* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download 485-291 - Wseas.us
Law of thought wikipedia , lookup
Jesús Mosterín wikipedia , lookup
Bayesian inference wikipedia , lookup
Propositional formula wikipedia , lookup
First-order logic wikipedia , lookup
Quantum logic wikipedia , lookup
Hyperreal number wikipedia , lookup
Curry–Howard correspondence wikipedia , lookup
List of first-order theories wikipedia , lookup
Non-standard analysis wikipedia , lookup
Mathematical logic wikipedia , lookup
Intuitionistic logic wikipedia , lookup
Interpretation (logic) wikipedia , lookup
Propositional calculus wikipedia , lookup
Quasi-set theory wikipedia , lookup
Model theory wikipedia , lookup
A Note on Expressive Power of Fixed-Point Logics GABOR SAGI Renyi Institute of Mathematics Budapest H-1053 Realtanoda u. 13-15 HUNGARY Abstract: Fix-Point logics play an important role in theoretical computer science, particularly in descriptive complexity theory. We investigate the expressive power FO(PFP) of First Order Logic endowed by Partial Fixed-Point Operations. We will show that on finite, unordered structures the expressive power of this logic is not the same as that of second order existential logic. In addition, we prove that the class of finite, unordered structures whose universe is of even cardinality cannot be axiomatized in FO(PFP), even if our language contains relation symbols with arbitrarily large arities. This result is a sharpening of Proposition 8.4.4 of [2]. Key-Words: Fixed-Point logics, second order logics, expressive power, cylindric algebras. 1 Introduction Fixed-Point Logics play an important role in theoretical computer science and in descriptive complexity theory because complexity classes can be characterized in terms of the expressive power of these logics. Such characterization results say that a set K of finite, ordered structures can be axiomatized in a certain Fixed-Point Logic if and only if the membership problem of K belongs to a certain complexity class (like PTIME, PSPACE, etc.). However, it is also interesting to investigate the expressive power of these logics on unordered structures. Let FO(PFP) denote First Order Logic extended with Partial Fixed-Point operations (this logic will be recalled later). Below, in Corollary 5 we show that on finite, unordered structures the expressive power of FO(PFP) is not equal with that of second order existential logic (cf. Corollary 7.5.9 of [2]). We also note that by Proposition 8.4.4 of [2] for a relational vocabulary containing unary relation symbols only, the expressive power of FO(PFP) is equal with that of usual first order logic. Therefore, if our vocabulary contains unary relation symbols only, then the class Keven of all finite structures whose universe is of even cardinality cannot be axiomatized in FO(PFP). As a strengthening of this result, in Corollary 5 we show that Keven cannot be axiomatized in FO(PFP) using any vocabulary (containing relation symbols with arbitrarily large arities). To prove the above, we will use the theory of Cylindric Set Algebras. Such an algebra can be regarded as the “meaning-algebra” of a given first order model. That is, a cylindric set algebra describes the complete algebraic structure of definable k-ary relations of the model (for a fixed natural number k). All the required details will be recalled below. For more on cylindric algebras we refer to [3] and [4]. Our another tool will be the theory of random structures, essentially in the form presented in Section 4 of [2]. The idea behind our results is that finite dimensional cylindric set algebras of random structures are isomorphic with each other with asymptotic probability 1 and these cylindric set algebras determine the action of the fixed-point operations as well. The main result of the paper is formulated in Corollary 5. Throughout ω denotes the set of natural numbers and for every n in ω we have n = {0,1,...,n-1}. Let A and B be sets. Then AB denotes the set of functions from A to B. In addition, P(A) denotes the power set of A, that is, P(A) consists of all subsets of A. 2 Construction and Proof We start by recalling the notion of cylindric set algebras. Definition. Let U be any set and let k ω. The full cylindric set algebra of dimension k with base set U is the algebra: P(kU); , -, CiU, Di,jU i,j k where and - are set theoretical intersection and complementation (w.r.t. kU), respectively, Di,jU = { s kU: si=sj} and CiU is defined as follows. For every X P(kU) one has CiU(X) = {s kU: ( z X) ( j k )j i sj=zj}. The class Csk of cylindric set algebras of dimension k is defined to be the class of (isomorphic copies of) all subalgebras of full cylindric set algebras. Thus, a cylindric set algebra is a Boolean algebra with further operations. In the above definition, the diagonal elements Di,j correspond to atomic formulas vi=vj while the “cylindrifications” Ci correspond to existential quantifiers “vi”. Roughly speaking, a cylindric set algebra is the algebra of first order definable relations of a given structure. More precisely, cylindric set algebras can be naturally associated to relational structures as follows. Let