Complexity of Logics with Bounded Modalities

Robin Hirsch, Evan Tzanis February 19, 2012

Abstract For n N, we dene the following modal operators: =n , n , n , , with intended meanings at every point n steps away, at every point within n steps, at every point at least n steps away, at every point from now on, and at every point, respectively. L( =n , : n N) is the modal language of modal propositional formulas built using and any =n , for n N, for example. Modal logics using these operators can be very concise, because the superscripts of the modal operators are binary encoded. We consider the complexity of the satisability problems over arbitrary Kripke frames and over linear discrete frames. We consider a number of variants of PDL where binary encoded program iterators =n , n are allowed, as well as program operators , , ; , for union, intersection, composition and reexive transitive closure, and consider the complexity of the satisability problem. We have obtained the following results. Language (Range) Arbitrary Frames L( =n , n : n N) L( =n , : n N) L( =n , n , n , : n N) L( =n , : n N) L( =n , n , n , : n N) Frame (N, <) L( =n , : n N) L( =n , n , n , : n N) Dynamic Logics P(, , n : n N) P 1 (=n , n : n N) P(, , ; , =n , n : n N) P(, =n , : n N) P(, ; , =n , n , n , : n N) P(, , ; , ) P(, , ; , =n , n , n , : n N) Complexity EXPSPACE-complete 2-EXPTIME-complete 2-EXPTIME-complete EXPSPACE-complete EXPSPACE-complete EXPSPACE-complete 2-EXPTIME-complete 2-EXPTIME-complete

Keywords: Modal Logic, Complexity, Satisability Problem, Alternating Time Automata.

Introduction

The topic of this article is the complexity of various modal logics with bounded modalities. These bounded modalities can express, for example, that a property will continue to hold at all points within n steps. Since we are interested in the complexity of such logics, it is convenient to represent this bound n as a binary number. This can make our formulas very concise, so we might expect a computational cost incurred by the use of these binary encoded bounds. This turns out to be the case, the logics we consider have fairly high complexities, and we will provide a detailed analysis. For each modal logic dealt with here we can dene a Kripke-like semantics and consider the complexity of the satisability problem for the logic over arbitrary frames. We will also consider the case of discrete, linear (non-branching) semantics. The use of binary encoded bounds on modalities is not new. In this section we will briey review the syntax and semantics of a number of well-known modal and branching logics which adopt this kind of binary encoded bounds, and consider the complexity of the satisability problem for each of them. In the following section we will introduce new modal logics which use bounded

unary modalities, making them in this respect more simple than some of the previously considered logics. We will establish the complexity of their satisfaction problems. Figure 1 summarises the main known languages and their complexities, but may need some illumination. First we explain the syntax of these logics. In each case, except for the nal two, a Name Modal Modal Universal CTL RCTL RCTL RCTL= PLTL TPTL Atoms Prop Prop Prop Prop Prop Prop Prop Prop, 1 2 , 1 d 2 Prop Prop Unary C. Binary Connectives State Based Semantics Complexity PSPACE-C EXPTIME-C EU, AU EXPTIME-C EU, AU, EUn , AUn EXPTIME-C EU, AU, EUn , AUn EXPTIME-C n n EU , AU EU, AU, EU=n , AU=n 2-EXPTIME-C Linear Path Based Semantics U PSPACE-C U EXPSPACE-C Citation [L77] [Hem96] [EMH82] [EMSS92] [EMSS92] [EMSS92] [SC82] [AH94]

, EX, AX EX, AX EX, AX EX, AX X x. X

CTL CTL

E, X E, X

State and Path Based Semantics U 2-EXPTIME-C [JL03] U 2-EXPTIME-C [VS85, EMJ86]

Figure 1: Summary of various modal, branching logics, their semantics and the complexity of their satisfaction problems. formula is dened recursively by := atom||(1 2 )|U n()|Bin(1 , 2 ) where atom is the set of atoms, which always includes a countable set Prop of propositions, U n is any of the listed unary connectives and Bin is any of the listed binary connectives. The nal cases, CTL+ and CTL are two-sorted logics, with state formulas and path formulas , dened by: CTL+ = Prop||(1 2 )|E = Prop||(1 2 )|X|U(1 , 2 ) CTL = Prop||(1 2 )|E = ||(1 2 )|X|U(1 , 2 )

We use standard abbreviations: F = U(, ), and G = F, . For TPTL there is more than one kind of atom. This language has terms = c|x + c where x is a variable and c is a binary string. An atom is either a proposition p Prop, an inequality 1 2 or a congruence 1 d 2 where d is a non-zero binary string. The intended meaning of 1 d 2 is that 1 is congruent to 2 modulo d. The binary strings c, d represent natural numbers. The superscripts of EUn , EUn and EU=n are any natural numbers written in binary. This means that the formulas of these languages can be very short. Thus, for example |EU=n (1 , 2 )| = 1 + log2 n + |1 | + |2 | i.e. we charge 1 for the modality EU= , log2 n for the binary representation of the bound n plus the lengths of formulas 1 and 2 . Note that throughout the paper we do not count brackets and commas when calculating the length of a formula. We will use standard abbreviations = , = , 2

etc. For the semantics, state based formulas are evaluated at points w in Kripke structures S = (W, R, V ). Here W is a set of worlds, R W W is an accessibility relation and V : Props (W ) is a valuation assigning a set V (p) of worlds to each proposition p. We write w to denote a countably innite sequence from W , and it is implicit that w = (w0 , w1 , . . .) W . If (wi , wi+1 ) R, for all i < , then we call w and R-path and we may write (w). To evaluate a formula, S, w S, w S, w S, w S, w S, w S, w |= p |= |= (1 2 ) |= |= |= EX() |= EU(1 , 2 ) w V (p) S, w |= S, w |= 1 &S, w |= 2 w W ((w, w ) R S, w |= ) w W S, w |= w [(w) (w = w0 ) S, w1 |= ] n N, w [(w) (w = w0 ) S, wn |= 2 i<n S, wi |= 1 ] w[((w) w0 = w) n N, S, wn |= 2 j<n S, wj |= 1 ] k N, k n, [(w) (w0 = w) w S, wk |= 2 i<k S, wi |= 1 ] w[((w) w0 = w) k N, k n, S, wn |= 2 i<k S, wi |= 1 ] w [(w) (w= w) 0 S, wn |= 2 i<n S, wi |= 1 ] w[((w) w0 = w) S, wn |= 2 i<n S, wi |= 1 ] k N, k n, [(w) (w0 = w) w S, wk |= 2 i<k S, wi |= 1 ] w[((w) w0 = w) k N, k n, S, wn |= 2 i<k S, wi |= 1 ]

S, w |= AU(1 , 2 ) S, w |= EUn (1 , 2 ) S, w |= AUn (1 , 2 ) S, w |= EU=n (1 , 2 ) S, w |= AU=n (1 , 2 ) S, w |= EUn (1 , 2 ) S, w |= AUn (1 , 2 )

Note that EU is treated as a single binary connective and EU(1 , 2 ) means there is a branch and along that branch 1 is true until 2 . We abbreviate AX = EX. With path based semantics, we can still use a Kripke structure S = (W, R, V ) but we evaluate a formula at an R-path w. For i N we write wi for the R-path (wi , wi+1 , . . .). PLTL formulas are evaluated by S, w |= p S, w |= S, w |= S, w |= X S, w |= U i i i i i w0 V (p) S, w |= S, w |= and S, w |= S, w1 |= j(S, wj |= and k (k < j (S, wk |= )))

