated by

I. We
will construct a model M

of approximated by

I
such that M(P(d
y
, d
z
)) ,= M(P(d
y
, d
z
)). Assume
M(P(d
y
, d
z
)) = t. Because

J(P(d
y
, d
z
)) = f, both

I(z P(d
y
, z)) = f and

I(v Q(d
y
, v)) = f. Hence,
there exists some d
v
such that

I(Q(d
y
, d
v
)) = f. Let
M

be the structure that corresponds to M, except that

it assigns M

(P(d
y
, d

z
)) = f for every d

z
D and
M

(Q(d
y
, d
v
)) = f. Clearly, M

is approximated by

I
and M

is a model of . Since we now have that both M

and M

belong to I [ I, I)
p

I and I [= , it holds
that O

I)(P(d
y
, d
z
)) = O

I)(P(d
y
, d
z
)) = f, as de-
sired.
The case where M(P(d
y
, d
z
)) = f is similar.
The case where

J(Q(d
y
, d
v
)) =

J(Q(d
y
, d
v
)) = f is
similar to the previous case.
Proof of lemma 11 The case where and have no pred-
icate symbols in common is trivial, as is the case where

I() = t.
To prove the case where

I() = t, assume

I() = t
and let

J be an arbitrary approximate structure such that

J
p

I. It is sufcient to prove that O

(O

(

J)) =
O

(

J). Indeed, then every renement for T above

I
can be transformed to a renement sequence for T

above

I by replacing each application of O

by an application
of O

followed by O

. To obtain a contradiction, assume

O

(O

(

J)) <
p
O

(

J). Then there exists a structure M
such that M, M)
p

J, M [= and M ,[= . Hence
M ,[= , which is a contradiction to

J() = t.
Proof of lemma 12 Observe that it is sufcient to prove
that O
T
(

I)
p
O
T

I), as it is obvious that O

T
(

I)
p
O
T

I) holds. Because

I
n

p

I, we have O
T
(

I)
p
O
T
(

I
n
) = O
T

(

I
n
). Since

I
n
)
0in
is a renement se-
quence for T

above

I, also O
T

(

I
n
) = O
T

I).
References
Ackermann, R. 1967. An Introduction to Many-Valued
Logics. London: Routledge and Kegan Paul.
Cort es-Calabuig, A.; Denecker, M.; Arieli, O.; and
Bruynooghe, M. 2006. Representation of partial knowl-
edge and query answering in locally complete databases. In
Hermann, M., and Voronkov, A., eds., LPAR, volume 4246
of Lecture Notes in Computer Science, 407421. Springer.
Cort es-Calabuig, A.; Denecker, M.; Arieli, O.; and
Bruynooghe, M. 2007. Approximate query answering
in locally closed databases. In Proc. 22nd National Con-
ference on Articial Intelligence (AAAI), 397402. AAAI
Press.
Enderton, H. B. 1972. A Mathematical Introduction To
Logic. Academic Press.
Fagin, R. 1974. Generalized rst-order spectra and
polynomial-time recognizable sets. Complexity of Com-
putation 7:4374.
Goubault, J. 1995. A bdd-based simplication and skolem-
ization procedure. Logic Journal of IGPL 3(6):827855.
Grahne, G., and Mendelzon, A. 1999. Tableau tech-
niques for querying information sources through global
schemas. In Proceedings of 7th International Conference
on Database Theory ICDT, volume 1540, 332347. LNCS.
Greco, S., and Molinaro, C. 2007. Querying and repair-
ing inconsistent databases under three-valued semantics. In
Dahl, V., and Niemel a, I., eds., ICLP, volume 4670 of Lec-
ture Notes in Computer Science, 149164. Springer.
Lenzerini, M. 2002. Data integration: A theoretical per-
spective. In Proceedings of the Twenty-rst ACM SIGACT-
SIGMOD-SIGART Symposium on Principles of Database
Systems, June 3-5, Madison, Wisconsin, USA, 233246.
Levesque, H. J. 1988. Logic and the complexity of reason-
ing. Journal of Philosophical Logic 17(4):355389.
Mari en, M.; Wittocx, J.; and Denecker, M. 2006. The IDP
framework for declarative problem solving. In Search and
Logic: Answer Set Programming and SAT, 1934.
Mitchell, D., and Ternovska, E. 2005. A framework for
representing and solving NP search problems. In AAAI05,
430435. AAAI Press/MIT Press.
Mitchell, D.; Ternovska, E.; Hach, F.; and Mohebali, R.
2006. Model expansion as a framework for modelling and
solving search problems. Technical Report TR2006-24, Si-
mon Fraser University.
Pawlak, Z. 1992. Rough Sets: Theoretical Aspects of Rea-
soning about Data. Norwell, MA, USA: Kluwer Academic
Publishers.
Schaerf, M., and Cadoli, M. 1995. Tractable reasoning via
approximation. Articial Intelligence 74(2):249310.
Selman, B., and Kautz, H. A. 1991. Knowledge compi-
lation using horn approximations. In National Conference
on Articial Intelligence, 904909.
Tseitin, G. S. 1968. On the complexity of derivation in
propositional calculus. In Slisenko, A. O., ed., Studies in
Constructive Mathematics and Mathematical Logic II, vol-
ume 8 of Seminars in Mathematics: Steklov Mathem. Inst.
New York: Consultants Bureau. 115125.
Ullman, J. D. 1988. Principles of Database and
Knowledge-Base Systems, Volume I. Computer Science
Press.
Wittocx, J.; Mari en, M.; and Denecker, M. 2008. Ground-
ing with bounds. In AAAI08. accepted.