A be a structure and let k ω be fixed. Suppose φ(v0,...,vk-1) is a first order formula whose variables are belong to { v0,...,vk-1}. Now let || φ ||A = {s kA: A╞ φ [s]}. Consider the set D = { || φ ||A: φ is a first order formula containing variables at most v0,...,vk-1}. This is the collection of relations of A first order definable by k variables. D can be extended to a cylindric set algebra by stipulating that Di,j = || vi = vj||A, A - || φ || = || ┐φ||A, || φ ||A ||ψ ||A = || φ & ψ ||A, Ci || φ ||A = || vi φ ||A. This cylindric set algebra will be denoted by Csk(A). Clearly, Csk(A) can be generated by the set { ||R||A: R is an atomic formula in the language of A}. There is some similarity between cylindric set algebras and the notion of ”s-invariants” in [2] (Section 3.3.2). However, in our opinion, cylindric set algebras are much more natural and traditional. For more information on cylindric algebras and connection with logic we refer again to [3] and [4]. Throughout k denotes a natural number. Theorem 1. Suppose A is a relational structure such that Csk(A) is finite. Then there is a first order formula φk such that A ╞ φk and for any structure B one has B╞ φk Csk(A) Csk(B). Proof. We will construct a kind of diagram of Csk(A). For any a Csk(A) let φa be a first order formula such that a = || φa} ||A and if a = || R ||A for some atomic formula R then choose φa = R. φk is defined to be the conjunction of the following formulas: • ┐(v0 ... vk-1 φa(v0,...,vk-1) φb(v0,...,vk-1)) for every pair of different elements a,b Csk(A), • v0...vk-1( ┐φa(v0,...,vk-1) φb(v0,...,vk-1)) for every pair of elements a,b Csk(A) satisfying -a=b, • v0...vk-1 (vi φa(v0,...,vk-1) φb(v0,...,vk-1)) for every pair of elements a,b Csk(A) satisfying Ci(a) = b, • v0...vk-1 ( φa(v0,...,vk-1) & φb(v0,...,vk-1) φc(v0,...,vk-1)) for every elements a,b,c Csk(A) satisfying a b=c. Since Csk(A) is finite, the above conjunction is finite, so φk exists. Clearly, A╞ φk. Now let B be any structure with B ╞ φk. For every a Csk(A) let f(a) = || φa ||B. It is routine to check the following: • f is injective because of the first clause of the definition of φk • f is surjective because the range of f contains the set { || R ||^B: R is an atomic formula in the language of B} generating Csk(B), • f preserves the cylindric operations by the last 3 clauses of the definition of φk. Thus, f is an isomorphism between Csk(A) and Csk(B). Below we will use relatively standard methods from the theory of random structures. Throughout Gn denotes the random model, in which the truth of atomic formulas on tuples are selected independently, each of them with probability 1/2. In more detail, such a random structure on n (labelled) elements can be “constructed” as follows. For each k-ary relation symbol R and k-tuple a let us associate a random variable ξR,a taking values 0 and 1 with probabilities 1/2. We assume that these random variables are completely independent from each other. Observe the values of these random variables, and declare A ╞ R(a) iff ξR, a=1. The resulting structure A is called a random structure. For convenience, we assume that the universes of our random structures are of the form n ={ 0,1,2,... } for some ω n. So for fixed n, our probability space is the set of all structures on n (labelled) elements (in other words, elementary events are particular structures on n elements). If φ is a first order sentence in the appropriate language, then Pr(Gn ╞ φ) denotes the probability, that an above described randomly chosen structure (on n elements) is a model of φ. By a classical result there is a structure such that a random structure on countably infinite elements is isomorphic to with probability 1. is called “the random structure” and can be characterized as the “universal countable structure”. Among others, it has the following remarkably nice properties. The first order theory of can be axiomatized by a very concrete set of axioms, such a is explicitly given in [2], Example 3.2.11. Here we don't recall the concrete form of these axioms because we need only some properties of them. Namely, forms a complete theory and, in addition, the asymptotic probability of each element of is 1, that is, if then limn Pr (Gn ╞ ) = 1. This can be found in Lemma 4.1.2 of [2]. Moreover, is an ω-categorical structure which has quantifier elimination (this can be proved similarly to the remark after Lemma 6.4.3 of [5], page 177). For more details on random graphs and random structures we refer to [1] and [2]. Let us fix an arbitrary relational vocabulary throughout the paper. The next theorem says that the cylindric set algebras of the countable random structure “can be approximated” by cylindric set algebras of finite structures, whatever the fixed vocabulary is. Theorem 2. For every k for every large enough n there is a finite structure Hk,n such that Hk,n has exactly n elements and Csk(Hk,n) Csk( ). Proof. As we mentioned before, the first order theory of has quantifier elimination. Therefore elements of Csk( ) can be identified by quantifier free formulas having at most k many variables. Since up to logical equivalence there are finitely many