The semantics of TPTL are slightly more complicated because we have variables and terms. TPTL is interpreted over linear, discrete models: (N, <, V, T ) where N is the natural numbers, < is the accessibility relation, V is a valuation, V : Prop (N), and T is an assignment that maps constants to themselves (recall that constants are binary encodings of natural numbers) and variables to natural numbers, clearly T can then be extended to all terms. Given a model S = (N, <, V ), an assignment T and an index i N we evaluate formulas as follows: S, T, i |= p S, T, i |= 1 2 S, T, i |= 1 d 2 S, T, i |= i i i i i V (p) T (1 ) T (2 ) T (1 ) = T (2 ) S, T, i |= 3

(mod d)

S, T, i |= S, T, i |= X S, T, i |= U S, T, i |= x.

i i i i

S, T, i |= and S, T, i |= S, T, i + 1 |= j (S, T, j |= and k, (k < j (S, T, k |= ))) S, T [x := i], i |=

where T [x := i] denotes the assignment that agrees with the assignment T on all variables except x, and maps x to i N. Finally, CTL+ and CTL have semantics for states and paths, so the semantics is based on mutual recursion. Let S = (W, R, V ) be a Kripke structure, let w W and let w be an R-path. Let 1 , 2 be state formulas and let be a path formula. Then S, w |= E i there is a R-path w = (w0 , w1 , . . .) such that w0 = w and S, w |= . And S, w |= U(1 , 2 ) i there is k N such that S, wk |= 2 and for all j with j < k we have S, wj |= 1 . It is clear from gure 1 that the more expressive the modal logic is, the higher the complexity of the satisfaction problem for the logic is, but also a logic with the same expressive power as another but with a more concise notation tends to have a higher complexity. It seems that there are several dierent factors which might potentially be contributing to the higher complexity. There is the use of the universal modality, the use of path based semantics and path quantiers, there is the use of the Until connective, which is strictly more expressive than and there is the use of a binary encoding for the restricted Until connectives U=n and Un . It would be helpful to separate out these issues in order to analyse the source of the higher complexity. The use of binary encodings for restricting modalities and the eect of this on the complexity of the satisfaction problem is the main focus of this paper. To this end, we concentrate on state based Kripke semantics and we use only unary modalities for the logics we dene. The outline of this article is as follows. We start by dening the syntax of some new state based modal logics which use a binary encoding for superscripts to . In section 3 we prove some easily derived upper bounds regarding the complexity of their satisability problems. In section 4.1 we give the denition and summarise the complexity results for Alternating Turing Machines and in section 4.2 we use our modal languages to describe the computations of Alternating Turing Machines thereby establishing lower complexity bounds for our logics. Complexity results for the modal logics introduced in this paper are proved in section 4.3. In section 5 we focus on linear and discrete structures and we prove that the satisability problem of a language between L( =n , : n N) and L( =n , n , n , : n N) are EXPSPACE-complete. In section 6 we establish the complexity of the satisability problem for variants of PDL where binary encoded iterations of programs are permitted.

Binary Encoded Modal Logics

In this section we introduce binary encoded bounded modalities and three modal logics logics which adopt them. Definition 1 (Languages) Given a sequence (i : i I) where each i is a unary modal operator (e.g. , ) we let L(i : i I) be the language dened by = p | | (1 2 ) | i () : i I where p is an arbitrary propositional letter. For n N, when we write n our convention is that n is written using the binary notation, so that | n | = 1 + log2 n + ||. The languages we consider in this paper use the following unary modal operators: { , =n , n , n , , : n N}. We use standard abbreviations , , etc. dened by ( ) = ( ), ( ) = ( ), = , n = n , =n = =n , n = n , = and = .

Definition 2 (Models) A structure S for any of our languages is a Kripke structure S = (W, R, V ) such that W is a non-empty set of possible worlds, V : P rops (W ) assigns a set of worlds to each proposition and R is a binary relation on W . For n N we dene the binary relation Rn over W inductively. Let xR0 y hold if and only if x = y. For n > 0 let xRn y hold if and only if there is z such that xRn1 z and zRy. The binary relations R , Rn and Rn W are dened to be R = n< Rn , Rk = kn< Rn and Rk = nk Rn . Formulas of these logics can be evaluated as follows: S, x |= p S, x |= S, x |= S, x |= S, x |= n S, x |= n S, x |= =n S, x |= S, x |= i i i i i i i i i x V (p) S, x |= S, x |= and S, x |= v W (xRy S, y |= ) y W, (xRn y S, y |= ) y W, (xRn y S, y |= ) y W (xRn y S, y |= ) y W, (xR y S, y |= ) y W, S, y |=

If S is a structure, is a formula and S, x |= then (S, x) is a model of . If S, x |= for all x W and all structures S then we say that is valid. Dually, if there is some x in some model S such that S, x |= then is satisable. Given any two formulas , we write if ( ) is valid. The table of complexities in the abstract summarises the complexity of the satisfaction problem for these languages. The results for arbitrary Kripke frames are all established in theorems 8 and 18 and may usefully be compared to those in gure 1. The slightly unusual result is the EXPSPACE-complete complexity of L( =n , n : n N)-SAT which does not correspond to any of the known results in gure 1.

Upper Bounds

Next, we quote some well known complexity results for the satisability problem for some elementary modal logics. We use these results to prove natural upper bounds to the satisability problems for sublanguages of L( =n , n , n , , : n N). Proposition 3 1. The satisability problem for L( ) is PSPACE-complete [L77]. 2. The satisability problem for L( ,

) is EXPTIME-complete [HM92, theorem 5.1].

3. The satisability problem for L( , ) is EXPTIME-hard (by theorem 16, or the proof of [Hem96, theorem 5.1]). The following theorem seems to be very well known, but we had diculty in nding a published proof. Theorem 4 The satisability problem for L( , ) is in EXPTIME. PROOF: { , , , } we have and Let a formula of L( , ). For . Hence, in linear time, we can replace by an equivalent formula , built from propositions using negations, disjunctions, conjunctions, , s or , s, where 5

negations only occur immediately above propositions and there are no modal operators above any occurrence of or . We check the satisability of by ltration over { , }-free subformulas of . Let X be the set of subformulas of not involving or and let Y by the closure of X under single negations. A maximally consistent set (MCS) is a subset S of Y such that exactly one of and belongs to S, for each X, if the conjunction is in S then so are and , and if the disjunction is in S then so is either or . If S, T are MCSs let (S, T ) R i for all formulas S we have T . In exponential time, we can replace by an equivalent , a disjunction of conjunctions of formulas of the form , or , where , , are subformulas of not involving or . Conjunctions (1 ) (2 ) can be replaced by the equivalent (1 2 ). To check the consistency of it suce to check the consistency of each conjunctive clause, and the number of clauses is bound by an exponential function. To check the consistency of one conjunctive clause i<k i , it suces to check that for each i < k, i is consistent and is consistent (if each of these formulas is satisable then a model of the original clause can then by found as the disjoint union of their models) and the number of such conjuncts is again bound above by an exponential function. To check if i is consistent in exponential time, let Z be the set of all MCSs that include and check whether i belongs to an MCS in Z. The case is entirely similar. Corollary 5 The satisability problem for L( , ) is EXPTIME-complete. PROOF: By proposition 3 and theorem 4. Definition 6 The translation g maps formulas of L( =n , n , n , , : n N) to formulas of L( , , ). It is dened by g(p) = p, g() = g(), g(1 2 ) = g(1 ) g(2 ), g( ) = g(), g( ) = g() and:
n

g( g( g(

=n

) = ) = ) =
in

...
n

g()
j

n n

... g( )

g()

The use of superscripts in =n , n was already introduced, as an abbreviation, in Ladners seminal work [L77] where he proved the PSPACE-completeness of the satisability problem for L( ) (i.e. the basic modal logic K). However, in Ladners work there is no binary encoding of the superscripts. Our use of a binary encoding for these superscripts in the current paper makes our formulas more concise and pushes the complexity up. Since the superscripts of =n , n , n are encoded in binary, the last three cases can produce exponentially longer formulas, but not worse. Lemma 7 If L( PROOF:
=n

: n N) then g() and |g()| 2|| .

