An Introduction to Formal Languages and Automata

Page 1

b d

An Introductionto FormalLanguages andAutomata Third Edition

PeterLinz Universityof Californiaat Davis

6l;l .rf;ti.

etig*

dtry'l-,tlti,tFI, hgfryfl6a

t' o

filru;;;:

';;'ut: \ n qf I" A\ ,r'f7 lA ,obi

IONES AND BARTLETT P,UBLISHERS .BOSTON

Stdlnry, Massnclrrsrtr TORONT'O LONDON

SINGAPORE


Workl Headquerters Iones and Bartlett Puhlishers

40 Tall PineDrive MA 01776 Sudbury, 978-443-5000 info@jbpub.corn www.jbpub.com

Jonesand Bartlett Publishers Canada 2406 Nikanna Road Mississauga,ON L5C 2W6 CANADA

Jonesand BarJlettPublishers International Barb House,Barb Mews London W6 7PA UK

Copyright O 2001 by Jonesand Bartlett Publishers,Inc. All rights reserved.No part of the materialprotectedby this copyrightnotice may be reproducedor utilized in any fonn, elcctronicor mechanical,including photocopying,recording,or any infotmation storageor retrievalsy$tem,without written permissionf'romthe copyright owner. G'

Library of Congress Cataloging-in-Puhtication Data Linz, Peter. An introductionto formal languagesand automata/ PeterLinz'--3'd cd p. cm. and index. Includesbi hliographicalref'erences

rsBN0-7637-1422-4

26+,3 . L5 4 LooI

l. Formal languages. 2. Machine theory. l. Title. QA267.3.Ls6 2000 5 | 1.3--dc2l

A

00-062546

Chief ExecutiveOfficer: Clayton Jones Chief OperatingOfficer: Don W. Jones,Jr. ExecutiveVicc Presidentand Publisher: Tom Manning V.P., ManagingEditor: Judith H. Hauck V.P.. Collese Editorial Director: Brian L. McKean V.P;, Dcsigir'and"Prodgction: \ Anne $pencer V. P., Salcs anit*ffrarket+rg-i.FauI Shefiardson V. P., Man uf aeturingjandilnhrr'trrry dpntrol : ThereseBriiucr SeniorAgquisitionsEditor; Michacl $tranz f)evelopment and Product Managcr: f,lny Rose Markcting Director: Jennifer.Iacobson Production CoordinationI Tri{ litrm -Pt'oject M anagcment Cover Design; Night & Day Design Composition:NortheastCompositors Printing and Binding: Courier Westford Cover printing: John Pow Cotnpany,Inc. ;..#F*F*.,.

Covel Imasc O Jim Wehtie This book was typesetin Texturcs2. I on a MacintoshG4. The fbnt families usedwere Computer Modern, Optima, and F'utura.The first printing was printed on 50 lb. Decision94 Opaque.

Printedin the UnitedStatesof Arnerica _. -'_ 04030201

lo987654321

I

lchL

4

I ,r./1,il.t!\


his book is designed for an introductory course orr forrnir,l larrguages, autornatir, txlmputability, and rclated matters. These topics form a major part of whnt is known as tht: theory of cornputation. A course on this strbitx:t rnatter is now stir,nda,rdin the comprrter science curriculurn ancl is oftrlrr ta,ught fairly early irr the prograrn. Hence, the Jrrospective audience for this book consists prirnrr,rily of sophomores and juniors rnirjrlring in computer scicntxlor computer errgirrwring. Prerequisites for the material in this book are a knowledge of sorne higher-level prograrnrning la,nguage (cornmonly C, C++, or .Iava) and fatrrilinritv with ihe furrdarn<lnta,ls of data structures and algoriihms. A colrr$e in discretc mathematics that irx:hrclesset theory, furrctions, relations, logic, and elernerrtsof mathematical reasorringis essential. Such a corlrse is part of the standard introductory computer sciencecurriculum. The study of the theory of cornputa.tionhas several purposc$, most imprortantly (1) to fa,miliarizestuderrts with the fbundations and principles of computer sciettce,(2) to teach tnaterial that is useful in subsequerrtcolrrres! rrnd (3) to strengtlrcrr utudents' ability tu t:ilrry out formal and rigorous rrrirthematical argurnerrts. The presentatiorr I ha,ve chosen for this text fa-

III


lv

F RHr-AciE

vors the first two purpose$r although I would rr.rguethat it a,lsoserves the thircl. To prt:sent ideas clenrly arrd 1,ogive strrdcrrtsinsight into the material, tlte text stressesintuitive rnotivation and ilhrstration of idcir.sthrough ex* a,m1llcs.When there is ir choice, I prefcr arguments thtr,t a,reeasily grer,sptxl to thosr.'tlnt are concisr,land elegant brrt rlifficult in concellt. I state clefinitiorrs ancl theorems llrecisely and givt: the tnotiva,tion ftlr proofs, brrt tlf'tt:rr k:ave out the rorrtirre and tediorrs rlctails. I believe tlrrr.tthis is desirnblc for peclagogir:nlrcasdhs, Many proofs are unexc:itirrgapplications of irrduction or contra,clit:tiotr,with diff'ererrt:esthat are sptx:ific to particuLrr llrobletns. Presenting $rrdr arguments in full detail is not orrly ullllecessary,lrtrt interferes with the flow of the storv. Therefore, quite a few of the proofs are sketchy irrrcl someone wlxl irrsists on complerttlrressInay consitlclr tltern lacking in cletrr.il. I do not seq:this as a clrawback. Mathematica,l skills are uot the byproduct of reading sorrreorleelse's argutttents, but comc frorn thinking atrout the essenrxlof a problem, disrxlvtlrirrg idea-ssrritatllc to make the poirrt, thel carrying tltetn out in prtruistldetail. The kr,tter skill certainly has to be lea,rnerd,arrd I lhink th.r,t the proof sketches irr this text providc very appropriir,tcstartiug points fbr such a practitx:. sornetitnes vi(lw a course in the theory of StudentS irr courputer sclit1rrce computation aa urlnecessarily abstract and of little practical con$(xpelrce. 'Ib convinr:c thetn otherwi$e, t)nc treeclsto appeir.l tcl their specific irrterests and strengths, suclt a,stena,t:ity and inventivttntlss itt clealing with hard-toof tltis, tny a,pprtlitt:h empha,sizeslea.rnirrgthrough solverllroblettts. Beca,user probletn solving. By a problem-solvitrg approa,ch,I rrteattthat students learn the material prirnarily througlt problem-type illustrative examplcs that show the motivation bohirrd the concepts, a^swell as their conncction to the theorcrns attd clefinitiotrs. At the sa,metirne, the examples rrriryinvolve a nontrivial aspect, for whir:h students must dist:ovc:ra solution. In such an approach, htlrnework contribute to ir,rrrajor part of the leartting procefJs.The exercises exrrrc:ises :rt the end of each sectiorr are designed to illutrftrate and ilhrstrate the matr:rial and call orr sttrdents' problem-solving ability a,t vtr,riotrslevels. Some of the exerci$csare fairly sirnple, pickirrg up where the discussiotrin the text Ieaves ofl and asking students to carry ou for antlther step or two. Other extrrcisesare very difficult, challenging evtrrrthe best ntinds. A good rnix of such exercisest:ilrr be a very eff'ectivt:teaching tool. Ttr help instructors, I have provitled separately an instructor's guide thrr.t outlines the sohrtitlrrs of the exercise$irrrd suggeststheir pcdagogical value. Students need not trrr asked to solvc all problems bqt should be assigned those which support tlte goals of the course and the viewpoint of the instnrt:tor. Computer sr:ience currir:ulir tliffer from institrrtiorr to iilstitutiorr; while a few emphasize the theoretir:nl side, others are alrnost entirely orientt:d toward practiclnl application. I believe that this tt:xt can serve eitlNlr of these extremes, llrclvided that the exercises a,re stllected carefully witli the students' btr,c:kgroundatld intertlsts in mind. At ttle same time, the irrstructor needs to irrform tlle


PRnrecn

students aborrt the level of abstraction that is expected of tlxrm. This is particularly tnre of the proof-orietrttxl exercises. When I say "llrove that" or "show that," I hrr,vein mind that the student should think about how a, proof rnight be cclnstrur:ted ancl then produr:e a, clear argurnent. How fbrrrtal srrch a, proof should bc needs to be deterrnined by the instructor, ancl stutltlnts should be given guitlrllines on this early irr the txlrrse. Tltc content of the text, is allllropriate for a one-sernestcrtxrurse. Most of the nraterial can be covered, although some choice of errrpha.siswill have to be rnirde. In my classes,I gencrirlly gloss over proofs, skilr4rv as they are itr tlte tcxt. I usually give just enough coverageto make the rcsult plausible, asking strrdents to read the rest orr their own. Overall, though, little can be skippexl entirely witltout potential difficulties later on. A few uections, which are rnrlrked with an asterisk, c:rr,nbe omitted without loss to later material. Most of tht: m:r,teria,l,however. is esscrrtial ancl must be covered. The first edition of this book wrr,upublished in 1990,thc:stxxrnda,ppeared in 1906. The need for yet another cdition is gratifying and irrtlic;atesthat tny a1l1lrorr,ch, via languagesrathcr than computations, is still viable. The charrgcs ftrr the second edition wercl t)volutionary rather than rcvolrrtionary and addressedthe inevitable itrirct:rrra,c:ies and obscurities of thtl Iirst edition. It seertrs,however, that the second r:dition had reached a point of strrbility that requires f'ew changes,so thc tlrlk of the third editiorr is idcntical to the previous one. The major new featurtl of the third edition is the irrc:hrsionof a set of solved exercises. Initially, I felt that giving solutions to exerciseswas undesirable hecause it lirrritcd the number of problerrts thir.t r:a,nbe a,ssignedfor hourework. However, over tlre years I have received so rrrany requests for assistance from students evt:rywhere that I concluded that it is time to relent. In this edition I havc irrcluded solutions to a srnall rrumber of exercises. I have also added solrro rrew exercises to keep frorn rtxhrcing the unsolved problems too much. Irr strlec:ting exercises for solutiorr, I have favored those that have signiflcant instructioner,l ver,lues.For this reasorr, I givc not onlv the answers, brrt show the reasonirrg that is the ba,sisfor the firml result. Merny exercises have thtl ser,metheme; often I choose a rupresentative case to solve, hoping that a studerrt who can follow the reasorrirrgwill be able to transfer it to a set of similar instances. I bclicrve that soluiions to a carcfirlly selected set ttf exercises can help studerrts irrr:rea"retheir problern-solvirrg skills and still lcave instructors a good set of unuolved exercises. In the text, {lxercises for whir:h rr,solution or a hint is g-ivcrrrr,rqridentified with {ffi. Also in response to suggcstitlns, I have identified sonre of ther harder exercist:s. This is not always easv, sirrt:e the exercises span a spectrrrm of diffic;ulty and because a problen that seems easy to one student rnay givr: considerable trouble to another. But thcre are some exercises that havcl posed a challcnge fbr a majority of my studcnts. These are rnarked witlr a single star (*). There are also a few exercisosthat are different from most in that they have rro r:lear-cut answer. They rnay crrll f'or upeculation,


vt

PRnrncp

suggest additional reading, or require some computer programming. While they,are not suitable for routine homework assignment, they can serve &s entry points for furtlter study. Such exercisesare marked with a double star (** ). Over the last ten years I have received helpful suggestions from numerous reviewers, instructors, and students. While there are too many individuals to mention by name, I am grateful to all of them. Their feedback has been in'aluable in my attempts to improve the text. Peter Linz


Chapter

1

fntroduction

to the Theory

of Computation

1.1 Matlrenratical Prelirrrirrrlricuar,ndNotation Sets 3 Functions and Relations 5 Craphs and l}'ees 7 Proof Techniques I 1.2 Three Basic Concepts 15 Lirrrgrrir,ges 15 Grarnrnilrs 19 Automala 25 +1.3 Some Applications 29

3

Chapter 2 Finite Autornata 35 2 , 1 I)eterrrrinistit: Finite Accepters 36 I)ctc:rrnirristic Accepters and'IIrrnsitiorr Grir,phs Languir,gcsand Dfa,s 38 R.t:gulil,rL:lngrrages 42 2 . 2 Nondeterrriinistit:FiniteAccepters 47 Definilion of a Nonrleterministic Accepler 48 Whv Notxlctt:rrninism'1 52

vll

36


vlll

CoNrnNrs

2.3 Equivalence of Deterministic and Notrdeterministic Finite Accepters 55 +2.4 Reduction of the Number of States in Finite Automata 62 Chapter

3

Regular

Languages

and Regular

Grammars

fl

3.1 Regular Expressions 7I Forma,l Delinition of a Regular Expression 72 Languages Associated with Regular Expressions 73 3.2 Connection Between Regular Expressions and Regular Languages 78 Regular Expressions Denote Regular Languages 78 Regula,r Expressions for Regular Languages 81 Regular Expressions for Describing Simple Patterns 85 3.3 Regular Gra.trrnars 89 Right- anrl Left-Linear Grammars 89 Right-Linear Grammars Generate Regular Languages 91 Right-Linear Grammars for Regular Languages 93 Equivalence Bctween Regular Languages and Regular Gra,mma,rs 95 Chapter

Chapter

99 4 Properties of Regular Languages 4.1 Closure Propertitrs of Regular Languages 100 Closure under Simple Set Operations 100 Closure under Otlter Operations 103 4.2 Elementary Qrrestionsabout Regular Languages 111 4.3 Identifving Nonregular Languages 114 llsirrg the Pigeonhole Principle 114 A Pumping Lemma 115 5

Context-Free

Languages

L25

5.1 Corrtext-Free Grammars 126 Exarrrples of Context-Flee Languages 127 Leftntost and Rightmost Dt'rivations 129 Derivation Tl'ees 130 R.elation Between Sentential Fttrms and Derivation 'llees 13? 5.2 Parsing and Ambiguity 136 Parsing and Mcnbership 136 Anlbiguity in Grarnrnars and Latrguages 141 5.3 Context-Ftcc Gramrnars and Programmirtg Ltr,rrgrrages 146


CoNrEr-rts

Chapter

6

Simplification

of Context-Flee

Grammars

ix

149

6.1 Methods for Tfansforrrring Grammars 150 A Useful Substitution Rule 150 Removing UselessProductions 15? Removing.\-Productions 156 Removing Unit-Productiorrs 158 6.2 Two Important Normal Forrns 165 Chomsky Normal Form 165 Greibach Normtr,l Form 168 +6.3 A Me:mbership Algorithm for Context--F]'eeGrarnrnrr,rs 1,72 Chapter

7

Pushdown

Automata

175

7.7 Nondeterrnirfstic Pushdown Automata 176 Definition of a Pushdown Arrtomaton tTti A LangrrageAccepted by a Pushdowrr Automaton I79 7.2 Pushdown Automata and Context-Free Larrguagcs 184 Pushdown Autorrrata fbr Context-Flee Languages 184 Corrtcxt-Floe Grammars for Pushdown Autorrrata 189 7.3 DerterrrinisticPushdown Autornataand Deterrrfnistir: ContextFr{lc Lirnglrri;r,ges 195 *7.4 Gramma,rsfbr Deterministic Corrtext-F}ct:Langua,ges 200 Chapter

8

Properties

of Context-Flee

Languages

205

8.1 Two Pumping Lemmas 206 A Purnpirrg Lcrnrna fbr Context-Flee Languages 206 A Purnping Letrrnil firr Linear La,ngua,ges 210 8.2 Closure Propcrtien and Decision Algorithrns for ContextFree Languages 213 Closure of Context-Free LangJuages ?13 Some Decidable Properties of Contcxt-Fre,'e Languages 218 Chapter

9

Turing

Machines

221

9.1 The Standard T\rring Machine 222 Definition of a Thring Machine 222 T\rring Machines as Language Accepters 229 Tlrring Ma,chines as Tlansducers 232 9.2 Combining Tlrring Machines for Cornplicated Tasks 238 9.3 T\rring's Thesis 244


(,'onrnrlrs

Chapter

10

Other Models of Turing

Machines

249

10.1 Mirxlr Virriatiotrs on the T\rring Ma,t:hint:Therne 25t) Eqrrivalcrrt:tlclf Classesof Autonrata, 250 Ttrrirrg Machines with a, Sta,y-Option 251 Thring Machines with Semi-Infinitc Tape 253 The Off-Line Tttrirrg Mat:hine 255 'I\rring 10.2 Ma,chineswith Morc Cotttplex Storage 258 Mullitape Ttrring Ma,chiners 258 Mttltidimensional T[rring Mtr.chirrt:s 261 10.3 Norrtletertninistic T\rring Ma,chines 263 10.4 A lJrriversal I\rring Machine 266 10.5 Liricar Bouttded Autotnata 270

Chapter

Ll.