such formulas, Csk() is finite. By Theorem 1, there is a formula φk such that for any structure B we have B╞ φk Csk(B) Csk() and ╞ φk. Recall that is a complete theory and ╞ so ╞φk. By compactness, there is a finite 0 such that 0 ╞ φk. By Lemma 4.1.2 of [2] the asymptotic probabilities of elements of are 1. Since 0 is finite, this yields limn Pr(Gn ╞ 0 ) = 1. It follows that there is a natural number N such that for every n > N one has Pr(Gn ╞ 0) > 0. Therefore, for each n > N there exists a structure Hk,n such that Hk,n has exactly n elements and Hk,n╞ 0. But then, by construction, Hk,n ╞ φk so by Theorem 1, Csk(Hk,n) Csk( ), as desired. By Theorem 2 for each k there are structures Fk and Gk such that Csk(Fk) Csk() Csk(Gk) and in addition, the number of elements of Fk is even while the number of elements of Gk is odd. Let K0 = {Fk: k } and let K1 = {Gk: k }. Definition. Let L be any logic (one can think to an extension of usual first order logic) and let C0 and C1 be classes of structures. We say that C0 can be separated from C1 in L iff there is a formula of L such that C0╞ and C1╞ ┐. Note that the above notion of separation is not symmetric because L need not be closed under negation of formulas. Theorem 3. K0 can be separated from K1 in second order existential logic. Proof. Let Φ be a second order existential formula expressing that “there is an equivalence relation such that every equivalence class has exactly 2 elements”. Clearly, K0 ╞ Φ and K1╞ ┐Φ. Now we turn to show that K0 and K1 cannot be separated by formulas of certain fixed-point logics. Below we recall from [2] the notion of Partial Fixed-Point Logic FO(PFP) which is an extension of First Order Logic. We start by semantics. Let A be a fixed structure. Suppose φ(R) is a formula in which R is a relation variable. One can associate an operation Fφ to φ by stipulating Fφ(R) = {s kA: A╞ φ(R) [s]}. Consider the sequence (*) , Fφ(), Fφ(Fφ()), ... and let fpR(φ) be the first element of the sequence which is equal with its successor (let fpR(φ) = if there are no such term in the above sequence). Now the formulas of FO(PFP) are those of first order logic together with stipulating that if φ is a formula and R is a relation variable then FPR(φ) is also a formula. The meaning (or interpretation) of such a formula in a given model A is || FPR(φ) ||A = fpR( || φ ||A). Observe that the terms of the sequence (*) are always belong to Csk(A) if φ contains at most k many individual variables. To prove this, it is enough to show (**) if x Csk(A) then Fφ(x) Csk(A). This can be done by induction on the number l of fixed-point operations in φ. If φ doesn't contain fixed-point operations then Fφ(x) is the result of a term function of Csk(A) (this term function can be obtained from φ by a straightforward way). So in this case (**) holds. Now suppose (**) holds for φ which contains l many fixed-point operations. Then, by induction, every member of (*) belongs to Csk(A) so || fpR(φ) ||A Csk(A) as well. Since Csk(A) is closed under ┐, & and under cylindrifications, it follows that (**) holds for any formula with at most l+1 many fixed-point operations. This completes the induction. Thus, we have proved that for each formula φ of FO(PFP) every member of (*) belongs to Csk(A). It follows that for each formula φ of FO(PFP) and for each finite structure A there is a first order formula t(φ) such that || φ ||A = || t(φ) ||A. Of course here t(φ) depends on A as well. Suppose φ contains at most k many individual variables only. Then one can show by a similar easy induction on the complexity of φ that if ||φ||A = || t(φ) ) ||A and Csk(A) Csk(B) then ||φ||B = || t(φ) ||B. Particularly, if Csk( A) Csk(B) and A ╞ φ then B ╞ φ. Theorem 4. K0 cannot be separated from K1 in FO(PFP). Proof. Suppose φ is a formula of FO(PFP) such that K0 ╞ φ. Let k be bigger than the number of different individual variables occurring in φ. Then on one hand, Fk╞ φ. On the other hand, Csk(Fk) Csk() Csk(Gk) so Gk ╞ φ. Thus φ doesn't separate K0 from K1. This completes the proof. Now we can prove the main result of the paper. As we mentioned, it is related to Corollary 7.5.9 of [2] and generalizes Proposition 8.4.4 of [2] to vocabularies containing relation symbols with arbitrarily large arities. Corollary 5. (1) The expressive power of FO(PFP) on finite unordered structures is not the same as that of second order existential logic. (2) The class Keven of all finite structures whose universe is of even cardinality cannot be axiomatized in FO(PFP) for any relational vocabulary. Proof. Immediate from Theorems 3 and 4 because Keven separates K0 from K1. Acknowledgment. Research supported by Hungarian National Foundation for Scientific Research grants D042177 and T035192. References: [1] B. Bollobás, Random Graphs, Academic Press Inc, London, (1985). [2] H.D. Ebbinghaus, J. Flum, Finite Model Theory, Springer-Verlag, (1999). [3] L. Henkin, J. D. Monk, A. Tarski, Cylindric Algebras Part 1, North--Holland, Amsterdam (1971). [4] L. Henkin, J. D. Monk, A. Tarski, Cylindric Algebras Part 2, North--Holland, Amsterdam (1985). [5] W. Hodges, A shorter model theory, Cambridge University Press, (1997).