The University of Warwick THTCRY OF COMPUTATON REPORT No. I DECISlON PROCEDURES FOR FAMILIES DETERMINI .} Department STIC OF PUSHDO|'IN AUTOMATA LESLIE G. VALIANT of University of Computer Warwick COVENTRY CV4 7AL ENGLAN D. Science August 1973 DECISION PROCEDI'RES FOR FA},TILIES OF DETERMINISTIC PUSHDOIIN AUTOI'{ATA Leslie G. Valiant Department of Computer Science, l.Ian"rick, Coventryl England. Itniversity of A dissertation submitted for the degree of Doctor of Philosophy. July 1973. PREFACE I ehould like to thank my superviaor, Michael Patergon for hie continued intereat discuseions and encouragement and for the numerous nith hirn that provided the etinulation for this research. The contente of Chapter 5 is the reeult of work done jointly with hin. I am aleo grateful to Albert Meyer for bringing the regularity problen to uy attentionr to the Science Research Council for financial support, this theeie. and to Jill Pladdys and Roeeuary llent for typing ABSTRACT The existence and corrpLexity fanilies of deterministic of decision pushdown automata procedures for are investigated, with special eurphasis on positive decidability results for those questions, such as equivalence, which are known deterninistic restriction is The equivalence problem to become undecidable when the removed. is proved decidable for the following three determini.stic families, all of which are already extensive enough to have undecidable inclusion (a) nonsingular automata problems: - a realtirne subfamily, which the largest corresponding classes with previously extends known equival-ence tests, (b) finite-turn aut,omata - characterised by having a bound the nrmrber of times the direction of the stack on rnovement can change, and (c) one-counter aLphabet tlhe prubLem automata to just - defined by restricting the stack one symbol. of whether a )-anguage defined by a rnachine in one fani.ly1 can be recognised by one in another, is a convenienE formulation of numerous decidabLe or potentially decidable queseions, l,Ie ehow that such questions as whether a detennisusc:Le conte:rt-free language can be recognised by a noachine named in any one of the above classes, muet be, if decidable at all, at least as difficult to decide as wheeher such a l-anguage regularity test of Stearns, is regular. We re-exasnine the and obtain an improved algorithm. We do this by reducing by an exponential order the upper possibLe state conplexity pushdoltn aut.omata bound on the of regular sets recognised by deterministic of a given size, to a level close to one knovn to be achievable. I{e pursue an application problem. We of this analysis to a schema theoretic coneider the succinctness with which certain functional schemas can be used to e)rpress equivalent large fLowchart schemas, and obtain closely matching upper and lower bounds for a measure of this. CONTENTS Abstract Introduction I Prelininaries 1: Chapter Determinis tic Pushdown Autonat'a Elementary ProPerties Subf aniLi.es Some Famil.y ProPerties 1.1 T,2 1.3 1.4 1.5 1.5 L.7 1.8 Derivations complexity Measures Tradeoffs in DescriPtion Famil-y RelationshiPs Chapter 2,L 2,2 2.3 2.4 2.5 II DEFINITIONS 2: Introduction Emptiness Finiteness A NotmaL Form Totality chapter 3 : A Nonsingutar Automata Al"ternate Stacking Main Results UndecidabilitY of Inclueion A Ccnjeeture about Rg 4.1 4.2 4.3 4.4 4.5 4,6 5: 2L 23 27 29 30 33 34 37 39 42 44 46 47 54 57 FINITE-TURNAIJT0UATA lntroduction Definitions Proof Strategy ParalLel Stacking Existence of Verifying Machine The Decision Procedure Chapter 5.1 5.2 5,3 5.4 4: 15 18 REALTIME FAMILY Introduction C:hepter t5 EASILY DECIDABLE PROBLEMS The EquivaLence ProbLeg 3.1. 3,2 3.3 3.4 3.5 3.6 9 L2 59 60 63 64 67 72 ONE COUNTER AIEOI'{AIA Introducrion PrelioinarY Results ProprietY Decision Procedure 75 77 82 87 III Containment Probleme Chapter l6,2 6.3 6. IV Asl Introduction 93 94 101 Conments 7: TIIE REGULARITY PROBLEM Introduction NuLl--transparency and {,-invisibiLity Main Theorem Bounds for SubfamiLies Time Complexity Appli"catlon to Chapter 8. 1 8,2 8.3 8.4 RELATIVE COMPLEXITY Results Chapter 7,L 7.2 7,3 7.4 7.5 6: 90 6 L2L Schenas e MONADIC FUNCTIONAI 104 105 111 1L5 119 SCHEMAS Introdr.rction Evatuating Pushdown Automata t23 t25 r!B-S.Autonata Bounds on Succinctness L27 L29 Conclusion L34 References 136 i. INTRODUCTION To understand the meaning have of an executable computer program rre to relate its finite specification to its possibly infinite distinct behaviours when applied general ways of doing to different inputs. Finding this is a prirnary goal. in the theoretical study of computaLions. The specific aims of such investigations are to gain insight into the computational behaviour of whole classes of prograns, and hence to be abl-e to analyse instances of various particular problems for any program within these classes. The recursive decidabiLity of the problems concerned is an important criterion of the practical viability of such analysis: the existence of an effective decision procedure capable of determining the truth of any instance of a particuLar question for a given cLass of prograns, uleans not only that this analysis can be automated, but usually atso that an a priori bound can be pLaced on its difficulty. In contrast, if the problenr is undecidable for the class, that is, there is no effective pr'rcedure for soLving it in general, then new instances of it can always be found for which the solution is more difficult than before and requires further creative effsrt" The courpuLaurons we shaLl" study are those that can be carried out by abstract deter,"ntinisti.e autocnieta rhat have onLy a pr:shdown stack and a fini.te*state control for storage. This is a formalisation of the stack coniellt lrhi-cbE i.s wi.dei"y for exampLe i.n used in practical wricing syntax anaJ-ysers, or in progrannning, implementing reeursion in compi.Lers. Consequently our resuLts will relate to sreas outside autoloata ttreory, such as the syntax of programring nanguages, and cert.ai.n f,crmal" wdels of recursive programs. 2 The observation primaritr"y rnr:ti."uating r:ur worit i-s these f,amiiies of autcmata are wide enough practical computations, behaviour can be shown tlrat, to be reievant whil-e t.o of the important questions atrout their to be dacid.abl.e. 'Ihis is a very rare rnany conjunction of, attrributes for comprrtational rnodels so far sttrclied. It is generally found that for abstract useful classes of shown programs e models designed to describe nearly all the key properties can be to be undecidable. For exarnple, if rse extend any of the families we strall consider by introducing nondetern:inism, or a1 bv lowing a second staek, then rve already lase the deci.dability properties we desire most. The actual nnachines we investigate are deterrninistic automata (dpda), which have been widely studied pushdown beforei'2'3, anci cerLain restricticns cf them. These are all accePtors in the sense lhat the input consists of a sequence of symbols on a Eape, and Lhe output of ei.ther a tyest or a tnc,t depending on Lhe configurations of the machine afrer the input tape has been scanned. Thus each naehine defines a set of strings of syrnbols, or a language, namely that consisting of exactly the sErings that are accepted, i.e. lead ro a ryes?. These l.anguages are all context-free, 'and have the additional useful prcperty that they can all be parsed easily in linear cime. We note that this input-outpur behaviour is not qrrite restrictive as i"t would appear. superficially syurbol for Problems more general rnachines each symhol read t0 do as with related but (e.g. transducers which output a ) can often be reduced to questions about acceptors and languages. correspcnding 3 The decision problem these families of automata with which rre are most concerned for is that of equivalence, i.e. is there effective test co determine whether an two rnachines perform equivalent computations? The existence of such a test has the practicaL sig- nificance that it provides a most convenient model-independent mechanism for verifying correctness. which is optinised in and we want some Thus if we have an automaton rray, and therefore perhaps complicated, to prove that it will perform exactl-y as it is then we could build a second automaton which intended, is nore perspicuous in structure but possibly othervise unsuitable, and test the Ewo for equi.vatence. Other criteria of correctness would, in contrast, necessitate that a new distinct language for describing the intended computation be introduced and The inclusion problem related to the automata formal-ism. (i.e. to decide whether one language is a subset of another ) is related to the equivaLence problem in that any procedure for deciding the former would lead directly to the Latter. Thus, as distinct from the existing analysis of one for such classes as the regular sets3, bounded Languages4, and parenthesis languages5, for all of which both inclusion and equivaLence are decidabi.e, positive sol-utions to the equivalence with an undecidable inclusion problem assume new significance. Our main results for the problem equival-enee problem for families are to show for three distinct families of automata that they are in just this category. The onLy comparable family we know of is that of Ewo-tape acceptors6rT. For rhe class of LL(k) languages8 (which include the sirnpleg languages), atthough equivalence has been shown to be decidable, the inclusion problem is currently open. that both Ewo-tape acceptors and We will show LL(k) languages correspond closely t+ to particular restrictions of two of our farnilies, and hence that the decidability of the equivalence problem for botl'r of them now also follows as corollaries to our results. In addition to proving dec_idability results we are also concerned with giving bounds for the complexity of the derived decision procedures. A bound on the time complexity of a problem (i.e. on the time required by the best possible decision in the worst case ) gives an indication not only of may procedure how much time be needed for an automatic anaLysis, but also ofthe possible difficulty in an informal senserof resolving parricular instances it. The complexiries are expressed in terms of functions r,rhich have the paraineters of the machine description as arguments. Thus if have a polynomial ti-roe bound sn a decision procedure for of we one problern, but know that an exponential time is necessary for deciding another, then this will indicate that for: sufficiently large machines,it will be easier to solve a problem of the first kind than one for the second. Valid measures of absolut.e ccmplexity for decision procedures are numerous, and may depend additir:nally on the nachine model on which one int,ends to execute them. For this reason we are usuai-ly eontent with answers speeifying cnly the nurnber of levels of exponentiationn if any, involve-d. ?hus tvpicaliy l,Ie may say that procedure takes polynomial, or perhaps exponential time, without giving further detaj-l-s. This kind of ciassification is known to robust enough not to depend on which one of the customary machine modeLs is chosen.lo More a satisfying results, which are now totally machine be 5 independent, can be given, however, for the relative complexities of the various decision probLems. Thus we show that, in a specified sense, certain decision problerns sust be equally some others must be at least as difficult difficult, and that as these. Such results, besides confirming our experience with known decision procedures, also throw light on several specific currently we show open problems, which to be at least as difficul-t to decide as a qLrestion for the best procedure ure have been able to find works in which double exponentiaL time. Although the decidability and complexity yield considerable information, the techniques results used themselves in the proofs and the properties on which they depend are equally important in givi.ng insight into the structure of trese computations. Our presentation will therefore attempt to highlight informally the ideas we believe are new and nost important. of We shall omit details arguments whi.ch are well. known and occur elsewhere in the litera- , for the sake of compleEeness, we mention resuLts not directly related to our main theme, Ire shal-l, just sketch the main ture. W?rere ideas from which a proof can be reconstructed. We suggest the fotr-lowing as appropriate both for preLiminary reading and for providing rrctivation to our work from diverse angles: A general exposition of the concepts of automata and decidability is given in Hopcroft and U11rnan3. Basic resul-ts about dpda are also to he found there, as weLl as in Ginsburg and Greibachl. Various grrrnrnatical eharacterisations and of fanilies of deterministic their relevance to parsing are described by Knuthl1, languages Korenjak and llopcroftg, Rosenkrantz and SteamsS, and Harrison and llarell 2. An introduction to progran schemas, and a description of the relation- 6 ship bemeen certain recursive schemas and pushdown automata can be found in Patersonl3. 7 PRELIMINARIES This first section introduces the definitions, basic notions, and prel-iminary results which we shall use in subsequent chapters. Fron the start we take a rDre complexity conscious approach to the is available subjact than elsewhere in the literature. Consequently the preLininary ideas we have to introduce, although rnostly well known, have had to be reformulated so as to give the rnore precise constructions which we now require. In particular we have to bypass those of the widely exponential explos We Our that lead to used standard constructions i.orrs . first define the class of deterministic criterion of acceptance differs from the most is the nore natural for our purposes. in the sense that for pushdown automata. The any dpda specified popuLar one3, but definition is very broad in one of the other standard ways, there is al-ways an equivalent one in our formuLation that is not much nore complex. Further, by placing sinrple restrictions on this class, sha11 Iile can also define several important subclasses which we later study. The size or compLexity of a dpda we describe by means of several of the parameters of its description, which \ile go on to specify. I.Ie then when a machine is economise on one In the investigate the tradeof fs that can be realised transf,ormed of these to another equivalent one in order parameters second chapter we at the expense to of the others. give results about the complexity of problens atready known to be decidable. I{e justify the clain that B these problens are 'easy'by exist for solving verifying that sinple decision them, which work in tirne procedures depending only poly- nornially on the paraneters of the tested machines. In the course of this ne define a normal form for dpda, deriving some and il-l-ustrate of the decidability results as trivial its use by consequences of its existence. Throughout we judge the relative pohrer of the various sub- fanilies of nachines by comparing the classes of languages to which they correspond. For sirnpLicity, therefore, when this is not otherwise confusing, we often identify a family of machines with the cLass of languages it defines. Ctrapter DETINITIONS 1 1.1 Deterministic Pushdown Autonata A pushdown automaton (pda) is an abstract device whose menory consists of a pushdom stack and a finite state control. It can read a string of characters from a finite aLphabet from its unique input tape, which can progress past the input head of the machine in only one direction, and only once. At any step of the computation the transitions which the machine can undergo are determined by the state of the finite state control, the contents of the top of the stack, and the character under the input head. An input word is either ttaceept,ed" or "rejected", d.epending on the "confi.gurations" the rnachine can reach after having read its l-ast character. such automaton defines a Language i.e. the set of words it The languages which are defined by the class of Each accepts. pushdown automata are exactly the context-free languages of Chomskyl4'3. We are here interested in the deterministic restriction of this cl"ass, namely the case where each coribination of state, top stack symbol, and input character defines a unique machine transition. The cLass of such deterministic pushdown automata (dpda) we caLl D. More forma11"n for a machine Me D, let I be its finite input al"phabet {a, b, c, ...}, f its finite stack_4-p@. {A, B, ...} and Q its finite state set {s1, szr...}. input characters by o, rrr2 etc. We B, etc., and We denote strings of strings of stack symbols by define e and A to be the null elements of I* and l* (,)I t 10 In addition we use the special symbol resPectively. $? to denote an eropty staek. c = (s, o) is an eLement of Q x ({n} v I+) and describes the state and stack content (starting from the enpty stack A configuration end) of the machine at some instant. is an element frour Qx ({A}uf) symbol of c. is (s, 0). Thus the mode of The machine has The mode of a configuration and describes the c state and top stack (s, oA) is (s, A), and that of (s, 0) a set F . Q* ({CI} vr) of distinguished accepting modes. The set of transition rules A is a set of rules of the form (s, A) -l* (st, where ei ther or ,r) r e x u {e}, with the additional restriction that i.e. occurs on tteleft in just t.here with an j* t.urr"ition. (i) an (i i) e-mode a reading upde i.e. has at most a e f,, and no i+ rule. The machine M makes the move one configuration (s, A) i* to the other if (s, oA) and only one 1I -+ if a one rule, rule for (st, ,rt) one each mode is and each frorn the of the rules is (s', ,t)rwhere n e X u {e}. If n e I then n, the input character currently under the input head on the tape, is eonsidered to have been read, and the head rpves to the next position on the tape. Otherwise the head does not nove. ct is " sequence of moves starting configuration c and ending with c t. rt is an a-derivation, A derivation c -+ from " *5 "t, 11 if the execution of it is on input tape. Ehe There is a distinguished starting configuration c", which for simplicity we assume t,o belong to Qx configuration ct i" rea"hable from c c by the reading of the word s accompanied -> cr . I^Ihen we say that cl (in) u r). we say iff there is some that a derivation is reachable we mean that it is reachable from c". An input string whose mde bel-ongs cr is to F, accepted by M and rejected iff "s S c for some c iff it is not accepted. configurations c, c' are distinguished by o if a-derivations take one to an accepting mode but not the other. The rank of configurations c and ct (denoted by rank (", Length and "t)r) Two can two is defined as the of a shortest string distinguishing the two, if - othervise. Tr,ro one exists, configurations are said to be equivalent, i.e. c f cf, iff rank (c, ct) = -. Two nachines M, Mt are equivalgnt iff their starting configurations are equivalent. I,Ie denote the set of strings which can take the machine from a configuration c to accepting rmdes, by L(c). Then clearly c = c' Ltt L(c) = i,(c'). For a machine M we denote L(cs) by L(M), and for a class X of such machines , we wi.1l abbreviate {L (M) | U . X} to L(x). ' Thus, to summarise, each dpda (as well as the language it defines) is completely specified by a sextuple M(I, l, Q, F, A, cs). The foll-owing iLlustrates the main points: f rl t2 ExampLe 1.1 For machine M let I = {a, b, c}, f = {Ao, A, B}, Q = {s1, s2, s3}, F = Q x {o}, cs = (sr, Ao) and A the set of transitions: j* ("r, A) b, ("r, B) (s r, A) (sr, A) *9* (sl, AA) (sr, B) b , ("r, BB) (sr, B) -3* (sr, (sr, Then Ao) 1s2, n) B) 5 j+ e, B) (sz, B) 3> (s2, A) (s (sz, A) 3* 1s2, n) ("s, A) 5 it can be (s3, A) r A) (s3r A) 1s 3 easily verified that L(M) = {anbman lr, r:il u {anbmcn lr, * ril. For our notation we follow standard conventions for sets. For a string x we denote the reversal of x by xR, the length of x by l"l, and the concatenation of n copies of x by xn. Fot Z a set of strings, z* its transitive closure under concatenation, and Z+ rLLL denote with and empty without the nuLL string respectively, and I will- denote the set. The syrnbol $ wilL always stand for an input character which, by convention, does not belong to the I being discussed. L.2 Elementary Properties We first note that the trDre customary definition of is by states rather than rnodeslt3. IE fied (Hopcroft and ULl-nan3, Lettuna can be however acceptance easily veri- I2.3) that the classes of languages defined by the two methods are identical. Further, as we shall now 13 indicate, all the key properties on which theorems particular to depend, are a consequence of determinism, and are independent manner of of dpda