2

The equivalence g() derives directly from the denition of our semantics and the denition of g. Let L( =n , n , n , , : n N) and let k be the number of occurrences of n (any n) in , clearly k < ||. We prove by induction over k that |g()| 2(k+1)|| . If k = 0 there are no occurrences of n in . When calculating
n

g() we replace each occurrence of

=n

(which has length 1 + log2 n) by

n

...

(which has length n) and each occurrence of n by ... (which has length || n + 1). It follows that |g()| 2 when k = 0, proving the base case. Now let k > 0 and suppose for j < k that if has only j occurrences of operators n then g() has length at most 2(j+1)|| . Consider a formula = n (some n N) where has only k 1 occurrences of operators n . We have g( n ) = in g( =i ). By the induction hypothesis, |g( n )| n (|g( =n |)) n (2k|| ). But log2 n || so n 2|| , hence |g()| = |g( n )| 2|| 2k|| 2(k+1)|| , proving the induction step. For an arbitrary formula L( =n , n , n , , : n N), the number of occurrences of operators n in is less than ||, hence |g()| 2|||| , as required.

Theorem 8 1. The satisability problem for L( 2. The satisability problem for L( 3. The satisability problem for L( PROOF: 1. To determine whether L( =n , n : n N) is satisable we rst compute, 2 in exponential time, the translation g() L( ). By lemma 7, |g()| 2|| . By proposition 3 there is a polynomial p such that the satisability of g() 2 can be solved in space p(|g()|) p(2|| ). Hence the satisability problem for L( =n , n : n N) can be solved in EXPSPACE. 2. For L(
=n =n =n =n

, , ,

n n n

: n N) is in EXPSPACE. , ,
n n

, ,

: n N) is in 2-EXPTIME.

: n N) is in 2-EXPTIME.

: n N) we know that g() L( , ) and we can

||2

solve the satisability of g() in time 2p(|g()|) 2p(2 by proposition 3. Thus the satisability problem of L( is in 2-EXPTIME.

=n

, for some polynomial p, , n , n , : n N)

3. Similarly, for L( =n , n , n : n N) we can solve g() L( , ) in exponential time, in terms of |g()|, i.e. double exponential time in terms of ||.

4
4.1

Lower Complexity Bounds and Alternating Turing Machines

Basic denitions and results

In this section we introduce the needed notation and denitions regarding Alternating Turing Machines of [CKS81], in order to prove tight lower complexity bounds for our logics. Our basic notation and denitions mostly come from [LL05]. An alternating Turing Machine M is of the form M = (Q, , q0 , qacc , qr , ), where Q is the set of states, is the alphabet which contains a blank symbol, and q0 , qacc , qr Q the starting, the accepting and the rejecting state respectively. Q is partitioned into Q = Q Q {qacc , qr }. The transition is a quintic relation:

 Q Q {L, R} We also write (q , b, m) (q, a) to denote (q, a, q , b, x) for q, q Q, a, b and x {L, R}. A conguration (w, i, q) of a Turing Machine consists of a word w , an integer i < |w| and a state q Q. It represents the state of the Turing Machine when w is on the tape (followed by blanks), the tape head is in position i and the machine is in state q. We write wj for the jth character of w. For each conguration (w, i, q) and each instruction (p, b, L) (q, wi ) we can obtain a successor conguration (w[i/b], i 1, p) where w[i/b] is obtained from w by replacing wi by b and for each (p, b, R) (q, wi ) we obtain another successor conguration (w[i/b], i + 1, p). No other congurations are successor congurations to (w, i, q). An initial conguration has the form (w, 0, q0 ) for some w , where q0 is the start state. A nite computation path of an ATM M on a word w is a nite sequence of congurations c0 , c1 , . . . , cn such that c0 = (w, 0, q0 ), if cn = (w , j, q) then q {qacc , qr }, and for i < n ci+1 is a successor conguration to ci and ci = (w , j, q) q {qacc , qr }. An innite computation path is an innite sequence of congurations c0 , c1 , . . . with the same properties as nite computation paths, except that none of the states occurring in the congurations belongs to {qacc , qr }. All the ATMs considered in this paper have the property that all computation paths are nite. Hence, we dene the acceptance for an ATM with nite computation paths, the general case was given in [CKS81]. Definition 9 Let M be an ATM with nite computation paths. The acceptance of a conguration (w, i, q) depends on the sort of the state q and: If the state q is qacc then the conguration is accepting (so if q = qr then the conguration is not accepting). If q Q , then the conguration is accepting i at least one of its successors is accepting. If q Q , then the conguration is accepting i all of its successors are accepting. Since all computation paths are nite this denition is well founded. A Turing Machine accepts input w if the conguration (w, 0, q0 ) is accepting. We write L(M) for the language accepted by M, that is, L(M) = {w | M accepts w}. We write APTIME for the class of problems that can be solved by an alternating Turing Machine in polynomial time, etc. The following proposition was proved in [CKS81, corollary 3.6]. Proposition 10 APTIME = PSPACE

APSPACE = EXPTIME AEXPTIME = EXPSPACE AEXPSPACE = 2-EXPTIME Definition 11 Given an ATM M with nite computation paths, an acceptance tree T = (T, F, ) consists of a set T of nodes, a binary relation F over T dening a tree with root r T , and a function mapping nodes to acceptance congurations, such that: For the root node r T we have (r) = (w, 0, q0 ), for some w . If (, ) F then ( ) is a successor conguration to ( ). For all T , it is not the case that ( ) = (w, i, qr ). 8

 For all T , if ( ) = (w, i, q) where q Q then there is T with (, ) F . For all T , if ( ) = (w, i, q) where q Q and (w , i , q ) is any successor conguration to (w, i, q) then there is T with (, ) F and ( ) = (w , i , q ). Lemma 12 ([LL05]) Let M be an ATM with only nite computation paths. Then there exists an acceptance tree T of M on w i M accepts w.

4.2

Reducing Alternating Turing Machines to Modal Languages

In this section we dene a polynomial time reduction of the inputs to an exponential space ATM M to formulas of L( =n , : n N), for various modalities { , , n : n N}. This will allow us to prove hardness results for various modal languages L( =n , : n N). For example, we will deduce that the satisability problem for L( =n , : n N) is 2-EXPTIME-hard. For the case where M is known to run in EXPTIME we will be able to prove that L( =n , n : n N)SAT is EXPSPACE-hard. Further restrictions on the computation of M yield lower complexity bounds on a number of other modal logics. Let M = (Q, , q0 , qacc , qr , ) be an an exponential space bounded ATM and let p be a polynomial such that the space used by M on any input of size n is bound by 2p(n) . Let w = w0 , . . . , wn1 be an input for M. In the following, all congurations will have length m = 2p(n) (congurations may include blank symbols). We will dene a formula M,w ( , m), where { , , n : n N} and m is the conguration size. We will substitute dierent operators for , depending on the particular complexity result we wish to obtain. The models of the formula M,w ( , m) (if any) will represent acceptance trees. Each world in such a model will represent a cell in one of the accepting congurations. The position of the cell within the conguration will be determined by the truth of the space propositions si (i < p(n) at that world. If the position of the cell is strictly less than m 1 then the only accessible world will represent the next cell in the same conguration. If the position of the cell is m 1 then the successor worlds will represent the initial cells in successor congurations. We will use the following propositions telling us everything about the ATM. Q {si : i < p(n)} {pf irst , pseen } The intended interpretations are: q Q means that the tape head is at the current cell and M is in state q. We will use formula h = qQ q, which means that the tape head is on the current cell. a means that a is the symbol in the current cell. sp(n)1 , . . . , s0 represent a counter in binary code for counting the m = 2p(n) tape cells of a conguration. The counter value will be 0 in the leftmost and m 1 in the rightmost tape cell, for instance. pf irst should be true at each cell of the initial conguration, but need not be true elsewhere. pseen should be true at each cell of any conguration to the right of the tape head position, but need not be true to the left or at the tape head position. We need formulas expressing particular space counter values. Let k < m and let ki be the ith bit of a binary representation of k, for i < p(n), so k = i<p(n) ki 2i . S=k =
i<p(n), ki =1

si

i<p(n), ki =0

si

Observe that S=0 , S=m1 express that the space counters represent the leftmost and rightmost cells of a conguration, respectively. Let Sn = S=k
kn

