Professional Documents
Culture Documents
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
Springer is collaborating with JSTOR to digitize, preserve and extend access to Synthese
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
MERLIJN SEVENSTER
1. INTRODUCTION
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
258 MERLIJN SEVENSTER
x = x\,... ,xn and y = y\,... ,yn. The syntax of IF logic allows for
independence, in that the variable y? is meant to be independent of
the variables in Yi9 although it appears under their syntactical scope.
Hintikka and Sandu (1997) propose game theory as verification
al framework for IF logic. That is, with every IF formula and suit
able model they associate a semantical evaluation game. These game
typically have the eye-catching property of being of imperfect infor
mation. Truth of an IF formula is defined in terms of a strategy that
wins the evaluation game, despite the imperfect information. From a
mathematical perspective, this notion of (uniform) winning strategy
resembles witnessing Skolem functions. In this paper we will mainly
ignore the game theoretical content, fascinating as it may be, and
adopt the mathematical terminology.
That is, evaluation of an IF formula <t> takes place relative to a
model (D, /), consisting of a domain and an interpretation func
tion. Let O be as in (1), then it is true on (D, /) iff
(D,I*)\=3fl...3fnVxi...Vxn l?(x,f),
where f?: DZi -> D and Zi = {x\,... , x?}\Yi. The functions f\,... , fn
are called Skolem functions and are given their interpretation by the
function /*, that extends /. I.e., / and /* assign every first-order
variable the same object, and /* also assigns functions to f\, ... , fn.
IF logic may be the common ground, with different disciplines
come variations on this theme, though. In natural language seman
tics, for instance, after a lively debate Barwise's (1979) account was
generally accepted, that held that some natural language sentences
had best be analyzed by means of branching quantifiers. In computer
science, Bradfield and Fr?schle (2002) introduced an independence
friendly modal logic to capture concurrency. In game theory, van
Benthem (2000) mentions that IF logic may be used as a calculus
for games of imperfect information. For the study of many-valued
logics, Sandu and Pietarinen (2001, 2003) introduce an IF proposi
tional logic. They show that adding an independence friendly con
nective, transjunction, to propositional logic yields a logic that is
functionally complete for three-valued functions.
In the present paper we concern ourselves with another IF prop
ositional logic, as introduced in Sandu and Pietarinen (2001). For a
change, our aim will not be to show that the IF variant is richer in a
revolutionary way - as in describing natural language sentences with
branching quantifiers or capturing three-valued logics - but rather
to reinvestigate the old notions. In particular, we will compare the
[4]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 259
[5]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
260 MERLIJN SEVENSTER
2. RELATED WORK
[6]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 261
(D, /*) f= V*u ... Vxhh ... V*?fi... V^,? 0(x, g),
for some suitable first-order formula <I>3_coi and the model MG cap
turing G. The authors proceed by investigating the computational
consequences of putting restrictions on the variables in the Henkin
quantifiers. Since the variables y\, y2 are supposed to be assigned
one of the three colors it comes as no surprise, that also
[7]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
262 MERLIJN SEVENSTER
quantifiers over what we call indices, that is, variables that range
over a binary domain: I and r. If 3 (or V) quantifies an index /,
rather than a standard variable, we write V (or /\) instead, and say
that / is disjunctively (or conjunctively) quantified. Using this nota
tion, Henkin quantifiers of the form
(Vxi V*i\
\Vx2Vi2)
cannot express NP-properties, unless coNL = NP.5
Vxi(3yi/Yl)...Vxn(lyn/Yn)K(x,y),
where Y? are subsets of {x\,... ,x?}.
To appreciate the difference between QBF and IFQBF con
sider the computation tree of QBF instances. In the case of O =
Vx3y (x**y), we find its depiction in Figure 1. In a computation
tree, every node corresponds to a (partial) assignment. In partic
ular, the leafs are complete assignments. We label them true, if
the corresponding complete assignment satisfies the propositional
k; otherwise, we label them false. The labels of the leafs can be
propagated all the way up the tree to the initial node, by means
of a backwards induction algorithm, well-known from game the
ory, cf. Osborne and Rubinstein (1994) and van Benthem (2000).