the acceptanee chosen. One effect of the deterministic restriction is to ensure that once a sequence is uniquely of e-moves has determined by the started, the subsequent original configuration, eomput.ation and can depend on the input tape again only when the E-sequence has Eerminated. Thus if, the latter never terminates, then it is and can be eliminated. If it is finite cl-ear1y redundant and gives rise to the deri- vation (s, urA) ", ("t, ootB; where (rt, B) is a reading mcde, then such e-derivations can be eliminated by replacing the e-rule for (s, A) by rules of the form (s, A) -3t (stt, r,rtrtt) for ("t, B) -g* (s", ,"). Thus we conclude that each rule e-moves can only be essential i"n a dpda for derivations which cause the stack to decrease in height" Similarly, but for other reasons, having a multiplicity of states cannot be essential in derivations where the stack is increasing. State information can always be coded into the top stack symbol and thus transrnitted is ever rernoved. Thus in indefinitely, as long as no top Example symbol f.i-, while the distinctness of scates s2 and s3 is essential, the information which sl carries' i.e. that the stack is in the stack symbols. We che increasing phaseo coutd be coded into could therefore nndify M to obtain another equivalent one with state set just {s2, s3}, but with an enlarged stack alphabet {A0, A, A', B, Bt}, if we replaced cs by (s2, A6), and replaced (s2, the transition rules involving st by the following: A9) a -> ("2, At) ("2, A') -g+ (s2, AAt) L4 ("2, A') 4 ("r, n') (sz, n'; (r2, B') -3+ (sz, A) A q"r, (s2, B') -9> (ss, Another consequence of determinism BB') A) that is useful for showing that certain languages cannot be recognised by any machine in a given class, is the following observation, of which a formal version is stated and proved by Ginsburg and Greibachl: Observation The f.i effect of input strings periodic in a given word a (i.e. of the form on for n > 1) on a dpda is either (a) to cause the stack to never grow larger than some eonstant hei.ght, (b) to creete etacks which are periodic. in or e..'erywhere except for parts of some stack word bounded length at the top and bottom. Using this we can formaLly verify that e-moves, multiple states, and a mriltiple stack alphabet, are all essential- features of any dpda reeognisi.ng the J.anguage of Example 1.L. Formal- arguments of this ktnri are given by Ginsburg and Greibachl, and by Harrison and Havell2. We note that the normai- form given by Greibach for context-free languages shows tha.t in the nondeterrninistic case, multi.ple state sets are inessential"iS. e-moves and 15 1.3 Subfamilies We can define some important subfamiLies of D by specifying sinple syntactic restrictions on the machine descriptions: One-counter machines (C) are those dpda with just one stack symbol. In these the stacks can only be used to count, and hence configuration is best described as (s, n) where s E Q, n : o. a Such counters are closely related to the registers of Minskyl6. ReaLtime machines (R) are those with recognise strings in rea1 time, as opposed e-modes. They merely linear time" Stateless machines (S) are those with just one state. This state can therefore be omitted from the description of the configurations or transition rules. A consequence of the restriction is that no information can be transmitted during stack decreasing derivations. Thar the languages defined by each of these three classes is properly contained in L(D) follows from the previously vation that the l-anguage urade obser- defined in Example 1.1 cannot be recognised by any machine in C, R or S. For alL such classes we can have the additional restrictions that (a)FcQx{CI} or that (b) no rule is defined for modes in Q x {521. For the classes D, C, R and Srwe respectively define the sub- 16 elasses Dg, C0, R6 and So by imposing both and D1, each Ct, R1 and St restrictions (a) and (b), by imposing just the first of these. That of the implied containnents is proper can be easily verified by observing that if X is any one of D, C, R or S then I* eL(X1) but * t" t L(Xo), and {anbn | *: n} e L(X) but I L(Xr). Alrhough X1 rhus properly extends Xg, it does not use the stack in any more general wa5 and consequently many of its properties can be deduced directly from those of the corresponding X6. For simplicity, therefore, we shall only study the classes X and X6. We note that classes equivalent to D6 and R9 have been studied by Harrison and Harzef 2 lwho caLl them strict, and realtime strict deterministic respectively), and the class Hopcroftg. To show that latter, we have to Sg is just the rsimpler demonstrate that This is obvious since stack symbols can So by Korenjak and machines of the in it. with the property that A 5 e-noves are redundant clearly be elj,minated from the rules, while e-moves which n cause increases in stack height can be removed in the manner indicated in 51.2. L.4 Some F_amil"y Pfoperties There are some and languages further restrictions on families of machines that are related to the definitions above. A language L c I* is said to have the prefix property iff F c X+ and it is a e L *uB I immediate that L" these ?rom the definition of D6, C0, R6 and 56 all have this property. For L(Do) we can 17 sayr further, that it contains all the languages of L(D) that this property. For consider erty, "a and an M e D s.t. L(M) has the prefix that no accepting derivation has an intermediate new proB- nodify it to an equivalent Mr e Dg as follows: rntroduce special 0-simulating symbol at the bottom of the stack to uration. have Then for each accepting npde of ensure empty stack config- to M introduce an e-move state frnm which any stack wiLl be emptied by further a €-moves. The modified machine clearLy has the required propertiesl2. The quasirealtimel2- property restriction. nurnber E-moves is a relaxation of the realtime are now allowed to occur, but only a of tirnes consecutively. rf this to n stack characters can be removed bound bounded is say n, then only in a single e-derivation. up Thus, by changing the sta.ck alphabet so that each stack is now specified by symbols encoding bLocks of n + I old symbols (with a block of possibly fewer synbols at the top of the stack), and changing the transition rules so that these blocks are manipulated correctly, old machine carr be simulated by a neh/ one with the same states, the but with no e-nodes. A npre formal proof of a similar statement is given by Harrison and Havel12. Thus we can conclude the forlowing: Observation 1"2 The power of of the quasirealtime machines is no larger than that R. Finall"y we mention a syntactic rui.es. A dpda is conservative iff in restriction on the transition any ru1-e (s , A) -+ (s t , to) .ltitnere luJl I Ir 1t is the case that the f irst syrnbol of o is A. 18 Lemna 1.3 will show directly that this is not a proper restriction on the power of D, C or of the quasirealtime machines. However and S9 are properly respectively. restricted to the symnetric l-anguages, An innnediate property S Sy and Syo of these symrnetric machines is while the stack is that they cannot transmit information even increasing. | " : 1} Thus the language {an$an cannot be recognised by such a machine since the latter could not distinguish between being in the upward phase from being in the downward phase of the computation. Consequently Ey6 g Sg and Sy g S. 1.5 Derivations We now introduce some terminology and notation for describing the geornetrieal rbvements of the stack, that are in addition useful concepts in talking about derivations. The height lcl of a configuration c is the length of its i.e. l"l = lrl if c = (s,,,r) for 3, stack some s. is a stacking derivation iff l"l . l"'l and every intermediate configuration in the derivation has height The derivation " "' It is then written as c +(o) ct. The derivation is a popping derivqtion iff l"l t l*'l and every intermediate con- greater than l"l. figuration in the derivation has height greater than lc'1. It is (N.8. The notation t, * is used by then written as c +(o) "t. Stearnd but with the converse rneanings.) It is a positive derivation, written as c f empty stack. "t, iff no internediate eonfiguration has 1Y We define the slggElqg sequjlnge r:f a stacking derivation as a of nndes ui, Each ug is the nnode of the unique configuration of height (i + 1) in the derivation that is not followed by any configuration of height < (i + f) subsecruently in the derivation. sequence Note that if (s, ur) +(o) (s', ,') then ui is not defined unless lrl 5i+f S lr'1, and may not be defined for atl intermediate values (e.g. if the derivation consists of a single move corresponding to a transition rul-e of the form (s, A) * (s', o1) where lrt | > 2.1 The significance of pi for the configuration corresponding to itn is that it contains all the information about the previous that may computation influence the subsequent part of the derivation. lJe def ine analogously the popping _sequeEe derivation to be a sequence of ,st.gte,s oi. Each for a popping oi is the state of the unique configuration of height i in the derivation that is not preceded by any configuration with height < i. * e I' , j " i: o we define oirj to be the substring of o An s tar ting fron the (i + f)th character and ending with the jth. For trr index seE N is an ordered subset of the positive integers. N induces a natural- segmentation and of o into words oirj, Thus anv where i <j i, j eN. w.r.t. index set I (Fig" 1) are just the subsequences of these corresponding to the elements of The stacking and popping sequences N. We def c +(a) "t ine ttre segmenf"_ation of u in the stacking derivation qr11rt, an index set N = {ii, ... ir}, as the unique of input words o\ , ,... , cin_', i' with the properties ! qip, that therc exist iI, oin, in+l whm ig +1 = l. | , inal+1= I "' | , s. E. sequence 20 Trace of c t(g) "t i3 - l-o -z Trace of c *(cl) lEI "t - I .. ct. ". I I .14 I *{ I I I I 1o- - L rig. L t13 27. oinrin+1' d - oiori10i1ri2 and that if cf has stack o, ther: for the stacking sequence {(sr, Ar)} w.r.t. inal}, {io, (sr, Aa) +(airrir+i) (sr+i, rrrirrir*rAr+r) f,or 1 : r : n. In other words cirrir+I is just that part of the input stri.ng which takes the derivation from the last configuration of height ir+L to the last one is fol lowed by c'enfigurations of height iral+l" provided that neither srnal"ler than themselves. We can similarly define ui,3 for i < j and i, j e N. Analogously the segmentation can also be of a popping deqivation w.r-jJ:- N defined. oi,j is then the part of the input string that takes the derivatiorr from Lhe first configuration r:f height in to the firse configrrratiori of height j. r+here now i > -j" 1.6 Complexity MeasursE We shall describe the size of a dpda in terms of the following paralneterg ! t = size of stack alphabet q = siae tlf state set p = size af input alphabet h = Length of longest stack word occurring in the transition rules u = number of transition rules. 22 We machines shall express the complexity measures relating to these in terms of functions with these parameters as arguments e.g. X(t, g). We shall be primarily values, as the arguments monotonic in each become parameter. concerned with the asymptotic large, of positive valued functions We introduee the fol-Lowing notation to describe such behaviour: Definition The function X(x1r..., xn) is of order En iff n is the largest integer such that Tffi 1og(n+-I)x(y,.. ., y-)rog y Then, for exampte, x2 , wil-l be El, etc. even t) , o. /*, *logx will be go, *y2/*2 For convenience rre shall refer to an expression X, if it is of several variables, as poLynornial if it is above by some When and xl bounded multinomial. we are interested in the leading exponent in an expression, we use the following more detailed notation; Def ini tion ".. xm) is of order En(Y(x1, ... *n,)) iff 3 k, k' > 0 s.g.for all sufficiently large x1, tnr The function X(x1, Expn(kY) < X < Expn(k'Y). [ner" rxp0(x) = x, Expr*l(*) = 2Expr(x) tor r: o, and simiLarly for Lo9, which we shall always take to the base 2.] 23 L,7 Tradeoffs in Desg:'iption The dpda definition allows equivalent machines to have very dissimiLar specifications. For example any language recognised by finite state some dpda We auto{naton with n states can clearly be recognised with q = n and t = 1, and also by some with Q a by = 1 and t = n. investi-gate an aspect of this flexibil-ity of description by asking to what extent an arbitrary alent one, so es to dpda can be modified econosiise on one of the to obtain an equiv- Parameters at the possible expense of the others. I n3) as the class of languages specified by dpda with q j nI, t j nz and h 5 ns. Then the questions rre want We to ask define LD(n1, can be elasses. We n2 naturally phrased as inclusion probl-ems for these note. howet'er, that only the f irst of the fcrllelwing lemras depends on deterrninism. Lema 1. I LD(qo t, h) # tD(fl , tlz, n3) for an] n2, n3. Harrison and Havel"l2 show that the set {a%karlbk il":t jn, 1<m} can he recognised by a dpda with n states, but not by one with fewer. I 24 Lernma 1.2 LD(q, t, h) c LD(q, th, 2) for t > 1. Proof We give a simple construction for converting an arbitrary to an equivalent one which simulates it closely, but dpda does not require long right hand sides in the rules, or extra e-moves. We leave the state set unchanged but enlarge the stack alphabet to consist of all strings in f+ of length less than h, each typically denoted by [as..C], in the following Replace each rul-e and n ,_ 2, by the set of (s, [rl] g > (s ' (s, A) i* way: (st, 81 ... Bn), .where n e I u {el ruLes , [rn1] [n2 Bn] ) , *, I s.t. lrl < h - 1, and make the analogous replacenents for the rules with n < 2. one for each o e ctearly the new stack al-phabet will be of size t + t2 + .. . + gh-l " th, and if acceptance is defined in Then obvious way, then the new nrachine riill be equivalent Note the to the old one. 1 The new stack syrnbols that are not redundant in the rnodified of the form [uA] where Bu-ror'appears in some rule of the old nachine for some B and ut. Consequently the nachine are exactly those Rew stack alphabet is essentially of size at nost thu, and so really depends only polynomially on the original machine description. Frorn 25 this it fol-lows that the of the description, cons number of transition ruLes, and hence all need only be increased poLynomially by this truc tion. Note 2 The construction prGservcs the conscrvativc synurnetric languages have a normal forrn with h - property. Thus even 2. Lema 1. 3 LD(q, t, h) . LD"on"(qth, r, 2) for t > 1 and LD(q, 1, h) - LD(qh, 1, 2). Proof At the ean reduce expense h to 2 and make the machine consarvative. The new Q Each u of introducing ncn e-rpves and incraasing q we state set consists of i[t, ,J | " . Q, o e r+, l.l < h] transition rule (s, A) -I* (rt, Bt ... Brr), with n > 2, we repl-ace by the set: (s, A) Jt 1["', Bz (["', Bi... (["', Br,] Br,], Br-r) nrr] , rr) , Bi-ilj* ([r', ri+l 5 ("', B"] , Bi-rBi) for r < i Bn-rrn) This construction merely ensures that instcad of a large stack word being added in one move, it is buiLt up in a succession of specially <n 26 created E-rnoves. I Note The non-redundant subset of the new state set depends only polynomially on the original- parameters and is bounded by uh. The rnachine polynorniai- description as a whole is again increased only ly. lem+e-L.S. lD (q, t, h) e LD(qt, 2, h.ftogtl ) . Proof We r1 llogt l. code each stack syrnbol into binary words of fixed length For each old state of the machine, Ehe nerr one has to be able to interpret the top coded word in the stack. To do this it traces down, via e-moves, a path of a binary tree of depth ftogtl while popping this top word. Thus the new machine needs a tree with about t nodes in its staEe diagram for each original state. Thus the state set is now of size qt, and h has increased to tr.flogtl. il I,le e,,:ncl"ude and h. By the determinist{c condition on the transition rulesr u is clearly mayr bounded above by qtp. On the other hand rhe input al-phabet in theory, be arbitrari.ly Large. llowever, the number of char- acters whieh ded by reJ-ating the other parameters p and u to q, t prardulee distinct behaviour from the automaton is boun- by the other parameters in the fr:l-lowing way: Each input char- acter can occur in at nost qt rules, each of which has one of rnore than qth+l different possible ri-ght hand sides. Thus the no 27 machine can ters, and only distinguish up to (qth+t)et aifferent input charac- if there are more, then at least identified with some of them can be each other without influencing the computations. 1.8 Family Relationships lJe conclude by su'nnarizLng in fig. 2 the retative. polter of the classes of automata with which we shall be most concerned., the foLLowing relations for arbitrary classes X and (i) (ii) X -> We use Y: t only if V M e Y, I M' e X s.t. L(M) = L(l'l') X ---+ Y only if V M e Y, I M' e X s.t. L(M)$ = L(M') where $ is a disti.nguished syn$ol. For the definitions of N0, T and dB - S, see chapters 3' 4 and 8 respectively. Al-1 the relationships in the diagram ae irnnediate from the definitions or well known, with the exception of the position of N9, which we prove in Chapter 3, that of 2-tape acceptors, explained in Chapter 6, and that of S, for which we mit the proof. 28 Family Relatiernships D dpda rI a Do f,l-acce rP I I I B-S a realt i \tI t I I t finite-turn I , I ce I c; I one-c0unter I '0 s t state ess r (two-tape acceptors) .- t Fsa ./ a finite-state automata Fig.2 simple 29 Chapter 2.L EASILY DECIDABLE 2 PROBLEMS Introduc-t,ion. Here we shal1 discuss some properties of deterministic languages which have been previousry known to be decidable. we give decision procedures for these, and show them to be easier than the procedures f,or equivaLence and regularity that we derive i.n later chapters, in the sense that they all require only polynomial time. These (i) "easily decidable" menbership - does problems for t e L(D) include a e L? (ii) eurptiness - is there sor€ o s.t. o e L? (iii) finiteness - is L finite or infinite? (iv) totality - is there some o s.r. a l. t? (v) equivalence with reguLar set Lt - is L (vi) has L the prefix = Lt? property? first three of rhese are decidable for context-free Langrrages in general. but the last three rely on the determinist.!"c The restric.tion. Techniques for proving such decidability results given by Bar-Hillel, Perl"es and ShanirlT and by Ginsburg and Greibaahl. I{ere we shalx concentrate on the complexi.ty of just of thern. we know of rro arguments are some giving interesting lower bounds for the time complexity of any of these decision probrems, and, for this 30 reason, we sha1l not Ilowever, measure for linger to give exact uPper each problem there bounds either. is often an alrernative natural of complexity. For some of these we can give upper and lower bounds which particular differ by only a modest arnount. measures Examples of complexity include, for the emptiness of such problem, the maximum length of shortest strings accepted by non-empty machines of a given size, and, for the reguLarity problem, the maximum size of the smallest finite-state automata equivalent to some dpda of a given size. 2.2 Emptiness We shaLL consider from two viewpoints the complexity problem, which plays a The proof crucial rol-e in of decidability of PerLes, and Shami-rI7 of this subsequent decision procedures. emptiness given by Bar-Hil1el, for context-free languages depends on the derivation tree3 of the gramnar for a shortest accepted string. If the grarrrar has x non-terminals and right hand sides never longer than h, then in any derivation, starting with a non-terminal, of a shortest string, no non-terminal need repeat along a path in the tree, which therefore can be taken to have depth < x. lrle now define the depth of each non-terminal to be the depth of the shallowest derivation tree of which it is a root. Then, knowing which terminal-s have depth < i, we can find all those that have depth < i + 1 siurply by looking for productions in which all the on the right hand side have depth < i. list of productions at most non- symbol-s Thus by going through the x times, and establishing the depth of 31 all the non-terminats, has a defined depth Thus (a) we can deterrnine whether i.e. whether the language is the starting symbol empty. it fcllows that there is a decision procedure for empciness which works in time depending only polynomially on the gramar description, (b) and the shortest accepted string can be no Longer than hx" Ttre standard constructionr for a pda with q states and t stack synbols gi.ves a granmar with qzt non-terminals. Further if (and only if) h is taken to be a constant, the nunber of productions in the graunnar wiLl depend However, from the notes rde can modify any pda only polynomially on the parameters. to Lernmas 1.2 and 1.3 we already know that to an equivalent one with h = 2, with only polyn'cmial g'rowth. Thus we can concLude that: Lenrna 2. L Emptiaess time. ft:r pushdofrtn automata can be tested in po1-ynomial tJ Since Ehe standard construction of the graflurnr from the nactrine does nst change h, it aLso foll-ows that: Lerrlna 2.2 If M has parameters h > l, q and t, tken 3 a e L(M) s.t. lol : hq?t. il and L(l'l) is not empty, 32 We can show that for q = 1 the order of this bound is achievabl-e. Example 2.1 Let M e S9 be the simple machine defined by qh Ai+AT+t lj -3* where the forl"<i<t n starting syurbol is A1 and acceptance is by empty stack. Let min(Ai) be the length of the unique string generated from A{. Then from the productions it is clear that min(Ai) = 1 + h.min(Ai..1) for Since min(Ag) = 1. < i < t. t-i 1, it follows that rnin(Ai) = f, hr. Thus M accepts r={ just one word, and this is of Length of order nl(t.logh). I For q > t we can find a class of rnachines which achieves bound of order fl(qt,logh). Ex_ample 2. 