Note that the length of Sn is bound by a polynomial (O(p(n)2 )) in terms of |w| = n. Let inc(s, i) = (si sj ) ( (si sj ) ((sj sj ) (sj sj ))
j<i j<i i<j<p(n)

For i < p(n), inc(s, i) says that if si is true but for j < i, sj is false then at any successor the last i + 1 space counters ip, but the other space counters are unchanged. Let inc(s) = inc(s, i) ( si si ) (1)
i<p(n) i<p(n) i<p(n)

inc(s) says that the space counters at a successor cell represent one more (modulo m) than the space counters at the current cell. Now we dene the formula M,w ( , m) L( =n , : n N): M,w ( , m) = start (w) (f irst seen inc(s) head conf acc ) where start (w) f irst seen conf head acc = S=0 q0 pf irst w0 (w1 = (pf irst (pf irst S=0 ))
q,q Q,q=q

(2)

(w2 . . . (wn1

(pf irst blank) . . .)

= (h pseen ) (S=0 pseen ) = ( a) (a b)

a a,b,b=a

(q q )

= (pseen h) = qr = [((h a)
a =m

a) ( (q a ) (q a )

((q a) ((q a) ((q a)

=m1

(q ,a ,R)(q,a)

(a

q ))]

qQ, a

= =

(q ,a ,L)(q,a)

=m1

=m (a q )) =m (a q ))

qQ , a

(q ,a ,L)(q,a)

(q ,a ,R)(q,a)

=m1 (q a )

qQ , a

(q ,a ,L)(q,a)

(q ,a ,R)(q,a)

The rst part start says that we start with space counter zero and the rst n cells on any path starting from the root denes the non-blank part of the starting conguration (w, 0, q0 ). The formula f irst ensures that pf irst holds throughout the initial conguration, so start in conjunction with f irst ensures that after the nth cell there are only blanks on the tape in the initial conguration. The formula seen ensures that if h is true at a cell of a conguration then pseen is true at each cell to the right within that conguration. In conjunction with head it ensures that the head of the Turing Machine cannot be in more than one cell of a conguration. inc(s) ensures that the space counters correctly represent the position of a cell. It expresses that the value of the space counter increases by one (modulo m) when you move to a successor. conf 10

expresses the fact that at each world there is exactly one tape symbol and at most one state. acc makes sure that no computation path ends in the rejecting state. requires that any successor conguration can be obtained from the current conguration by some -transition. ensures that all successor congurations to a universal conguration can be found. And nally, says that some successor conguration to any existential conguration can be found. Theorem 13 Let M be an exponential space ATM, then M,w ( , m) is satisable i w L(M). PROOF: For the right to left implication suppose w L(M) and let T = (T, F, ) be an acceptance computation tree. Let W = T m and R = {((, i), (, i + 1)) : T , i < m 1} {((, m 1), ( , 0)) : (, ) F } Let the valuation V be dened by V (pf irst ) = {(0 , i) : i < m} where 0 is the initial conguration

V (pseen ) = {(, j) : = (w, i, q) T, j > i} V (q) = {(, i) : T, ( ) = (w, i, q) some w} V (a) = {(, j) : ( ) = (w, i, q), wj = a} V (si ) = {(, j) : T , bit i of binary code for j is 1} Now it is just a routine exercise to check that (W, R, V ) |= M,w ( , m), so M,w ( , m) is satisable. For the left to right implication, suppose S, r0 |= M,w ( , m) for some structure S = (W, R, V ) and some r0 W . Since S, r0 |= inc(s), for all y W we can dene the position pos(y) = 2i
i<p(n),S,y|=si

Since S |= inc(s), for any x, y W if xRy then either pos(x) = m 1, pos(y) = 0 or pos(y) = pos(x) + 1. Dene T W m by { } (x0 , x1 , . . . , xm1 ) : pos(xi ) = i, xi Rxi+1 for i < m 1 T = (3) i < m, M, xi |= h We write x for (x0 , . . . , xm1 ). Dene F T T by letting (x, y) F i xm1 Ry0 . For each x T and each i < m, since S |= conf there is a unique symbol a such that S, xi |= a and there is at most one state symbol q Q such that S, xi |= q. Since S |= head and by denition of T , there is exactly one i < m such that S, xi |= h. It follows that each y T denes a unique conguration (w, i, q) where S, yj |= wj , for j < m, and S, yi |= q. We denote this conguration as conf (y). We claim that T = (T, F, conf ) is an acceptance tree for M, w. For any x T with x0 = r0 , since S, r0 |= start (w) f irst we have conf (x) = (w, 0, q0 ), the initial conguration. Since (W, R, V ) |= , if (x, y) F then conf (y) is a successor conguration to conf (x). Since (W, R, V ) |= , if x T, conf (x) = (w, i, q) and q Q then there is y T such that (x, y) F . Finally, since (W, R, V ) |= , if x T conf (x) = (w, i, q) where q Q , and is any successor conguration to conf (x) then there is y T such that (x, y) F and conf (y) = . Hence (T, F ) is an acceptance tree. By lemma 12, M accepts w. Theorem 14 M,w ( , m) is satisable if and only if M,w ( PROOF:

, m) is satisable.