[8]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 263
\/x3y (x ^+ y)
Vx(3y/{x}) (x <r+ y)
[9]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
264 MERLIJN SEVENSTER
where / g {V> A} and ^ is a function mapping {I, r}1 onto the set
(<2,-"<2 I Q^v)- So negations only appear directly in front of a
predicate. P will be called a predicate pointer. ?\y(o) is obtained
from ?(ct), by applying the following procedure a finite number
of times to any 0 g ?(<?):
If \? js occurs in <j> and Js ? {iu ... , is}9 replace it by (\J js/Js).
? and ?if are not defined recursively. We like to point out that a
recursive version is easily provided for ?. As to ?if, and IF lan
guages in general, one might wonder what is the value of a recur
sively defined IF language, as its semantics (given below) is not
recursive. This question touches on the discussion whether or not it
is possible to give a (natural) compositional semantics for IF logic, a
discussion we are happy to avoid in this writing. We refer the inter
ested reader to Hodges (2001) and Sandu and Hintikka (2001).
We will not set out the usual game theoretical apparatus to define
the semantics for ?tf- Instead, we stick to defining truth in terms
of Skolem functions. In order to do so, we introduce the Skolemi
zation of a formula 0 g ?if; (</>)Skolem is obtained by applying the
following sequence of steps to every disjunctive quantifier in (p:
- remove (VisM)
- put 3fs in front
- replace js in 0's predicate pointer with fs(is), where is is the string
of conjunctively quantified variables containing it iff js occurs in
the syntactical scope of it and it g Js.
[10]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 265
Note that the order in which these steps are applied is immaterial,
due to the semantics that will follow in due course.
For the sake of illustration, observe that (/\i\/ j P(i, 7))Skolem =
3f /\i P(i, /(/)). Here, j is dependent on i. This is not the case in
(f\i(\/ j/{i}) P(i, y))Skolem = 3f/\i P(i, f). Recall that formulae in
?if are evaluated with respect to {l,r}. So in the latter formula /
is a constant function, either assigning ? or r. As such, / functions
to the effect of an index, hence the formula can equivalently be put
asViA? P(iJ)
DEFINITION 2 (Semantics ? and ?IF). Consider ?iF(cr) and a
suitable interpretation function /. Then, for every 0g?if(oO
ni]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
266 MERLIJN SEVENSTER
(J P(c,f).
ce{l,ry
LEMMA 4. ({?, r}, /*) \= 3/i... 3fn /\ ix... f\ in P(i, f) iff for some
witness g = gi,... , gn of the variables f\, ... , fn9 it is the case that
for every QePg,I(Q)=true; and for every ->QePg9 I(Q)=false.
Proof of Theorem 3.
Membership. Consider 0 g ?if and a suitable interpretation function
/. It is convenient and harmless to assume that 0 be of the follow
ing form
[12]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 267
P (\ i r>=? A ? MG\=?2-co\(Vri^,Vrh-i,Vrhj^,VrhJ^)
/-3-COHI1.I2,? |_A if MGfc?3- l(Vrir,Vri2,,VrJ[j2.,Vrhj4n).
We claim that (4) is true iff {{l,r},I) |= 0, where / is such that
1(A) = true. The point is that (4) is true iff there is a coloring F:
V -> {1, 2, 3} iff there are two Skolem functions f\, fi mimicking F
iff the power set P/-,,/2 contains only As and no -"As.