2 sq}, stack symbols Let M e Rg have states {rt, {Ai, ... At-1, B}, and transition rules: (si, A3) e* ("t, a!+r) 1 j i : q, I j j . t - (si, Ag-t) -3* (si+r, BAI) 1J i < q (sqn As-t) -3* (sq, A) (si, B) -1* (si-1, A) I . i : q 1 a 33 Let the starting configuration be (s1, A1) and the unique accepting mode be (sr, $l). This is a generalisation of the previous example. occurrence of Ai proliferates to Each produce an exponentiaL number of At-lts, but now instead of these being renrcved by a single input character, they are each allowed to proliferate themselves all over again for a different state. For any configuration the depth reacheci in this sequence of renewal"s is given both by the state and the nurnber of r:ccurrences of a B in the stack. By a second induc- tion process, similar to that given in the previous example, it follows that the unique string accepted by M is of lenglh of order nl (qt. logh) We . I leave unresol-ved the gap between this lower bound and the previousLy derived upper bound of El(q2t.logh). 2.3 Finiteness Bar-Hillel-, Perles, and ShamiriT show that a context-free Language is i,nfini-te if and onLy if there is some derivation tree which has a path atong which some non-cerminaL repeats and in so doing generates distinct terminal strings. Thus to test for fin:,te* ness nre produce a polynomial size grarunar, eliminate redundant non* terminal-s (i.e. 'those that do not generate terminal words, or cannst be derived frorn f.he starting symbot), and then embeddi.ng test for the proper properEy. This last test can be done by looking at each of the remaining prodr:ctions that have some terminal on the right hand 34 side, and searching through the other productions to see whether a self-enbedding cyctre of productions can be found for this given non- terminal. As all this can be done in polynomial time it follows that: Lerma 2.3 Finiteness for pushdown automata can be tested in polynomiaL time. I 2.4 A Normal Form There is a well known3 normaL form for dpda, which is very useful for our purposes and fits in naturally with our definitions. Def ini tion A dpda finite is nunber of loop-freel iff every input word can be read in machine moves. This condition excludes the possibilities of either (i) from some reachab,Le configuration being undefined, or (ii) nnves there oceurring an infinite sequence of consecutive €-moves, thus preventing further inputs from being read. T,enrna a 2.4 Every dpda M has an eguivalent loop-free Mt stack alphabet, and just one extra state" with the sane 35 Proof I^Ie f,iret au$nent 1"1 by introdr-rcirrg I transition rules (s', A) ..3t (st, A) for without changing each aeI mode we add and new and A e iin the rules (s, A) ", (u' , (s, A) --4, does not previously oecur. stopping, the s' F. For each reading whenever new staEe maehine will now continue A) Thus instead of to read the tape and reject it. To exci-ude the otber possibility mentioned above, we to note that in any infiniie e-derivation there must be first some configu- ration which has a smallest stack. If the mode of this is (s, then hy replacing the Eransition (s, A) :S' (nt, A; for each have A), rule for this by the set a, we again ensure that bire r*:sL of, the tape is read and rejected. These rmdes can be effectively found by appropr!"aL'. emptiness teses. il A11 the neru rul-es introduced serve the same do not create additional comptexity eonsequently r*e role in any significant and therefcre sense. shall frosi now on assume that the loop-free machine i.s nf eseentialXy the same size, in every way, as the original one. Def ini tion A dpda i-s in pormal_rEgmrr accepting mede i"e a reading iff it is loop-free, mode, and every 36 Lenuna 2 .5 For each M e D with q states and t stack syurbols, there is equivalent Mt in normal form with the same stack alphabet an and 2q + L states. Proof }Je "t first make M loop-free by introducing a distinguished state as in the previous leqma. For each other state si E Q we intro- duce an sitt. The role of si" is the same as that of si, except that the former indicates in a derivation that an accepting been passed since the Thus for mode of M has tast time an input was read. each accepting e-rnode of M, we replace (si, A) - r* (sj, o) by (si, A) - 3+ (sj' , u:) " For each readiug rule (si, A) -3* (s5, ur) we add the extra rule ("i", A) -3+ (s5, rrr), and for the extra rule (si", A) -:' we (sj, M and accepts the (si, A) ur) we add (s3", ,l). This new machine clearly simulates strings if €-n each e-ruLe define the new set of accepting same modes by F' = 1(si", A) | (si, e) is a reading mode) u {(si, A) | (si, A) e F and (si, A) is a reading mode}. I 37 2.5 TotaLity A,s corrol-lari"es to Ehe lemna above, decidability proof,s for properti.es whi.ch are known to be undecidable in the nondeternninistic case3, can now be easily derived. CorolLarv 2.I For each M e Dr3 M' e D with about twice as many states, which accepts exactl"y the cornplement of L(D). Proof We change M to M" in normaL form. Then we redefine the accepting:,rnod,eg to be jr-lst those reading modes which are not aceepting nodes in U" " [J Corollarv 2.2 Total"ity ferr dpda can be tested in polynonial time. Proof ConstrucE ttre eomplement machine and tesE cqrq.l legy !. for emptiness. I 3 For M e D, and Mr a finite state automaton, the following can 38 be tested in time depending polynomially on the parameters of M and M': L(M) = t(M'), L(M) n L(Mr) = /, L(M) c L(M'), L(M')cL(t{). Proof rf Mr has state set Qt of size qt, and M state set Q of size then both machines can be simulated together by a dpda which that it has state set Q * Qt and is able to mirnic Mr in the finite state control at the same time as simulating M. Provided that M is in normar form, each of the above questions resembles M, except can be tested by defi.ning acceptance and testing for emptiness. I suitably in the new machine q, 39 rr THE EQUMLENCE I.Ie have aLready observed of PROBLEU that an equivalence test for a class automata can provide a method of verifying the correctness of a given machine by eomparing it with another of the same kind. In addition, we can attribute other useful properties to classes for which equivalence is decidable, that give us increased confidence in the possibility of systematicalLy analysing and handling particular machines contained We in them. first notice that for aLl automata with a decidable membership problem, inequivalence can enumerate is partially decidable, For all strings over the input alphabet and test we whether the two machines behave differently in each one. Then if the machines are inequivalent, we are sure to recognise this when we reach the first offending input string. We then observe that all the automata in a syntactically defined class, such as D, can be enumerated in some lexicographic manner. It foLlows that if we can decide equivalence, then for automaton in Conversel-y, such a class we can enumerate if we can enumerate all all other equivalent machines equivalent to any ones. any partieuLar one then, since inequivalence is partially decidable, can decide equivaLenee by running uri.eh ehe this we enumeration simultaneously partiatr decision procedure. Thus we arrive at the following intuitive interpretation of the signif,icance of this problem: equivalence is decidable if and only if for any automaton in the classr r{e can en!.merate al-1 other 40 \rays of doing equivalent computations. Undecidability then means that the relationship between the automaton descriptions and computational behaviour has a and we can no longer hope different ways their distinctly higher order of complexity, to be able to describe adequately all in which the freedom avaiLable within the Lhe class can be used to perform a given task. The partial decidability of inequivalence has also the following imnediate consequence for all the classes we shal1 study: Observation If equivalence is partially decidable, then it is decidable. The decision procedure consists of simply running the decision procedures, for equivalence and partial inequivalence concurrently. Thus, except when we can derive meaningful results about the complexity of an equivalence test, we shall, for the sake of sirn plicity, be content with proving decidability by showing partial decidability. This section consists of three chapters in each of which distinct subfamily of D is shown a to have a decidable eguival-ence probl.em. There is an essential unity in the methods used for aLl three, for which the basic inspiration StearnsS. They show that for comes from Rosenkrantz and dpda recognising the LL(k) languages, there is an easiLy derived canonical forn with the property that two equivalent machines must, after reading the sarne input rdord, any 41 have stack heights differing by less than a ce.rtain constant. such machines can therefore be simulated by a single dpda Two with a t'{ro track tape, and equivalence verified by a suitable emptiness test" Wtrat we sha1l show is that even for m<rre general classes, where such close relationship between the st.ack movements need occur, suitable machines. pushdown automata can be devised to simulate equivalent no - 42 Chapter3 A REALTIIfl FAI.{ILY 3.1 Introduction The syntax of progranrning languages is customarily defined in part as a general context-free granmar. Most frequently, however, the languages generated by such granmars are not only deterministic, but also expressible as Rg languages, i.e. those recognised by realtime deterministic pda accepting only by empty stack, and even subclasses of these. such subclasses as the simple languagesg, and the LL(k) languages8 which generalise them, have been studied with exactly this nnotivation, and equi"valenee tests have been found for these particular ones. principal restriction on LL(k) recognisers is that they require essentially only one state, and therefore cannot transmit The information during stack decreasing derivations. Thus, for example, in Al9o1 60, after some expression'((.. (( Expression ))..))'with arbitrarily nested matching brackets is remenber whether the contents of the gcanned, no such machine can innermost brackets was an arithmetic or a boolean expression. since'if ((..(( Expression ))..))' has distinct valid sequeLs depending on just this condition, we con- clude that no essentially one state machine can recognise arithuretic expressions as specified Here we in that language. shall define a fanily N6 of quasirealtime dpda such that L(No) properly contains the tl(k) languages. N6 machines may have arbitrary number of statcs, and can easily accormodate the above an 43 mentioned example. languages In addition, however, they extend the LL(k) in a much more significant wave thus indicating that generalisation of the equival.ence test is an important one. procedure for LL(k) our The languages given by Rosenkrantz and SEearnsS relies on the fact that nin(c), the length of the shortest string accepted from a configuration c, is a very well behaved function of the stack c(rrents. deduced The value of this function for any configuration can be to within an additive constant from the contents of the stack atone, and, further, can only change a bounded arnount from one configuration to the next in a derivation. This property does not hol-d for concept scme languages in L(No), such as {anbcn} u {ande2n}. The of tthicknessr of stack words, central to the proof of Rosenkranta and Stearns, is no Longer applicable here. ldhile the inclusion pr6blem for sinnple and LL(k) l.anguages is at present open, we can already prove this to be undecidable for Ng" In fact we strali show that L1"L2 is undecidable where L1 is a simple language and !,2 e L(Ng) l'ras ja"rst two states" We eonjecture that the property on which the validity of rhe equi-velence test fnr Ng depends hoLds betieve t-hat a n:r*of of this ?he enphasi,s 'i"n this "*rc,r-rtd c.'ixapter itse1"f,, rathex'tlran the elass Eire iucreased generality sf l-{60 also for R6 in general, and be illuminating' is ern the derision procedu::a which !ile use merely o,+rr E,eehnique to demnsEraee over previous ones. 44 3.2 Nonsingular We Def Automata defi"ne the cl-ass Ng of nonsingular machines as fol1ows. inition M e D9 Vr^,, is nonsingular iff 3m>O ot , f*, s, st e Q where lrl s.t. > L(s, urto) = L(st, ,t) + L(st, ,rr') m, = y'. Theorem 3.1 (i) LL c L(No) whcell, is rhe class of all LL(k) (ii) languages. L(N6) e L(Ro). Proof (i) For a configuration c let rnin(c) be the length of a shortest string in L(c). Then from Lernma 8 of Rosenkrantz and stearnsS, we can deduce for eertain canonical LL(k) recognisers, when translated into our terminology, that there exist positive eonstants .c, s.t. for all s, st, o, and rrtt, (a) | nin(s, o'o) - min(st, (b) 0', lrl t *itt(", r,r) rrrt) - min(s, ,^,) | : l, and 2 lrl, provided that min is defined throughout. It follows directly that any dpda L' with these properties , 45 belongs to N6, with nonsingularity constant m equal to 1,. that for the sinple9 We note machines this constant can alwavs be taken to be sero. is proper, consider the following To show that containrnent the dpda M recognising A: a (sr, A) i, A) --> (sr, A) (s cs = (s1, A), In M l-anguage {anbcn} u {ande2n} for n > 1. (sl r AA) (s2 r A), (s2, A) - r'+ -$ (ss, A), (s3, A) - 3'+ (s4, -) b d (sz, A) A), (s4, e,) 5 (s3, A) F=Qx{CI}. for any i, j and m, n where m f n, t (sir An) # L(s5, An). Thus !{ e N0. However, it is inplicit abova: that :\e in conditions (a) and (b) value of min can only change by up to a bounded amount: from one cnnfiguration to the next in a derivation. For any dpda accepti.ng che above language, ubves this cannot be true for the set of in which a synbol d is read. [,Ie therefore conclude that r.(M) g LL" (ii) nf a rnaetline is nonsingular, then the appropriate constanc lri clearn"y sees an !"ipper bound raeion can eharage in two eonfigr"u:ra&iones in erra Lhe arount the height clnc csurse of any sequence of of a configu- e-moves, for any sueh a derivation must necessarily be equiva- Lent. Thus, from an argunen8 given in 51.4, it follows that the same l"anguage can be aeeepued by a realtime machine. 46 To show that the converse is not true, consider the language ianbcn) u {andczn}, which clearly be1-ongs to L(Ro). Let a dpda M recognising this language reach the equivalent configurations (s, and (str rt) after reading a2nb n. Then from Observation and and respectively 1.1 (gf.2) ve for .rr) some large rhar lrl is abour rwice loltl , and further that the effect of ck for some "small" k is to reduce (st, ,t) to a configuration whose stack is periodic right to know the top. Thus for an infinite set of distinct values of n, string ek takes the two equivalent configurations to some new ones with stacks stil1 differing in height by about n, but with one srack being a prefix of the other. Thus M I No. 3.3 Alternate We Mn, I Stacking now describe a way for simulating Erro of constructing a single stack machine, realtime dpda M, fr together. A configuration of M' can be specified by a pair of states, one from each segments of M and il, and a stack which consists of words from the stack alphabets two machines. The simulated configurations recovered by taking the corresponding of distinct) of M and fr the can be state component, and concaten- ating the appropriate set of alternare to obtain the original stack. (assumed of alternate segments On reading each of the stack of input symbolrM' simulates M and M by sirnultaneously manipulating the two topmost segments in its stack according to the respective transition rules of the component nachines. Mt n 47 More formally, distinguishing the notation for M and ll by over-lining everything concerned with the tatter, so that M' has stack alphabet f u T and state set Q * Q, we describe a typical configuration of Mt by t | -1 ([s, sj, where only the olo2o3or] ... uozn-262rr-lo2n), last segnent, i2n, may be enpty. hrithout loss of generality we shalL assume that the tno topnost non-nuli. segments of the stack ar€ o2n-t = oA and i2n = ;f. If the traneition'rulee of M, il specify that (s, A) t, ("tr lrof ) and (1, f) tr (6r, ii'), I if rrrt is non-nul.! and nefu{A} otherwise, and sinilarly ror i) then for input a in M' , rhe segment urAil will change ro ,niir'-.t, and the state rrom [s, i] to [rt, ir]. (rhere n e An iuoportant inplication of this notation is that if, for exauple, ii = A, then the words un and r,lt wilL merge into one segnent si.nce they are from the same alphabet. 3"4 Main tr('esults The alternate stacking eonstruction can ensurc is only useful to us if we that the simulating naehine is itself a pushdown automaton" Definition Alternate stacking for a pair of nachines, for a given set 48 of input strings is said to succeed iff the simulation of them together for these strings by the above construction produces stack segments We of only bounded size. define an input- string o to be live for a dpda M iff it is the prefix of some accepted string, and a configuration c to be live iff L(c) * l. Theorem 3.2 For M, M equivalent nonsingular dpda, alternate stacking succeeds for all live inputs. Proof We shall shsw that in the simulating machine described in F3.3, no input a that is Live in M, il M' constructed "at as lead to a configuration ([s, s] , oriz ... u,;) with lrl > z(22 + m). Here z is the maximum of the lengths of the shortest strings accepted by dpda of size no larger than M and M, and m is the larger of the nonsingularity constants of these two machines, which are taken to be realtime, Let us assume the contrary. Consider the configurations e and E of the M and fr respectively at the moment when the bottorn symbol of segrnent il was actually placed there. Let s be the input read since that time, which has taken u, fr to the present live configura- 49 tione cr, Er respectively. By our choice of E and a, all the configurations in the a-derivation in fr ha.re height 3 l;1. Also, by virtue of the alternate stacking construction, all the configurations in the a-derivation in M have height > l"rl, for what is finally the i segrent in Mt has been in existence throughout the o derivation, and hence the segment below it cannot have been increased in the neantine (Fig. 3). Let ;l B be a shortest string in l"rl = l;1. f.(ir), and let B = BtB3 where lBrl : lrl uy the realtime property. 'Also B nust be a shortest string in L(cf ). But then B is +(Br) E2 and Then the concatenation of segments induced by the popping sequence in each one taking some (q,A) to sorne (s2, A). If such segment must also be nininaL and Thus Bl consists of a sequence of B is minirnal, of length no rDre than such nininal M, each z. segments terminated possibly by a proper prefix of another such segment. Let Bzt a prefix of 83, be the completion of this last segment. A^ A^ ca and -c2 4 cg. ClearlY c2 4 Then 1et l"rl-l"gl:lsrll". Now consi"der a shortest string y taking M from cto ca. that lyl : " + l82l 7 the def initions of E2, E3 Let c --Lt ct g. (a) "tt",rte Then 22. Then l.i - l"'si :22. But if M and But from (a), (b) we conclude that l'1 and (b) are eguivalent then by construction the observations that l"l : l"tl ca and _l - e ler | 3. : l;l 50 M: It: fig. 3, Derivations in Proof of Theorem 3.2 51 l"rl : l"'.1 + 2z - lillz. Thue if lil t " - 2z > m then we have a contradiction to the non- singularity condition. l{e therefore have to eoncLude that for equivalent nonsingular autonata, alt,ernate stacking succeeds for alL Live inputs. n Theorem 3.3 If alternate stacking succeeds for all live inputs for all pairs of nachines M=E in sore class Ic D, then the equivalence problem is decidable for X. Proof ig permitted by Lema 2.5, that M, M are in Assume, as form. First can grow llr for that we know the bound to which stack in the alternate stacking sinulating nachine for ll Then we can t(ld') is Buppose normal segments and il. effectively construct a pda Ut' with the property that ernpty iff l,t = il. M" nimics the M, E by encoding the top segrent alternate stacking nachine of Mf in its f inite state controL. As Long as this top segment never gets larger than the given bound, M" accepts iff exactly one of the configurations it is eimulating is in an accept mde. I.lhen the bound is exceeded, M" proceeds nondeterninistically to ninic one of M or il, and acceptsas the appropriate machine would. By assr:mption for live inputs, , if M= Ii then alternate stacking will and the stack segment bound succeed will only be exceeded 52 once nothing rncre can be accepted by M and M. 