A Hierarchy

of Formal Languages and Autornata

11.1 Recursive and Reclrrsively Euurnerable Languages 276 Languages That Art: Not R,tx:ursivelyEnumera,ble 278 A Language That Is Not R,t:cursively Enumerable 279 A Language That Is Rer:rrrsivr:lyErlrrrterable But Not Recursive 28.l 11.2 Uurestricted Grarnmars 283 11,3 Context-Sensitivc (]rarnrna,rsarrd Lirnguages 289 Conterxt-Srlnsitivc Languages and Litrear Bounded Aulomata 29t) Relation Betweeu Recursive and Ctlrrtt:xt-Setuitive Languages 2gz 11.4 I'he Chomskv Hierarchv 29Ir

Chapter

12 Limits

of Algorithrnic

Cornputatiorr

299

12.1 Some Probletns That (ltr,rrnotBc: Solved By l\rring Machines 300 The T\ring Machine llalting Problem 301 H.etlucitrgOne Undecidable Problem to Another 304 12.2 Uritlt:c:itlrrble I'robletns for Recursivelv llnrtmertr,ltlrr Languages 308 12.3 Tlte I'osL CorrespondencePtoblem Sl2 12.4 [Jndccidable Problems for Context-Free Lir.nguages 318

278


Cot-{reNrs

xi

Chapter L3 Other Models of Computation 323 13.1 RecursiveFunctions 325 P-rimitiveRecursiveF\nctions 326 Ackermann'sF\rnction 330 13.2 Post Systern$ 334 13.3 R.ewritingSystems 337 Markov Algorithms 339 L-Systems 340

Chapter 14 14.1 14.2 14.3 I4.4

An Introduction to Computational Complexity Efficiencyof Computation 344 Ttrring Machinesand Complexity 346 LanguageFamiliesand Complexity Classes 350 The Complexity ClassesP and NP 353

Answers to Selected Exercises References 405 Index

4OT

357

343



INTRODUCTIO N TO T H ET H E O RO YF COMPUTATION

otrprrterr science is ir,pra,ctical discipline. I'hose who work irr it oftcn hirve ir,mrlrked pref'erencefbr useful and tangible problerns ovt:r theoreticrrl spt:c:ulirtion. Thiu is certa,inly true of computer science studcrrts who rrru interested rna,inly in working on difficult applicatious from the real world. Tlteoretical qucstions arcrinteresting to them only if they help in finding good solutions. This attitude is appropriirte, sinr:e without npplications there would be little interest in cornputers. But givcrr ihis practical oritlrrtir.tiorr, onr: rnight well a,sk "why study theory?" 'Ihe first arrswer is that tlrrxrry provides concepts and principles that help us understand tlrtl gerrcral rrirturt: of the discipline. The field of computer scienceincludes a wide rarrgr:of sper:irr,ltopics, f'rom machine design to progratntrtittg. Tlte use of cornputtlrs irr thel rea,l world involves a wealth of specific detail that must lre lerirrrrcxlftrr a uuccessfirla,pplication. This makes computer science a very diverse arxl lrroarl rlis<:ipline. But in spite of this diversity, there are soure colrtlrlotr urrclcrlyirrg prirrt:ipltrs. Tcl strrdy these basic principles, we construct abstract rnodels of corrrllrtcrs and comprrtation. These ruodels embody the important features tlnt are cornnron to both harrlwarc and softwtr,re,rr,ndthat a,reessential to many of the special and complex corrstructs we crrcourrtrlr while wclrking with computers, Even


Chopter I IurnorrucjrloN To rHE Tsr:enY ol' Col,tputarlott

whertrsuch moclelsa,retoo simplc to be applicable immediately to real-world situations, the insiglrts wt: gain frotn studying them provide the foundations on which sptx:ific; rlevelopment is ba*sed. This tr.pproach is of course not unique to rx.rrnlxrtcrscience. The construc:titlrrclf rnodels is one of the essentials of any sc:iurrtificdisciplitte, and the usefiilnessof a discipline is often clependenton the exi$ttrrrt:cclf simple, yet powerfirl, thtxlric:satrd laws, A second, tr,rxlllcrhaps not so obvious answer, is that the ideas we will discuss have srlmt: irnrnediate and itnporta,nt applit:atiorrs. The fields of digital design, prograrntning laugua,ges,tr,ndrrirnpilt:rs are the most obvious erxarnplcs,but there are rnanv othcrs. The cotrcepts we study hert: nrrr like a thread through mrrr:h of txrrrrputer sciettce, from opera,ting systerrrs to pa,ttern rtxxrgrritiorr. The third irlrswer is oue of which we hclpc to txlrtvittce the reader. The srrtricc:trnatter is intellectually stimrrltr,tirrg atrd furr. It provides ma,ny crha,llenging, prrzzle-likeproblems that can lead to ir()rrrcsleeplessnights. This is probkrrn-solvittgin its pure essence. In this hook, we will look at models that represcrrtfcatures at the core of all c:ornputers and their applica,tiorru. Trr rrrodel the hardware of a comprrtt:r, we introcluce the notion of iln automaton (plural, automata). An automaton is a, construr:t thir,t possessesall Lhe indispensable f'eatrrrt:stlf a digital computer. It :r.rxxlptsirrput, produces output, may have somtl tcrnporary utorilgrl, and can make decisions in tra.nsformirrg the input into tlte output. A formal language is arnir.bstractiorr of the general characteristics of prograrnming languages. A ftrrmal lirrrgrrage cotrsists of a set of symbols irrrd some rules of forma,tion by whit:h thcse sytnbols can be cotnbined into crrtities called sentences. A f'ormell lirnguage is the set of all strings permitted by the rules of fi)rrnirtiorr. Although sorne of the formal langrrirgcs we study here are simplt:r thirrr prograurmitrg langua,ges,they have rnarry of the same esserrtial features. We cau learn a great deal ir.bout programming lir.rrguirgesfrom formal languages. Fina,lly, wtr will forrrralize the concept of a rnechanical computation by givirrg a precise clefinition of the term algorithrn and study tlrt: kittds of problems that are (and tr,renot) suitable fbr solution try srrclt trtechatrical Ineans. In the cour$e of orrr stutly, we will show the clo$er(xlrrrrc(:tiotr between these abstractions and irrvc:stigate the conclusions we carr tlcrive from them. In tlx,'first chapter, we look at these ba,sicidea,sin a vcry broad way to set thtl stagc for later work. In Section 1.1, we revit:w thc rrrain ideas from ma,tlrttrnaticsthat will be required. While intuition will frcquently be our guide irr exploring idea,s, the conchrsionrr wu draw will be based on rigorous arguments. This will involve sclmelrnilthernatical machinery, although these requirementn alrelnot t:xterrsive, Tlte reader will need a rea^sonably good gra,spof the terminology and of the elementary results of set thtxrry, ftnetions, anrl rclatiorrs. T!'eesand graph structures will be rmul f'requently, a,lthoughlittle is needed beyond the definition of a,lir,beled,directed graph, Perhaps the rnost stringent requirement is thu rrtrility to follow proofs aud


J 1.1 MarrrnlrATrcAl PRr:r,lnrwnRrES AND Norauou

atr utrderstarrdingof what constitutes proper rnathcrnirtical reasoning. This includes farniliarity with the hasic proof techniques of dcrluction, induction, ancl proof by c:clrrtrir.diction.We will assurne that thc rcirrlrlr ha,sthis necessarybackground. Sectiott 1.1 is induded to review some of the rrririrr results that will be used arrrl to entahlish a notational colrurrorr grourrrl f'rrr subsequentdiscussion. In Section 1.2, we take afirst look at thc r:entral concepts of languages, gralrllnar$, trrrd a,utomata, These cortcepts oc{:rrr irr rnarry specific fbrms throughout the book. In Section 1,3, wc givc some simple a,pplica,tionsof tlrr:sc gerrera,lidea,sto illustrate that thesc c:tlnr:rrptshave widespread uses itt cornputcr ur:ience.The discr.rssionin these two scc:tionswill be intuitive rather tltirrr rigororrs. Later, we will make all of this rmrr:hrnoro precise;but for lhe ntotttettt, thtl goal is to get a,clear picture of tire corrceptswith which we are derrling.

Mothemoticol Preliminoriesond Nototion WffiWmHilW Sets A set is a collectiott rtf t:lclrno'rrts, without any structure olher tharr rnr:rnhership. To indicate that r is arr clcrnrrnt of the set 5, we write r € ,9. The sta,tementthat r is not in S is written r f 5. A set is specified by cnr:losingsome description of ils elernentsin curly bracxrs;fbr exa,mple,the set of irrtt:gers0, 1, 2 is shown as 5 : {0,1,2}. Ellipses are usetl wltcncvc:rtlNl rneir,ningis clear. Tltus, {a, b, ... ,z} slands for all the lower-caseletters of thc Engliuh a,lphabet,while {2,4,6, .,.} denotes the set of all positive everr irrtcgrlrs. When the neecl arises, we use rrrore explicit notation, in which we write S= {i:i >0, ziseven}

(f,l)

frrr the ltr,st example. We read this as tt,9is sc:t of irll ri, srx:h thrr,t rl is grea,ter tltatr zero, a,ndrj is even," implying of course that z is irrr irrteger. The usual set operations arc union (U), intersection (n), and difference (-), defined as 51U52 : {z : r e Sr or r €,92}, 5 1 1 5 2 : { z : r € S r a r r r rl E , 9 z } , 5r - Sz : {z : z € Sr arxl r fr 52}. Anothttr bir,sicopera,tion is complementation. The cotrplerntlnt tlf a set ,9, denotecl by F, consists of a,ll elernenls not, in S. To rnakc this


Chopter I

4

llqrnooucrroN To rrrn THnoRv cln Cor,tpu'rn'rtou

rnerarrirrgful,we need to know whir,t the universal elements is. If U is specified, thcrr

set U of a'll possitrlt:

F:{*:rr-(I,n#S}. The sct witlt no elements, called the ernpty set or the null set is denoted by #. Flom the definitiorr rtf a set, it is obvious that

Sus: S-fr:5, Snfr=9, fr =U, 5::5. The following useful identities, known a.sthe DeMorgants

laws,

Srl3z, ffi: F l n S z : , 5 rU S z ,

(t 2) (1.3)

are needed orr $eivc:raloccasions, A set ,9r is said to be a subset of 5 if every element of 5r is also atr element of S. Wc write this as 5rc5' If St C S, hut 5 rxrrrtilirrs ir.rrelernetrt not in 51 we say that Sr is a proper subset of .9; wc write tltis as $rCS. If 51 a,ncl5'z have IIo coilllrron elemeut, that is, Sr n 5'2 = fl, tltett the sets are said to berdisjoint. A set is said to be linite if it contains a finite nlrmbcr of elemenls; 'Ihe otherwise it is infinite, size of a finite sct is tht: rrurrtberof eletnents in it; this is denoted bV l5l. A given set norrnally has marry sutrsets. TIte set of all subsets of a, set 5 is callecl the powerset of S ir,nd is denoted by 2's. Observe that 2s is rr, set of sets.

Exottplq l'.f I

If $ is thc set {a,b,c}, then its powersetis z s : { f r , { o } ,t b } , { c } , { a , b } ,{ n , r : } {, b , c } , { o , b , " } } . Here lSl : 3 and lZtl :8. finite. then

This is arr instirrrce of a generalresult;if 5 is lrsl - ,rlsl I


\

AND Norauolt 1.1 MnrHntutnrtcAL PRt:t,ltvttw.q,n.IEs

In rnany of our exa,mples,the elements of il stlt irre ordered sequencesof elements frorn otJrer sets. Srrr:h$ets arc said to be the Cartesian product of other sets. For the Ca.rtcsiarrproduct of two sets, which itself is a set of orclerednairs. we writer S:Sr

x52:{(*,:,/) :râ‚ŹS'1,Ee Sz}.

2 , 3 , 1 ' r6,| . T 6 c ' S r x 5 ' z : { ( 2 , 2 ) , ( 2 , 3 ) , ( 2 , 5 )( ,2 , 6 ) ,( 4 , 2 ) ,( 4 , 3 ) ,( 4 , 1 ' r()4, , 6 ) } . Notc that tlte order in which the elements of a, llnir are written matters, Thc pair (4,2) is in 51 x 5'2,but (2,4) is not. The nolation is extendeclin a,nobvirlrs firshiorrto tlte Cartesian product of rnr)rt)than two sets; generally S r x 5 ' r x ' . ' x 5 r : { ( r 1 , T 2 , . . . , n , , )r: , ; â‚Ź S r } .

I

Functions ond Relotions A function is a rukt that assignsto elements of one set a,unirptl cletrtetrtof another set. If / dcrxrtt's a futrcLion,then the flrst set is t:ir,lltxlthe domain of /, and the serxrndsct is its range. We write / : , 5 1- . $ 2 to itrdicate thal the doma,in of / is a strtrsc:tof ,51atrd that the ra,ngeof / is a subset of 52. If tht: tlornirirr of / is all of 5r, we say thrlt / is a total function on 5r; otherwist: ,f is said Lo be a partial function. In ma,ny applir:rrtiorrs,the donaiu and rauge of the firrrt:tiottsinvolved are in the set of positive integers. Furthermorel we il,rcoften interested only in the heha,virlrof tltese functions as their arguments btlclotttevery large. Itr arr urrrlerstanding of the growth rtr,tes is oftetr sullicient and a such c:asers corrrrrrorrorder of magnitude nota,tion carr be used, Let / (n,) and .q(n) be functions whose doma,in is a, subst:t of the positive itrtegers. If thcre exists a, positive constant c such that for all rz

f(n)tcs(n), we sav that .f ha,sordcr

at most g, We write this ir,s

f ( n): o ( s( n) )


'r'o'r'Hu'l'Hnonyor,'ConrurArrorv ChopterI IurR,onuc;'l'roN

If

, l/ (")l l c ls(?z)l then / has order at least g, I'rrr whit:h we tuJe

/(") :o(g(")), Finar,lly,if there exist constants c1 and c2 such tlnt

cr lg(?l)l< l/ (")l ! czle(rl)l, / and g have the same ordcr of magnitude,

expresseda,s

I(n)*o(g("))' In this order of magtritude trotittitlrr, wtr ignore multiplicative constants and lower order tertns that becotnencgligibkl as n increases.