[13]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
268 MERLIJN SEVENSTER
[14]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 269
QT=AQ
(-,<2)T=^4Q
[15]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
270 MERLIJN SEVENSTER
/\h/\h(\/j/{h})/\h p&n>
that is true on a model (D, I) iff (Z), /*) |= 3f Ah Ah Ah P(l
f(?2?). Now, we can push \J j to the right, so that it appears at
the right-hand side of /\?3. We must be careful here, not to make j
dependent of 13. This we prevent by adding i3 to the set {i\}9 yield
ing the ?if formula Ah Ah Ah (V j/{iu h}) P(lj). Again, this
formula is true on (D, I) iff (D, I*)\=3f Ah Ah Ah P(l f(h))
[16]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 271
For future reference, we define the notion of a clause for CNF for
mulae of ?if- In propositional logic a clause is disjunction of liter
als. In ?if, we rather consider clauses as multisets of literals. For let
0 g ?if be in CNF as in (8), then for every c g {I, r}1 let the multiset
Ocn be the rc~1th clause of 0, where
[17]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
272 MERLIJN SEVENSTER
Proof
Membership. Fix a formula 0 g?if- Obviously, the number of atoms
in 0 is smaller than |0|, since it can impossibly exceed the size of the
extensive encoding of 0's predicate pointer. Therefore, one can non
deterministically guess the interpretation function I in polynomial
time. This leaves us with the problem of deciding whether I satisfies
0, i.e., a model checking problem. Theorem 3 shows how to do this
in non-deterministic polynomial time.
Hardness. Since ? is contained in ?if, it suffices to reduce
CNF4-SAT to ?'s satisfiability problem. This is easy, recalling our
discussion on CNF for ? and ?if from the previous section. We
showed, namely, that every instance k of CNF4-SAT has an equiva
lent ? formula 0. 0 can be obtained from k in polynomial time. D
Proof
Membership. Fix 0g?if. 0 is not a validity iff there exists an inter
pretation / that does not make it true. The latter proposition can
be checked by first non-deterministically guessing an interpretation
/ that will falsify 0, if any. Next do the model checking with respect
to / and 0 doable in NP, due to Theorem 3.
Hardness. We rather give a proofsketch, as the details are lengthy
and tedious. We reduce from the problem QSAT2, that is, given a for
[18]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 273
[19]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
274 MERLIJN SEVENSTER
/\jl/\jl(\/j3/{jlj2})(\/j4/{jlj2})P(l?)
and P is defined in such a way that ?rP and x(0 are satisfiability
equivalent.10 If clause Drp is of the form (L *>--//), x(i) be speci
fied as in Figure 3.
It suffices to prove that Drp and x (0 are satisfiability equivalent.
This can be evaluated by checking all of x(0's power sets, due to
Figure 3. Depiction of x(i)> in case ?'s clause Z)rp is of the form (L^>-^L').
A is a new atom not occurring outside /(i).
[20]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 275
Lemma 4. Per string cg {I, r}\ there are four different pairs of Sko
lem functions fo, f\ that determine the values of 73, 74, respectively.
Let f? be the function such that f?(i) ? d9 for d e{l,r} and m e
{3,4}. The power sets read as follows:
PftJt(i) = {-A,^A,L9^L'}
PftJ;(i) = {-A,^A,^L,L}
Pfrj?(i) = {^A,Lr,^L',-^A}
P/?t/?(i) = {L,,-A,-A,-L}.
Obviously, any interpretation that makes all literals in Pftfe(?) and
Pfrjr(i) true also makes Drp true. Furthermore, no interpretation
can satisfy all literals in the other two power sets.
If Drp is of the form Kv--Mv^iV), x(\) works to the effect
of the tree depicted in Figure 4. In this case, it is easy to see that
Drp and x(i) are equivalent, since no new atoms are introduced.
Remains to note that whatever form a clause in ? has, every
clause in the reduced ?tf formula 0 contains at most one positive
literal. Hence, 0 is an instance of HORN-?if-SAT, that is satisfi
ability equivalent to ?.