'Thus l-t" will be empty by construction" Conversely, if M" is empty, then clearly no input can produce different behaviour in M and il, which are therefore equivalent. Thus, if we have an a priori bound, \Je can test equivalence by constructing this pda and testing it for emptiness. However, if we do not know this bound, by enumerating and even testing for empti- ness the possible candidate machines, we can obtain a partial decision procedure, That is,we construct pda of the form M" for assumed segment bounds of L, 2, successively. If M, M are not equivalent, then none of these constructed nachines can be while if they aren then one of We them must be. therefore have partial decidability and hence, by earlier observation, decidability. Corol empty, an tl lary The equivaLence problem for nonsingular machines is decidable. Proof Immediate from Theorems 3.2 and Note 3.3. D l. Nondeterminism in the siurulating rnachine is an inessential convenience, which can be avoided. M and fr couLd have been preprocessed to recognise instantLy whether a configuration reached is 53 live, and to enter a distinguished dead state if it is not. A simulating deterninistic pda could then be easily constructed. Note 2 I{e can show that the order of the bound on the segment growth derived in Theorern 3.2 is achievable for simpl-e urachines. For these we have already observed that the nonsingularity constant m is zero, and, in 92.2, that z is El(t.Logh). Thus the bound z(22 + m) is of order El (t.logh) also. Let A1 be the starting Example ht. syrnbol for the simple grarmar of 2.1 generating the singleton string {an} where n is of order Then consider the simpLe languages generated by B and C resPectively: B -5 nrBz n2 -3, n1f2 c j'lrcrcz C2 3, Cp2 n2J*41 Czj*n Bl-3+A C1 3n set {arban+r I r 3 1}. However, if s - srx$6n, then t l* E , but C 3t Al. Thus in an alternate sEacking rnachine fon these tvo, the top segment can gro\t to size n, Both B and C generate the which is of order EI(t.logh). Note 3 To illustrate a s}ightly different way in which the success of alternate stacking can be used, we outline a proof that the equivalence problem for symetric dpda, Syo ( 1.4)rcan be solved in polynomial tine. 54 We recall from the notes to machine can be transformed Leruna 1.2 (91.7) that any such to one, only polynomially increased in size, with h = 2. Since alternate stacking clearly succeeds, and the enptiness test is itself polynoniaL, it remains only to prove that the simulating rnachine is itseLf of only polynomial size. The key observation here is that, instead of having to mernorize the ruhole top segnent, to remenber it is now sufficient for the finite state control just the top symbol of the next-to-top segment, and to treat the top segment itseLf in the normal stack-like way. The property of Syg machines from which this can be derived is that any rcve that causes the stack to increase, also eauses the value of the function min( ) to increase. We note, honrever, that for the class 561 and therefore also Syg, algebraic properties can be derived9 which give nnre detailed insight into these restricted classes than our more macroscopic approach. Thus for Sy6 grarnr,ars with a fixed terminaL alphabet with h = 2, an equival"ence test on the lines of Korenjak can be obtained, that works and and Hopcroftg in tirne cubic in the number of non- terminals. 3.5 Undecidability gE_Inclusion FriedmanlS prolr"s the undecidability for dB - S schemas (chapter of the incl-usion problem 8) by showing that for each instance of the Post Correspondence Prob1em3, a pair of appropriately related dpda M1, M2 con be constructed iff it has no solution. with the property that L(Mr) c The construction is valid for certain L(Mz) 5.'., classes (e.g. dB - S) which accept b:y enpty stack'buc are essentially not real-time, and also for rea.Lcime fanil-ies that do not tiave this empty stack We restriction (e.g. use a S). similar formulation of the problem but, by taking a refinement of the correspondence prnblem that is implicit in the customary proof of its unsolvability, undecidable even Because we can show for a real-time class with of the slight novelty in'roLved, we that inclusion is enrpty stack acceptance. shall digress here to give this undecidability proof. Theorem _3..4 For M1 e S0, M2 e Ng ir is undec,idable whether i"(M1) . L(Mz). Proof The instances of the correspondence problem just those obtained directly from Turing standard proofs shord that for of its undecidability. each Turing machine that we take are machine computations in the Thus Hopcroft and lJl1man3 a correspondence problem with the foll"owing properties can be constructed: Let X = xl, .". xk, Y = yle .,r yk be the words over a finite alphahet X. kle def{ne a sequences sequence of non-nulL of positive 5 k) to be a solution (or partial sotution respectivei.y) for X, Y iff x:xil .". (or a proper "in is equal to integers i1, io (a3"1 prefix cf,, respectively) )llli1 ... yirr. Then the construetion is sr.ech that a solution exi.sts for X, Y if and only if the !:he stri"ng; Turing nachine has a terminating eomputation, but no partial solution 56 for Y, X exists under any circums tances. lle construct two dpda M1, M2 which both reject strings not of the form aRB$ where c is a string over the integers {1, ... k}, * 8eX and$is a distinguished terminating character. Both initiall-y a special syrnbol A at the bottom of the stack. M1 crRB$ place machines is a simple (Sg) machine which accepts exactly the strings where B is the word indexed in Y by the sequence o. Thus M1 records aR in its stack, matches it with B, and accepts iff the natching never fails and the $ is read when there is only the symboL A Left in the stack. It rejects al-1 other strings by placing a special symbol on Lhe stack which permanently freezes its motion once matching M2 fails. also tries to match o and B, but If natching is successful reached and the $ specified by X. is read when the syrrbol A is at the bottom of the stack, then rejection occurs. Other- wise, if matching fai1s, M2 goes effect of any X-input is to Wtren now as into a second state, in which the pop the stack without changing the state. the A is reached in this second state, all input synbols leave the configuration unchanged, except for $ which pops the A. If acceptance is defined by empty stack then clearly the length of a shortest string accepted frorn a configurafion is just the height of the configuration. Thus it follows that M2 e No. If there is a solution to the chosen instance correspondence problen then the inpuE of the string which specifies it is, by definition, accepted by Mr but not by Mz. Thus L(Mr) - L(Mz) implies that it has no solution. Conversely, if it has no solution, 57 and oRB$ is any string accepted by Mt, then there can be at most partial solution specified by an initial segment of a. a However, by our particular choice of X and Y, we have ensured that at the point where the syrnbols * c, B faiL to match in M2, the stack of M2 has fewer left than that of M1. Since M2 proceeds to accept any string from I-$ that is long enough to empty its stack, it must accept in particul-ar, since "Rg$ the remainder M1 will take aE least as long to accept of this string as M2 will take to reach the bottom of its stack. Thus if there is no solution then L(Mr) . L(Mz). Since for any Turing machine rde can construct an instance of the correspondence problem, and hence also the machines M1, Mz with the above properties, it follows that if we could decide inclusion for such dpda, then we could decide whether Turing nachines had terminating computations, which, howeverr Ite cannot. I 3.6 A Conjecture about R0 We conjecture that Theorem 3.2 hol"ds for M, fr e R9, and hence, by Theorem 3.3, that the equivalence problern is decidabl-e for $ere we shall- outline the proof of a property of not Btrong enough to prove the eonjecturer maY R9 R9. which, though throw some light on it. Lema There exists a function F(Q, in 9, with the property that for t, h, 9), aslmptoticall-y linear t:tro machines M, il e Rg rdith 58 par:rmeters appropriately bounded above by q, t, and h, if c, i tt" equivalent configurations of M, il respectively, then if c t(a) cI nhere cris live Jcr' s.t. c f(cr') c1 and C--> and (l l"tl - l"l t 11 where 2 e., then l"t | - l;l > F(q, t, h, Outl-ine Proof Choose the crr to give the shortest stacking derivation from c to c1. If ct, c" are any two configurations occurring in this | 7 zt (where z is as in then it is easy to verify that min(c') > min(""). derivation with l"' | segmented segment 1"" Theorern Thus o,t can be into lengths of no npre than 22, such that in i.t takes c, and therefore also with increased values of min. The i, to 3.2), each further new configurations nain argument then is to show, using this segmentation, that no subderivation of the or derivation in fr can produce a stack drop of nore than G(q' t, h), El function. This is where G is an done by assurning the opposite and inducing on the number of states that can be reached at this lowest level by popping derivations from the previous configurations in the sub- is that if min(s1, ,r'; ' nin(s2, rrrott) for some sI, s2, o, ot, ,tt, such that ltt l/lr"l 2 zt then 3s s.t. (s, ,rl) is not reachabLe from (s2, otl"). Since the reading of the successive segments of qt must lead to a set of pairderivation. The observation used wise inequivalent configurations, the statement of the lemma then easily follows. I P.\ " 59 Chapter 4.L FINITE.TURN 4 AUTO},IATA Introduct,ion We consider a family of deterninistic pushdown automata on which the only restriction is one on the movement of the stack. Using a technique related we show to the that the equivalence one given probLem in the previous chapter, is decidable. fuain a pda to try to simulate the tested machines, but now we this is build an intrerently nondeterninistic one. Since two equivalent machines in this family may have totally unrelated stack movements, a deter- ministic simrlation by a pda is no 1-onger possible. A speciaL case are the l-turn machines. the inclusion problem for these is in our proof of Theorern The wel-l- known, and undecidability of is also implicit 3.4. A further restriction gives a class which is intimatel-y related to the tr*o-tape acceptors of Rabin and Scott6, for which equivalence has already been proved decidabLe Bird7. From by our proof, therefore, another equivalence test can be abstracted that is not directly related to that of Bird and, while possibly less efficient, involves a technique of apparently general- more applieability. Although the finite-turn property is essential to the proof of the effectiveness of our main construction, we know of no pair of equivalent dpda for which a simulating machine of a broadly sinilar nature cannot be found. Thus it is possible that, when more detailed knowledge about the structure of dpda in general becomes 50 availabLe, our methodology nay be extended to prove decidability for the unrestricted case. 4.2 Definitions We istic define the class T c D of finite-turn dpda as a determin- anaLogue of the nondeterministic class studied by Ginsburg and Spanier I 9. Definition A derivation it in a dpda is a stroke if either no single nove in decreases the stack (i.e. an upstroke) or if no single move increases the stack (i.e. a downstroke). Definition A dpda M belongs to the class n - T, for some n > O, iff every derivation in M fron the starting configuration can be no nore ttran segmenred into n + I strokes alternating in direction. Definition A dpda M e T iff M e n - T for some n > In other words, the restriction O. we impose on T is that in the set of all derivations from the starting configuration of a machine, there is a bound on the number of times the direction of the stack movement can change. 51" Clearly the language {arbr I t: l}m belongs to r,(l2rn - t] - rl Further, -in general, L([2"] - 11 = but not to L([Zr - Z] - rl. r,([Zn - f] - T). This is because after an even number of Lurns the stack is inereasing, and so after its last turn a [2"] - T machine can only proceed to recogni.se a regular such a machine cannot be essential set. Thus the last turn in for recognising the language. In particular we note that L(O - T) are just the regular sets' lJe give Erilo properties of f inite-turn autornata which can be derived from analogorls properties given by Ginsburg and SpanierI9. Ttrey both depend on l'lt the fact that for any pda M, an equivalent way and, in addition, to in its state set lthether the derivation it has been doing can be constructed remernber one to minic M in every is in an upstroke or a downstroke. Observation 4.1 It is decidable for M e D whether M e T. Proof For M we eonstruct a nondeterministic pda M" wtrich mimics the stack movements of M by e-moves instead of reading the norrnal inputs. In addition M" remembers in i"ts finite-state control direction of the current stroke, this changes. Then M e and reads Thus the a character $ whenever T iff L(M") has strings of only length (i.e. iff L(M") is finite). all to test turn property, we test L(M") for finiteness. D M bounde<l for the finite- 52 Definition AmactrineMen-Tis ordered iff (i) its state set ie the disjoint union of sets Qg, (ii) a state reached is in Qi iff the Qnr and in a derivation from the starting configuration derivation has undergone exactly i turns. Definition A state s of an ordered nnachine is of order i iff " e Qi. Observation 4.2 For any M e n - T we can construct an equivalent ordered Mt e n - T. Proof l{e first nodify M so as to be abte to remennber the stroke direction, and then make n + I copies of the transition rules, each referring to a distinct state set Qi, for O : i _. n. We obtain by slightly nodifying this ners set of rules. Mt starts in the starting configuration corresponding to way except that whenever QO, and mimics M in M' every it is in a state s e Q1 and the current stroke is terminated by a nove in the opposite direction, the appropriate transition occurs to a state of Qi+i, instead of Qi. The computation continues rules, with changes within the new copy of the transition acceptance defined as in M, until further directional occur. Mr is cl-early deterministic properties. u and has the required 63 We notice that we can test easily whether M e n - T for a particular n, by buil,ding the candidate ordered M' t [n + t] - t equivalent to it, and testing whether any derivation in Mt can reach a state of order n + 1. The latter can be done by suitably redefining acceptance in M' and testing for emptiness. 4.3 Proof Srrategy We shaLl prove that equivalence is decidabl-e for T by showing that it is partially decidable. In particular, for each pair Mi, M2 E Tp a faurily P of we shal1 show that pushdown automata can be constructed with the properties that (i) if L(MI) = L(Mz) then for some M (ii) if L(Ml) * L(M2) then for all Thus the enunneration and M I t e P, L(Mt) 6, = and eP,L(M)*1. testing for emptiness of all the in P constitutes the required partial decision machines procedure. For sinplicity of presentation, we shaLl- form a machine M from M2 as f ollows. initially reads a character from the input tape and, depending on the dis joint, uni.on of the two tested machines M1, M what this is, moves to the starting configuration of one of M1 or which it then proceeds to simulate precisely. Clearl-y, if then M e T also. now rre need The advantage we gain by MI 1 ll2 e T, this construction is that only talk about the equivalence of configurations in single finite-turn M2r a machine. In our constructions we shal1 further assune that this rnachine 64 M is: (i) in normal. form (ii) ordered (SA.Z1. can be transforned We 4 and 2.5 and Observation 4.2 ensure that any finite-turn Lenma 4. (52.4), into this autornaton form. shall present our proof in a nurnber of stages. Paral'lel Stacking We pda Mt shaLl first outline the general form of the nondeterministic that is to sinnrlate M. A configuration of Mr has a stack which can be thought of as having a Left track and a right track, the top of each one being associated with a state of M. of size by special syrnbols, called 'rceiLings", occupying bounded The stack is segmented into lengths both tracks. The finite-state control is able to manipulate directly the top segrpnt (i.e. the part of the stack above the topnost ceiling). rn each segnent both tracks contain stack words of length greater than one. rnto each ceiling is encoded the following information about the previous history of the computation: (1) a quadruple (st, At, s2, A2) which states that at the tine the ceiling nas created, the two tracks (sz, Lz) respectively. had nodes (st, At) and 65 (2) an inlic,ator IgLr from {L, R}2 specifying the connecEion of the tracks above the ceiling to the ones irnmediately below. For example, (L, L) will mean that both the tracks above are to be associated with the left track Each configuration to two configurations of below. of Mt is to be interpreted as corresponding M in the obvious way, i.e. the M configu- rations can be recreated by taking each track in the top segment of M' and concatenating it with the appropriate words in the below as specified by the indicator pairs in the ceilings, segments and by adopting th'e corresponding state. The basic operations of Mr are to try to urirnic simultaneousLy for all inputs the transitions appropriate to the two simulated con-' figurations of M. rn order to be able to do this, while at the same time maintaining an upper bound on the length of the segments that can arise, the machine Mr can also on occasions, depending on the contents of uhe top stack segment, do one of, the f,ollowing additional operations without reading inputs: (a) if one of the tracks in the top segment is empty, and the other contains a word from a certain set of tshortf words, then the ceiLing below it is rernoved, and the tracks formerty irunediately below and above this ceiling are fused tc form one segnent, in the manner specified by the indicator pair. (b) if both tracks have more than one symbol, then a is placed to be jr:st below the top symbol of The indicator pair (L, R), and the quadruple ing to the modes of the tracks are ceitins " encoded ceiling each track. correspond- into this 66 move (a): gt st (1 c c ) (RR) rl ll A B A B A c A A A A A A ') ( ) (rR) ) (RR) rcve (b) : st lc (sBsfC) (tR) B __-:_l A A A ( C )( ) s4 rpve (c) : Ar (r) t 4l (t (s A s2A2) CtE (erAt e2A2) (LR 1A1s2A2) (LR) Fig.4 67 (c) if one track is eurpty or has just one syrnbol, while the other contains a word from some. speci.fied set of tlongl words, then Mt has the foLlowing nondeterministic choice of moves: A replacerent stack word, with the same mode and indicator as the large track, is introduced to reptace either one of the tracks in the top segrent. The simulation is then to be alLowed to continue to compare the newly introdueed configuration with which- ever one of the old ones is left. The replacement word is tshortr in the is uniquely specified above sense, and by the Mr configuration. Examples in fig. 4. We its quadruple trpveg of each of these three kinds of moves are illustrared notice that once a ceiling has been created, cannot be modified, of type (a) or (c) its indicator can be changed by . I{e define acceptance to occur in Mt when both tracks of top segment are in reading nndes, exactly one of which is accepting rcde of 4.5 though the an M. Existence of a Verifying Machine l{e have to shor, that if M has two equivalent configurations, then there exists a simulating pda M' of the kind just described, which starts with these configurations in its tracks and accepts input string. no 58 Clearly any pair of eonfigurations of M, reached via the input string from a pair of equivalsnt configurations, will equivalent. we have of step in Mt, back to show be that when the tracks get too much out we can, by uraking appropriate replacements, get them within a finite bound while still naintaining the equivalence of the pairs of configurations simulated. Further, that the replacements, and the whole simulation, by a same tre have can be to verify carried out pda. [,le nors Q define the function Rep over * f xQ x f xQ x (fu {n}) * Q' f x {1, 2} Eo determine these replacements. Definition Rep(sl, Al , s2, A2 : s3e A3 3 s4, A4, 0) = ar,where ur, if defined, is a shortest non-nulI stack word with the properties that (i) (s r, Ar) --t (sq, (ii) yol, t^l2 . f*, if (s1, then (sa, arlurA4) , oA4) and tr:1A1) = (sf , = (s2 , w2A2) rr.rUA3) . I,lithout having to construct this function or to determine the arguments for which it is defined, we a finite domain, ne can define p to be can observe that since it has the maxirnrm of the lengths of the stack words in its range. , if the top ceiling contains (st, At , s2e A2) and (L, R), and the right tsack above contains (s3, A3), and the left one some rn Mr 69 (",., ,'Aq) where (sq, Aa) is a reading mode, then we make the nondeterninistic replacement by (s,*, oAr+) whgeRep(s1, Al, sz, A2 : 83r A3 ! 