ExompleI.3

Let f (n) : 2nz + iJtt,, I (rt) : "'t,

h , ( n ) : 1 o r z+2 1 o o . Thcn / (t) : o (s (rl)). s (n) : o (ft.(ru)), / (n) : o (h (r,)). In order of rnagnitude notatiorr, tlrrl syrnllol : should not be interpreted irs txlra,lity a,nclorder of magnitude expressiorrsr:annet be treated like ordirrirry cxl)r{}ir$ions.Ma,nipulations such as O (rz)+ i) (n) = 20 (n) a,renot sensibleand catr lead to irtr:clrr(lctrxrnclusions.Still, if used properly, the order of magnitude argurnents tlrrrr tlr: effective, a*swe will see in later r:hirpturson the a,nalysisof algorillurs. I Some functiorls can be rtlprt:srlrrttxl by a set of pairs { ( " r , y r ) , ( r z , u z ), . , . } , wh{:rc il; is a,n element in the clornain of t}re furrc:tion, and gti is the corresportdirrgvilhrelin its ra,nge.For such a set to delirrc a firnt:tion, ea,ch11 can occur at rno$t on(:ea,sthe first element of a pair. If ttris is not satisfied, the


l.l

M.q,uml,rATrcAL PnnLrNrrNanrESAND No'r,t'l'tor'r

set is called a relation. Relatious are Inore general thtlrr firrrt:tions: in a function each element of the doma,in ha,sexir.ctly orrclitssociated eletnent itt the ra,nge; in a relir,tion tht:re miry trcl scvtlral such elernenls in the range. Orre kirrd of relatiott is that of equivalence, a generalization of thc concept of equality (identity), To indica,tethat a,pair (r:,37)is arr crpivirlcrrce relation, we write :I:='!J. A relatiori rlcrrotexl lry : i. consiclered atr equivalence if it satisfies three mlcs: the reflexivity rule r: = r: for ir,ll z,

the syrnrnetryrule if

'':

r , ll rhr a v un

o y , :-

dr t'

and the transitivity nrlc If fr:

I

f xsnrslF f ;4

' '

A a n c ly :

a, then n # E.

Consider the rela,tion on thc sct of rrotrttegativeiulegers defined hy 'l':

l'

if and only if rmocl 3 - 37mod3. 'Ihen

2 : 5, 12 = 0, and 0 = il6. Clearly this is atr equivalencerelation, irs it satisfies reflexivitv, syrntttetry, and l,ransitivity,

I

Grophsond Trees A graph is a construct consistirrgof two fitilte sets, the set V : {tt1,'tt2,...,'Dn} of vertices and the set E: {e1,e2,...,err} of edges. Ea,chedgttis a pair of vertices fiom V, frlr irrstance e.i:

\U.i,Ltk)

is an edge from ui to tr4. Wc srry that the edge e,; is a,n orrtgtlirrg edge for ?ri and an incoming edge forr.'r.. Such a construct is actually ir. directed graph (digrrr.ph),sirrce we associate a direction (fiorn ui to u6) with each edge. Graphs miry bc labeled, a label being a ntrme or other itrformation with parts of the graph. Both vt'rtices atrd edgesmay be lahclctl. a*ssor:iated


Chopter I

Itrrn,onuc'l'roN'l'o tsr: Tttr:rtHv ou Cotr,tt,u'r'n'r'tow

Figurc 1.1

f \ \'otl ---t-------------

Graphs are conveniently visualized by diagrarns in which the vertirx:s are represented as circles and the edges as lines with arrows corrrrt'cting tho verti<x,'s.The graph with vertices {u1, u2,u3} and edges {(u1, u3) , (u3,u1) , ( t r , u r ) , ( u 3 , u 3 ) | i s d e p i c t e di n F i g u r e 1 . 1 . A s e q u e n c eo f e d g e s( a t , ' u i ) , ( u i , u * , ) , . . . , ( ' , , , , , , r r " i,s) s a i d t o h e t r ,w a l k fiom rri to urr. The length of a walk is the total nurrrber of rxlgcs travr:rscrl ,in going from the initial vertex to the final orre. A wrrlk in which no eclge is repeated is said to be a pathl rr path is simple if no vertex is repeated. A walk fron ui to itself with rro rcpcir,trxltxlges is ca,llerla,cycle with base u4. If no vertices other thatt tlte base are rrlllc:itttxliri ir r:yr:le,then it is sa,id !o be simple. In Figure 1.1, (z1,ur), (rr,u2) is a simple perthfiom ?rrto ??. The sequenceof edges(ut,rr), (rr,rr), (r*,rt) is ir cyc:le,llrt rxit ir,sirnple one. If the edges of a, graph are labeled, we can talk about the label of-a walk. This label is thc scqucrrr:cof r:dgo ler,belsencorrnteredwhen the path is traversed. Fina,lly, a,n eclgefrom a vertex to itself is calk:d a loop. In Figure 1.1 there is a loop on vertex u3. On several occasiotts, we will refer to atr algoritlun for lindirrg all sirnpkr paths between two given vertices (or all siurplc c:yrlcs bn^sedon rr,vertex). If we do not concern ourselves with efficiency, we carr llsc tlrrl following obvious method. Starting frotn tlte giverr vcrtcxr say ?ri, Iist all orrtgoing txlgt:s (u;,116),(ui,ur) ,.... At this point, we have all paths of length orrt: startittg at u4. For a,llverrticesuk1,t)t,1 ... so rea,ched,we list all outgoing edges a,slong as the.y do not lead to arry vcrtclx alrtlirdy rrsed in the pa,th we are rxlnstnrcting. After we do this, we will have all sinrple paths of lerrgth two origirrrrtirrgat a,. We r:ontinue this until all possibilities are accounted for. Since there ate orrly ir finite number of vertices, we will eventually lisi all sirnple paths beginning at rr,;. Flom these we select those ending at the desired vertex. Tlees are a particular type of graph. A tree is a directed graph that has no cycles, and that htust)ne rlistinct vertex, called the root, such that there is exactly one path frorrr the root to every other vertex. This definition implies that the root ha^srro irrcoming edges and that there &re some vertices without outgoing edges. These are called the leaves of the tree. If there is an edge from ua to ui, then ua is said to be the parent ()f rrj, il,nd ui the child of u1. The level associated with each vertex is the nunber of edgesin the path from the root to the vertex. The height of the tree is the Iargest level number of any vertex. These terms are illustrated in Figure 1.2.


1.1 MarHu,rATrcALPn,nulumnn,rEs AND Norauoru

Figure 1.2

Root

Height = 3

I

I

- L"'"13--t

At times, we want to a*ssocirr,te an ordering with the nodes at each level; in srrch ciL$e$we talk aborrt ordered trees. More details on graphs and trees can be found irr rnost books on discrcte mathematics.

ProofTechniques An important requirement for reading this text is the ability to follow proofs. In mathema,ticalarguments, we employ the accepted rules of deductive reasorring, ilnd rnilny proofs artl simllly a .sequenceof such steps. Two special proofteclufques are used so frequently that it is approprintc to rcvielwthem briefly. These are proof by induction and proof by contradiction. Induction is a technique by which the truth of a number of statements can be inf'erred from the trr.rth of a few specific instances. Suppose we have a sequenceof statements Pr , Pz, . .. we want to prove to be true. Furthermore, $rrpposealso that the following holds: 1. For some fu ) [, we know that Pt, Pz, ... , Pk are true. 2. The problern is such that for any z ) A, tlrt: trutlm of P1,P2,...,P,,. imply the truth of P,,-1, We can then use induction to show that everv statement in this sequenceis tnre. Irr a pro<lf by irrclucticln, we rrrguo as follows: Ftom Condition 1 we know that the first k statements are true. Then Condition ? tells us that P611 alsrl rmrst btr tnre. Brrt now thir,t we know that the first h * 1 statements are trlrc, we r:arr allply Contlitiorr 2 agairr to tlaim that P61z must be true, arrd so on. \Me need not explicitly continue this argument because the patterrr is clcrrr. Thc cltairr of rcit*sorrirrgt:itrr btl cxtended to any strrtcrnerrt. Therefclrc, every statement is true.


FIN ITE AUTOMATA

tho",t;l

ur introduction in the first cha,ptcrto the basic cotrceptsof comprrtation, particularlv tIrc tlist:ussiotrof autornata' was hrief antl irrftlrrnirl. At this point, we harvt,rorrly ir general understanditrg of whir.t irrl automaton is and how it c:itrrbe representedby a gra,ph. Ttl llrogress, we must be more prct:istt, provide formal defiuitions, ir,rrrlstirrt to clevelop -- rigorous results. Wt: llcgitr with linite accepters, whit:h ir,r't:a sirnple' specia,l case of thc gcrreral scherneiutroduced in the ltrst drapter. I'his type of autclmittotr is characterized hy having no tt:rnporary storage. Since a,n input file cannot be rewritten, a firritc itutornittott is severely limitcd irr its ca,pacityte "rcmcmbcr" things during the comprrtir,tirlrr.A linite atnoutrt of inf'ormir,tiorrcarr be relained in the control rrnit tly placittg Lhe uttit itrto a, sptx:ifi<:state. But sitrcethe number of srxrtrstir.tt:sis firriLe,a finite a'utomat(lrr (:arronly deal with situations in which tlrt: infortnalion to tre stored at arry tirrre is strictly bounded. The a.utornittorrirr -Exatnple1. 1ti is tt,ninstarrt:t: clf ir firrite acceuter.

35


Chopter2

AUToMATA

DeierministicFiniteAccepters

ffi

The first type of automa,ton we study in detail are firrite accepters that are deternfnistic in their opera,tion. We start with a precise frrrrnir"ldefinition of deterministic accepters.

ond Tronsilion Grophs Accepters Deterministic sniiit$lil]tnir tiftffif[j! f,ffifrtll A deterministic finite accepter or dfa is defined by the quintulller M : (Q,E,d,qs,.F),

where Q i s a firtite st:t of internal states, E i s a flnitc sct of symbols called the input alphabet, d ; Q x X - Q iu el tota,l function called the transition function, t J O â‚Ź Q is the initial

state,

T -C Q is a set of final states. A deterministic finitt: ir.rx:epteroperates in the following tttanrrt:r. At to be irr thc initial state q0, with its input thc initinl time, it is a,ssurnecl mechanisrrrort the lefhmost symbol of the input strirrg. Drrring eelchmove of the automaton, tlrtl irrlxrt meclha,nhmadvancesone position ttt tlrc right, so each rnove colrsurrrrlrione input symbol. When the etttl of tlxr string is reached, the string is ircx:clltedif the automaton is in one of its Iirtr.l stir,tes. Otherwise the string is rejer:tcxl.Tln input rnechauismcan Inove otrly frorn 'I'he tratrsitiorrs Icft to right and reads exactly ontl symbol on each step. govcrned futrctiorr internal are the transition frorn one bv state to another if d. Ibr exirtrllle, rI (qo,a) : gr , then if tlrtl rlfa iu in sta,te4u ancl the current, itrlrut syrnhol is a, the dfa will go itrto stirtc' q1. In dist:ussirrg arrtomata, it is essential to havc a. clt:iu and intr"ritive picture to work with. Trr visutr,lizeand represetrt fittitr: irtrt{lrna,tir,we use transition graphs, in whir:h the vertices represerrt stattls arrtl the edges represent transitiorrs. Thc Inhels on tlre vertices are Lherrirlrr()sof tlx: sttr,tes, while the labels orr the eclgcs art: the current values of the input sytnbol. Rrr tlxtr,mple,if qp and qt are interlral strr.tesclf some clfa M, then the graph associatcd with M will have one vertex lal-rclcdq0 irnd another labeled q1. An edge (qn,qt) la,bclcda represetrtsthe [ratuitiorr d (gs,n,) = q't. The initial


2.1 DnrnnluNrsrrc Frurrn Accl;plr:RS

87

Figure 2.1

I

arrow ttot origiuating a,t sta,tc will be identilied by a,trincoming rrnlir,bt:lt:r1 arry vertex, Final sta,tesa,redrilwn with ir,double circle. More forma,lly, if M : (Q, X, d-,,lo,.f,')is a cletertnitristic:finitt: ttt:t:cpter, then its associatedtrtr,nsitiorrgraph Gna has exactly lQl vcrtitr:s, t:itch otte Iabeledwith a diff'erentq,i â‚ŹQ.-For every transitiott rrtle d(rJ,,a): qi, the 'Ihe with q6 is called graph has a,nedge (qr, qi) Iabeled a. vertex assocria,tecl the initial vertex, while tltose labeleclwith qy â‚Ź F arc tht: final vertices. It is a trivia,l ma,ttcr to converl from the (8, X, d, q0,1I) tlc:firritionof a dfa to its tra,nsitiorrgraplt represeutatiou and vice veruir..

represents tlrc dfh

IUI= ( { g u ,q t , r 1 z ,} { 0 , 1 } , d , 4 0 {, q r } ) , whcrc d is given by

:;i:i;; :i;li:i ii=r,

This dfa accepts the strirtg [Jl. Stnrting irr state gs, the syrnbol 0 is rcird fi.rst. Looking at the edgesof thrl graph, we see that the arrtrlmatott remains in state qs. Next, the l is relirrlirntl the autotnaton goerJintcl state 4r' We are now at the end of thrl strirtg and, at the same time, irr a fltral state q1. The dfa does not a,cccptthe string 00' Therefore, the string 01 is a,crt:cpted. c:orrsecutive 0's, it will be in uttr.tctlo. By similar since after reading two that the autouraton will accept thtr strirrgs 101, 0111' and reasoning,we $ee

11001, hrrtnot 100or 1100, tlf

T

- #

b . 6 r"L-+Q It is convenient to irrtrorlut:e the extendecltra,nsition firrx:tiotr d* : Q x X* - Q. The ser:orrdargutrtettLof d* is a string, rathrlr tltatr a single symbol, and its va,luegivtls tlre stale the autotna,tonwill llc irr after reading that string. For exermplc,if d ( q o .n ) : q '


38

Chopler 2 FIt'rIrp AuToMATA

and 6 ( q t , b ): q z , then d* (qn, ab): qr. Formally, we can define d* recrrrsively by

d. (9,A) : q,

(2l) 1' r)\

&t-+!-q-F Ir u â‚Ź Ej, s -EA To seewhy this is appropriate, let us apply thesedefinitions to the simple caseabove. First, we use (2.2) to get

U-_!no' ab]:_d (d*(qs,a), b)

(2.3)

But d" (q0,0) : d (d- (qo,l) , q) d (go'a)

:

!'!''