11

The formula is valid, so any model of M,w ( , m) is also a model of (M,w) ( , m), proving the left to right implication. Conversely, suppose (W, R, V ), r0 |= M,w ( , m). Let W0 W consist of all those worlds which can be reached from r0 by a chain of R steps and let R0 , V0 be obtained from R, V by restricting to this set. Since the deletion of inaccessible worlds cannot aect the truth of L( =n , n , : n N) formulas, we have (W0 , R0 , V0 ), r0 |= M,w ( , m). Since all worlds in W0 are now accessible from r0 , we have (W0 , R0 , V0 ), r0 |= for all L( =n , n : n N), hence (W0 , R0 , V0 ), r0 |= M,w ( , m). This proves the right to left implication. Now we consider the case where the run-time of M is bounded by m = 2p(n) for some polynomial p (we lose no generality by taking m to be a bound on both the space and the time used by M). Theorem 15 Let M be an exponential time ATM and let w . M,w ( i w L(M). PROOF: By theorem 13 we know that M,w ( , m) is satisable i w L(M). Any model for M,w ( , m) is also a model for M,w ( m , m), since k is valid, for 2 any k N, so if M,w ( , m) is satisable then so is M,w ( m , m). Conversely, 2 2 let S, r0 |= M,w ( m , m) be any model of M,w ( m , m), say S = (W, R, V ). 2 As before, since M,w ( m , m) does not involve , we can assume that for each x W there is an R-path from r0 to x. Because the run time of M is bounded by m = 2p(n) , the maximum possible length of a computation path for w is m. Hence the maximum length of an R-path from r0 is at most m2 (m congurations, each of 2 length m). Hence S, r0 |= m , for all L( =n , n : n N), and 2 therefore S, r0 |= M,w ( , m). We have now shown that M,w ( m , m) is satisable i M,w ( , m) is satisable. This proves the theorem. We continue by supposing that M runs using only polynomial space. Theorem 16 L( , ) is EXPTIME-hard. PROOF: By proposition 10, APSPACE=EXPTIME. Let M be a polynomial space ATM, let q be a polynomial and suppose a computation of M on an input of size n uses at most q(n) space. We reduce the language accepted by M to the satisability problem for L( , ). Let w have size n. We map w to g(M,w ( , q(n))) L( , ). Consult the denition of M,w ( , q(n)) in (2). Each occurrence of =k occurring in M,w ( , q(n)) has superscript k q(n). Hence the translation g given in denition 6 runs in polynomial time. By theorem 13, M,w ( , q(n)) is satisable i w L(M). By lemma 7, this holds i g(M,w ( , q(n))) is satisable. Hence the reduction is correct and can be computed in polynomial time. If M is known to run in PTIME (say time p(n)) we can modify the reduction of theorem 16 so as to eliminate all occurrences of and produce a formula in L( ). Theorem 17 Let M be an polynomial time ATM, let p be a polynomial bound on the run-time of M, and let w . Let m = p(|w|) this is a bound on both the time and space of any 2 M-computation. Then g(M,w ( m , m)) L( ) can be computed in polynomial time and it is satisable i w L(M). PROOF: 12
m2

, m) is satisable

Similar. This shows that the satisability problem for L( ) is PSPACE-hard. Looking back to gure 1, we have now gone in a complete circle by reproving Ladners result that the satisability problem for the basic modal logic K is PSPACE-complete.

4.3

Complexity Results for Branching Modal Logics

: n N) L L(
=n

Theorem 18 1. If L( =n , complete. 2. If L( =n , complete. 3. L( PROOF: The upper complexity bounds for all cases are established in theorem 8. We now prove the lower bounds. 1. To prove that L( =n , : n N)-SAT is 2-EXPTIME-hard, let A be an arbitrary 2-EXPTIME decision problem. By proposition 10, A can be solved by an ATM M using exponential space. Let w be an instance of A. In polynomial time we can compute M,w ( , m), where m is an exponential bound on the space used by M on input w. By theorem 13, the map w M,w ( , m) is a polynomial time reduction of A to L( =n , : n N)-SAT. Hence this problem is 2-EXPTIME-hard. 2. By theorem 14, M,w ( , m) is satisable i M,w ( , m) is satisable. Hence the map w M,w ( , m) is a polynomial time reduction, so L( =n , : n N)SAT is 2-EXPTIME-hard. 3. If A is a decision problem in EXPSPACE then by proposition 10, there is an exponential time ATM M that solves A. By theorem 15 the map w 2 M,w ( m , m) is a polynomial time reduction of A to L( =n , n : n N)SAT, so this problem is EXPSPACE-hard.
=n

: n N) then L-SAT is 2-EXPTIME: n N) then L-SAT is 2-EXPTIME-

: n N) L L(

=n

: n N)-SAT is EXPSPACE-complete.

Linear Discrete Time

So far we have been considering unrestricted Kripke semantics. We now turn our attention to linear discrete frames. We show that the satisability problem for languages between L( =n , : n N) and L( =n , n , n , : n N) on the linear frame (N, <) are EXPSPACE-complete. To prove that, we use the same approach as we did in the last section but work this time with Deterministic Turing Machines (DTMs). Note that DTMs can be considered as a special case of Alternating Turing Machines (ATMs): a deterministic Turing Machine is an ATM where all its states are existential and the transition function is single valued. The semantics of L( =n , n , n , : n N) on (N, <) follow: Definition 19 Formulas of L( =n , n , n , : n N) are interpreted with respect to linear discrete models: (N, <, V ). Given S = (N, <, V ) and an index i N we evaluate formulas as follows:

13

S, i |= p S, i |= S, i |= S, i |= S, i |= S, i |= S, i |=
=n n n

i i i i i i i

i V (p) S, i |= S, i |= and S, i |= S, i + n |= k, i + n k (S, k |= ) k, i k i + n (S, k |= ) k i (S, k |= )

Definition 20 The translation f maps formulas of L( =n , n , n , : n N) to formulas of TPTL and is dened by f (p) = p, f () = f (), f (1 2 ) = f (1 ) f (2 ) and f( f( f( f( ) = x.G(y.(y = x + n f ())) ) = x.G(y.(y x + n f ())) n ) = x.G(y.(y x + n f ()))
=n n

= Gf ()

where x, y are new variables, not occurring in f (). Proposition 21 ([AH94]) The satisability problem for formulas of TPTL formulas over (N, < ) is EXPSPACE-complete. Lemma 22 Let L( 1. |f ()| 8 ||. 2. (N, <) |= f () . PROOF: Substrings n , n , =n of of length 1 + log2 n are replaced by at most 8 + log2 n characters in f (), so at most 7 additional characters for each of these operators. We count any variable x. as single character and we charge 3 for G = F. The equivalence of and f () over (N, <) is direct from the denition of the semantics. Theorem 23 The satisability problem for formulas of L( solved in EXPSPACE. PROOF: To determine whether L( =n , n , n , : n N) is satisable in (N, <) we rst compute, in polynomial time, the translation f () TPTL. By lemma 22, |f ()| 8 ||. By proposition 21 there is a polynomial p such that the satisability of f () can be solved in space 2p(|f ()|) 2p(8||) . Hence the satisability problem for L( =n , n , n , : n N) over (N, <) can be solved in EXPSPACE. For the lower bound, we take an exponential space DTM M and an input w to the machine. In order to avoid prematurely terminating runs, it is convenient to assume that the transition function : Q Q {L, R} of M is always dened (if q is a nal state and a then we may let (q, a) = (q, a, L) except at a marker for the leftmost cell where (q, ) = (q, , R), so M stays in this nal state forever). Note that although runs of M are innite, computation paths are still nite, according to the denition we gave just before denition 9, because on any innite run, M enters a nal state after nitely many steps. In polynomial time we compute the formula M,w ( , m) of (2). 14
=n n n =n

: n N),

: n N) can be

Theorem 24 M,w ( PROOF:

, m) is satisable over (N, <) i w L(M).

If M,w ( , m) is satisable in (N, <) then of course it is satisable. So by theorems 13 and 14, M has an accepting run on input w (in fact, since M is deterministic, there is a unique run of M on w and M accepts w). Conversely, suppose M accepts w. By theorems 13 and 14 there is a structure S = (W, R, V ) and a world v W such that S, v |= M,w ( , m), though this model may not be linear. As in the proof of theorem 13 we can dene the position pos(x) of each world w S, we can dene T W m using (3) and we can make T into the base of an acceptance tree for M. Since M is deterministic, the conjuncts and of M,w ( , m) demand at most only a single successor conguration for each conguration. We can now restrict S to S , say, by starting with m worlds representing the initial conguration and then appending only a single successor conguration to each conguration, and M,w ( , m) will still be true in S . This restricted structure S is clearly isomorphic to a structure (N, <, V ). Hence M accepts w i M,w ( , m) is satisable i M,w ( , m) is satisable over (N, <). Theorem 25 L( =n , n , PROOF: We have already seen in theorem 23 that satisability of L( =n , n , n , : n N) over (N, <) is in EXPSPACE. We prove that satisability of L( =n , : n N) over (N, <) is EXPSPACE-hard. Let A be an arbitrary EXPSPACE decision problem, that is, A can be solved by a DTM M using exponential space. Let w be an instance of A, coded into the alphabet of M. In polynomial time we can compute M,w ( , m). By theorem 24, the map w M,w ( , m) is a polynomial time reduction of A to the satisability problem for L( =n , : n N) over (N, <). Hence this problem is EXPSPACE-hard.
n

The satisability problems for any language between L( , : n N) over (N, <) is EXPSPACE-complete.