84, Aq, 2) E o, provided that lrl . lr'1. This is illus- trated in fig. 4(c). If the indicator were (L, L) then the vatue of for 0 - Rep 1 would have been appropriate, were interchanged we would treat them while if the top trackg in a sinilar It is easy to see that under the way circr-rmstances by sp@try. specified for the replacement, the function Rep nrst be defined, for ro' is itself a guitable candidate for its value: I{e first notice that the sinpleet way of realising the necessary conditions is if there is Some c 8.t. (sr, Ar) * (sz-, Az) Then tor t"u, ,,,'Aq) and * t"s, As). clearly has the properties required of a value of (except poesibly ninimality). In addition, all other Rep ways of reach- ing the specified situation in M' (i.e. if nondeterministic replacementa have been rnade in the meantime) invotve only valid eqrrivalence preserving replacetrpnts which are defined completely by the ceiling concerned and subseguent computations, irrespective of the stack contents below. Thus by induction on the nuniber of such replace- rnts made during a derivation in repLacement is available whenever Mf , we can deduce it is that a valid necessary. A finite-state control can carry out these replacements since Rep is finite in every senae, provided that it can always appropria- tely nanipulate the whole of the top segtrent. Thus it renains to prove that a borurd x exists such that no track segment can become 70 longer than x in any computation. It is here that che finite-turn property is required. I^Ie f may irst consider in turn the various eventualities that arise: (1) If both tracks are steadily increasing then new ceilings will be frequently (2) added and no segment will become 1-arge. If one track is increasing steadily in an upward stroke, while the other one is stationary (i.e. with stack height not chang- irg) then after a stack increase of q2thin the one, a valid ment musf be possible. For there repLace- must be two intermediate points in this derivation at which the mode of the growing track and the configuration of the stationary one both repeat. Thus a replacement word can be obtained by cutt,ing out a segment from the growing track. Therefore, for however long this track is trying to grow, the effect on Mt will be to repeatedl-y rnake nondeterministic rpves so as to keep the segment lengths bounded. (3) If one track is decreasing then situations can arise in which the length of the segment created depends in a bounded way on the size r:f the prerriously existing segnents, and is not a priori bounded by p + 1". There are just tlro ways in which a Lrack can grotl "out of bounds" without imnediately being arrested by a nondetermin- istic replacement: (i) If one track in the top segment is empty, but the other one is not Eqrite i"ong enough tcl qualify for a replacement, then the ceiling below may be rennoved by atr (a)-move. Thus the resulting fusion may create a top segment which is suddenly longer by y than before, where y is the previous bound on the segments occurring in VL the configuration. (ii) A track may grow larger replacement being possible, and larger without if the other track does not become enpty or a singleton in the meantirne. This second track can assured to be steadily decreasing since, as a be in (2), stationary periods do not contribute to growth. Thus if the decreasing track is initially of height y, then this may cause to arise a segnent of length no more than q2t2hy. The crucial observation for both (i) and (ii) is rhar rhe gain in length achieved by each nethod can only be exploited to achieve further gains by the sarne method, once the simrlated configurations of M the order of at least one of has increased (i.e. after a turn has been nade). More preciselyrwe define the order of an M' configuration to be the pair (i, j) when i, j are the orders of the states of rhe left and right track respectively. Thus the effect of a replacement on a configuration of Mf of order (i, j) is to create one of order (i, j) or (i, i) or (j, j), while an ordinary simularing move would lead ro one of order (i, j) or (i + f, j) or (i, j + 1) or (i + 1, j + 1). we have to enslrre that the sinulating nachine cannot enter any ttloop" which could cause the segrnents to increase in size indefinitely. The ease we have to consider is that of derivarions from configura- tions of Mt of a given order to others of the same order, via of differertorders. clearly these must involve kind that substitute for the empty or singleton track. definition the rcdes of the tracks are the repl-acement, repLacements same ones of the Since by after any such either both tracks will be in an upstroke, or both in downstroke. rn the former case ir is obvious that before the two a 72 tracks can again extribit dd.fferent behaviour, turns must be rnade by both tracks. In the latter case the height of the top segment can be exploited by method (ii) for further growth, after a turn been made by just one of the tracks. Howeverrthe new gains has will only be achieved in the track that has undergone a turn, and cannot be exploited for further gains untiL a turn has occurred in the other track also. Thus the times when successive gains can be made of segments in excess of the bound p, in the size can be regarded as occurring at periods"during which M' has configurations whose orders form a npnotonic increasing seguence under the ordering defined by: (i, j) r (i', j') iff: 1r J > ,l l" , and j or i > it If M can paLrs. >J .f t j',orjtj'andi:i'. Hence we conclude size of a segment that Consequently a that there is a bound on the may occur in any computation of Mt. finite-state conErol is sufficient to specify for each stack word moves J only make n turns, then this sequence can have no more than about n2/2 maximum or 1, that nay arise in a track of the top of type (a), (b) or (c) are appropriate segment, whether and what form these should take, and to carry out the normal simulation with inputs otherwise. 4.6 The Decision Procedure If a machine Mt is constructed for the tested automata Ml, M2 73 as described above, with an arbitrarily and segment bound, and guessed repl.acement function if it is found to be enpty, then we have a verification of equivalence. For let us assune to the contrary that the starting configurations of M1 and each point in the conputation of M' M2 are inequivalent. Then at we can folLow the shortest string distinguishing the sinulated configurations. Then as each input character is read, the rank of the sirmrlated pair Wtren one decreases. nondeterministic replacements are made then, clearly, at least of the new pairs has rank no greater than that of the old Further, the rgplacerents one. all occur in reading modes to ensure that further progress along the shortest distinguishing string can be nade imediately. The only exception to this is the situation that occurs when successive replacements without intermediate inputs have to be made because of a long stack decreasing e-derivation. this mr.rst once come conclude to an end since the stack is finite. that if l\ I 142 However Thus we then for some input string we will reach an accepting configuration of Mr after a finite nurber of steps. Conversetlr !t€ have already shown that if Mt = Mz then some M exists which aLways simrlates pairs of eguivalent configurations of M, and therefore accepts no string. Thus enumerating such siutulating machines for M1, Vl2, for aLl possible replacement functi.ons and appropriate fi.nite state controls, and testing each f,or emptiness is a partial decision f,or equivalence. The mai.n result therefore follows. procedure 74 Theorem 4.1 Equivalence of deterministic finite-turn pushdown automata is decidable. I Corollary Equivalence of deterministic Ewo-tape acceptors6 is decidabLe. Proof For a Ewo-tape acceptor M' with input aLphabet Xt, and (cl, g) on its input tapes, construct a dpda M, with f = X', X= is X'u {$}, that accepts exactly the strings accepted by Mf" The oR part is read by M, oR$B such words and that (c, B) and stored syrnbol by synbol on the pushdown stack. Once the $ is read the finite state control treats th.c input tape and the stack exactly as M'treats its pair of input tapes. whichever M can be made of the various criteria of and Scott6, nird7l to accept whenever M' accepts, acceptance is taken (e.g. . l,t is evidently a one-turn machi.ne. ! Rabin _/{ Chapter 5. 5 ONE -COIJNTE R AI'TOMATA 1- Introduction Counters have heen snudied in a variety of computat,ional contexts as simple and naturat mechanisms for unbounded storage. A counter can be convanientl-y regarded as a stack with an aLphabet of just one syml"'ol". The gain :i.n simptric!"ty achieved by this restricfii"ar. is that 6 eouat€r llae eesentiall"y just two behaviours, depending on whether it is enapty or not. Despite the apparent severity of this Lirnitation, it is well known that severaL basic probLems that are undecidabl.e for automata with staeks, remain so even when these are restricted to counters. Such ',:nderi-dability resuLte r,sa be- deduced from rhe theorem cf 1A Minsky'" that any Turing machine can be sinnulated by a 2-register machine" 'fhisr tratter can be regarded as a deterrninistic automaton with twc counters, a finire sEate controL, and transitions reading no i,nputs. The computation executed by such a machine can be described by a sequence empriness *ond{ei.ons configurations of triples, cf fhe resrchecl . each giving the state and the trvo counterso Si.nee of the successive a deterrninistic one-counter automlton (doca) can check whether successive triples correspond to valid transi.t!.ons, and alsa whecher the ,:veralt aetion induced in one of the tg{o colmters f s *ons.i.stent with the sequence, the terrnination prcbleu for Z*register mactrlines (arad hence Turing machines) can reduced to bhe nullity of -{ntersection this latcer probJ.en musfi he r-s+decidabLe probLem for doca. Thus be 76 Since the class of doca CC) can be sholin to be closed under corylenenration by the siune arguments as used in 92.41 52.5 for dpda, it can be easily deduced that totality, and hence equivalence, for nondeterministic one-counter autonnta, and, more pertinently, inclusion for the deterministic case, are alL undecidable. Against this background we are neverthelees able to develop a detailed analysie of the structural properties of the computations of doca, and hence derive a decision procedure lJe have seen that any for doea can be transformed equivaLence. into normal form (52.4) with h = 2 (91.7) with no more than a polynonial increase in q We shall assume that the machines are all in thie form, so that rre can e)rpress all the derived properties in terms of just rhe one paraneter q. The decision procedure takes the form of a nondeterninistic sinul-ation, as in the previous chapter, but now the properties deduce are eufficient to speci.fy exptricitly the nondeterministic replacenents regr-rired, nondeterminisrft here Vire note, however, that, as in Chapter 3, is a convenience that is not essential. procedure could also have been obtained number we of deterru.i.nietic sinmrtations. A in the form of a finite 77 5.2 Preliminary Results The function S(q) (which also gives rhe maximal order in the syrunetric group on q elerents) will. be the factor in the bounds we derive. element of an dominating Definirion S(q) = nrax {{,.e.m. Lenora 5. inr} I Xri = {, ni > 0} 1 -_ s(g) = "6Toaq'1(q) where 1(q) * 1 as q * -, Proof (outri.ne) uaing etandard nuurb.er theoretic resulte, this aslmptotic functi.on can be easily obtained for the case where } in the definition is additionally reetricted to be a set of prines. To show that the sane bornd is valid for arbitrary {nr}, it remains to prove that the possible extra contributions from {n. is not significant. Details of this given in Valiant and patersonzo. I prime porrers argument are For a deternninistic one-counter automaton M, we describe a configuration c with state s, and counter contents n (n > 0) , by (s, n). For this c, we u6e c + m to denote the configuration (s, n + m)r provided that n + m > o. 78 Definition The input word B is a stan{$ffS,eqg for the configurations c, ct iff (i) B ie a shorrest string (ii) g = Blg;83 (iii) where B such that c ? lgr8sl. q2, llzl c', = q, and r > o, Forgomestatese andwrd>0, forall-vs. RnV -L*2 c ---:+ (se , w - t. O and <vSEr vd). Legrna 5.2 For a doca with q states there is a positive integet Z, greater than s(q), such that c? if l"l ct, then there is a standard loop drop d divides l"'l ,sequence q,2, l"'l > q2 no "r,a for c, cr in which the Z. Proof We define the efficieqcy of a state s to be the maximtm vaLue (possibly infinite) ot d/ fyl, where, for all sufficiently large n, the derivation (s, n + d) * (s, n) existsrbut repeats no state except I at the beginning and end. Clearly d, l1l S q. Suppose o ie a shortest string such that " * "t. We mark the laet occurrence of one of the states with greatest efficiency af\ in ttr^is derivation, Let this state be sur y, simpl-e loop be generated by and ret its efficient and carrse a dr:op of d, where d > c, excise from this derivation a set ot (not necessarj.lv sirnple) Now disjoint loops of to due is a multipre r:f d, thern preserved" maximal tertal length, suctr that the totaL ar:.rl that rhe trte ean show drop fhe rnarked occurrence of, so is l-ength m of the remaining derivation is no more ehan qi * t* brl firsfr rgs4lpring that at 1,east t = [(m + Z)iq - el disje,int simple loops, nor containing the markecl su in thei.r interiors must occur it. some But k > d would imply that non-null subset of these 1"oops accounts f,cr a total drop ttrat is a nultiple of d. This subset eould therefore have been removad in the originsl exe.i.sion contrary to the maximality conditiou" (nc + 2) I .q * 2<d- 1-, anel $o Tll < d < q. By a simitar arglffient ir qi * i follows c,en Ilence inrnedi.an:elv {.f be shsr,n that t?le Length of the derivssli.on *f,ter the malr,.*C s* i"s i*-cs than q(q - 1), if d < q, Let g* *$* be Ehe input stri,ngs fnr the parts of this remaining derivation before and after ihe choseR oecurf,enee 'thene cLea::?"y, for some integer r of se reepectivel.y. rtR P1i',* e ..''--'-' ' ,"1 ' + l-ci , + Eut l*l* l*'i > r'? > m ::rn ! i e{ r.ha;: '!t I i:-, r :, {";. t f. since the bounds derive'C f;r the rarts of this derivation in the d . q, and eiraple opeciaL argtlF"le::lrs that the eounter is at na in the trivial case d = e' case €osur€ strlge e'*pt:y. Also, since we have repl-aced 80 arbitrary loops by ones of at l-east the sane efficiency, BrttB, nust stil1 be of ninimal length. I,Ie not,e that Ehe caee of a null Y has not been excLuded this argunent. t{e aLso note that a sinrpler if the bounds of q2 were relaxed To obtain a value to argument would in suffice 2q2. for Z we investigate the set of possible values of d in the above construction. For each state s we sel-ect, if possible, a maximally efficient sinple loop Ehrough s' and denote the set of states in this loop by Loop(s). ClearlY' if st e Loop(s), then the efficiency of st is greater than or equal to the efficiency of s. Alsor any standard sequence whose principal loop is based on s couLd be replaced by one based on s' , by applying the construction to the derivation of the old sequence, in which st must occur sincer by definition, r > O. Let s t > I be the transitive cl-osure of the relation defined by en e Loop(a). Defining s, sr to be equivalent iff st > I and s 2 st, the relation > becomes a partial ordering on the equivalence classes" Let one from each clase slr..r8k be a selection of representatives, that is maxinaL in this ordering. It is easily verified that the corresponding loops must be disjoint, standard seguences can always be based on some such drope due z = 1,. c,n { <la a. is the drop due to I Thus Z toop. to atrl these must add up to no more than q, divide Z where < s(e). I Loop (sr) } . and that Then the and each one must 81 That thie bound for Z is achievable can be verified by examioing Exanrple 7.3. As we are concerned only with asytnptotic bounds, and as SCq) clearly dominates any fixed polynomial in q as q becomes large, it will be sufficient for our rather we etran derive. technique purposes to prove the existence of, obtain specific expressions for, the various polynomials The proof of the foLlowing Lerma introduces a useful " Lemg 5.3 Ihere is a polyaonial F3 such that for any configuration c with l"l > nr(e), (i) and any positive uank (ii) nultiple Y of Z, (er c*Y) +Y/q<rank {c+Y,, c+2Y) <rank Cc, c+Y) +Yq, c = c+ Y iff c +Y = c + 2Y. Proof Assming that c and c + Y can be distinguished, there nnust be nrinimal dietinguislaing sequellee S6* a whene .: bt (s, q-) ;- for eooe ensures s. Provided that p, is sufficiently large, Lema 5.2 that I noay be taken to be in the forn of a standard OUOfOr. Let the drop due ta 3, be d, where d > o. Since Bd sequence 82 dietinguiehes c and c + Y, clearly grul * t'u uro distinguishes c + Y and c + 2Y. since lgrl < q and d > 1, the right hand inequaLity follows. r > 'Ild, to be a rninimal string distinguishing c + Y and c + 2\. Then 8rl'r- t'uuro distinguishes c and c + Y. But lBzl > 1, for In a similar fashion lre can choose OrO!016, where otherwise, since d divides Z and thus also Y, the sequence would not distinguish c + Y and c + 2Y, The Left hand inequality then foLlowssincedsqalso. Statenent (ii) is an imrndiate consequence of Ci) . n 5.3 Propricty I.le now eetabLish some retationships that hoLd for periodic sets of configurations, Definition A configuration c is impgrper iff c = c + mZ for all integers m (not necessarily positive) such that l"l Len'na 5. + nz > nr(e). 4 If c = c + rnZ for funproper. eome m > O and l"l > nr(e), then c is 83 Proof It ie easy to ' eee that for any set of configurations {crr...1crr}1 rank (cor Hence, t .ili "t) ' 1s5" {rank ("3, if rank (c, c + fl) = or "3*1)}. then rank (c, c + z) > min {rank (c + iZr c * (i + l)z)i. 13i<m By Lemra 5.3(i) it foLloqrs that these ranks ngst all be infinitet and therefore atso, by Lerma 5.3(ii)' that c is inproper. D Definition A conf iguration is propel if f it is not iryroper. That the period Z is optimal for the propriety condition can be seen from Example 7.3, where for any configuration c in the starting Btate, c r c + i iff i is a nultiple of Z = S(q - 1). Lenma 5.5 is a polynomial p5 such that if l"l and c = ct, then c is improper" Ttrere > Pr(9).2, l"'l t q.2 Proof rhat c is proper, and fet ErBlBr6 be a string distinguishing c and c + Z, constructed exactly ae in the first part of the proof of Suppose 84 Lenrna 5.3, but for the case Y = Z. We define "rr, ".1 for n > 0 by B. B: g', "r, and ct Bl "ir, where, in the case of e{oves, naxinal derivations ci.e. to readi-ng Eaken. If c = cr, then cLearly .r, = "rl for all n' The polynomial p, is chosen to ensure that r is sufficiently large for rnodes) are the following argr:ment to work: Either in cip...pc'(rr|) sourc configuration repeats, or else eome cl in this set has height not less than 2q3. In the latter case it is easy to verify that for some for we i1 j such that i , 3 < 2q4, oj -i -i+JJ1 cl and cl=c-*w, "'.'?, w> some o. In either case, for some i, j such that i ' J = 2q4, have' putting g = (j-i).21 sl+wZ "i*1, = where now w > rank o. By Lemna 5.3(i) (ci, clas) < rank ("i*o' However, if r is large Leurna 5. (i) 3 and trivially if w = or cj.