Substitutingthis into (2.3),we get

d*(qo, ab): d(o:,!: oz as expected.

longuogesond Dfq's Having made a precise definition of an accepter, we a,renow ready to define formally what we mean by an associated language. The associatiorr is ob. vious: the language is the set of aJl the strings accepted by the automaton.

lsfi.ffl$ifii ,

r

The language accepted hy a dfa M : (8,E, d, go,F) is the set of all strings on X accepted by M. In formal notation,

L (M):

{ru e E* : d* (q6,ur) e F} .


2.1 Dnrurl,llNIS'tICiFtNtre Accr:p'tuns

39

Note that we rt:cpire that 15,and trlrrstltluenlly d*, bc trltal functions. At ea,chstep, a urrique tnove is dtlfirrcd, so t'hat we irr'(l justifiecl in ca,lling such a,narrtornatott deterministit:. A dfa will pror:tlsstlver.ystritrg in X* ir.rrd either ir.cc:cptit or not a,ccolltit. Nonaccepta,nt:ttrnealls that thtr tlfa stops in a nrlrtlirral state. so ther,t

L(M):

{ r ue X * : t 5 *( q o , w f) F I .

Consider the dfir,irr Figure 2.2 In drawirtg F-igure 2.2 wt: allowecl the use of two labels on a sitrgle edge. Sut:h rnultiply labelerdcdges are shorthand for two or mor(r distittct trirrrsitiols: the trelnsition is taken whenr:vcr the input syrrrbol nratches any of the edge labels. The automaton irr Figure 2.2 reurtrins in its initia'l strr.teq11until the first b is etrcountered. If this is also the la"rt syrnbol of the input, then the string is accepted. Tf nqt, the clfa goes into trtir,teq2, frotn which it can never 'I'ire sla,te q2 is ir trap state. Wtr see clea,rlyfiorn tlte graph tha,t e$(rirpe. thg autorlatotr a,cr:ttlttsall slrings clclnsistirrgof au arbitrtr.ry number of c,'tt, followed by a single b. All other irrput strings are rejtN:tt:d. In set nota,tiorr, ttre langua,gea,cr:cptedby the a'rrttlrlatotr is

f, * lu,''b:rt,> 0l .

I These exarnples sltow htlw t:ottvenient transitiott graphs artl for workiug with finite irutotnata. Whilc it is possible to hase all arguments strictly on thc properties ofthc transition functiorr attcl its extensiorrtlrrough (2.1) arrcl (2.2), the results are hard to fbllow. In our discrrssiorr,rMeuse gra,phs,which are more intrritive, as far i.r.rptlssible. To do so, wc tttust of cotrrse irave sonre assllrirnce that we a,renttt rtfsled bV the rtlpresentation antl thal arguments ba,sedon graphs are emvalid as those that use the fbrrrtrl properties 6f d' Thel following preliminary result gives us this assura,nt:tl.

Figtrre 2.2


40

Chopter 2 Frr,rr.ru: Aurovarn

I,et M: (Q,X,.l,qo,F) be a deterrninisticfinite accepter,irrrtl k:t Gna be its a,ssociatedtrarrsition graph. 'l'herr frrr every q?.,ei € Q, arrtl .ri.,€ X*, 6* (rqi,w) : qi if antl only if there is in G,1air walk with label zu frorn q, to qi. Proof: This clairn is fa,irly obvious frorrr irn exa,mination of srrr:h simple carrcsir$ Ilxample 2.1. It r:a,nbe proved rigorously using an induction on the Icrrgth of ur. Assurne thnt the claim is truc frrr a,ll strings u with lrl < n,. Corrsirlr:r then any ,urof lr:rrgth n * 1 and write it ns ,u :,DQ.

Supposenow that 6* (qi,u): q6. Since lul: n,, there rnust be a walk in Gy labeled u from qi t,() qk. But if d* (r71,ur): qi, then M rmrst have a tranuition d(qr,a) : {i, so that by construction Gy has arr cdge (Sn,qi) with label a. Thus there is a walk in Gnr labeled ua: u between {a and qi. Since the result is ohviously true for n: l, we can clainr by induction that, f'or every 'ur€ I+, 6- (qi,w) : qt

(, t\

implies that there is a walk in Gy from qi to qj laheled u. The argurnerrt r:irn be turned around in a, straiglrtforward way to show that the existence of such a path irrrplit:s (2.4), thus completing the 1rr{lof. I lJt-_lr.

Agir,in, the result of tlx: theorem is so itfrtititfr' obvious thrrt a formal proof seems unrlecessary. We went through the details for two rt:a^sons.The first is that it is a simple, yet typiur,l example of an inductive proof in connection with automata. ilire secorrdis that the result will be rrsed over and over, $o sta,ting and proving it as a theorerrr lets us argue quitc confidently usittg graphs. This rnakes orrr exa,mplesand proofs more transpa"rent than they worrld be if we used thc properties of d'*. Whilt: graphs are convcrrir;nt fbr visualizirrg irutomata, other representations art) also useful. Fbr exa.mple, we can represent the functiorr d as a table. The table in Figure 2.3 is equivalent to Figure 2.2. Here the row Iabel is thc r:urrent state, whilcl the column label representsthe currerrt input symbol. The errrtrv in the table tlcfines the next state. It is apparent f'rom this exarnplc that a dfa can easily be implemented a^s ir {xrmputer prograrn; fclr example, as a simple table-lookup or a,sa sequence of "if" statements. The best implernentation or representation depends on tlte specific applicatiorr. Tra,nsition graphs are very corrvcrrient for the kinds of argurnents we warrt to make here, so wr) rr$e them in rnost of our discussions. In construc:ting automata ftrr la,nguagesdefinerrl informally, we errrllloy reasoning sirnilirr to that for prograrnming in higher-level languages. But the


2.I

Figure 2.$

DETERMINISTIC FINTTU ACCEPTERS

4l

b

b'

tl

!o

?o

?t

q,

Iz

{t

9z

!t

?z

;",J

programming of a dfa is tediou$ and sometimes conceptually complicated bv the fact that such an automaton has few powerful fbatures.

Find a deterministic finite accepter that recognizes the set of all strings on X : {a, b} starting with the prefix ab. The only issue here is the first two symbols in the stringl after they have beerr read, no further decisions need to be made. We can therefore solve the problem with an automaton that has four states; an initial"state, Itwo statesrfor recognizing ab ending in a final trap state, and one nonfinal trap state. If the first symbol is an a, and the second is a b, the automaton goes to the final trap state, where it will stay since the rest of the input does not matter. On the other hand, if the first symbol i$ not an a or the $econd one is not f, b, the automaton enters the nonfirlal trap state' The simple solution is shown in Figure 2.4. I

Figure 2.4


q

Chopier 2 }'rllrn Aurolr.tra

Figure 2.5

liol*r# riilIr-,rrr

.

,

f

i

(

1 ' ( o t lc ) o

I

fixotrtpl*

[.4

Find a dfa thtit tr.cceptsall the strings on {0, 1}, except those corrtaining the substring 001. In decidirrg whether the substritrg 001 ha,soccurred, we need to krrow not only the currcnt input symbol, but wt: also need to remember whethrlr or not it has been preceded by one or two 0s. We can keep track of this hy pr.rtting the autornaton into specific slates arrrl la,heling them accordilgly. Like variable narrr()sin ir prograrnming languager,state names are arbitrary arrtl r:an be chosen for ttttternorrit: rea^rons. For example, tlx: sta,te in which two 0s were the immediately prrrcding symbols can be labcled simplv 00. If tht: string starts with 001, tht:n it must be rejected. This implies that therc must be a path labeled 001 lrom the initial state to a rronfina,l state. Fcrr cxrnvenience,this nonfi.rral sttr,te is labeled 001. This state mtrst be a trap stelte, hecause Iater synrbcllsdo not matter. All other stateri are acceptirrg state$. This gives us the ha"sicstructure of thc solrrtion, but we still rnust add provisions for the srrbstring 001 occurring in the middle of the input. We must define Q and d so thnt whatever we need to rnirkn the correct decision is rernerrbered by the autornaton. In this ca-se,whett a synrLrol is read, we need to know sclrnelpart of string to the left, for example, whether or not the two previous syrnbols were 00. If we labcl the states with the relevarrt symbols, it is vr:ry ea,syto see what the trrrnsitions must be. For exantple, d ( 0 0 , 0 )- 0 0 , becausc this situation arises only if there are three consecutive 0s. We are only interestcd in the la,st two, a fact we remember by keeping the dfa in the state 00. A complete solution is shown in Figure 2.5. We see frorrr this exarnplc how useful mnemorricrlabels on the states are for kceping track of things. Tbar:ea fbw strings, such as 100100 and 1010100,to see that the solution is irrdt+edcorrect. I

Regulor[onguoges Every finite arrtomaton accepts sorne larrguage. If we consider all possihle with thertr. We will call finite autornatal, we get a set of larrguagcs ir.ssocia,ted such a set of Ianguages a family. The family of languages that is accepted lty detcrministic finite acceptersis quite limited. Thc strrrcture and properties


2.1 Dulnnl,trNrsTrc F'tt'tt'l'FlAccpprpR,s

43

in this family will bet:ome clearer irs our study proceeds; of the la,ngua,ges filr ttxl trorn(lrrt wtl will sirnply attach a rrarrreto this farnilv.

ffi$$lliRtl$ri tffiR,fi uil,!.r.qftfr

f,*

1", " L.o,,

F

A language .L is called regular if and only if there exists sornedetertnitristic flnite accepter M such that

L: L(M).

Exolmple f .5

Show that the lirngrrtr.gt: 1, : {utL,tt: 'u e {o, b}* } is regular. To show that this or any other la,ngualgcis rcgular, all we have to clo is find a, dfa fbr it. Thc txrrrstructiorrof a dfa for this language is similn,r to Exrrrnpkr 2.3, but a little more complicated. What this dfn, must do is check whethcr a striug begins and ends with au a; wltat is between is immateritr,l. Tlrl sohrtiorr is cotnplicated by the fact that there is no explir:it way of tr:stirrg tlrtl errd of the string, This clifficulty is ever(xrrnc try sirrtply prrtting thrl rlfir irrto a final state whenever the second a is enrxlrrrtcrtxl. lf this is rxrt tirc end of the string, and another b is ftlrrrd, it will take the dfa out of the final state, Sca,nningcontinrresin this wiry, caclt a taking the autorna,tonba,ckto its finir.lstiltc. Thc: cornplete solution is shown in Figure 2.6. Aga,in, trace a few exa,tnplesto see why this works. Aftcr orre or two tests, it will be obvious tha,t the dfh, a,cceptsrr,strirrg if atttl otrly if it begins and ends with an a. Since we have txlrrstnrc:tc:drr dfa for the lauguage, we is regular. can claim that, by definition, tlrc lir,rrgrrirgc I

Let tr be the language in flxa,mple 2.5. Slxrw that -1,2is regular. Aga'in we show tha,t the langrra,gcis rr:gular by corrstructing a,dfa for it. We carr writc a,nexplicit exprussirlnf<tr L?, nanely, 7,2 : {rz.nytil,Lr.tza i ,LL)1,,11t2 e {a, b}. } . Therrcftlrc,wcl rrt:cxlrr dfa that recognizestwo conseclttivc strings of essen'Ihe diagra,m tially tlrc sartreforttt (but noi necessa,rilyidentica.l irr value).


44

Chopter 2 FtuIrn Aurolvlare

Figure 2.6

\------=-.-

tl--/

in Figure 2.6 can be used as a starting point, but the vertex gs has to be modified. This state cafl no longer be final since, at this point, we must start to look for a secondsubstring of the form awa. To recognizethe second substring, we replicate the states of the first part (with new names), with q3 as the beginningof the secondpart. Sincethe completestring can be broken into its constituent parts whereveraa occurs,we let the first occurrenceof two consecutivea's be the trigger that gets the automaton into its secondpart. We can do this by making d(gs,a) : qa. The complete solution is in Figure 2.7. This dfa accepts,L2,which is thereforeregular.a

I t

t Jt bJL

r i f+_!4 " * ' A ' * t ' r A (fitl,r aa YJ aa 9F{ r?,F j * -_.-v |

v

tnlr

,w , q^

l

,-T.u.rp"o u r t f t - i l

={ tu f)J^ 41-11 61aujLft

Figure 2.7

t?rfl'b,'

t L,ilat" t

t

r\t

r

,,1,86 4."

ri,,-


?.1 DETERMTNTSTTC Frr-rrrn Accnprnns

46

The last examplesuggeststhe conjecturethat if a language"Lis regular, so are .L2.,L3. .... We will seelater that this is indeedcorrect,

1. Which of the strings 0001,01001,0000110 are accepted by the dfa in Figure

2.r? E = {a,b}, constructdfa's that acceptthe setsconsistingof

f)t*

(a) all strings with exactly one a, (b) all strings with at least one a, (c) all strings with no more than three a's,

ffi

(d) all strings with at least one a and exactly two b's. (e) all the strings with exactly two a's and more than two b's.

3 . Show that if we change Figure 2.6, making qs a nonfinal state and making qo1qt1 qz final states, the resulting dfa accepts Z. 4 . Generalize the observation in the previous exercise, Specifically show that if -nI

M : (Q,8, d, go, F) and

"(fr) A*

= (Q,8, d, qo,Q - F) are two dfa's then T(6

:

dfa's for the languages

G"ive

(a) I=

{abswba:tue{a,b}-}

(b) fI : {wpbwz:

ffi

tor â‚Ź {a, bI* ,wz â‚Ź {a,b}-}

Give a set notation description of the language accepted by the automaton depicted in the following diagram, Can you think of a simple verbal characterization 6f the language?

b

g

I (T

( z) fi"a

dfa's for the following languages 61 5 = {a, b}.

(a) I=

{ta: ltalmodS=0} ffi

(b) Z, = {tr : ltrl rnodSI 0} ( c ) I = { w : n o ( t r r ) m o d 3> 1 } (d)

{w : n* (tu)moct3 > n6 (w)mod 3}

W


46

Chopter 2

FtmIrn Aurouare

-nu(tu))mod:} >0} {ur I (n*(ur) (f) , : {u : ln* (w) - nr, ('u)l rnotl:l < 2}

(e) r,:

q$

Nole

tl^t

* 7 rna:/ :t"- !-

e run in a string is a silbstrirrg of length at least two, as long as possible anrJ conrristing entirely of the satre symbol, For instance, the string a,bbhaa,lt contains a rul of b's of length three and a nrn of n's of lertgth two. Find dla's for the following languages on {4, h}. (u) t : {tr : ru cxrntairlrr IIo rtllls of length less than four} (b) .L : {'ur : every rtrrr rtf a's }ras lengt,ir either two or three} (c) I : {'ur : t}rere are at tnost two runs of a's of length three} (.1) f : {tl : there are exactlv two ruus of a's of length 3}

t.}tg

Consider the set of striflgs on {0, 1} delined hy thc rcquircments below. For each construct an acccptirrg dfa. (a) Every 00 is fbllowed irnrnediatel.y by a 1, For cxample, the strings 101, 0010, 0010011001 are in thc languagc, but 0001 and 0010t) are rrot. ffi (l-r) all strirrgs containitrg 00 but not 000, (c) The lef'tmost' symbol diffcrs frorn thc rightrnost one. (d) Every substring of four symbols has at most two 0's. Ftrr exarnple, 001110 arrd 011001 are in the latrg;uage,but 10010 is not since one of its substrings, 0010, contains three zeros. {il (e) AII strings of length five or rnore in which the fourth syrrrbol lrom the right erxl is tliffererrt frorrr the leftrnost sytrbol' (f) All strings in which the leftrnost two syrnbols a,rrd the righttnost two syrnbols are iderrtit:al.