-nL -^M -.JV -iTV -.L -.M -.N -.AT -.L --M -.JV -.N -.L ^M ^N -^N
Figure 4. Depiction of x (i), in case ?'s clause Drp is of the form (->L v ->M
[21]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
276 MERLIJN SEVENSTER
where K does not occur in k'. For the new atoms Q\, 02, the
formula
(Kv(QiaQ2)) or (KvLvMvN).
We argue that there exists a CNF formula in ?if in which every
atom appears only twice, that is satisfiability equivalent to X. Our
argumentation proceeds along the lines of the proof of Theorem 8,
as we consider the formula
[22]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 277
x(i)
Figure 5. Depiction of x(i)> m case ^'s clause CrP is of the form (Kv(Q\ a
g2)). A, A! are new atoms, not occurring outside xO)- Note that any pair of Sko
lem functions fi, f, where f3(\) = r, cannot be a witness, as it would result in a
power set containing both A and ->A or A' and ->A'.
x(i)
[23]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
278 MERLIJN SEVENSTER
/\h-.-f\h(\/j/j) P(U),
where J = {j\,... , jm} ? {ix,... , /?}. We can manipulate the order of
the /\s in 0, without affecting its truth-condition. In essence this is
the same trick, now applied conversely, as transforming every for
mula of ?if into an CNF equivalent. So the formula
l\i\-f\i'kl\h-.-/\jm{\/j/j) Pd',lj),
is equivalent to 0, where i[,... , *? = i' are the indices in the comple
ment of JU{j}. Now, we push the disjunctive quantifier to the left
until we meet /\/? and remove /:
do Ai'?---A?'?V;'A?---A?? win,
being a formula without independencies! The formula (10) is satisfi
able iff there exists an / such that for every cg {I, r}1', there exists a
/(c) G {I, r}9 such that no matter what de {?, r}J we pick, it is always
the case that /(P(c, d, f(c))) = true. We show that whether such /
exists can be cast as an instance k of CNF2-SAT, as follows: Intro
duce two new atoms Ac?9 Acr per ce{l,r}1'. Eventually, if k is satis
fied by an interpretation /, / can be read of from it in the following
way:
t if I(Aci) = true
/(c) = r if I(Aci) = false and I(Acr) = true
undefined otherwise.
[24]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 279
Proof
Membership. Fix a formula k of propositional logic. Due to n, the
number of atoms in k is less than 1(\k\). Now, collect these atoms
and do the model checking with respect to every different interpre
tation that can be defined over them. There are less than 2/(|/c|) =
?(\k\c) different interpretations for some constant c, hence, this
computation can be performed in a polynomial number of steps
with respect to \k\.
[25]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
280 MERLIJN SEVENSTER
Proof.
Membership. Follows directly from the membership-part of the proof
of Theorem 6.
Hardness. Satisfiability requires model checking. The hardness
part of the proof of Theorem 3 can be copied here, as it introduces
only one atom. D
7. CONCLUSION
We summarize our results below. Note that all problems are pin
pointed up to completeness, where appropriate.
[26]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 281
ACKNOWLEDGEMENTS
I thank Peter van Emde Boas for his continuing support, guiding
remarks and insights. I thank the anonymous referees and Francien
Dechesne for their comments, that improved the paper considerably.
I thank the organizers of the Knowledge and Games workshop,
for hosting my virtual presentation. I thank Victor de Boer for edit
ing my video presentation.
NOTES
NLcpc ^ ccoNPNP
- - coNP ~
[27]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
282 MERLIJN SEVENSTER
REFERENCES
[28]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms
INDEPENDENCE IN PROPOSITIONAL LOGIC 283
ILLC
Universiteit van Amsterdam
UK
E-mail: sevenstr@science.uva.nl
[29]
This content downloaded from 131.100.243.254 on Sat, 17 Sep 2016 16:00:22 UTC
All use subject to http://about.jstor.org/terms