+2g). enough then from the propriety of c' and ' rank (cr, T1li-s if w > 0, "i*.Q,) > rank ("i*0, "i*2t). contradicts the assrmtption that "* = "rl for all n' Thus c be inproper. D must 65 l{e can now derive, as a consequence of this result, the property of equivalent configurations on which our decision procedure depends. Definition Integers m, n are (x, :r)-rational,ly related iff there exist integers a, b with 0 < a, b { xr such ttlat lna-nbl sy. Lerrma 5.6. exist polynonialr p6, il6 such that if c = cr, l"l t FU(e).2, and c is proper, then l"l, l"'l are (c2, e5,(q).2) - rationally There related " Proof that c = ct, that c is proper, and that l"l , FU(g).2, sufficiently large for the fotr lowing argument to work. Choose Suppose for f. 'b 81rg2r$316, and define "rrr ni as in the previous lemma. Let !, be leaet n such that miur c' I cr, + lc;l) . q2. infinite strictl-y c* ia proper for all n s l, and therefore l, mugt exiet, for otherwise decreasing sequence. (i*,r1, ti"*l] Z if l"rrl > nr(e). would be an the 86 rf l";l . q2 tb.n by Le,r,rna 5 and the propriety of co, we have rbar l"gl = p5(q).2. Alternatively auppose l"cl . q2. For O some i, kt < i < i + kt 3 q, the states of clr clalt are the .i.*t, * ci such that same' l"l = q'. Then if k is a large enough nultiple of kf to ensure that l"o-nrl t pg(q), the propriety of the cn-cives e Bay' where t,-kzl tu-t, + dkZ = tL-zkz' Hence ,L-vz t .L-nr, that is "i,*"tz*ei+Zekz' By lernrna 5, l";l < nr(e).2, for othemise lcil woula be proper, contrary to the previous statement. Thus for a suitable choice of p; we have in both ll-l - ,rsl . pi(q) .z and ll"'l - ellkl ' Since O < d < q and 0 < e/k n q2, lIcl"e/k it follows for - l"'l.al < nu(e) ni(q)'Z soroe ,2. n cases p, that af 5.4 Decieion Frocedure Using the result of Le'smE one-counter automaton Mt which 5.5 'are construct a nondeterminiscic is abie to sirnulate, in a certain sense, the computations of a pai.r of equrvalent doca. By taking the disjoint union of the star.es and transition rules of the two rnachines, we can regard the simulation as maintaining a representation of pairs of equivalent confi"gurations, c and co, of the cornbined machine. point in the si"mulation alcl - Ul"'l = f , for some arbrf such that 0 < a, b. q?o arrd if l . pg(q).2. Then We ensure that at each Mt can represent c and cu h;l hol-ding alcl in i-ts counter remernbering f ,arb, and the states of c and and co, in its finite state controL. that The action of n' is as wheneveE m or n is greaier than po(q),2, f'ollows " Let Fo be some polynomial such and they raLionaLl"y reX-ated, then they are rel-ated wittr respect , are (g-, lU(!.).2)' to only one admissible rational ratio a/b" Whenever l"i, l"'l are both Less than lo(e).2, their values are stored in tire finite state control. When a simuLation step is about to exceed tlris bound, the finite state control det.e;mines the coefficie&8s arb, ri any, and sets up the counter for the appropriate represenfiagion. ia]"rten a sj.mulation step would reach a pair of configura;i.r:ns not ratr-!.l.a.lii lr:X.:rr*erln s.!)'c ie jusC too large for ctn then by Lerflm.l5.6, i"f q1 :: {i:'1 ':[len c must be improper. Instead of, cont"inuing the sinulaticn wiLti either to (c, c 'Z), tr: o *t), a nondeterministic step is ic * Z, e') g to the simulation of Ttren, i-f c = r{ a.:ld eo also c = c - Z, tl:.e simulation Ehe simulatior: c",t continues to be one for equi'v'atent cgnfigurations in either case. made 88 Since the originaL doca are in normal form, we can easily define acceptance in llr to occur if and onl-y if exactly one of the simulated eonfigurations is in an accept mode. Thus if the starting configurations are indeed equivalent, our discussion shows that no string is accepted by Mr. 0n the other hand, if they are inequivalent, we can show, as in the previous chapter, that stri.ng must be accepted, For if some some d distinguishes the starting eonfigurations, then either both derivations wiLl be simuLated directly to their different concLusions, or else the rational relationship nust fai1" In the latter case, if (c, ct) is reached c I ct, then the remainder of a distinguishes one of the new pairs created. The assumed normal fora guarantees that any where long e-derivations wiLl steadily reduce the counter. This, in turn, ensures that further progress a finite number of a1-ong moves, and therefore o can always be made in that o will- eventually be accepted. The construction and testing for enptiness of the sinnrlating machine described therefore constitutes a decision procedure for equivaLence. The nuniber of states of this nachine need be no more than p(q),2* for some polynoniaL p, where q is the totaL nuriber of states of the tested machines, and Z is bounded above by '__---;-q. S(q) * evg'roge Assr,rrning a fixed input al-phabet, and recalling Lenma 2.1 we conelude that E9 Ttreorem 5.1 The equivalence problem for doca is decidable, and there is a decision proeedure whieh, for q state machines, has a running time bounded above by ,r,G.-ffi for some constant k. fJ 90 III CONTAINMENT PROBLEMS Certain restricted classes of deterministic pushdown automata have gome important basic properties which do not hold for the whole cLass. For exanpte, finite-state machines recognise a class of languages that are closed under the Boolean operations, one-counter nachines, we have seen, have a rigid periodic structure, while for simple machines eaeh configuration can be related very directly to the language it generates. Furthermore, within these and some other subclasses we know how to test for equival-ence. It is also plausible that even for for which decidabil,ity problems can be proved in the unrestrieted case, easier decision procedures ean be found for these subclasses than for the whole c1ass. For any of these reasons we may want to determine whether for a given dpda there exists an equivalent one beLonging to a particular subcLass. Formally we ask the foLlowing containment problem: If X, Y are Ewo elasses such that L(M) = of L(M')? automata and M e I,Ie denote X, then is there an Mf e Y the containment problem for X, by (x:Y). Emptiness, f i"niEeness and containment probLems, Testing seen total-ity can all be phrased for the prefix property as we have in 51"4 to be equival-ent to the problem (D:D6). 0f the difficult probLems mentioned above, the onLy one known to rnore be deeidable is that of regularity, i.e. (D:Fsa). A proof of this hae been given by Stearns2. Without resolving the remaining open prnblems we shall nevertheless throw sone light on their expected Y :t difficulty by relating following them !. ro the regularity problem in the way. I{e shaLl define a very generar notion of reLative complexity with which one can compare the inherent difficulty of various containment problems. ResuLrs expressed in terms of this are of wide applicabiLity. For example, our result that testing for enptiness, and for the prefix property are, in che defined sense, equally diffieult, irnplies that if the time conrplexity for deciding these two probLems on any machine modeL are polynomiaLs in the parameters polynomials of the tested automata, then the leading terms of E i11 differ by on1-y a multiplicative consLant. The result$ we then prove using containment rroblems these this notion are that such as (D:C), {n:t1, and (n:Sg) must be, if decidable at arL, at least as difficult to de*:ide as regularity. with this as one source of urotivation, we then pr.oeeed in the following ehapter to investigate the reguJ.arity probLem in detail. For a naturai particular nreasure of complexity used by stearns, improve his upper bound from rde a trebLe to a dou"nle exponential level, which now closely approaches a knorn"n double *xponential l-or*rer bound" As a consequence r/* can a1"so si,g*l'r"ficantLy improve the upper bound on the tine complexi.ty of ttris prob!.em. we observe that all the eontainment problems in which rde are here intereeted become undeeidairl* if D is replaced by ND, the class 92 of nondeterninistic pushdown autonata. This can be deduced from theorem of Korenjak of the Post and Hopcroft9*ho Correspondence Problemr a construct, for each instance a context-free grailnar over a terminal alphabet r u {$}, with the properties that (i) if the PCP has no eolution the language generated is ,*$, (ii) if the PCP has some solution then the tr anguage is not deterministic. From this we conclude that for any class (ND:Y) Y c D such that [*$ is undecidable. It therefore folLows that (M:Y) is undecidable if Y is DorSorTrCrFsa, etc. e L(Y) ' 93 Chapter 6 RELATM C0MPLEXITY 6.1 Introduction I{e e:(press wilL shall define a partial ordering on decision problems to the relative difficulty of solving them. Thus Pt > P, nean that ifr on any rnachine model, the probLem P, needs Eine x to be decided, then P, will require at least about the sane time. Furtheraore, similar concLusions can then aLso be made about sPace reguirenents, and various other measures as weLl. We could define guch an ordering very sinply by saying that tl = P2 iff any procedure to decide Pn is effectiveLy able to decide P, also. For our appLications to dpda problems, however, it is convenient to relax this condition slightl-y. We denote by D(nrrnrrn3rn4rn5) the cLass of dpda whose parameters grtrhrPru are respectively bounded above by nlrtr2rt3rnO and nr. t'le define a transformation of a machine description to be direct iff it can be carried out by an aLgorithm that transition rules, a finite amount of (and nodifies memory makes only one pass of the each one as necessaryr) requires on1-y additional to the capability of recognising accepting modes, and increases each parameter at nost linearly. We then say that a procedure can decide problem P directly for a cLass X, iff there is a direct transformation which takes all rnachines in X to a form in which the application of the procedure effectively decides P. Definition Pl > P2 iff 3 k > O s.t. any procedure to decide P, for 94 D(knrrknrrknrrknOrknr) can directly decide P, for D(nrrnrrn3rr4,r5). Definition P1 = P2 iff P1 > P, and t2 = Pl. It is iarnediate from the definition that for the classes of autooata XrXr,Xrr the foLloning reLation holds among the containment problems: (i) Xf c Xrr I{e obeerve + (11":X) > (Xf :X), that, in order to decide any global property of machine, lre require operations a at least as difficult as a direct transformation. Thus the relaxation of our definition of the ordering does not endanger its validity for our 6.2 ResuLts purposes. As ie per:mitted by virtue of Lewra 2.5, we shall aseune here, for convenience, that all dpda are in normat form. Theorem 6.1 (D:T) > (D:Fsa). Proof Let L = t(M) c X* where M e D, and let L'= (Lf)* where 95 $ f t. We shall show that Ll = L(l,tt) for some Mt e D which is only slightly larger than, and easily obtained frorn M, and, further, that Lt e L(T) iff L is regular. Thie is clearly guff,icienr, for then any M can be tested for regularity by testing the appropriate t(X') for containrnent in t(t). create Mr to recognice Lr by modifying M in the following We Iray" For every accepting mde of to a nelr epecial state that M we introduce a move on causes the stack $ input to enpty and the starting configuration of l{ to be reetored, all via an e-derivation. If we make the starti:rg and accepting configurations of Mf to be the starEing configuration of M, then clearly L(Mt) = O$l*. To do all thie and we we need add no tpre than one will at lmst have doubled the state, nunber Furthermore, Mr can be obtained from M by a To shon that L regular + L regular + (tf)* and one input s;rmbol, of transition rules. direct transformation. Lt e T, we sinply observe that regular'+ Lt e O-T =' Lf e T. To ehow the converse, rre recall a coment made in 84.2 that L(O-T) = L(Fsa). Ttrue, if L is not regular then in any recognising machine for (tf1*, turna may have to occur during the parsing of each eubetring between eucceesive $ machine cannot be finite-turn. D characters. In that case the 96 Theorem 6.2 CD:C) > (D:Fea). Proof Let L = L0{) c E*, where M e D, and ter L, = il9,$1"$r", n=1 f, $, I r. we shall shon thar there is an Mr e D recognising L' that is only slightly larger than M, and that any procedure for testing Lr e t(C) will autornaticalLy decide whether L is regular. where I{e construct Mr to simulate M repeatedln and to count the number of stringe from Lf read, by keeping a string bottom of the stack, where A is a new symbol. An at the Thus as in the previous theorem, whenever an accept mode of M is reached and a $ inmediately follows, the staek is emptied, but now onty up to the toprrost A, an extra A is added, and the starting on top of this. oode is restored $, characters are read, they are checked one by one against the Ate, and acceptance occurs iff they are equal in nrsber. liltren the Such an Mr can transforrEtion, and It renaine to clearly be produced from M by a direct will not be much larger. show that Lr e L(C) iff t is regular. Clearly if L ie regular then a recognieing rnachine for Lt exiets which only uses ite stack to store the Ars. suppose L is not regular. Thus Lr e L(C). Any nachine recognising Conversely Lr must have an infinite set of pairwise distinguishable configurations reached 97 via input strings terminating with a $. If it is a l-counrer machine and L is not regular, then to recognise some words from L from any such configuration, the machine will have to enupfy iCs stack, for othenrise, regularity would be inplied. Ilowever once the stack is enpcied, all but a finite anount of inforrnation about the nrsber of inst:urces of L$ atready parsed, is tost. Thus Lf c LcC)" 0 Definition ForalanguageLcX*: Sinit(t) =E*-LE+. Lenrna 6.1 If L has the prefix property then: L regularc+ Sinit(t) is regular. Sroof Ci) Since SinitCL) is defined by regularity preserving operations, if L is regular then so is SinitCt). (ii) If L has the prefix property then it consists of just those lrords in Sinit(L) ttrat are not proper prefixes of other worde in Sinit(L). Thus a finite state automaton for L can be obtained from one for Sinit(t) by sirnply removing all- states from wtrich any further strings accepting states. 0 can be accepted, from the set of 98 Theorem 6.3 For any cl.ass Xo c Ro such that L(I's.)$ . LO(O), - CD!x^) > (D:Fsa) Proof Let L = t(M) c X* wtrere M e DOr and let Lr = Sinitct)$, wtrere $ I l. x-transitions from accepting I,rle rnodify M by reptacing to rnodee all ones co a special reject moder ard by adding a $ traneition for all reading nodes (other than the reject nnode) to a nelr node which is trcde. Then this new machine, say now declared the sole accepting Mr, recognises Lt, and so L(|,tr) is regular iff t ie regular (from Leuna d.1). If L(Mr) is regular rhen, by definirion, LCM') e LQ(o). However' if L(uf) is not reguLarr then sinit(L) is not regular so requires arbitrary large st,acks to occur during recognition. But from any such live configuration we e:(pect acceptance to be possible in M' with a further input of only a single character Thue L(M') not regular =t and $. L(Ur) I L(Xo). we can therefore concLude that L(ltt) e L(xo) iff t is regular, and hence that @o :xo ) > (Do :Fga). 99 Ilowever, from property (i) in 96.1 we know that CD:xo) > (Do :xo). Also, eince any test for the regularity of t$ for any L c E* can be used as a tegt for the regularity of L by the now farniliar arguuent, we al"eo havc that (Do By the :Fsa) > (D rFaa). transiti.vity of the ordering the result follows. I Theorem 6.4 (O:enpty) = (Dltotal) = @l:Do), where enPtYr total refer to the class of nachines accepting nothing, and x* reepectively. Proof From the argument in Corollaxy 2.1 to Lenma 2'5 it is imediate that a direct transformation exists to nodify any machine M in normal form, to one that recognises exactly the conplenent of L(t{). It follows that equally difficult To sholr emptiness and totality are to decide in our sense. that (D:enpty) > (o:Dg) we recalt that the latter is equivalent to teating for the prefix property. For any M e D 100 lte can construct en Mr consisting e88entia11y of two copies of the transition rules for M with distinguished state sets. These rules are mdified so that transitions from accepting rnodes in the first to the appropriate states of the second copy' The starting configuration of ilt is defined to be that of the first copy of M, while the acceptingnodes are those of the second copy. copy lead Clearly Mr accepts just those strings of t(M) that trave proPer prefixes in t0{). Thus testing Mr for emptiness is equivaLent to testing wh.ether L0{) e L(Do). To see can be that (D:D') > (D:erytY) we observe that any M e mdified by replacing all- the traneition rules from D each aceepting node by a reading rute leaving the mode unchanged. Then testing this nachine for the prefix Property will effecCively test M for enPtiness. I Ttreorem 6,5 (DrFsa) > (D:finiee) >(D:enptY) Proof Let L - L(l{) c X* where }t e D, and let L' = L$*. clearly M can be transformed to hecome a recogniser for Lr with in size. since L' is finite if and onLy if I is little change emptyr L can be tegted for emptiness by testing Lr for finiteness' 101 inequalitlr tre use the observat,ion that for a dpda M', LOi') is infinite iff Mt has sone live derivation To show the other that repeats a configuration, or has a repeated node in its sEacking sequence. I{e introduce some nen stack synbols, an E, for each Ai . I, 1T (s,Ar) and a special B.For each mode (srAr) bI Cs{) 3 (sl,w) we replace each rule $_ rute (s,Ar) i 6r,{), $., and also, if it is an acceptingrgd", the rule (srAi) -* G*rn) Cst,w) and add rhe s-a i.s a nelr epecial state, €rnd $, $., are nelr input characters. $" " 'l I^Ie also add the transitions (srA) -5 (srl) for any mode with where t = r., or A = B, Aseuming can be and add (surCI) to the set of accepting rnodes. that Mt is in a fo::sl never requiring to eupty its stack, it verified that the mdified nachine jusr described will recognise a regular set iff L(M') is finite. I 6.3 Comentg We have only obtained theorems However numerous corparable neans for (X:y) results can be derived by sinilar for X equal to various subsets of Thus we can where X = D. D. relate the realtime property, and the stateless property to regularity. If we define T1 to be the class of ordered deterministic one-turn machines, with only one state of ord,er zero, then we can show that (T:R) > (TI:Fsa) and that 1 (T:S) > (T-:Fsa). To do this, the construction we need for Mr e tl is of an M e T that accepts the language each LO2 } u oy' { "$o$r1 | "$s$v I ey' rthere each cr, B is a turn*free prefix of LOtf). { L(M') L(M') We } note that despite the restricted nature of tl, our analysis of regularity (Chapter 7) gives no indication of this easier to decide for T1 than for The equivalence problem problern being substantially D. is not a containment problen. Ilowever the above mentioned kind of argument shows imediately that it is no easier to decide for D9 than for D. For to test for LL = L2 in D, we could test for tr$ = lr$ in Do. Even more trivial-ly we notice that since enptiness is a particular instance of equivalence, it cannot be more difficult to decide. I{e conclude by sumarising our results in the following diagram, in which Pl* P2 inplies Pl = P2, and each language is assumed to be specified by a dpda in normaL form. 103 finite-turn? ieL one-counter? sinple? I ieL regular? isL finite? J Fig. 5. ReLative Complexity of Containment problems, 104 Chapter 7 THE RXCULARITY PROBLEM 7,L lntrodgction Ste"trr"2 shows that the problen of whether a dpda accePts regul-ar set a is decidable. To do this he proves that if a dpda of a certain form, with q states and t stack eymbols' accePts a regular language L, theo L is also recognised by some finite-state the statee of which nrnber no more than I'leyer and Fisch.t2l gi.r. an example to circumstances a some E3 expression show automaton' in q and t. that under these finite-state autonaton of n2 size nay indeed be necesSary. reeult in this Chapter is to reduce the upper bound given by Stearns for this, by an exponential level, to an E2 Our main function which differs from the lower bound of Meyer and Fischer by only a rnultipticative factor in the leading exponent. As a consequence ne can also derive an E2 tine algorithrn for testing dpda for regularity. By siuilar analysie for the regularity of and S6we obtain each of the classes R, 1 distinct E' expressiong, the orders of which in case rre can show each to be valid as both upPer and lower bounds. Alnoet all the ideas we sha1l use can be for.nd in stearns' paper. llowever, in addition to the improvenent in the final results' the following differences are notertorthy. Our proof right is fot the general case allowlng arbitrarily hand sides in the transition rul-esr while Stearns long considers C 105 only conservative mechines with h - 2. Thus, although for dpda q may have to grow et(ponentially when it is reduced some to the h = 2 form, we shal.1 show that this is not a oource eontributing an extra e:cponential in the main result. Furthermre r our analysis will directly applicable to' for example, stateless (Chapter g), for wtrich equivalent consetltative be and dB-S machines machines do not exist in general. Ttre notions of nuLl-transparent and C-invisible segments, introduced by Stearns, and the proofs of their existence in sufficiently large stackc, remain at the centre of the argurnent. tlowever, we have r.rnified the proof,e of theee theorems Cl,errma 7.2) by applying the technique used by Stearne for the one Cnulltransparency) to obtain the now improved reeult for the other (f,-invieibility) also. Our definition of the latter is a generalisation of that of Stearns, that contributes a further enaller iryrovement. In addition lre are rather mre e:cpLicit about the that eonespond to regularity and phenomena irregularity respectively. ![e exhibit the fact that our main construction picks out a fanily of conputations that reeernblee a one-cotmter autonaton in structure and tras the sane behaviour We vie-a-vis regularity. shall use the notation introduced in 51.5. 