Corrstruct a clla that accepts strings on {0, 1} if and only if thc value of the string, intcrprctcd as a binary representation ofan integer, is zero morlulo five' lbr example,0101 and 1L11, representing the integers 5 and 15, respectivel.y, are to be acceptcd' 11. Show that the language 7 : {uwu : 'u,w E {o, b}* , lrl : 2} is rcgular.

*10.

L2. Show that tr : {.a," : ",> 4} is regular. 13. Show that the language L: [iQ ."

) is rcgular. ffi (1,1,2, "'} is regih,i,,k fixerl, j :

It^: rt.) 0'n I

Sf]o* that the langua,geL : {a|" : n,: i, I ula,r.

15. Show that the set of all rcal numbcrs in C is a regular lauguage. L6. Show that if -L is regular, so is Z - {I}. Qf

Use (2.1) and (2,2) to show that

d " ( s ,t u u ): d " ( d " ( q , w ), u ) frrr all tr,u â‚Ź E'


?.2 NorurErEFMrNlsTrc Flrulrl

Accnpr'trrs

47

1 8 . Let -L be thc language ar:cepted lry thc autorrraton in Figure ?.2. Find a dfa that at:cepts L2.

1 9 , I,et .L be the langrrage acceptcd by t,he automaton in "F-igure2.2. Firxl a dfa for tlre larrguage Lz - L,

2 0 . Let I, be the language in Example 2,5, Show that .L* is regular, 2L. Let G,r.rhe the transition graph for some dfa M, Prove the following, (a) If f (M) is infirrite, then G,y must have at least one cycle lbr which there is a path fronl thc initial vertex to some vertcx in the cyclc and a path frorn some vertex in the cyr:le to some final vertex. (b) If , (M) is finite, then no such cycle exists. ffi 2'

Let rrs define an opcration trun,t:a,te, which removes the rightmost symbol lrorrr arry utring. l,br example, trurtt:a,te(aaliba) is aaab. The operation can be extenderl to languages bv :ate(L):

{truncate (tu) : trr â‚Ź I}.

Show how, given a dfa for any regular langrrage 1,, onc can construct a dfa for truntate (tr). Flonr this, prove that if -L is a regular language not containing ,\, then truncaLe (.L) is also regular. L e t r : a o e , r , , ' a m t a : b o b r . . . b n r z : c o c l ' , , c " , b c b i n a r y n u r r l b e r sa s defined in Exarnple 1.17. Show that the set of strings of triplets

(il)(li) (r) where the di, lt.i, c; are such that ;r * U : z is a regular languagc, 24. Wltile the language accepted by a given dfa is unique, there are normally man.y clfa's tirat accept a language. Find a dfa with exactly six states that accepts the sarne larigrrage as the dfa in Figure 2.4. m

N o n d e t e r m i n i s tFi ci n i t eA c c e p t e r s Finite act:rlptersa,rernorecomplicatedif we allow them to act rrondeterministically.Nondeterrninisnr is a powerful,but at firs{Er.Shunusualidea..$,3 gormally thirrk of computersas contplettilydeterministic,and the elernt:rrtof chijiEedeeihsout of piritie. Nevertheless,rioridctermintsmis a useful notion, as we shall seeas wc prot:eed.


48

Chopter 2 .l'l.rrtr: Aurounre

Accepter Definitionof o Nondeterministic Norrdeterminism mea,nsa choice of moves for atr automaton. Ratlter than prescribing a,uniqrrt: rnttve itt each situtr.titlrr, we allow a set of possible tnoves. Formally, we a,chievcthis by delining the trarrsition function so thirt its ratrge is a set of possible states.

onfilWin A nondeterministic finite accepter or nfa is definedbv thc quitrtuple M : (e,X, rl,r/0,F) , whereQ, E, {0, F are deiineclas for deterministicfitrite accepters,but d : Q x ( E u { A } )- - 2 Q .

Note that there are three major difli:rerrces between this definitiott ancl the definitiorr of a dfa. In a, nondctcrrninistic accepter, the rarrge of d is in the powerset 2Q, so that its vtr,hreis ttot a single element of Q, but a subset of it. This subset defines the set of all possible states that can be reached try the transition. If, firr irrstance, the current stattr is q1, the symbol a is read, atrd ,5(qr, *) : {qo, qz} , therr either 8o or 8z could be the uext state of thtl rrfa' AIso, we allow .\ as the second irrgutnetrt of d. This mcirns that the nfa can mtr.kt: a transition without corrsutningau input symbol. Although we still assurnethat the input rrrechanismcan only travel to the right, it is possible that it is statiorrary on some Inove$. Finrr,lly, in an nfa, the set ,) (ql, *) ttray be empty, mearrirrg that there is no trarrsition defined for this specific situation' Like dfa's, nondeterrnirristic accepters can btl representedby transititlrr graphs. Ihe vertices are rlt:tertnined by Q, while arr edge (q1,qr) with labcl a is in the graph if arrd otily if d(qi,a) contairrs {i. Note that sintxr n rllay be the empty string, there can be some edges labeled .\' A string is accepted lry au ltfa if thcrc is sone sequenceof possitrle rrroves that will put tlte machiue in a,firral state at the end of thc string. A string is rejcctecl (that is, not accepted) only if there is no possible sequenceof moves by which a firral state c:arrbe reached. Nondetcrrnitrism can therefortl be viewed as involving "irrtuitive" insight by which the best move ctrn be chosetr at every sta,te (assurning that the nfa warrts to accept everv strirrg)'


2.2 NoNnnTERMINIS'r'tc -t'rrurlr:AccnrrnRs

49

Figure 2.E

i.WWMorrsidertItetrarrsititl'g-pr'inFigure2.8.Itdescribesarrtrrrrltttttrministic acceptersirrcctlx)rc irre two transitionslabeleda out of gg. I

WtrorrcIeterrrrirristit:a,rrtorn-atonisshowninFigure2'9.Itisrrorrd(lterministic not otily bcc:au$cseveral edgeswith the same label origirratr: from one vertex, but also bct:iurser it has a A-transition. Some trarrsitiorr, suc:has,i(q2,0) are unspecified irr tIrc graph. This is to be interpreted as a transition to the empty set, that is, d-(g2,0) : g. The automaton acceptsstritrgs .\, 1010, and 101010,but not 110 and 10100. Notc tha,t for 1.0there are two alternative walks, one leading to qe, tlte other to q2. Even though q2 is not a final sta,te, the string is acccpted hs:ause one walk leads to a final stattl. Again, the trarsiticlrr firrrr:tion can he extended so its second argurnrlnt is a string. We require of tlrt: t:xtended transition function d* that if d. (q1,tu) : Qr, then Qi is the set of rrll possible sta,testhe automaton may be in, having uttr.rtedin state {1 and }ravittg rearl rir. A recursive definition of d*, analogous to (2.1) a,nd (2.2), is possible, but rrot particularly enlightening. A rnore easily appreciated defirrition t:an be made through transition graphs. I

Figure 2.9

'r^Y-

0

-Y,_

1'-=l*--^-7 l.

\

0,r

Iz


50

Chopter 2

FINITE Aurolvrrrrn

Definilion

?.5

Fbr an nfa, tlre extendctl tratrsitiou function is dt:firred so that d. (qr,r) contains 4i if rr,nrlorrly if there is a walk in thc: transition graph fiom q, to qy lir,bclt:d'u. This irolcls for all r7n,t1,e Q and ru e E*.

Figure 2. l0 reprtlserrts att nfa. It has severtr,lA-trarrsitiotrs aud some undeflned tra,nsitirlrrssuch as d (q2,a). 'fhere is a walk labelod wc wattt to find d* (qr, a) nrxl d* (,1r,.\). Sr,4rpost: n, involvirrg two .\-transitions from q1 to itself. tsy using some of the A-trdgtls twicur, wr: see that there are also walks irrvolvitrg .\-transitions to q11arrd 92. Thus , ) : { g o 'q r i q z } d* (,1ta Sincethere is rr .\-edgebetween {2 alrd q0rwe have irnrnediately that d. (q2,A) contains gs. Also, since any state can be reat:hetlfrom itself by making no move, a,ndrxlrrscquentlyusing no input symtrcll,tl* (q2,.\) also contains q2. There,ftlrtl 6* (qr, A) : {So,q'i} . Using as mir,ny .\-tratrsitions as needed, you (rirrl also check that d * ( q r ,a a ) : { q o ,Q t , Q z I

I The definition of d* through labeled wirlks is somewhat informal, so it is rrscful to look at it a little more t:loscly. Definition 2.5 is proper, sincc' between any vertices ui and r.r; there is either a walk labeled tu or there is not, inclicating that d* is cottrpletely defined. What is perltaps a little harder to scc is that this definition urn irlways be used to flnd d. (gi,ru). Itr Section 1,1, we descrihcd art algorithm for finding all sirnple paths bctween two vertices. We crr,nrrotuse this algorithm directly flirrce, as Exilrnple 2.9 shows, a labeled walk is not always a simple path. We catt ttrodify the sirnple path algorithrn, rernovirrg the restriction that no vettex or edge

Figurc ?.10


2.2 NoNDHTERMrNrsrrc Frr-rrrnAccEprERs

51

can be repeated. The new algorithrn will now generate successivelyall walks of length one, length two, length tlrree, arrd so on. There is still a difliculty. Given a ru, how lorrg can a walk labeled trr be? This is not immediately obvious. In Exa"rrrple 2.9, the walk labeled a between {1 antl q2 has length four. Ihe problem i$ caused by the .\transitions, which lengthen the walk but do not corrtritnrte to the label. The situation is saved by this obscrvation: If between two vertices r.',.and ui thercl is rr.ny wa,lk labeled ,ar, t,herr thr:rc mrrst be sorne walk labeled u.' of lerrgth rro more than A + (1+ A) lrl, whcrc A is the number of .\-edgcs in tlte grir.ph. The a,rgurnent for this is: While A-edges may be repeated, there is always a walk in which every repeatexl A-edge is separated by arr edge labeled with rr nonempty symbol. Otherwise, thc walk contains a cycle labeled .\, whir:h can be replaced by a sirnple path without changing the Iabel of the walk. We leave a fbrmal proof of tlfs clairn ir,$an exercise. With this observation, we have a,urethod for computing d* (q,;,ru). We elvalua,teall walks of lerrgth at trxrst A + (1 + A) ltul originating at tr,;. We select fiom them those that are labeled zr. The terminating vertices of the selectedwalks are the elements of the set 6* (qi,,ut). As we have rt:rnarked, it is possible to define d* irr ir, rercursivefashion as was done for tlte tleterministic case. The result is urrfcrrturrtr,telynot very transparent, arrd arguments with the extended transitiorr fun<:tion defined this way are hard to follow. We prefer to use the rnore irrtuitivc and more manageablealternative irr Definition ?.5. As for dfa's, the larrguagc acr:cpted by an nfa is definecl forrnally by the cxtended transition function.

lrR,Ffin',tllllellrii ,ti The language -L accepted by a,rrnfa, M : (Q,X,d,qo,F) is defined as the set of all strings accepted in the abovtl scrrsel.Fcrrma,lly, L ( M ) : { r i r â‚Ź X * : d * ( q o , w )n I ' I n } . Irr words, the language consists of all strirrgs ur fbr which there is a walk labeled 'u I'rom the initia,l vertex of the transitiorr graph to some final vertex.

,

i

t

Exumple 2.10

;

What is the latrguage accepttxl by the a,utomaton in Figure 2.9? It is car,sy to see from the graph that tlrtl orrly way the nfa can stop in a final state is if the input is either a repetition of the string 10 or the empty strirrg. Theref'orethe automaton accepts tlr: larrgrrer,ge I = {(10)"'; n > 0}.


62

Chopter 2 FINITEAurouil,rt

What happens wherr this automaton is presented with the string tu 110? After reading the prefix 11, the automaton linds itself in state q2, with the transition d (q2,0) undefined. We call such a,situatiotr a dead configuration, and we r:an visualize it as the automaton sirnply stopping without further action. But we must always keep in mind that such visualizations are imprecise and carry with them some darrger of misinterpretation. What wc carr say precisely is that ) o. d * ( q 0 , 1 1 0= Thus, no final state can be rttached by processingu, - 110, arxl hence the string is not accepted.

I

Why Nondeterminism? In reasorritrgabout nondeterministic mat:hirres,we should be quite cautious in using irrtuitive notions. Intuition c:an tlasily lead us astray, and we mrrst be able to give precise arguments to substarrtiatc our conclusions. Nondeterrrrinisrn is a difficult concept. Digital conrputers are completely deterministic; tlteir state at anv time is uniqucly predictable from the input and the initial state, Thus it is natural to ask why we study nondeterttftristic rnachifies at all. We are trying to rrrodel real systems, so why includc such nonrnechartical features as choice? We ca,n an$wer this question in various wiiys. Many deterministic algorithrns rcrluire that one make a choice at $omc) stage. A typical example is a game-plarying progrant. Flequently, the best move is not known, but ca,nbe f'rrrrndusirrg arr exhaustive search with backtracking. When several tr,ltt:rnatives are possible, we choose one arrd follow it until it becomes clear whcther or not it was best. If not, wâ‚Źl retreat to the last decision point and explore the othtrr <:hoices.A notrdeterministic algorithur that can tnake the best choice would bc able to solve the problern withoul backtra,cking,brit a,dtltcrrtritristicolre can simulate nondeltcrmirristtt with some extra wrlrk. F<rrthis reason. noncleterministir:rnat:hirrrlscalt serve rrr morlt:ls of search-and-backtracka,lgorithms. Nondctcrrnirristn is sometimes helpfirl in solvirrg probletns easily. Look at the rrfir itr F'igure 2.8. It is clear that thertt is a cltoice to be made. The first alternaLive leads to the acceptarr:t:of tlte string ail, while the seconrl accepts all strings with an even mrmhnr of a's. The Ianguage accepted bv the nfa is {a3} u {ot" : n > 1}. While it is possible to find a dfa fcrt this Ianguage, the nondeterminisrn is quite natural. The language is the urriotr of two quite difftrrcrrt sets, and the uondeterminism lets us decide at the olrtset whir:h case we want. The deterministic sohrtion is not as obviouslv


2.2 NounptrrRMtNrsTrc FrnrrrnAccnlrnns

5S

rclated to the definition. As we go orr, we will seeother and more convincing exarnplt:sof the rmefulnessof nondeterrnirrisrn. In the sarrrcvein, nondeterminism is an effective rrrcr:hani$mfor describing some cornplicated ltr,ngua,gesconcisely. Notice that tlrc definition of a gralrlrrlar invtllves a nondeterministic element. Irr ,9 -

a,9bl.\

we can at any point chooseeither tlrc first or the second production. This It:tu rrs specify many different strittgs usirrg only two rules. Firrally, therreis a technica,lreason for irrtroducirrg rrondctcrminism. As we will see, tltlrtirirr results a,re more easily established for rrfats thtr,n for dfats. Our rrext maior resrilt indica,testhat there is rro essential diffcrcnce betweetr tlrt:sc two types of automata. Consequently, allowing rron(lcterminism ofterr sirrrplifies f'rrrmrr,larguments without affecting the gerreralitv of the conc:lusiorr.

l.