=n

: n N) and

PDL with Bounded Modalities

In this section we introduce bounded iteration operators in the context of Propositional Dynamic Logic (PDL) [FL79]. There are two main dierences between these dynamic logics and the modal logics we have previously considered. Firstly, dynamic logics usually include more than one modality. Note that the satisability problem for the multi-modal generalisation of K remains PSPACE-complete [HM92], so the extra modalities by themselves do not account for the increase in complexity. The other dierences is that dynamic logics include operators that produce new modalities from old ones. In the context of dynamic logics, the modalities are usually called programs. With respect to complexity, we prove in theorem 35 and in corollary 34 that the complexity for several dynamic logics using binary encoded operators is raised by one exponential when compared to their variants without these abbreviations. But most importantly, we rise an exception of this phenomenon. In theorem 37, we prove that the complexity of the tree satisability problem (denition 36) for IPDL (denition 26) using binary encoded programs is the same as its variant using standard unary programs. We shall focus now on the technicalities. Definition 26 (PDL) Let Prop be a countable set of propositions and let A be a set of atomic program names. Formulas and programs of PDL are dened by the following syntax rules: := p (p Prop) | | | [] := a (a A) | 1 2 | 1 ; 2 | 15

Operators P(, , ; ) P(, ; , ) P(, , ; , )

Complexity of Satisability Problem PSPACE-complete EXPTIME-complete 2-EXPTIME-complete

Reference [F01] [FL79, PR79] [LL05, D84]

Figure 2: Complexity of Propositional Dynamic Logics with bounded modalities where A is a countably innite set of atomic programs. We write as an abbreviation of []. We write + for ; . Formulas of IPDL are dened similarly, but the denition of a program becomes := a (a A) | 1 2 | 1 2 | 1 ; 2 | Formulas of PDL/ , IPDL/ are obtained from PDL, IPDL (respectively) by omitting from the denition of program. Formulas of these languages may be evaluated at a point of a multimodal Kripke structure S = (W, Ra : a A, V ), where each Ra is a binary relation over W (corresponding to the program a A) and the valuation V maps propositions to subsets of W . For each a A let aS = Ra and let (1 2 )S , (1 2 )S , (1 ; 2 )S , ( )S be dened to be (1 )S (2 )S , (1 )S (2 )S , (1 )S |(2 )S , ( S ) , respectively, where | denotes composition of binary relations and denotes the reexive, transitive closure of a binary relation. Formulas are evaluated at a world w W of a structure S = (W, Ra : a A) by, S, w |= p S, w |= S, w |= 1 2 S, w |= [] w V (p) w |= S, w |= 1 and S, w |= 2 v ((w, v) S S, v |= )

A formula is satisable if there is a multimodal Kripke model such that is true at some world in the model. Proposition 27 1. The satisability problem for IPDL/ is PSPACE-complete ([F01]). 2. The satisability problem for PDL is EXPTIME-complete ([FL79, PR79]). 3. The satisability problem for IPDL is 2-EXPTIME-complete ([LL05, D84]). We now augment these dynamic logics by adding programs =n , n , n : n N. As before, the superscripts are encoded in binary, so | =n | = 1 + log2 n. For the semantics, we let ( =n )S = S | S | . . . | S , ( n )S = in ( =i )S and ( n )S = ni< (( =i )S ). Let U n { , =n , n , n } and Bin {, , ; }. We write P(U n Bin) for the dynamic logic whose programs are dened by := a (a A) | u (u U n) | B(1 , 2 ) (B Bin) where A is a countably innite set of generator programs. We write P k (U n Bin) for the case where the set A of programs has at most k elements (so P (U n Bin) = P(U n Bin)). Thus PDL = P( , , ; ) and IPDL= P( , , , ; ), etc. Figure 2 summarises known complexity results for dynamic logics and the table in the abstract summarises our new results, we prove the new results in theorems 30, 33, 35, 39 and corollary 34.
n

16

Definition 28 The translation t maps programs of P(, , ; , =n , n , n , : n N) to programs of P(, , ; , ) = IPDL. Given a program of P(, , ; , =n , n , n , : n N) we obtain t() by ren n

placing each occurrence of =n by ; ; . . . ; , each occurrence of n by ; ; . . . ; ; and each oc currence of n by in =i . We can apply this translation t to formulas of P(, , ; , =n , n , n , : n N) by replacing each program occurring in a formula by t(). Lemma 29 Let P(, , ; , =n , n , n , : n N). We have |t()| 2|| and t() . Let X {, , ; , }. If P(X {=n , n , n : n N}) then t() P(X).
2

PROOF: Similar to the proof of lemma 7 Theorem 30 The satisability problem for P(, , ; , =n , n : n N) is in EXPSPACE. The satisability problem for P(, ; , =n , n , n , : n N) is in 2-EXPTIME. PROOF: For a formula P(, , ; , =n , n : n N) we know that t() P(, , ; ) = 2 IPDL/ , and we can solve the satisability of t() in space p(|t()|) p(2|| ), for some polynomial p, by proposition 27 and lemma 29. Hence the satisability problem for P(, , ; , =n , n , n ) is in EXPSPACE. Similarly, if P(, , ; , =n , n , n , : n N) we know that t() P(, ; , ) = PDL, and we can solve the satisability of t() in time 2p(|t()|) = 2p(2 double-exponential time.
||2

, i.e. in

Definition 31 Let a be the single generator for P 1 (=n , n , n , : n N). Let L( =n , n , n , n N). The formula h() P 1 (=n , n , n , : n N) is obtained from by replacing each occurrence of =n , n , n , by [a=n ], [an ], [an ], [a ], respectively. Lemma 32 1. h can be computed in linear time. 2. A formula L(
=n

: n N) is satisable i h() is satisable.

=n

3. Hence h is a polynomial time reduction of L( n N). The proof is obvious and we omit it.

: n N) to P 1 (=n , n , n , :

Theorem 33 The satisability problem for P 1 (=n , n : n N) is EXPSPACE-hard. PROOF: By theorem 18(3) and lemma 32. Corollary 34 Let k 1 and let {=n , n : n N} X {, , ; , =n , n , n : n N}. The satisability problem for P k (X) is EXPSPACE-complete. PROOF: By theorems 30 and 33.