106 7 .2 Null-transparency dnd Sinvisibil-ity It will be convenient here to say that, in a derivation c i (c) ct, r*rere q = o1ct2o3, "the o, subderivation segment for pcips the urU" in c iff c * (of) ,"ruoJ) and (srtlo,5) * (oZ) (sr 'oqi) some B, B t. Definition (srto) + (q) ct is a i-deritation w.r.t. index set N iff there are fewer than j pairs (nrn) of consecutive elements of N with the property ah.t ,rr' is popped by a non-null subsequence of d. Definition the segrnent trtf i" $!1g!g!!!g in (srutttttJ") w.r.t. index set iff for any sr, and any l,-derivation (srutt^ttto") + (cl) (sr rtttlt), it is the case that (strttot) * (e) (s',o). N In other words, the existence of the seguent or can only be detected in the configuration by derivations which pop by non-null input etringe at least t, of the segments of o" induced by N. Definition iff for all- s e Q' o it @ (s ,o) + (e) (s' ,0 ) * (s' ,o) * (e) (s' ,o) . The stack word LO7 A null-transparent seglnent therefore has the property that if (ero) + (e) (et$) I then for alL n > l, (rrrt) + Ce) (st,n). Thue e-derivatione which PoP sequences of a null-transparent word are incapable of distinguishing different numbers of occurrences of thern. Lema 7.1 ie null-transparent, then for all sr 01, no string a of length n can distinguish (srrtfl frorn (srttrt ) for mrm' > n. If rrl Proof s derivations from the t\to configurat,ions. If these have at no stage popped the top n + 1 copies of r,r in the stack, then clearly they cannot be distinguishing derivations. However, if they Congider have, then at least one copy of trl must have been popped by an e-subderivation. But then by null--transparency' the rest of the o seg@nts'rould al.so have been popped at the same time, without leaving any trace of their number. Thus we conctude that no a of length n can distinguish the given configurations' I I{e now prove by an inductive argument the existence kinds of segnents in sufficiently tr of both arge configurations. Lema 7.2 For a configuration c with stack r.l and an index set N of etements all lese than lrrrll fr' 108 (i) fr't q! + some segrnent of o induced by N is null-transparent (ii) F t l(gq)q + , some segmenr of ur induced by N is 0-invisible in c, provided that Lrj. > 2. Proof For each part ne produce an inductive aesertion of the form A(Pn'rNm) N, . N. for We show m= 0r1r2..., wherep* c a finite set p, that A(PorNo) is true for Po = / and No and = N, that, it -% is eufficienrly large and A(prrNr) holds, then either Nalready induces the required segment, or else we can m and find Pm+l and N*, such that A(P*lrNrn+l) is true also. that the induction terminates, ensure We and thus guarantees to produce the required segrnent, by showing that Pm qE t*L that -N--' Then by m+Iis greater than sone given function of F_. mpicking N large enough initially, lre can ensure that, though the and sete N, nay get successively snaller, they will always be large enough to enable the induction to continue untiL P ie We take P to be Q, and the assertion A(PrrNr) to be: irj e Nr, Then A(PorNo) (i) exhausted. " . Pr* (sro-) * (e) (srn), is trivially true. We agsume that A(PrrNr) is true, and that k r the srnallest and largest elenents of N, respectively, are distinct. and kr 109 Thcn if e_ r_r KrK' is null-transparent, the result is proved. Othemise, by definition there erist srsr such that (g) (s,r,rn,n, ) * (a) (s' ,0) , but (b) not (c' ,\,1r ) + Ce) (e' ,o) " Thcn let N*, be the eubset of N,o indexing the most frequently occurring etater Bay stt, in the popping sequence for (a) w.r.t, P_ No state in this popping sequence N , and let P_", lEtI = m u {e"}. $q4n belong to Prr for that would iryly, by the inductive assertion, that I| = srr . Pr, which would contradict (b). Thus it follows that Pr€ Pm+l, and then aleo that Fr*1 =F*/(C-m), for m ( g. A(u*rrNr*1) ig then clearly true. Also, if fr'= Fo t Q!, the induction can continue, if necessaryr until P, erhausts Q, without "r"t F-m < 2 occurring. This (ii) now We assume completes the proof of Ci). that A(Pr,Nr) is true, and thst k and kr, the srnelleat and eecond smallest elements of N, respectively, are dietinct. Then if utrk, is 0-invieible in c, the result is proved. Othenriee, by definition, there is some a-derivation that i3 an .0-derivation rcndering it visibl€1 i.€. (a) c * (o) (stroorkf), but (b) not (s t ,rrto ,U, ) * (e) (s t ,t^ro ,n) c derivation extract the e-subderivation that poPs the rc8t Begrnnts induced by Nrr and let Nt*l b" just those indicee inducing theae segmnts. Then, by definition, Ft*, > tF -flll. From the 110 in the popping sequence induced by fr'rr*t can belong to Prr for that w,ould contradict (b). Let s" be the nost frequently No etate occurring Btate in thie sequence, let N*, be the set indexing theee occurrences, and l-et Now P*, = pr, {s"}. if fr'= f^o > 0([q)Q irriti"lly, Then Nntl > Ntr*l /(g-a), the induction will continue eucceggfully until Q is extrausted. [l Note 1 The bo'nd can be ghown for (i) of q! is the same to be optimal by looking at as that of stearns. rt Begnents that are e-popped fron all stateg thereby performing peruutation operations on them, and regarding these as elements Note of the symetric group on q e1ement6. 2 rn the main theorem we sha1l be interested in c-invisible segnents, where.Q, is e:(ponential in g. rt is here that we gain our eignificant iryrovenent, by obtaining a bound of order (gq)q as compared with one of oc+9 given by stearns. rn general, if.0 is most of larger order than q, we can ehow that our bound is of optimal order tn the fotlowing sense. A dpda can be derived from rhe proof of 7.2(ii) with the properties that for some configuration and index set of size (llq)9, no .t-invisible segment can be found. t11 7.3 Main Theorem Theorem 7.1 rf l'{r a dpda in - stack words and of norrnar. l-ength form, has q states, t stack syrnbols, at most h in its transition rules, if t(M) is regular, then L(!"1) is recognised by sone finite with fewer than X(qrtrh) states, where X is of order E2, and automaron Proof shall prove that there is a function ycq,t,h) of order E1 that, if any reachable eonfiguration of M has height greater I{e such than Yr then either we can cut a segment out from the stack to obtain a small-er equivalent reachable configuration, or el_se there are input strings 6r$, s.t" the configurations reached after inputs of orol for m = Lrzr... are all pairwise inequivaLent, if L(M) is regular, the first possibir.iry musr always hoLd for configurations larger than y, and consequently M can onLy Thue be using up to qtY painvise distinguishable configurations in recognising the language. fhis gives us the required result, To prove the existence of Y we consider an arbitrary derivation cI -*c= (errrr), lrhere l*l = n > y(qrtrh). tle let N be the set of f"ntegers indexing the most frequently oecurring rnodes in the staeking sequence of this derivation. Then cLearly f > n/qth, tL2 I I I I I I f t L I I I c2l t I j I I I I crJ I I nu11-transparent segment r clt I I ( t I I I I t I cir rY I I o I I I I t I I I o I ., I I I I I I I I I I I I ) 1 I I invisible t segment r I ( I I I I I I I a, K TIII o. KrE iterations of nul 1-transparent ct K TID ct K c, K Fig.6. segment rlll rlll Constructions in Proof of Theorem 7.1 I 113 rf N is large enough, then by Lenma 7.2 (ii) , we can find a segrnent o.:, in it that ie (qgl)-invisibl-e in c w.r.t. N. Furthermore, 1J by the choice of N, the configuration ct = (sroorirjrr) is reachable from c^ via the input string o^ 10: _. I^le shall_ prove that if t(M) I ura jrD is regular, then c lct. For siruplicity we shaL1 not nodify the indexing of the segments of the stack in the translation from c to cl (e,g we shall still c I cr, and let I be the shortest string distinguishing Suppoee them. Then refer to the top s)mbol of ct m rrr-r,rrr. by the construction of ct, for some yrr1 s.t. B = yl, c * (V) (strt'lori) thie is not a (qq!)-derivation. It foll-ows that where an Nf c N of at least ql + t integers between j and n we can pick with the properties that (a) no segment of y, (b) the of tir induced by Nt is popped by an e-subderivation and elements sequence of N' all index identical states in the popping for y. lle now pick a null-transparent segment ,Or, irdrced by Nr in ,jrrrr as guaranteed by Lemra 7'2 (i). We to in c do with the popping of this c segment * (trrrr) "2, .z * (v*rn) "1, c' * Since yn define sone nerr configurations and cr: (trrro,) .i, .) * (v*,n) ci. is a minimal distinguishing string for c, ct, yr,jn must be and ykrjn for ca, .i. Since, by the "2, "Zr, construction of N', yrrk is non-nu1l, ylrj .rU t*rj rr"a be of a mininal one for different lengths. Ilence it is impossibLe rhar both c, = c, and ci = cj. 114 I.lithout loss of generaLity we shall assume that c, * c2, define the farnily of configurations'{rclr > oi by ,: ) r". - c-s f 1o '*o ,k*k rrn By the choice of NrNt the top part6 of the stacks of these will consist of We iterations of the null-transparent word. If we let {, = orrryrrr{,n forr=0r1r..., then Er r+1 '-c-]c2 Thus . rc *"1 and Er forr)0. c, I c, =t t" l t*1" for all r > O. transparency of ,krr, From thie, and the nu1l we can now deduce that t" I l*t" for all r > o, I > o. othemise for some t, L, and conaider the effect the configurations x'c+r" of inputs olr, and onrr, fot successive For if ate aseume values of x of Orlrz..., x'c*t" = then lte are led to on deduce that t*1" for all x ) o. rc rnd *l*t*l" = 7.1, the shortest distinguishing string for t", t*1" cannot distingui"h *0+r", xf,+r+l" for sufficiently large x. Howeverr by Lenma c f ct has forced us to the conclusion that the configurations'{rc} (or a corresponding set To sr:cnmarise, the assuqtion constructed from ct r)are all reachable and pair*dse inequivalent. Thus for regularity it must be that c = ct, which is the result we want. From Lema 7.zGD we recall that for the above construction it is sufficient that fr b" gr"ater than (e2q!)q*l. Therefore 1L5 Y(q,t,h) - qth(q2q:)qnl is the function and hence Note we requirc. This is of order fI(tog 6 + 1og h * q2 lcg q), X(qrtrh) is of order r21tog t + log h + q2 log q)" il 1 Meyer and Fisch.r2l q = t and h = 2t states. for the utl "hoo, that for a certain one-turn equivatent finite-state automaton requires r2(q) The bound obtained by Stearns case dpda wittr for X is of order s3(q * log t) of h = 2. tilote.l We notice that the family of configurations {rc} exhibits phenomena strongly reminiecent of the notion of propriety in one- counter automtte. In particular, they are either all equivalent, or all pairwise inequivaLent, 7,4 Bounds for Subfanilies !,le now give irnproved upper bounds for the function X for three restricted families of dpda, hound and show in each case that the order sf ttee ie achievable. For the classes Sg and R, since eltpves no longer play a part, the above analysis becomes trivial. From the definitions in 57.2 it is imediate in such cases that N > L is sufficient to induce a nuLl- 116 transparent segment, while N one. Substituting these in Corol-lary -* 7. > l, + 1 guarantees an l,-invisible the above argument gives the fol.lowing: 1 For the class Sg there is a bound X of order fl1ht.l"og t). CorolLarv 7.2 For the class R there is a bound X of order nl(nq2r.1og r). From Len'rna 5.4 it can be easily rcsniction h = 2, the bound dominant deduced q. (S(q) + pf that for (q)) suffices, factor, S(e), is of order El(/alog-q ). a eimilar bound rnore C We with the where the can obtain directLy as fol.lows. Corollarv 7.3 For the class C, with h = 2, there is a bound X ^, q2.S(q), **:ngre in- = q}, S(q) = lrrx {i"c,ru in,}l 11' Proof Ler {xri be the set of net staci( drops due to loops of e.ru*ves in the di.agrann for non*enpty counter transitions of the LL7 M. It follows i&nediately that any stack segment of machine length x = max {g.c.m.{x.}, will be nuLl"-transparent, where, q} rnoreover, x S S(q) since the e-loops must be disjoint. can then obtain the claimed bound from the construction in I^Ie 7,L Lt we also observe that any jderivation (w.r.t. the Theorem positive integers) popping a segmenr longer than jq must fini.sh within an e-loop" I We now ehow that the order of each bound is achievable in both of the genses defined in gl-.6. Exanple 7*L Let M be a sirnple machine with t = aU, fB, where fA =' {A.lr = i < m}, l, = {Brlo * i < m}, with x = tailr s i < rn] u {arlo e rf,}, wirh ", = Bo, and rransirions a B, $ sBr., Bm -9--t A J J+T a, A" *l-+ A J C< to occur. But there s1 lsj<n Clearly the stack has to grow to height and the turn j ( nm + L for B, to be reached reachabLe "r" rtt distinguishable configurations of rhis height. Since L(M) is evidentLy regular, l_ this shows 18 that X of order fI1nt.log t) is achievable. ! Exanple 7.2 I.Ie construct an Mf e & by generalising the previous in the following way. eugment We I by a synbol g, {(s"rfl)}, l"eave exarnpLe introduce a state set Q = {.or...rk}, repLace c, by (sorB.) and F by the input alphabet unchanged, and replace the transition" rules by a (s'Br)*,tr*l,r'rBl) osi<k, a o < i < k, 0 < j * r, (ri,Bj) --l!-* (si,oB5+l) a (s*rB*) tr (so,Br) (sorcorBl) a1 --:> (sk,A) 6. Th:i,o (sirA3) *-J* (si,A) O<i(krO.j.m, (s1C) o<i<k. ---9'* (s.-1,4) again is a l-turn machine. After the turn is made in ctrer:h"ation, the rernaining a string i"s accepted iff it natches the staek contents, and there are exactl.y k occurrences of C in the stae[a,, Thus the turn has to be made when there are about k2rur symbols in the stacko since each e is added after successive segments length kurr" fhus L(Mt) is regular, and Mr has order f11nq2t.1og t) distinguishable eonfigurations " I of 1" 19 Exanple 7.3 Let {xrr.".**} be the Fartition of q * 1 with the greatest least comnon nnulti"ptre. Let M e C read a string frcrn a* and increment the counter by one starting state" e-Loop wi.th Then x. $rates for each a, while staying in the if a character i frorn {1,"".n} is read, an causi-ng a stack drop of x' is entered. A.ceeptance occurs iff the contents of the counter at the turn was divisible Such an M wi.th bY xn' q states clearly exists, and recognises a regular set. lloweveru after each input of n(nod am it needs to know y) for y = *tn..rxn. This requires S(q) distinguishable configurations. fi 7 "5 Tigq Camql.qxitv. To test- a language accepted by given dpda for regularity we can constrrrct the candidate finite-state automaton speeified in the proof of fheorem last test 7.Lr say Mr, and test for their equivalence. Since this can be done E- size, to show ^2 timeo it E- ::earei:r,s .*2!; I.n M polynomial time (92.5), and since M' is of that the regulari.ty test i,tself takes no more than to show that the construcf ion of I'lt can be done tlme. The states of in of Mr *orresponcl to the no nore than of height less we need are those Lhan Y. To construce Mr the on1-y new transitions that specify for Yr the (qq!)-invisible E2 configurations segments each configuration just larger than with respect to the appropriate index sets (57.3)r that can be rerreved from their stacks. For each segment L20 or of o, lre can determine the set of pairs (srst) with the property that (sro') * (a) (st rO) for o f e, and also those for which cr = e. The poasibl-e ways of reaching each point in the stack by (qq!)- derivations can then be deduced, and hence the invisible segments found. To do all this for all the tr2 configurations requires only In a eimilar way, the bounds for R. E2 time. and C give single exponential time tests. However for Rg1 and hence aLso for Sgr regulariry can be tested much more easily, and in only poLynomial- tine. If in M e Ro an aecepting derivation goes through some configuration c with height a greater than hq't, then for some pair of levels repetitions in both ehe stacking and popping sequences respectively, of must occur the derivations before and after the occurrence of c. This would irnpLy that accepting derivations can go through arbitrarily Large configurations, and hence' by the restrictions particular to the class R0, that L(M) is not regular. Since the converse of this is obvious, we conclude that bte can test for regularity by testing whether, once a stack leveL ., exceeding hq-t has been reached, any further inputs lead to acceptance. This requires only an emptiness rest on a polynomiaL size machine. 121 IV A}I schemas A?PL]-CATION TO SCTM},IAS are direct formaLisations of coinputer prograns, and closely resemble them in syntax. Their essential characterisation is that the meaning of the of schemas conrnands is left undefined. The theory relates the description of such a formalised program to its possible computational effects once interpretations of various kinds are given to its uninterpreted comands. The relaticnship between certain schemas which have register3 and automata with one-way input tapes, is now just one working well known. t, Rutledge" has established a elose connection between one-register flowchart (Ianov) schefiras, and fi"nite-state automata. When such schemas are augmenteetr deterministic by a pushdown sta.ck, lhe correspondence transfers to pushdown autornata" This corresponclence is such Lhat, as il"l.ustrated by Patersonl3, an equivalence test for a subfamily of this class of automata leads directly to a test for strong equivalence (in the sense of Luckham, Fark and Faterrorr23; f,or the corresponding class of schemas. Thus, for exampler our result in Chapter 5 implies that strong equivalence is decidable for lanov trn an analogous way, the of whether a to one regularitl' tranov schema with an auxiliary counter. problem corresponds to the question with a pushdown stack is strongty equivalent without a stack. As the translatabiliny autonata to we schemas schenas of is direct, such decidability results and depends on a shall not pursue these further here" from well estabtrished techni{r"l3, t22 In contrast, holrever, the complexity problems do not necessarily'translate a canonical dpda for measures for directly. For some such schema may these exampLe, require to be of exponential size in terms of the parameters of the schema description. For this reason rre shal1 investigate the coupLexity of just one of these problems, for a case which is of special schema fheoretic interest. Monadic functionaL (deBakker-scott) ."h"r.r13 ,24 t25 are a f,ormalisation of recursive programs with a single working register. Faterson26 h", shown that for there does not exist some such schema an equivaLent flowchart scherna with any finite nuriber of registers. This ean be interpreted as corroboracing our intuitions about the increased poner of recursive notation. Other examples are known whieh can be flowcharted, but not with a single now elaborate on these ftrnetional schemas register. relationships by considering those that do have equivalent ranov schemas, tr{e shall monadic and giving a measure to the subetantial succinetness with which soup large Ianov sehemas can be re-enpressed In particular, by equival-ent srnall ftrnctional ones. we shaLl show that sorre functional schemas require an equivalent fLowchart to be of a double exponential (r2) size in terms of {es original parameters. Further, we shall show that they may not require larger flowcharts than this order. Thus we shalL be our f.ntuitive giving a theoretical resuLt to correspond to knowLedge about the considerable econony of description that can sonetines be gained by recursive notation. L23 Chapter UONADIC FUNCTIONAL SCHE}{AS 8 8.1 Introductisn In our definitions we shall- in the main follow Ashcroft ' Pnueli24, tho give iLLustrative examples and rather Manna and more detaiLs. e d!:S_gg@_ has a finite set F,of, t., monadic fungFion variabl-es {F-. t. } (with a distinguished initial fqnglrg4 F^), a finite set f, of t monadic function constants {f,}, a finite set of rncnadic !f,e4:Sgggg. {pi}, and an g individual variable x. A term is a cornposition of functions applied to x' e.g. ttFZ(f ,(x))) which is for short, omitting -the brackets and the x. A written aB f,F^f. i_23 conditional ter:n is any finite expression of the forrn if pi then t, else t, where 11 specif ied , ,2 are terme or conditional terms. The schema bya set of function definitions, one for itself is eac.h F. , of the form F.1 € r where t ie some tern or conditional. term. It is useful LU reserve the nane 1 for the idenrity funttion (I(x) = x) r and l.