Prove in detail the claim made in the previous section that il in a trarrsitiorr graph there is a walk labelerl rl, there must be some walk labeled tu of length rro rrrore tharr A + (1 + A) l,rrrl.

,

Fitrd a dfa that at:r:eptsthe langua,gedefined by thc nfa'in Figure 2.8.

3.

I n F i g u r e 2 . 9 , f i n d d * ( q 6 ,1 0 1 1 )a n d d * ( g 1 , 0 1 ) .

4 . In Figure 2.10, Iincl d- (qo,a) and d* (r;r,l)

ffi

5 . Fbr the nfa,in Figurc ?.9, find d- (qo,1010) and d* (t71,00).

6

Design at nfa with rto rnore than five states for thc sct {abab" ; rr.} 0} U {a,ha'o:rr,>0}. C.rrr"t.,rct an nfa with three statcs that accepts the language {tr,b,abc}-. W

O

8 . Do yorr think Exercise 7 can be solvccl with fewer than three states'l ffi 9 . (a) Firrrl an nfa with three states that acccpts the language L : {a" : rz > 1}' u {I,*aA' : rrr } fi, fr;> n ) t (b) Do you think the larrgrragein pa,rt (a) can bc a,cccptcd lry an nfa with fcwcr than three states'/ \lpt'

l,'ind an nfa with lbur states lbr -L : {a" : rr > 0} U {h"u.: n } I}.

@

Wtli"tr of thc strings 00, 01001, 10010, 000, 0000 are arceptetl by the following rrfa?


54

Chopter ? Flt{trr Aurolrere

0 /^\f -\?o/

\__.\

12. What is the complement of the language accepted by the nfa in Figure 2,10? 13. Let .L be the language accepted by the nfa in Figure 2.8. Find an nfa that accepts I u {a5}. 14. Give a simple description of the Ianguage in Exercise 1.2. (rs.)fina an nfa that accepts {a}* and is such that if in its transition graph a single edge is removed (without any other changes), the resulting automaton accepts i") W 16. Can Exercise 15 be solved using a dfa? If so, give the solution; if not, give convincing arguments for your conclusion. 17. Consider the following modification of Definition 2,6. An nfa with multiple initial states is defined by the quintuple M :-(Q,E,d,Qo,F),

where 8o C Q is a set of possible initial states. The language accepted by such an automaton is defined as L (M) : {tr.' : d* (q6,trr) contains gy, for any q0 € Qo,St € F} . Show that for every nfa with multiple initial states there exists an nfa with a single initial state that accepts the same language. Nft 18. Suppose that in Exercise 17 we made the restriction Qo fl F : fi. Would this affect the conclusion? ( rg./Use

Definition 2.5 to show that for any nfa

d- (q,uu) :

U

d* (p,r) ,

pE6* (tt,u)

for all q € Q and all trr,u € D*, 2o.

An nfa in which (a) there are no tr-trartsitions, and (b) for all g e I and all a e E, d (q, a) contains at most one element, is sometimes called an incomplete dfa. This is reasorrable sirrce the conditions make it such that there is never arry choice of moves.'


2.3 EQUIvaIENCE oF DptpRturt'ttsrrcANDNoNnnrnnN.rrNrsrrc FInrIre AccnRrnns

DD

For E : {a, b}, convert the incomplete dfa below into a standard dfa.

W

E q u i v o l e n coef D e t e r m i n i s t o i cn d N o n d e t e r m i n i s t iFci n i t eA c c e p t e r s

We now come to a fundamental question. In what sense are dfats and nfa's differerrt? Obviously, there is a difference irr their definition, but this does not imply that there is any essential distinction between them. To explore this question, we introduce the notion of equivalence between automata.

lIM Two finite accepters.Ml and M2 are said to be equivalentif L(Mr): L(M2,, that is, if they both acceptthe samelanguage.

As mentioned, there are generally many accepters for a given language, so eny dfa or nfa has many equivalent accepters.

.11 is equivalent to the nfa in Figure 2.9 since they both accept the language {(10)" : n, } 0}.

Figttre 2.11


56

Chopter 2 Frrurrp Auronara

When we colrlpare different classesof automata, the question invariably ariscs whether one class is more powerful than the other. By rnore powerful we mean that an automaton of one kind can a<:hicvesomething that cannot be done by any rrutomaton of the other kintl. Let us look at this cluestiott for finite accepters. Since a dfa is in essencea restricted kind of nfa, it is clear that any language that is accepted by a dfa is also accepted by some nfa. But the converse is not so obvious. We htrve added nondeterminism, so it is at least corrceivable that there is ir larrguage accepted by some nfa for which we cannot find a dfa. But it turrrs out that this is not so. The classesof dfa's and nfa's are equally powerful: For every language accepted by sorrrerfa there is a dfa that accepts the same language. This result is ttot obvious and certa,inly lms to be demonstrated. The argument, like most argurnents in this book, will be constructive. This me&ns that we can actually give a way of corrverting any nfa into an equivalent dfa. The corrstruction is not hard to understand; once the principle is clear it becomes the starting point for a rigorous argument. The rationale for the construction is the following. After arr rrfa has read a string ?u, we may not know exactly what state it will be irr, but we can say that it must be in orre state of a set of possibltrstates, saY {qz,Qjr...,QnI.An equivalent dfa after reading the same string rrrust be in some definite state. How can we make these two situations correspond? The answer is a nice trick: Iabel the states of the dfa with a set of states in such a way that, after reading ru, tlre equivalentdfa will be in a sitrglestate labeled {Uo,Ui,...,qa}. Since for a set of lQl states there are exa,t:tly2lQl subsets, the corresponding dfa will have a fi,nite number of states. Most of the work in this suggested construction Iies in the analysis of the nfa to get the correspondence betrveen possible states and inputs. Before getting to the formal description of this, Iet us illustrate it with a simple example.

Convert the nfa in Figure 2.12 to arr equivalent dfa. The nfa starts in state gs, so the initial state of the dfa will be labeled {qs}. After reading an a, the nfa c:an be itt state q1 or, by making a .\-transition, in state q2. Therefore the corresponding dfa must have a state labeled {gr, qz} and a transition d ( { s n }, a ) :

{qt,qz}.

In state q0, the tfa has no specified transitiort wltetr the input is b, therefore d ({sn},b) : @. A state labeled g represents an impossible move ftrr the nfa and, therefore, means nonacceptance of the string. Conseqrrently, this state in the dfa mrrst be a nonfinal trap state.


2.3 Eeurvar,ENCE oF DnrnRrr,uNrsrrc ANDNoFrnErrgRMINrsTIc FrNrrn AccBnrens

57

Figure 2.12

We have now introduced into the dfa the state {h, qz}, so we need to find the transitions out of this state. Remember that this state of the dfa corresponds to two possible states of the nfa, so we must refer back to the nfa. If the nfa is in state Q1and reads an a, it can go to q1. F\rthermore, from 91 the nfa can make a .\-transition to q2. If, for the same input, the nfa is in state q2, then there is no specified transition. Therefore d ( { q t , q z }, a ) :

{ g r ,q r } .

Similarly, d ( { q t , q z }, b ) :

{qo} .

At this point, every state has all transitions defined. The result, shown in Figure 2.13, is a dfa, equivalent to the nfa with which we started. The nfa in Figure 2.12 accepts any string for which d* (q0,ru) contains 91. For the corresponding dfa to accept every such w7 araystate whose label includes 91 must be made a final state. I

Figure 2.1S


Chopter 2 FIruIrn Aurovete

Let .L be the language accepted by a nondeterministic finite accepter M1" : (8r, E, dry,qo,.Fry). Then there exists a deterministic firrite accepter Mn : (8o, E, d;l, {qo} , Fp) such that L:

L (Mp).

Proof: Given M1y, we use the procedurc nf a-to-tlfa below to construct the trarsition graph Gp for X,tIp. To understand the constructitln? relllefiIber that Gp has to have certain properties. Every vertex must have exactly the lxl outgoing edges, each labeled with a different element of E. During continues procedUre construction, some of the edgesmay be missing, but the until they are all there. procedure:

nfa-to dfa

L. create a graph Gn with vertex {qs}. Identify this vertex as the initial vertex, 2. Repeat the following steps until IIo lrlore edges are missing' Take any vertex |qo,qj, '.', {t} of Gn that has no outgoing edge fbr sorne aâ‚Ź8. C o m p u t e d * ( g 1 , a ), d * ( q j , a ) . . . , d * ( q * , o , . Therr form the uniotr of aII these d*, yielding the set IU,q*,"',QnI' if it does not already Create a vertex for Gn labeled {qt,q*,"',{'} exist. Add to Gp alr edge from {qr,qi,...,4t} to tqt, Q*,...,{'} with a.

and label it

3. Every state of GIr whose label containu a.ny qy â‚Ź Flv is identified as a final vertex. 4. If MN accepts ),, the vertex {So} i" Gsr is also made a final vertex' It is clear that this procedure always terminates. Each pass through the Ioop in Step 2 adds arredge to Gn' But Gp has at mo$ 218'^/llXl edgcs, so that the loop eventually stops, To show that the constru<:tiorr also gives the correct aIISwer, we argue by induction on the length of the input strirrg' Assume that for every u of length less than or equal to n, the presence in G1y of a walk labeled u from qo to qi implies that in Gn there is a walk labeled tr from tsol to a state 8l : {...,gi,--'}- Consider rrow any u : ua and look at a walk in G1r labeled r,rrfrom qo to q. There must then be a walk labeled u frorn qo to Q,iand an edge (or a sequence of edges) labeled a liom qt to qt. By the irrductive assurrrption, in Gp there will be a walk labeleclu from {qo} to Qa. Bui by construction, there will be an edge from Qa to sone state whose label contains ql. Thus the inductive iutsumption


2.ll

EeuIv.q,t,HN(rEoF DnrnnMlr'usrrc AND Nolotrpn.rr,rrNrsTr(l !-rNrrp AflfletrnRs

59

lnlds fbr all strings of length rz * 1 As it is otrviously trrrt: for z : 1, it is true for all n. The result then is that whenever dir(So,trr) contains a firml state 91, so tl(xruthe label of di(gs,tt). To complete the proof, we reverse the argrrtrent to show that if the label of di(qs,tu) contains g/, so rrrust dh (qo,ru). r

Ihe argumerrts irr this proof, tr,lthough correct, are admittedly somewhat terse, showing only tlte ntajor steps. We will fbllow this practice in the rest of the book, emphasizing the basic ideas irt ir proof arrd omitting minor deta,ils, which you may want to fill in yourself. The construction in the above proof is tedious but irnportarrt. Lnt uu rlrl rr,ncltherexa,mple to make sure we understand all the steps.

2.14intoatlequivalentdetennitristictrraclrine. : {qo,qt}, we introduce the state {qo,qr} in Gt and add Sittcc t)-1,,(q0,0) an edge labeled 0 between {q6} and {,lo,sr}. In the $irrnewity, rnnsidering dlu (qo, t) : {qt} gives us Lhe trew state {q1} arrtl an txlgc labclud 1 between it and {(ro}. There a,re now a number of missing edges, so we continue, using the (:orrutructionof Theorem ?.2. With o : 0, ri : 0, J : 1, we compute di' (Sn'0) U div (qr' 0) : {,It,q,qzI . This givus us thc new state,'{qo,g',ez} and th.etransition d n ( { q o ,s r } , 0 ) : { { 0 , e r , Q z I. T l r c r r ,r r s i n ga , : I , i :

0, .l : I, k = 2,

dfu(so, 1)u (ti/(q',1)u dfu(sr,1): tsr,sr] makes it necessaryto introduce yet another state {q1,qz}. At t}ris point, we have the partially constructed autornaton shown in Figure 2.15. Since there a,restill some missing edges,we continue until we obtain the complete sohrtion in Figure 2.16. I

Figure 2.14

,n


Chopter 2 Frutre, Autottl'rn

Figrue ?.15

(--lrl

Figure 2.16

@

v'l\

lrrJ)

@

't\ V;/') \,

-r. ___,/

{

,/

,


2,3 EeUIVALIINCEoF DErnRuIuIsuc

AND Nounu'rpRltlNISTIC FINITE AccEP'I'ERS

61

One important conclusiolr we can draw from Theorem 2'2 is that every Ianguage accepted by an nfa is regular.

1, Use the construction of 1lheorem 2.2 to convert thc nfa in Figrrre 2.10 to a dfa. Can you see a sinrpler an$wer rrrore directly? 2, Convert the nla in Exercise 11, Sectiorr 2.2 into an cquivalent rlfa. ffi /ftor'rve.t

the ftrllowing nfa into an equivalerrt tlfa.

\0 4, Carefully cornplete the arguments in the proof of Theorcm 2.2. Show irr tletail that if the label of d-i;(q6,ur) contairrs 97, then dfr (So,tr.')also corrtairrs q1' it true that ftrr arqynfa M : (Q,x,d,{o,F) the cornplcment of I(M) is 4 (DIs * @}? If Bo, prove it. II not' give a equal to thc set {u e X* : d* (qo,u) nF: J counterexample.. '? fr it trr.re that ftrr every nfa M : (Q, X, d, r7o,F) the complement of I (M) is ft? * f eQual to the set {tr e E- ,d" (qo,ur)r-r(Q - F) # s)'l If .so,prove it; if not' give a r:outtterexantple. |. / \... tXZ/ p.o"* ttrat for every nfa with an arbitrary numbcr of final rtfates there is an eqrrivalent nfa with only one final state. Can we make a siurilar claim for clfh's'l ffi

#)

Firrd an nfa without ,\-trarrsitirrns and with a single firral state that accepts thc sct {a} u {b"' : n, > 1}. ffi

{a

9J Let Z be a regular langrrage that does not contain,l. Show that thcre exists arr rrfa without ,\-transitions and with a single final state that accepts I.

L0, Define a dta with multiple irritial states in an analogorrsway to thc corresponding nfa in Flxercise 17, Section 2.2, f)oes there always exist au equivalerrt clfa with a single initial state'l t-l t i" +

@)

j

Provc that all finite languagcs are regrtlar. ffi Show that if /- is regular' so is IE'

13. Give a sirnplc verbal descriptiot of thc la,nguagear:r:eptedby thc dfa in Figure 2.16. Use this to find another dfa, erpivaletrt to thc given one, but with fcwer statreB.


Chopter 2 FIuIrn Aurov.arm

{i+\ -

f,"t .L be any langrrage. f)efine euen (trr) as the string obtained by extracting from tu the letters irt even-mrmbe.red positions; that is, if w :

at.a2o"3a 4,,.,

then e . t t e n\ w ) :

tt'2a4....

Corresponding to this, we can definc a language eaen(L):

l e u e r t( u . , ): ' r r re I ) .

Provc that if Z is regular, so is euen (I),

ffi

15. Frtrm a Ianguage .L we create a rre;wlanguage chopL (l)

by renxrvirtg the two

'efrmosr svmbors ffiL;'Ti ; :TjTl',,,:,,

Show that if .L is regular then cft.op2(tr) is also regulat. ffi

R e d u c t i o no f t h e N u m b e ro f $ t o t e si n F i n i t eA u t o m o f o * Any dfa defines a uniqrte la,nguage,but the colrvortteis not true. For a givcrr language, there are rnany dfa's that accept it. Thcre may be a considerable diflerence in the numlrer of states of such equivalertt inrtomata. In terms of the questions we have corrsideredso far, all solutiorrs arc equally satisfactory, but if the results a,re to be applicd in a practical settitrg, tht:re may be reasolls fcrr preferring one over antlther.