17

Theorem 35 The satisability problem for any language between P(, =n , : n N) and P(, ; , =n , n , n , : n N) is 2-EXPTIME-complete. PROOF: The satisability problem for P(, ; , =n , n , n , : n N) is in 2-EXPTIME, by theorem 30. By theorem 18 and lemma 32, the satisability problem for P(, =n , : n N) is 2-EXPTIME-hard. An important diculty for proving upper bounds for IPDL and its variations is their lack of the tree model property [D84]. A closer look at the 2-EXPTIME-complete lower bound proof of IPDL (see theorem 1 in [LL05]) shows that the combination of iteration and intersection is strong enough for expressing the program operator =n over particular tree structures that represent an exponential space Alternating Turing Machine. Note that in [LL05], the program operators n and n are not explicitly expressed in IPDL. What happens if we enrich the syntax of IPDL with these operators? Is it still possible, over arbitrary tree structures, to express the program operators =n n , and n in IPDL? Clearly, the technicalities for answering this question are dierent from those ones developed in [LL05]. In what follows, we strengthen the result of [LL05] by showing that over arbitrary trees there is a polynomial time reduction from the satisability problem for P(, , ; , n , n , =n , : n N) to that of P(, , ; , ). Due to that reduction, the complexity of the tree satisability problem of P(, , ; , n , n , =n , : n N) has the same complexity as the tree satisability problem of P(, , ; , ). This result contradicts the so far obtained results; we were expecting the complexity of P(, , ; , n , n , =n , : n N) to be at least one exponential higher than the complexity of IPDL. For clarity now, we formulate the following decision problem: Definition 36 Let L be some modal or dynamic modal logic, interpreted over rooted Kripke models. Tree satisability. Given a formula of L, does there exists a tree like model M and point w such that M, w |= hold? for tree like models read [BRV01] (page 62). Theorem 37 1. There is a polynomial time reduction from the tree satisability problem for P(, , ; , n , n , =n , : n N) to that of P(, , ; , ). 2. There is a polynomial time reduction from the tree satisability problem for P(, , ; , n , =n : n N) to that of P(, , ; , n : n N). PROOF: 1. For the rst part, let P(, , ; , n , n , =n , : n N). We will dene the reduction by dening a formula = 1 [ ]1 P(, , ; , ) (4)

where is the union of all the atomic programs occurring in . Since the deletion of worlds not reachable by in any structure does not aect the truth of a formula, we will restrict our attention to generated structures. Let be a new program. 1 will include a conjunct ([]), so will be interpreted as the empty program, in any model of [ ]1 . Let e = . This program e will be interpreted as the identity over the domain of any model of 1 . Suppose n occurs in , for some program , some {, , =} and some n N. We can assume that n is even, as we can replace n by ; n1 , n 18

by e ; n1 and =n by ; =n1 if n is odd. For each n and such that n occurs in , for some , let k = log2 n . We introduce propositional space 2 counters r, si : i < k and programs , , i , i : i < k, unique to n, . Loosely speaking, the propositions si : i < k (together with corresponding programs i , i : i < k) suce to count the distance from one point to another, modulo n , 2 but we need the other proposition r (and corresponding programs , ) to dispense with the modulo n and measure the true distance from one point to another, so 2 long as the distance does not exceed n. In more detail, the propositional space counters dene the position posS (v) of any s world v in a structure S, by posS (v) = {2i : i < k, S, v |= si }. Let m < 2k and s dene s=m = si si
bit(i,m)=1

s<m

bit(i,m)=0

[si

si ]

bit(i,m)=1

i<j<k, bit(j,m)=0

Clearly, S, v |= s=m posS (v) = m and S, v |= s<m posS (v) < m. s s 1 will include, under the scope of [ ], a conjunct inc(s), dened to be (s= n 1 i<j<k (sj []sj sj []sj ))] j<i si ) j<i si ) ([](si i<k [(si 2 s= n 1 []s=0 ) 2 Observe that if S, v |= inc(s), posS (v) < n and u is any successor of v then s 2 posS (u) = posS (v) + 1(mod n ). s s 2 1 will include under [ ] a further conjunct Alts = (r []((s=0 r) (s=0 r)) (r []((s=0 r) (s=0 r)))) If S, w |= s< n Copys Alts and w is any w-path of transitions, then the 2 position posS (wi ) must increment by one modulo n every time you increment i, s 2 the proposition r must stay constant (either it must stay true, or it must stay false) every time you increment the position, except when the position goes from n n 2 1 to 0 and in this case the truth of r is ipped. We call a sequence of 2 worlds n S v0 , v1 , . . . v n 1 a block if (vi , vi+1 ) (all i < 2 1) and either S, vi |= r (all 2 i < n 1) or S, vi |= r (all i < n 1). In the former case we may call it an 2 2 r-block and in the latter a r-block. Let i = e, = e, = e and = e, for i < k. The reexive i programs i , , i , : i < k are intended to mirror the corresponding propositions si , r : i < k. 1 will also include under [ ] a nal conjunct: Copys = ( (si i si [i ])(r r )[ ]) i i
i<k If S, v |= Copys then propositions si , r are true at v i i , hold at the reexive edge (v, v) and negative literals si , r hold at v i i , hold at (v, v), for i < k. Thus these reexive programs exactly match the corresponding propositions and their negations. Let 1 = s< n inc(s) Copys Alts 2 s=s(,n)

19

Dene programs Same = Lef t = Right =

i<k ((i ; ; i ) ( ; ; )) i i [(i ; ; ) i [( ; ; i ) i

i<j<k

[(i ; ; i ) ( ; ; )]] i i [(i ; ; i ) ( ; ; )]] i i

i<k

i<k

i<j<k

We will see that if (u, v) SameS ( Lef tS , RightS respectively) then posS (u) = posS (v) (posS (u) < posS (v), posS (u) > posS (v)). s s s s s s We are now ready to replace occurrences of =n , n and n in . For {=, , } we dene a program n using only {, , ; , } and of polynomial length (in terms of the length of n ). In models of inc(s) Copys Alts , n will evaluate to the same binary relation as n . Let =n n n = [[( ; )+ ; ( ; )+ ; ( ; ) ; ] [( ; )+ ; ( ; )+ ; ( ; ) ; ]] Same = [ [( ; ) ; ( ; ) ] [( ; ) ; ( ; ) ] ] (Lef t ([( ; ) ( ; ) ; ( ; ) ] [( ; ) ; ( ; ) ; ( ; ) ])) = [([( ; )+ ; ( ; )+ ; ( ; )+ ] [( ; )+ ; ( ; )+ ; ( ; )+ ]) Right];

Claim 1: Suppose T , w |= s< n [ ](inc(s) Copys Alts ) and suppose every 2 T world of T is -reachable from w. Then n = ( n )T . We prove the rst case of the claim, is =, the other cases are similar. Consider T =n . Suppose (u, v) =n , both u and v are -reachable from w. Since (u, v) T Same we have u si v sT for i < k, hence posS (u) = posS (v). s s i By denition of =n , there is a -sequence = (u = u0 , u1 , . . . , up = v) from u to v and since all worlds in satisfy inc(s) Copys Alts , must consist of a nal segment of an r-block concatenated with a r-block and an initial segment of an r-block, or similar with r, r reversed. Hence v is two blocks after u. Since posS (u) = posS (v) it follows that the length of is exactly n, s s so (u, v) ( =n )T . Similarly for the converse, suppose (u, v) ( =n )T and let = (u = u0 , u1 , . . . , un = v) be the -path from u to v. Since each world satises s< n inc(s)Copys Alts , as we increment through the sequence the positions 2 increment by one each time, and we pass from a block to the next block but one, T hence (u, v) =n . This proves the (rst case of the) claim. Claim 2: If S, w |= and S is a tree with root w, then there is a structure S such that S , w |= s< n [ ](inc(s) Copys Alts ). 2 Proof of claim 2 Dene a structure S with the same worlds as S and all propositions and programs have the same interpretation in S , except the new programs , , i , i : i < k and propositions r, si : i < k. Hence S , w |= . The space counters si are given a valuation so that whenever (u, v) S we have posS (u), posS (v) < n and posS (v) = posS (u) + 1(mod n ). Since S is a tree, this s s s s 2 2 can be done. r is given a valuation so that -paths form blocks, i.e. if (u, v) S and posS (u) < n 1 then u rS v rS and if posS (u) = n 1 s s 2 2 then u rS v rS . The new programs are interpreted as reexive S relations such that (u, u) i u sS (u, u) S and similarly i i (u, u) S u rS (u, u) S . From this denition, we have S , u |= inc(s) Copys Alts , at each world u of S . The claim follows, since S , w |= . We are now ready to dene . Let 1 be obtained from by replacing each occurrence of =n , n , n by =n , n , n respectively. is now dened by (4). To see that this reduction is correct, suppose S, w |= , where S is a tree 20