-, for the undefined iunction (F* e F-) The schena fo(x), " is evaluateC in the expected way by starting from and appLying the rightmost function variable each time to the argument (i.e. to the string frorn f*x to its right). The only 1"2t+ further information lre need to define completely such an evaluation is to specify the Ehe course of val"ues taken by the predieates each tirne a conditional term has to be evaluated. Thus an for interpretatiog specifies the values taken by the predicates each argument, and we assume that it does so uniquely. Thus in eval.uation steps in which no function constant is applied to the argument, the truth values of the predicates cannot change. In other words, hre ean regard an interpretation as a function from f*x to 6 =' {-, +}9, where - and + indicate falsity and truth respectively. Since each interpretation an evaluation, for a schema uniquely specifies it defines a string (possibly infinite) of the 60111 f. 6. t. o1 where form: .... 6, is the truth vector defined by the interpretation for the J ergunoent t.tj-t .... f.tl t..to x. Conversely, each such string describes the step by step eval-uation of a function, as well as a set of interpretations. The schexna for a partieular int,erpretation is said to have a defined vaLue if and orrly if the eval-'-retion terminates producing a germ containing no more function variables. This term, consisting of only function constants applied to x, is then the defined value. L25 8.2 Eval-uating Pushdown $rtomatg A dpda M can be easily eonstructecl to evaluate such functional schema. At each step M keeps all the a unevatuated functions (i.e. everything to the Left of the argument) i"n its pushdown store, with the rightrnost function variable at the top. It reads input words from (6f)* i.e. strings of al,ternating truth vectors and function constants. Wtren there is a function variable at the top of the stack a truth vector has just been read, the variable is replaced in and the staek by the term specified by the corresponding function definition for the truth values read. state control, and these The vect,or is remembered repl"acements continue it, until a function constant first This is checked against the next via a F- replacement if it does appears symbotr- in in the finite accordance with at the top of the stack. on the input tape, and rejected not match. Otherwise, the function constant is popped (and can be regarded as being output and appLied tc the previous argumentr) and the next 6. is read. If further function *en-stenes then appear c.hecked at the top cf the staek, ttrey are similarLy against the input tape! with the intermediate truth vectors on the tape bei.ng ignored" The dpda M by enpty stack. ctarts with F 0 i"n its stack, and accepts strings Then elearly any tape accepted by the vaLue taken by the the tape. Also, the schema it wi1l specify for all interpretations consistent with language recognised from an arbitrary configuration L26 of M wilL relate in the corresponding term Fron this it in the way same to the values taken by the in the schema for different interpretations' fol-Lows that two configurations of automnton sense if and only M are equivalent if, for any interpretation, trro corresponding terms in the schema either evaLuate to the value, or are both undefined. Thus if M has only a finite X, of pairwise distinguishable reachable configurations, the scherna can same number then be rewritten as a lanov schema with X boxes. By examining the dpda M we have it the can be constructed just described, we find that to have the following parailEters: t=t+ttc v' where trr is the l"ength of the l-ongest term in the function h=hr definitions, p=28*t", q = 28 + 1' I{e notice that the only kind of that of being able to suhclass of Dg remember memory M M needs the Last input character read. with thi.s property we shall that the evaluating machine capacity Ehat denote by dB-S. I^Ie is The note is a special form of a dB-S machine, sinee a subset, f, of its input alphabet need not be distinguished in nernory, and also no element can be e-popped. That dB-S verified by examining the of f, regarded now as stack syutbolst is itsel-f properly contained in language { anb"t }u { anac?n }, Dg can be L27 8.3 dB-S Autonata The above mentioned that for dB-S nachines, restriction on memory implies directly states eaRnot change in the courge of e-derivations. A consequenqe of this is that a simplified analogue of Lenma 7,2 car. be derived. Lema 8.1 For a configuration c of a dB-S machine, if c has stack and N is an index set of (i) (ii) Any segrnent fr- elements all Less than of o is nulL trensparent, til, l.,rl, ttet and If F > 2.Q,t, where l, > 4 then some segment of (,) induced by N is .Q,-invisible in c. Proof (i) Trivial. (ii) Using the sarne notation as in the proof of Letma 7.1' I{e apply the induction chooee P principle there stated. = I and the inductive aseertion However, no!il A(PmrNn) to we be irj e Nrr A e Pr+ A does not occur it rirj. Ttren A(PorNo) is trivially true if Po =6 ar.d No = N. hre then assume that A(PrrNr) is true, and that rrrt, the smallest and second snallest elenents of N, are distinct, tt ,rrr, is L28 0-invisible, then the desired result is proved. 0therwise there must be some o-derivation that is an l-derivation rendering it visible i.e. (a) c * (o) (s t rr^ro., r) but (b) not {e|,oorr,) * (e) (st,oorr). We notc that this second condition now necessarily iurplies that o-ErE ' ort'-,). -r cannot be e-popped at all fron the configuration (strur^ Let N*, be the Largeet subeet of N* induced by the popping sequence of an e-eubderivation of the o derivation. Then fr.*, > (Nm-l-)/1,. This e-derivation must eventualLy terninate when some Af in the stack is reached. Furtherr this occurrence must be at a higher level in ur than r, whieh means that Ar I Pr" Atso, Ar cannot occur in any Begrnnt induced by Nnrl, for then that occurrence could not have been popped in the sane e-derivation. Thus, if Pr*l = P*, {At}, then A(Pr*lrNrn+l) must also be true. To ensure that the induction can proceed, if necessary, untiL l ie exhausted, it is now sufficient that fr' > zr,t. n In the light of this we can rework Theorem 7.1 to obtain the fol"loring. Lerma 8.2 A dB-S dpda recognising a reguLar set can have no more than 7 X = E'(t.1og q + 1og h) pairwise inequivalent reachable configurations. L29 Froof Wc use the same notation as Flav$.ng ehosen l{, we now in the proof of Theorem 7.1. require oni.y a q-invisible segnerrt indrrced by !i. Fcr then, usi"ng the shortest distinguishing string g, assuming 8rlaf. i"t exj-sts, we can find a nutt*transparent segrnent induced by N such rhat the B-subderivation nsf lre'e's cause a require* anei, of net change that pops it reads sone inprrt, but in steee. This is the construction we frenn Lenmna 8.1.n it clearLy works for any confi.guration hei-ghe greater than YrwhereY is of order 2hqtqt. This gives the required trotnd for X. I m-s" Th{"c shows that the dB-s restriction reduces the q-dependence *f, the l:o'urd x from a double to a singl"e exponential expression. 8,1* Bgqnllq*g rsuccinctness Freim Lenuna 8.2 and the observations of 58"2 we imediateLv e'l.1.Vg Sbes.gss" -E-*L For a dE-S terrns sohema wi,th t funeti.on syrnboLs, g predicates, and of !.ength no more than h f,n its fr,nction definitions, if l-anov schema strongly equivatenc to it (i.e. under all interpretations) exisBsr then the !.atter howss, [,-l a need have no more than order n2Gg + 1og h) 130 It remains to show that an E2 order of size rnay indeed be neceSaary. Theorem 8.2 For each positive integer n, there is a dB-S schema with 3(n+1) function synbolsr f, * 1 predicate symbols, h = 2, and with total description linear in n, that has strongly equivaLent schenas, lanov hut only of size at least n2(rr). Proof I{e use the same idea as Meyer and Fisch.t2l corresponding resuLt even in this more for dpda, and show that it restricted ' {f*, f-} can be made to work framework. For each n lte construct the following variables {For ,rr" for their schema tir...tlr ti,...ti, !t,...1r,}, with fr.nrction fr:nction constants and predicate"' {prp1r...prr}: Fo € lf pt then (if p then tl to f+ el'se ti to f+1 else 'f p2 else if p' then (if p then F+ to t* else F- Fo f*) eLse ++ F-t f . + F: € i.f p, then I else "f p2 . . else if pi-t then I else if pi then (if n then \ eLse if pi*l th"t Felse I. f+ else I) . . . . else 't Pn then F- 131 F. € if pl then I else "t pZ . " el"se if Pi-l then I else if pi then (if p then I el-se F. f ) . eLse 't Pn then F etrse if pi*1 then F* . el-se L. &. * if p, then I'- eLse 't pZ . . else 'f pi then F else I. Ler *= nin' i j I rrttf*tj) is false for 1 < i < n ] for a particular interpretation. Then the corresponding evaluation will consist of taro partsp definition, of the F_ o The m first part the first one consisting of and the second only of appLications of the others, produces an unevaluated term which over rhe function variables irj,...tl, m appLications is a string of F;,...t;), l-ength whiLe the second nonotonically shortens this, unless an F- occurs. Changing tG our dpda eerminology, we notice machi"ne that we have a l-turn which, depending on the inputl cdrl read any word from -+ Flr...trr}* into the stack on the uPstroke. Denoting by {t;.r...Frr, b *ny trr-lth vec!:or such that p1 = Erue, but p. = false for j ' k, we observe that each fl1-Ior Fr- as e-popped by k if k < i, to rejection if k.'i. If k = i bgt rhe sign *i F. and leads does directly not rnatch the tru8h vslue *f p, then the syr:bol is again e-popped. However, if there J"s mlat,ching, ,;hen the apprcpriate one of f+ of f is "evatruated" and the next trurh veetor is read from the input tape (i.e. the "interpretat{.or:"'} ' Leaving q *t the top of the stack, The rol.e of li. it to insist that the next vector be k for k > i.. Thus in the downstroke no more than n truth u-reeeffr's, aud n furrction rii:rrr$tiri"its e,nn he readu from whieh we deduce that L32 there are only a finiCe number of inequivalent configurations in such a machine. To see show that this number is, however, of order E2(rr), t" that for any of the 22t r,rbr.ts of {r*, f-}o, there is some configuration which can evaluate, depending on the interpretation' to everything in the subset but to nothing in its that any complement. We observe such subset can be represented by a binary with branches marked by a sign from {+, -}. tree of depth n Furthermore any such tree can be represented by a string from {t;r"..F*, F;r.'.F-}* in the following polieh notation: Each Fl represents a Left branch from rhe (i-l)th to the i-th leveI of nodes, and f] a siurilar right branch. fire string itsel"f is the expansion of the tree from the O-th l-eve1 (the root) by the foi-Lowing recursive process: <tree.> + <branchl> <branchlt, O < i < n 111 <tree > + n A <branchl> + <branchl> L Any A I tl*, <tree.+l>, o s i < n * A I' rl1+I- <tree,1+I-), O< i <n string generated from a!t€e6> defines the set of paths in tree that go from the root to n-th that the eval-uation of such l-eveL the nodes. It is easy to verify a string by our schema corresponds exactly to tracing a path in the tree it specifies, and that the possible final values such a string can take correspond to the paths in this tree. I 133 trle have therefore esrablished rhat the sr.lccinctness we are i-nvestigaEing involves two l-evets rueasure of exponentiation' lle note, however, that, in the above example g and t cannot be varied independentLy. Thus to obtain more deuailed results for leadi-ng exFonent for the different combinations of uhe parameter value$, furEher analysis is necessary. For exarnpteo in the schema above, we eould have economised on the ntrnber of predicates aE Lhe expense erf greater comptications used, in description' to obtain a sii.ghtly better result for one partieular such c:t6€. 134 COI{CLUSION l,Ie have shown pushdown autonrata that the various deterministic families of are rich in decidable and potentiall"y decidable properties. In doing so hre have also indicated areas outside automata theory to which our results relate. Several weLl motivated decision problems have been left tmresolved. Moreover, even for those shown to be decidable, the derived procedures usually require at least exponential time. we do not know of any arguments to aLgorithms do not show that exist for these, important lrle can, however, single out from As polynomiaL time gaps remain here al-so. among these open questions the ones which appear the most inurediate. Finding an equivalence test for the unrestricted class of deterministic pushdown automata nas the primary unachieved goal of our work. Although the existence of one can perhaps be now conjectured with considerabLe confidence in the light of our results, a proof of this would still be of great interest, for additional insights it techniques related may pro'.ride. It appears plausible that to our parallel and alternate stacking constructions, and our simulations by nondeterministic automata, wi1L play a suggests the pushdown part in settl-ing the probLern. Our work that the resolution of our conjecture about aLternate stacking for the class \ r and the finding of an equivalence test for the class S, My be significant next steps to that end. A positive solution to the latter problem, which we have not investigated, rj5 also appears to be a prerequisite for finding a test of strong for deBakker-Scott eguivalence schemas - For the regularity problen, since we have improved Stearns I test, to a near optirral level, to achieve further improvements new approach between this is necessary. and a number cf The a relationship we have established containment problems which are currentl-y open, can be interpreted as a trto sided challenge. In the abse:ree of a more efficient regularity test, it hints that if these other them problerns are decidabl-e, then it may be difficult tc prove to be so. We because have not given much attention to the inclusion probtrem of ttre welL-known negative results concerning it. However, since sre ha"rq proved its undecidability for even a very restricted case of seems R6 , a resol-ution of this probLem for the sinrple machines Sg mcst t.imeLy' Thus our work suggests that there remain numerous distinct fealures of fhe strlrcture of these classes of automata yet to be uncovered, and indicates some speaific directions along which these might be sought" The rer'rards E:f this search will be' we betieve" to increase ottr understanding of these computations, and to i:ender particutar instances of them more susceptible to practical analysis. 136 REFEREI{CES tt"l GINSBURG, S. and GREIBACH, S. A. Deterministic Context-free Languages. Inf. and control, 9 , 62A-648, (1966). l2l STEARNS, R. E. A Regularity Test for Pushdown Machines. Inf . and Control, !!, 323-340, (L967). t3l HOPCROFT, J. E. and ttLLl4AN, J. D. Formal Languages and their ReLation to Automata. Addison-Wesley, Reading, Mass. (L969). ' t4l GINSBURG, S. and SPAI'IIER' E. H. Bounded Algol-like Languages. Trans. Amer. Math Soc., L13' 333-368' (L964) t5l McNAucHToN, R. Parenthesis Grammars. JACM, 14, 490*5oo! (L967). t6l RABTN, M.O. and SCOTT, D. 17) BrRD, M"R. The Equivalence Problem for Deterministic Two-tape Automata. JCSS 7, 218-236, (L973). t8l ROSENKRAI.ITZ, t9l KORENJA,K, t10l Finite Automata and their Decision ProbLems. IBM J. Res. 3 : 2' LL5-L25, (l-959). D. J. and STEARNS, R. E. Properties of Deterministic Top-Down Grarunars. Inf. and Control. L7,226-255' (1970). A. J. and II0PCROFI, J. B. Simple Deterministic Languages . IEEE 7th Synrp" on Sqri.tching and Automata Theory' Berkeley, California' (1966). I(ARP, R. Reducibil-ity Anong Combinatorial Problems, in Complexity of Computer Computations (R. E. Miller and J. I,f. Thatcher, eds.), Plenun Press, N'Y. (1972). t 11.1 [12i lt. i{t{urlr the" translation of Languages from left to right. Inf. and Control, 8, 607-6390 (1965)" On HARRTSON, Or, M. A. and HAVEL, r. a Fanily of Deterministic M. Grarnmars, in Aueomata, Languages and Prograrming (1"1. Nivat, ed.), brorth-trtc1land, (1973) . [elso three more detailed reports; Department of flornputer Science, University of Cslifornia, Berkeley.i il-31 [14] M. S. Decisinn Problerne in Cornputational Models. Pror:. of ACM S:rrnp.on Froving Assertions about Programs, Las Cruces, New Mexico , (1972) . PATERSON, ciroMSKY, N. Context-free Grammars and Pushdown Storage" Quart. Prog. Rept" No. 65, MIT Res. Lab. Elect., LB7'L94, (te62) [15-1 . GREfBAc]i, s. A" Normal. Fonn Theorem Grancnars. JACM, U-, 42-52, (1_965) " A- New for Context-free Phrase Strrrcture t16J ii.j-l{SKY, M. L. t17l tsAR-IITLLEL, Y", PERI.ES, M" and SI{AMIR, R. On Formal- Properrties of SirnpLe Phrase Structure Grantnars, in Y. 8ar-Hi1Le1,, Language and Information, Aeiuisuri*WesLey, F,eading, Fla$rr., (1964i " tl.8:l Comprrtation: Finite and Infinite Machines. Frentice-Ha13., New Jersey, (L967). F R"TESMA$I, E. R. Tn* k*. ].'L;s:;,eir Prerbienr f*r Mor:tarii-e B"*pnrt-u Center for Res*, ar:ch Ilar."varr: Ljni.versitl'. r (1i73) . iri R*,:i:rsion Sc.hemes' Conputing Technology' t19l 4TNSBURG, S. snd SFAN{ER, E. iair'ite * Lurn Fuehelor*n Autcngta " SIAI'I J. on Control 4, 423'434, (i.966), tzO-l vA{,I.A}iiTo L. G. and PA'rERS0}I, M. S. Deterministic One-Counter Automata. Iroc" ,lli Conf , otri &trtnnrata T'heqrry and Formal Languagest Bon.ro Gernany, (i-qi3)r. 138 i21l 122) A. R. and FISCHER, M. J. Economy of Deecription by Automata, Grarmnars, and Fornal Systems. IEEE 12th Symp. on Switching and Autonara Theory, (1971). MEYER, RUTLEDGE, J. D. On lanovts Program Schemata. JACM 11, 1 - g, (1964). l23l LUCKIIAM, t24l ASHCRoFT,8.,l,tANNA, t25l PARK, D. M. R. and pATERSON, On Formalised Computer Programs. JCSS !, 220-249, (1970). Z. M. and PNIIELI, A. S. Decidable Properties of Monadic Functional Schemas. Int. Symp. on Theory of Machines and Conputation, Haifa, Israel, (1971). DE BAKKER, J. t{. and SCOTT, D. A Theory of Programs. Memo. 1261 D. C., , L969. PAIERSoN, M. S. A Sinple Monadic Recursive Schema which is not Equivalent to any Program Schema. Memo. , 1"970. 139 APPENDI X TO SECOND TI"IPRESS ION ;.j.trce this iaesl-i.r1,.s rcport was ha',.e been ob';ained that first Lo its relai.e directly f.iere we drarv att.ention conLents issued several to some of these: - frr Chapter 2 iL was shown that an arbltary dpda can be converted to one in normal form by Al'r-rilhnr q-.rv t!\JrltL<aL yvt.:/lrrnaai-l ]_>lliL q^rrrv ;lnalogous exponential 1 U1.l man- ). It *iqa urr:is -a- uPpvrEu ^nnncod time construction l-n Lv tho Lrrs of Hopcroft and has since been shown by the author that the conversion can be done in linear nachine. This strenqthens given in Chapter 6. I^le note that Rosenkrantz (Computational and context-free an algoriLhm requiring by H.B. Hunt and D.J. paralleIs between the regular languages, Proc. 6th ACM symp. on Theory of Computing, SeattIe, I974) The construction irr r-h-rnlsy time on a random access t.he force of the reductions quadr:atic time has been reported ,arl\r)r \/ an 3 (i.e . used in the undecidabilitv Theorem3.4) has been used by I ii.P. l'riedman (Inclusion Eth Princeton LC)7 4 Problem for Simple Machj-nes, Conf. on Information ) to show that the inclusion and Systems Science, problem f or simple lanorracres is also undecidable. A modified 4 can be found in: for deterministic Control, formulation Valiant finite-turn 25, June L974, L23 L.G., of the contents of Chapter The equivalence problern pushdown auLomata, Inf. -f33 and 1/^ The results wiI I appear as : Val iant. L. G. , and Paterson one-counter Determinisuic automata, The nain reductions substance of ChapLer 7 will i;or:rrla-ii-rr anrl rol:ied r.uJuLqL Lvl ;ru Loma t-a , proofs in Chapter 5, with revised JCSS M. S . , , L975. of Chapter 6 and t-he appear in: L.G-, Valiant nroblems for deterministic pushdown r!v+ , L97 5 . .j-AC14 In relation of chapter B it to the monadic functional schetnes has been shown by E. P. Friednan t.h.at t-he decicl"ibil i ty of equival-ence f or these would imply tlie of equivalence decidability for the whole class of pushdown automata. deterministic L.G.V. tlnirznrsitv of :je:pLcnber, 1974. I.eeds