The two dfa's depictcrl in Figure 2.17(a) and 2.17(b) are equivalent' a$ ir fcw test strings will ctuickly reveal. We notice sottre obviottsly unnecessary fcatrrres of Figure 2,17(a). The state q5 plays absolrrtelv no role in the autornrrton since it can nevet tlt: retr,chedfrom the initial state q0. Such a statt: is inaccessible,and it carr btr removed (along with tr.ll transitions relalilg to it) without affec:tingthe lrrrrguageaccepted by the irrrtomaton. But even aftcr the removal of q5, the first tlrttomatotr has sollre redlrndant parts. The states rcachable subsequent to the Iirst move d(S0,0) rrrirror those reachable frotn a first move d (So,1). The secorrdeurtomaton cotnbines these two options. I


2.4 RpnucuoN oF rnn NuunnR on Smrus IN FINITEAurouare

63

Figure 2.1?

@

From a strictly theoretica,l point of view, there is little reason for preferrirrg the automaton in Figure 2.17(b) over that in Figure ?.17(a). However, in terms of simplicitv, the second alternative is clearly preferable. Representation of an automatorr for the purpose of computation requires space proportional to the number of states. For storage efficiency, it is desirable to reduce the number of stir,tes as far as possible. We now describe an algorithm that accomplishes this.

rii-0-it iltlLfi

i

Two states p and q of a dfa are called indistinguishable

if

,5*(p, r) € F implies d. (9, ro) € f', and d* (p,u.') f f' irnplies 6* (rt,u) f F, lbr all tu € E*, If, on the othcr harrd, there exists some string u e E* such that d* (p,r) € F and 6* (q,ut)f F, or vice vcrsa, then the states p and g are said to be distinguishable slrrng ?{).

by a


Chopter 2 Fu-rtre Auroneta

Clearly, two states are either indistinguishable or distinguishable. Indistinguishability has the propertie$ of an equivalence relations: if p and q are indistinguishable and if q and r are also indistirrguishable, then fro are p and r, and all three states are indistinguishable. One method for reducing the states of a dfa is ba^sedon finding and combining indistinguishable states. We first describe a mothod for finding pairs of distinguishable states, procedure:

mark

1, Remove all inaccessible states. This can be done by enumerating all simple paths of the graph of the dfa starting at the initial state' Any state not part of sonte path is inaccessible. 2. Consider all pairs of states (p,q). If p e F and q fr F or vice versa, mark the pair (p, q) as distinguishable, 3. Repeat the following step until no previously unma,rked pairs are marked. F o r a l l p a i r s ( p , q ) a n d a l l a e X , c o m p u t e5 ( : p , o ) : p o a n d 6 ( q , a ) : eo. If the pair (po,eo) is nmrked as distinguishable, mark (p,q) as distinguishable. We claim that this procedure constitutes an algorithm for marking all disiinguishable pairs. The procedure nlurlr. applied to any dfa M : (8, E, 6,q0,F), terminates and determines all pairs of distinguishable states. Proof: Obviously, the procedure terminates, since there are only a finite number of pairs that can be marked. It is also easy to see that the states of any pair so marked are distinguishable. The only claim that requires elaboration is that the procedure fiuds all distinguishable pairs. Note first that states q,;and qj are distinguishable with a string of length ru, if and only if there are transitions 6 ( q r , , a ): q n

(2.5)

6 (qi,o) : qu

(2.6)

and

for somea â‚Ź X, with q6 and g1distinguishableby a string of length n - lWe usethis first to showthat at the completionof the nth passthrough the loop in step 3, all states distinguishableby strings of length rz or lesshave beenmarked. In step 2, we mark all pairs indistinguishableby .\, so we have a basiswith rz : 0 for an induction. We now a,ssumethat the claim is trrte


2.4 RbDUCT'roN oF rsp Nulrnnn, op Srarns ru Frlrrrn Aurotrlnrn

65

for all i : 0, 1, ...1n - 1. By this inductive assumption, at the beginning of the nth pass through the loop, all states distinguishable by strings of length up to rl - 1 have been marked. Becauseof (2.5) and (2.6) above, at the end of this pass, all states distinguishable by strings of length up to n will be marked. By induction then, we can claim that, for any TL,at the completion of the nth pass, all pairs distinguishable by strings of length rz or less have been marked. To show that this procedure marks all distinguishable states, assume that the loop terminates afber rz pas$e$. This means that during the nth pass no new states were marked. Flom (2.5) and (2.6), it then follows that there cannot be any states distirrguishable by a string of Iength rz, but not distinguishable by any shorter string. But if there are no states distinguishable onlv by strings of length n, there cannot be any states distingrrishable orrly by strings of length n * 1, and so on. As a consequence,whett the loop terminates, all distinguishable ptrirs have been marked. r

After the marking algorithm has been executed, we use the results to partition the state set Q of the dfa into disjoint subsets {qn,qj,...,Qr}, {qr,q^,'..rQn},'.., such that any (t E I occurs in exactly one of these subsets, that elements in each subset are indistinguishable, and that any two elements from different subsets are distinguishable. Using the results sketched in Exercise 11 at the end of this section, it can be shown that such a partitionirrg can always be found. Flom these subsets we construct the minimal automaton by the next procedure. procedurer

reduce

Given

fr:

(8,E,d,q0,F), we construct a reduceddfa as follows.

1 . Use procedure mark to lind all pairs of distinguishable states. Then from this, find the sets of all indistinguishable states, say {ft, Qj,...,Qnl1, {qr,q*r.'.tQn}, etc', as describedabove'

2 . For each set {qa, Qj, ..., enl of such irrdistinguishable states, create a state labeled i,j . . -h for M.

3. For each transition rule of M of ihe form 6 (q,,a) : qe, find the setsto which q' Ttd qo belong. If q' â‚Ź {U,ei,...,qr} and Qpâ‚Ź {qt,Q*,'..,8'},add to d a rule

TU,i'"k,a):Im...n.


66

Chopter 2 Fmrrn Aurorr,rnra

4 , The initial state f11is that state of M whose label includes the 0. o.

F ir th* set of all the states whose label contains i such that ga â‚Ź F.

Consider the automaton depicted in Figure 2.18. In step 2, the procedrre mark will iclentify distinguishable pairs (qo,g+), (qr,qn), (qz,q+), and (g3,ga)' In some pass tlrrough the step 3 loop, the procedure computes ,l (gr, 1) : q+ and d ( q u ,1 ) : g r . Since (qs,g+) is a distinguishable pair, the pair (qo,qr) is also marked. Continuing this way, the marking algorithm eventually marks the pairs (qo,gt), (qo,qz), (qo,gt), (qo,q+),(qr,q+), kh,qn) and (qs,qa)as distinguishable, leaving the indistinguishable pairs (q1, qz), (h,q3) and (qr,m).Therefore, the states q1,{2,n3 &râ‚Ź all indistinguishable, and all ofthe states have been partitioned into the sets {So}, {qr, qz,qs} ancl {ga}. Applying steps 2 and 3 of the procedure reduce therr yields the dfa in Figure 2.19. I

Figure 2.18

Figure 2.19


2.4 RpnucuoN oF rHn NuNaenRon Smrns IN FINTTEAuroraarn

67

Given any dfa M, application of the procedure reduce yields another dfa

fr suchthat

L(M):"(fr) F\rthermore, M is minimal in the sense that there is no other dfa with a smaller number of states which also accepts L(M). Proof: There are two parts. The first is to show that the dfa created by reduce is equivalent to the original dfa. This is relatively easy and we can use inductive argument$ similar to those used in establishing the equivalence of dfa's and nfa's. All we have to do is to show that d. (g;,Tr,): qi if arrd only if the label of F(ga,ur) is of the form...j.... We will leave this as an exercise. The second part, to show that M is minimal, is harder. Suppose M has states {ps,pt,pz,...,Fm.l, withps the initial state, Assume that there is art equivalent dfa M1, with transition function d1 and initial state gn, equivalelt to M, but with fewer states. Since there are no inaccessible states in M, there must be distinct strings 'trr,'trz,...,wn srtch that

, t ): P t , i : 1 , 2 , . . . , m . d " ( p ow But since Mr has fewerstates than fr, there must be at least two of these strings, sa,/?116 and ur1,such that d i ( q o ,w n ) : d i ( s 0 , 1 , , , ) . Sincepl andp; are distinguishable, there must be some strin&z such that d. (po, upr) : 6* (pn,z) is a final state, and d* (qn,.r*) : d* (pt'z) is a nonfinal state (or vice versa), In other words, rurr is accepted by M and rurr is not. But note that

di (so,tupn): di (di (qo,wk) , n) * dl (di (qo,ut),r) : .li (qo,trrr) . Thus, M1 either acceg$ both tlrz and ur;r or rejects both, contradicting the assumption that M and M1 are equivalent. This contradiction proves that M' cannot exist. I


Chopter 2 FIullp Aulor,tn'r'e

1. Minirnizethe rturrrberof statesin the dfa in Figure 2.16,

/-

(d

,l

ltrrq

minimal dfa's for the larrguagesbelow. In eachcaseprove that the rcsult

rs mlnlmal,

(*) fI:{o"b"''n}Z,mlll (b) r : {a"h: n> 0} U {b'a : n } 1} (c) I:

{ o " o: n ,> 0 , r l l 3 }

(d)r:

{a^:nt'2 andn+41.

ffi

is deterministic. 3. Showthat the automatongeneratedby prot:etlurere.du,ce fr)

Virri*i"e the states in thc dfa depicteclitr the ftrllowing diagram.

ffi f1/

Strow that if .L is a norrernPty langrrage ,such that any r.l in .L has length at lea^strz, then any dfa accepting -L rnust have at least n, f 1 states.

6. Prove or disprove the following conjecture. If M : (Q, X,,t, q6, F) is a' minimal dfa for a regrrlar la,ngua,gc.L, then M : (Q,E,d,go,Q - f') is a minima,l dfa,

for Z. ffi f- Z) Sl.,* that inrlistinguishability is an equivaletrcerelation but that rlistinguishabilitv is not. 8. Show the explicit steps of thc suggested proof of the first part of Theorem 2,4, namely, that M is equivalent to the original dfa, **

g, Write a cornputer pr()grarn xhat produce.ga rninimal dfh for any given clfa. 10. Prove the fbllowirrg: If the states g* and qa are indistinguishable, and if q,, arxl g,, mrtst be distinguishablc, ffi and q., arc distinguishable, therr r71,


2.4 RnnucrloN

oI,' rxn Nuvel;R oF STATESIN F'INITE AUToMATA

69

11. Consider the following process, to be done after the corrrpletion of the procedure rnar,b. Start with some state, saY, {0. Put all states rrot marked distinguishable from ge irrto an equivalerrce set with qcl. Then take another state, not in the precedirrg equivalence set, and do the sarne thing' Repeat until there are no more states aw,ilable. Then formalize this suggestion to make it an algorithm, and prove that this algorithm does indeed partition the original state set into equivalence sets.



R * g u l q rL q n g u q g e s ond Regulqr Grtt m mq rs

ccording to our definition, a Ianguage is regular if there exists a finite acce.pter for it. Therefore, every regular language can be described by some dfa or some rrfa. Such a description can be very useful, for exarnple, if wr: vrant to show the logic by which we decide if a given string is in a certairt language. But in many instances, we need more (roncise ways of desr:ribitrg regular languages. In this chapter, we Iook at other ways of representinpJregular languages. These representatiorrs have important practical applications, a rnatter that is touched on in some of the examples and exercises.

ffiM RegulorExpressions One way of describing regJular languages is via the notation of regular expressions. This notation involves a combination of strings of symbols from some alphabet E, parentheses, and the operators *, ., and +. The simplest case is the language {a}, which will be denotetl by the regular expressiona. Slightly rrrore complicated is the language {a, b, c}, for which,

71


72

nruDReculAR GRAMMARS Chopter 3 Reculan LANGUAcE$

usirrgthe * to denote uttiott, we hirve the regular expressiona+b+o We use ' for cont:eltenation aud + for star-closttre in a, similar way, The expression (o,1-b. c)* sta.rrdsfrrr the star-closure of {a} U {bc}, that is, the language }, {.\, a, bc,aa, ubr:,hco,,hcbc,aaa, a,abc,,,.

FormolDefinitionof o RegulorExpression We construct regulrrrexpressionsfrom prirnitive constituentsby repeatedly applying certain recursiverules. This is similar to the way we con$truct f amiliar arithnteticerxpressions,

it$,q,firril.ii,g- rfltltil

,lr,ril

Let E he a,given alphabet. Then l . #,.\, and n â‚Ź X are all regular expresrtions.These are called prirnitive regular expression$.

2 . If 11 a.nd12 are regular expressions,so are rt*rz,rt-rz,

ri, and (r1).

3 . A string is a regular expression if and only if it can be derived from the primitive regular expressions by a firritc mrmber of applications of the rules in (2).

(a+b'c)'.(c+o) is a regular expression, since it is constructed by application of the above rules. For example, if we take 11 = c and rz : fr, we find that c * fi and (c * o) a"re also regular expressions. Repeating this, we eventually generate the whole string. On the other hand, (a + b+) is not a regular expression, since there is no way it carr be constructed from the primitive rcgrrlar expressions, I


3.1 R,nculan ExpRpssror-ls

7I

Longuoges Associoted with RegulorExpressions Regular expressionscan be used to describesomesimple languages.If r is a regula.rexpression,we will let .L (r) denote the languageassociatedwith r. This languageis definedas follows:

The languagetr (r) denoted by any regular expre$sionr is defined by the following rules. l. fi is a regular expressiondenoting the empty set, 2. .\ is a regularexpre$sion denoting{A}, 3. for everya â‚Ź E, a is a regularexpre$$ion denoting{a}. If 11 and rz are regulaxexpressions,then 4. L (r1 r r2) -.L (rr) u L (rr), 5 . L ( r r , r r ) : L ( r 1 )L ( r 2 ) , 6. I((r1)): tr(rr), 7. L(rfi:

(I(rr))-.

The last four rules of this definition are usedto reduceI (r) to simpler componentsrecursively; the first three are the termination conditions for this recursion. To seewhat languagea given expressiondenotes,we apply theserules repeatedly.

$*qmpfq 5,f,

Exhibit the languageL (a* . (a + b)) in set notation. L ( o * . ( a + b ) ): L ( a * )L ( a + b )

: (r (a))-(r (a)u r, (o)) : { t r , a , a a r o , e , a ) . .{. o} , b } : { a , a a , e , e a r , . ,b, ,a b ,a a br , , , }

I


Chopter 3 Ruculnn Laucuecns aNo RBctiLA.n, GR.c\,I\,IaRS

to*bK\"^"{

Exomple3.3

There is otte problem with rules (a) to (7) in Definition 3.2. They define a Ianguage precisely if 11 and r? are given, but there may be some amhri;Euity in breaking a complicated expression into parts. Consider, for example, the regular expression a ' b + c. Wo can consider this as being made up of , e find tr(a'b+ c): {ab,c}. Brrt there r L : & ' b a n d r z : c . I n t h i s c a $ ew is nothing in Definition 3.2 to stop u$ from taking rt : a and 12 : b * c. : {ab,ac}. To overcomethis, We now get a different result, L(a-b*c) we could require that all expressiolrs be fully parenthesized, but this gives -* crrmbersome results. Instead, we use a convention familiar from mathematics artd programrning languages. We establish a set of precedence rules for evaluation in which star-closure precedes concatenation and concatenation precedes union. Also, the symbol for concatenation may be omitted, so we can write r1r2 for rL .rz. With a little practi<re, we can see quickly what language a particular regular expression denotes.