with root w. By claim 2, there is a structure S such that S |=

S n n S

s=s(,n)

s< n 2

[ ](inc(s) Copys Alts ). By claim 1 we have = ( ) , hence S , w |= implies that S , w |= 1 , so S , w |= . Conversely if T , w |= and T is a tree with root w then by claim 1 T , w |= . This proves that the reduction is correct. 2. The proof of the second reduction is similar. Let P(, , ; , n , =n : n N). The reduction maps to = 2 [ ]2 P(, , ; , n : n N). This time, the task is to eliminate all occurrences of =n from . Here we let k = log2 n. As before, we use space propositions si : i < k and corresponding programs i , i : i < k, for each occurrence of =n in , but we do not need r, or as we can use n to restrict the distance to a point. This time we dene the identity program by e = 0 , where can be any program and we let i = i e, = i e, i k for i < k. For m < 2 , let s=m , s<m , inc(s) be the same as in the previous part,though we have changed the denition of k here. The denition of Copys is slightly simpler, as we are not using proposition r, Copys = [(si i ) (si i ) [i i ]] Let 2 = s=s(,n) s<n inc(s)Copys . We now dene a program =n , intended to replace =n . =n = ([i ; ; n1 ; i ] [ ; ; n ; ]) i i
i<k S Observe, for any structure S where all points satisfy 2 , that if (v, w) =n then all space counters have the same value at u and v, and there is a -path of length at most n and length at least one, from u to v, hence the length of the path is exactly n and (u, v) ( =n )S . The other containment ( =n )S (=n )S also holds, similarly. Let 2 be obtained from by replacing each occurrence of =n by =n . The reduction maps to 2 [ ]2 . By the foregoing, if all points of a structure S satisfy 2 then ( =n )S = (=n )S , hence the reduction is correct.

ip(n)

Corollary 38 Let X be a signature containing {, , ; , } and contained in {, , ; , =n , n , n : n N}. Then the satisability problem for P(X) is double exponential time complete. PROOF: By proposition 27 and theorem 37(1). Theorem 39 The satisability problem for P(, , ; , n : n N) is EXPSPACE-complete. PROOF: P(, , ; , n : n N) is in EXPSPACE, by theorem 30. To prove EXPSPACEhardness, take an exponential time ATM M and an input w to the machine, let the run time (and space) of M on input w be bound by m = 2p(|w|) , for some polynomial 2 p. By theorem 15, M,w ( m , m) L( =n , n : n N) is satisable i w L(M). 2 Recall that the reduction of denition 31 replaces by [a]. Let M,w ([am ], m) 2 P 1 (=n , n : n N) be the formula obtained from M,w ( m , m) by applying this 2 reduction. Observe that M,w ([am ], m) does not involve , but it remains for us 2 to eliminate occurrences of =n . By the foregoing, M,w ([am ], m) is satisable i M accepts w. Let h be the polynomial time reduction from P(, , ; , n , =n : n N) to P(, , ; , n : n N) of theorem 37(2). To summarise: M accepts w i 2 h(M,w ([am ], m)) P(, , ; , n : n N) is satisable, proving that P(, , ; , n : n N) is EXPSPACE-hard. 21

Open Problems
Consider the satisability problem of P(, , ; , n , n , =n , : n N) over arbitrary structures. We conjecture that P(, , ; , n , n , =n , : n N) is double exponential time complete. In theorem 37, our proof relies heavily on the notion of distance between two points in a model. Models for IPDL and its variations are digraphs, thus the notion of distance between two points is not well dened any more. The technique introduced in [LWW06], might be useful here. Find the complexities of L(
=n

: n N) and L(
n

: n N)-SAT over Kripke frames.

=n

We conjecture that both logics are PSPACE complete. Find the complexities of L( =n , SAT over the linear frame (N, <). : n N), L( : n N)-SAT and L(
n

: n N)-

Find the complexities of P(, ; , n : n N)-SAT and P(, ; , =n : n N)-SAT. Find the complexity of P(, , ; , =n : n N)-SAT.

References
[AH94] [BRV01] [CKS81] [D84] Rajeev Alur and Thomas A. Henzinger. A Really Temporal Logic. Journal of the ACM 41:181-204, 1994. Patrick Blackburn, Maarten de Rijke and Yde Venema. Modal Logic. Cambridge University Press, 2001. A.K. Chandra, D.C.Kozen, L.J. Stockmeyer. Alternation. Journal of the ACM, 28(1): 114-133, January 1981. R. Danecki. Non Deterministic propositional dynamic logic with intersection is decidable. Proceeding of the Fifth Symposium on Computation Theory, Poland. LNCS vol. 208, Springer 1984, pp. 34-53. E. A. Emerson, J. Halpern. Decision Procedures and expressiveness in the temporal logic of branching time. In Proceeding of STOC 1982, pages: 169-180, ACM Press, 1982. E. A. Emerson, C.S. Juttla. Tree automata and the logics of programmes. SIAM-JC, 29(1): 132-158, 1986.

[EMH82]

[EMJ86]

[EMSS92] E. A. Emerson, A. K. Mok, A. P. Sistla, and J. Srinivasan. Quantitative Temporal Reasoning. Real Time Systems, Volume 4: pages 331-352, 1992. [F01] Fabio Massacci. Decision procedures for expressive Description Logics with intersection, composition, converse of roles and role identity. In Proc. of the 17th Int. Joint Conf. on Articial Intelligence (IJCAI-2001), pp.193-198, 2001. Fischer, M.J. and Ladner, R.E. Propositional dynamic logic of regular programs. Journal of Computer and Systems Sciences. v18. 194-211, 1979. E Hemaspaandra. The price of universality. Notre Dame Journal of Formal Logic, 37(2):174203, 1996.

[FL79] [Hem96]

22

[HM92]

Joseph Y. Halpern and Yoram Moses. A guide to completeness and complexity for modal logics of knowledge and belief. Articial Intelligence archive Volume 54 , Issue 3 (April 1992) Pages: 319-379, 1992 Jan Johannsen, Martin Lange. CTL + is complete for double exponential time. Proc. 30th Int. Coll. on Automata, Logics and Programming, ICALP, volume 2719 of LNCS, pages 767 775 , 2003. R. Ladner. The computational complexity of provability in systems of modal logic. SIAM Journal On Computing, 6: 467-480, 1977. M. Lange and C. Lutz. 2-ExpTime Lower Bounds for Propositional Dynamic Logics with Intersection. Journal of Symbolic Logic 70(4), pages 1072-1086, 2005.

[JL03]

[L77] [LL05]

[LWW06] C. Lutz, D. Walther and F. Wolter. Quantitative temporal logics over the reals: PSPACE and below. Journal of Information and Computation, 2006. [PN77] A. Pnueli. The temporal logic of programs. In Proceedings of the 18th Annual Symposium on Foundations of Computer Science (Providence, RI.). IEEE, New York, pages 46-57, 1977. V.R. Pratt. Models of program logics. In Proc. 20th IEEE Symp. Foundations of Computer Science, pages 115-222, 1979. A. P. Sistla and E. M. Clarke. The complexity of propositional linear temporal logics. Proceedings of the fourteenth annual ACM symposium on Theory of Computing, pages: 159 - 168, 1982. M. Vardi and L. Stockmeyer. Improved upper and lower bounds for modal logics of programs. In Proceedings of STOC 1985, ACM Press, 1985.

[PR79] [SC82]

[VS85]

23