For E:

{a,b}, the expression r=(&+b)-(a+bb)

is regular. It denotes the Ianguage L ( r ) : { a , b b , a a , a b b , b a , b b b , .. . . } We can seethis by consideringthe various parts of r. The first part, (a * b)*, stands for any strirrg of a's and b's. The second part, (a * bb) represents either an a or & double b. Consequently, -L (r) is the set of all strings on {a,b}, terminated by either an a or a bb. I

The expression y : (aa)" (bb)- b denotes the set of aII strings with an even number of a's followed by arr odd number of b's; that is L(r) : {a2"b2*+t : n } o, * > 0}. Goirrg from an informal description or set notation to a regular expression tends to be a little harder. I


ts.1 Rl;cur,nnFlxpREsSroN$ , D

r H*Wnrple3.S

For X = {0, 1}, givea,regula,r expression r suchthat L(r):

{rrr e X* :ur ha,sa,t lea,stone pair of consequtivezeros},

One catt arrive at alt arrswer by relirsclrrirrgsorntlthing likc this: Evcry utring in I (r) must conlaitr 00 somewhere, but what rxlrnc:sllc:frrrr:irnrl what gorrs after is completely arbitrary. An arbitrary strirrg orr {0, 1} carr bc dr:rrotrxl by (0 + 1).. Putting these observations together, we arrive at the solution

r .: ( 0+ l ) . 0 0( 0+ l ) . I

r the la,'guage 1 : {tu â‚Ź {0, 1}. : rl }rirsrro llair of rxrnser:rrtivozeros} . Even though this looks similar to Exanple 3.5, the answer is harder to construct. One helpful observation is that whenever a 0 occurs, it rnust be

lg[g*-gg-"gi.,-ly

H_Li.^suctra subsrrirffi

by atr arbitrary nutrrber of 1's. This suggcsts thrr.t thr: irrrswcr irrvcllvr:stlrrl repetition of strings of the form l .' . 101 . . ' 1, that is, the language denoted by the regula,rexpression(I-ilI I+)". However, the a,nsweris still incornplete, sirrcrrthe utringu ending in 0 or c:onsistingof all l's a,re una,ccollntedf'or. After takirrg care of these special caseswe arrivc at thc arrswcr

7 : ( r * 0 1 1 *()0- + A )+ 1 -( 0+ A ) . If we reasorr sliglrtly differently, we rniglrt corrrc up with arrother arrswer. If we see .L as the repetition of the strings 1 arrd 01, the shorter expression

?':(1+01)-(0+I) rniglrt llu rua,t:Ircrl.Although tlrc two erxpressionskrok rliflilront, both ir,rrswcrs are corl'ect, as they denote the sarne language. Generally, lhere are an r.rnlimited nunrber:of regula,rexlrressionsfor any given langua,ge. Notc that this la,rrguagcis tlrc txlrnlllcmt:nt of thc languagc irr Exarnple 3.5. Howtlvtlr, tlrc rc:gtrlir,rcxllrcssiorrs ar(l rrot verv sirnililr irrrrl do ttot suggest clearly l,he close relationship between the languages. I The last example introduces the notion of equivalence of regula,r expressions.We say the two regular expressionsare equivalent if they denote lhe same la,ngua,ge.One ca,n derive a variety of rules fbr simplifvirrg rtrgrrlirr


76

F,lculan. Lauc;uacns aNn R,ncuLeR GR.q.N'lN{,\}ts

Chopter 3

expresuirlrrs(st:t: Exclrt:isc 18 irr the following exercisesection), but since we have littkr rreed for such rnanipulations we will not pursr"tethis.

1 . t'ind a,ll strings in I ((a + t?)-b (a * ab)-) of lerrgth less tirart ftrur.

,

Does the expressiorr((t) + 1) (0 + 1)-)- 00 (0 + 1)- denotc thc languagein F)xarnple 3.5'l ffi Show that r : (1 +01 ). (0 11.) also denotes the langrragein Exarnple i1'6. Find two other equivalent expressions.

(,4)

Firxl a regular expressiotr fbr the set {a'b"' : ('rr*'rn)

is even}.

L s , Give regular e.xpressionsfor the following langrrages. (a) Ir : {1t"ll",rr}

4,'ntg 3}, O

(b) ,I, : {u,'"b"' : Tr.< 4,rrl { i1},

(r:) The corrrplelnentof Ir,

ffi

(rt) The trrrnlrlernetrt of -L2. @ 7,

Wfrot languagesdo thc cxprcssions(#.). ancla# clcnote? a

Give /

\ q

LL \aLL )

verbal

sirlple I

/

\ + \

Dil, \(n,

)

of tlre larrguage L((au)r b(aa)- +

).

8. Give a regtrlar expression fo, L", where .L is the larrguage in Exert:ise 1. jl,n'rn> ( 9 i C i t * a r c g u l a r c x p r c s s i o t rf o r L : 3]. $ fanb^:rtlL,m, {o,h"t.r.t:n,} i},tu e {a,b}+}. L1, Find a regular exptessiorrfrrr the cornpletnent of the language in Example 3.4'

fO,) tnir'raa, regula,rexpressiol for I:

12. Find a regular expressioufor L:

{uut : t,, w e {a,b}-,lol

: ?}, ffi

13, Firxl a regular expressiorrfor f :

,D

Iu

E {0, I}-

: ur has exactly one pair of consecutive zeros}

Ctu" rcgular expressiotrsfor the following languages on

( " ) all

{ o , b ,c } .

strirrgs .,.rrrt,rirrirrgexactly one a,

(l') all strings corrtir,itritrg IIL) Illore than three a's, ( c l all strings that contain at least ollc occurreltce of each symbol in

r,s

( d ) all str:ings that contain no run of a's of length greater than two,

* (") all

strings in whir:h all nrrrs of n's have lerrgths that are rrrultiples of three.


IJ.1 R,EGULAREXPRESSIONS

Tf

,'1i)Write regular expressionsfor the ftrllowing languagcs on {0, 1}. (a) all strings cnding in 01, (b) all strings not entlirrg irr 01, (c) all strings txrrttaining an even number of 0's, S (rl) all strings having at least two ot.:curreltces of the substring 00 (Note that witlr the rrsrtal interpretation of a substring, 000 corrtains two such oct:trrrences). (e) all strirrgs with at most two occurrences of the substring 00, * (f) all strings not corrtainitrg the substring 101. 16J Find regular expressions for the following larrguageson {4,6}. (u) /::

{ru : lurlrrrodS= 0}

{ffi

(b) f : {tn:no(to)mod3:0} (c) L = {w : no(ur)mod5 > 0} 17. Repeat parts (a), (b), and (c) of Exercise 16, with 5 : {4, b, c}. 18. Determine whether or not the following clairns are true for all regular expressions rr arxl rz. The svmbol : stands for e<luivalence of regular expressiotts itr the scuse that both expressitlttsdenotc the same language. (a) (ri)- : ri, (b) ri (rr * r:)* : (" + r'r)*, (c) (rr *rr).

= (rirfi).,

S

(d) (rrrr)- : riri. L9. Givc a general method by wirich auy regular expressiortr can be changed into frsuch that (I("))" : L(F). 20. Prove rigorously that the cxpressions in Exarnple 3.6 do indeed rlenote the specified larrguag;e. 21, For the case of a regular expression r that doelt rrot itrvolve ,\ or @, give a set of necessary arrl sufficient couditions that r rrrust satisfy if L (r) is to be infinite. S 22. Formal languages carr be used to dcscribe a variety of two-dimcnsional figures. Chain-code languages are dcfinccl on the alphabet X = {u, d, r, l}, w}rere t}rese symbols starxl for unit-lcngth straiglrt lines irr the dircctions up, tlrrwn, right, arrd left, rcspectively. Arr exarnple o1 this notation is 'urdl, wirich stands for the squarc with sides of urrit length, Draw pictures of the ligurcs denoted try the expressions (r'd)-, (urrLdru,)-, alrd (ruldr')'. 23. In Exercise ?2, what are sullicicnt conditions orr the exprcssion so that the picture is a r:losed cotrtour iu the sensethat the beginning and erxling point are the sarne? Are thcsc conditions also rrecessarv'l S


78

Chopter 3 Rnculen Laucuncns .+runRr:crJr.nRGRAMMAR$

6+)i"a

an nfa that accepts the larrguage L(o,a,- (a+b)).

a regular expression that rlenotes all bit strings whose value, when inCE)l'ina terpreted as a binary integer, is greater than or equal to 40. ffi 26, Find a regular expression for all bit strings, with leading bit 1, interpreted as a binary irrteger, with values not bctween 10 and 30.

Connection B e t w e e nR e g u l o rE x p r e s s i o nqsn d R e g u l o rL o n g u o g e s As the terrninologv srrggests,the connection betwct:rr r{rgrrlar languages and regular expressions is ir close one. TIte two concepts artl e$sentia,lly the satne; for every regular ler,nguagethere is a regular exprcssion, and fbr every regular expression there: is rr regular language. We will slmw this in two parts.

RegulorExpressions DenoleRegulor[onguoges We first show that if r is a relgular expression, then .L (r) is a rcgular la,ngua,ge. Our definition says that ir.ler.ngua,ge is regular if it is accc:ptrrrlhy some dfa. Bet:arrseof the equivalence of nfats ir,nddfa's, a language is also regulirr if it is rr,cceptedby some nfa. We rrow show that if we have arry rcgular expressiorrrr 'we (:irn construct an nfa that arx:eptsI (r). The constructiorr for this relies orr the recursive definition for I (r). We first construct simple automata for parts (1), (2), and (3) of Definition 3.2 on page 73, then show how they can Lrecorrrbirx:dto implement the rnore complir:a,tedparts (4), (5), and (7). Let r' bc ir regrtla,rexpression. Tlten thrlre exists some nondeterrrrirristir: finite accepter tlnt ir.{rceptsL (r). ConsequentlS I (r) is a regular language. Proof: We begirr with automata thal, accept thc ltr,ngua,gesfor the sirnple regula,r expressiorrso, .\, and fl, â‚Ź E, I'hese are shown in Figure 3.1(a), (b), and (c), respectivcly. Assrrme now that we havc irutomata, M(rr) and M (rz) tha,t accept larrguagcsrlenoted by regular expressions 11 and 12, rcsptx:tively. We neecl not explicitly r:onstruct these autornata, but may represerrttlx,.m sr:henra,tically,as in -Figurtt3.2. In this schema, the graph vertex at the lcft representsthe initial state, thc one on the right the firral state. In Exercise 7, Srxrtion 2.3 we claimed that frrr every nf'a there is an cqrrivalent one with a sirrgkl fina,lstate, so we lose rrothirrg in rr,ssumingthat tlxrrr: is only one final statc. With Jl,f(r1) and M (rr) rt:prcst,'ntedin this way, wo then construct autorrrata frlr the regular expressiorrs,r.1| 12, r1t s, and rf. The constructions are showrr in Figures 3.3 to 3.5. As irrrlit:a,ted


3.2 CoNNEcttot'r BnrwpnN RncuLaR ExpREssIONs Arrrn RuculaR

Figure 3.1 (a) rrfa accepts g. (b) nfa accepts {tr}, (c) nfa accepts {a}.

Figure 3.2 Schematic representation of an nfa accepting Z (r).

6)

(e)

M(r)

Latqcuecns

79

(c)

'/-_\. t

,\_./,

)

in the drawings, the initial and final states of tlte constituent machines lose their status and are replaced try rrew initial and final states. By stringing together several such steps, we call build automata for arbitrary complex regular expres$ions. It should be clear from the interpretation of the graphs in Figures 3.3 to 3.5 that this cotrstruction works. To argue more rigorously, we can give a formal method for constructing the states and traruitions of the combined machine from the states and transitions of the Jrarts, then prove by induction on the number of operators that the construction yields an automaton that accepts the language dettoted by any particula.r regular expression' We will not belabor this point, as it is reasonably obvious that the results are always correct. I

.L (r), where

, : (o+ bb)-(ba.+ A).

Figure 3.3 Automaton for L(rt * rz).


80

Chopter 3 Rncul,c,n Ler-rcuacns ar'ro Rnculan Gn-q,rr,rrr,raRs

Figure 3.4 Automaton for L (rrr2)'

Figure 3.5 Automaton for 't (ri).

Automata for (a * bb) and (ba* * A), constructeddirectly from first principles, are given in Figure 3.6. Putting thesetogether using the construction in Theorem 3.l., we get the solution in Figure 3.7 I

Figure S.0 (a) Mr accepts L (a + bb). (b) Mz accePts f, (ba. {- tr).

d

,

r-\t fln ,-]\ *\*-/*\--,rV (a)

Figute 3.7 Automaton accepts .t ((a * bb). (ba- + I)).

M

!

M

r /--\

b

z

1-J 6)


3.2 CollruncrroN Bnrwnnr-r Rncur,AR ExpRpssIols arul R.nculeR LANGUAGE$

81

for RegulorLonguoges RegulorExpressions It is intrritivt:ly reasonable that the rxtrrvurst:of Theoretn 3.L should hokl, and thrr,t for clvtlry regular language, thero should exist a corresponding regular exprussiort. Sitrce any regultr,r lir,nguagc:ltas an a"ssociatednfa a,nd hence a transitiotr graph, all we neecl to rlcl is to lind a regula'r expresuitlrr captr,hlcof getrerating the la,bels of all thc walks from q0 to anY fintr,l state. This dot:s not Iook too clifficr.rltbrrt it is corttplicated by the existerx:c of cycles thaL catr oftetr be travt:rsc:tlarbil,rarily, itr any order. This creates a bookkeeping probklrn that rnusl be hanclled carrcfully. Tltere are several ways to do this; one of the tnore intrdtivtt approaches requires a sidel trip into what are called generalized transition graphs. Since this idtlir is used h<":rein a limited way and plirys rro role in our further dist:trssiorr,we will deal with it informally. A getreralizedtransition graph is a transition graph whosc edges aret lub*l 'Ihe label of a,rrywrrlk from the initial sta,te to ir lirral state is sition graph. tfi---e concatenation of several regular expressions, and hent:tl itself a regular I expression. The strings denoted by such regular exprtlssiotrs are a subset of the language accepted by the generalized transition graph, with the full languagc being the union of all srr<:hgenerated subsets.

Figure 3.8 represerrts a generalized transitiorr graph. The language acrxlpted by it is L(o* + o* (o+b) c*), as should be clear ffom an inspet:tiorrof the graph. The edge (qo, go) Iabeled n is a cycle that can genertrtc any trumber of a's, that is, it represents L(o.). We could have labelad this edge a* without changing the language accepted by the graph. I The graph of any nondeterministic finite accepter (:an be considered a generalized transition graph if the edge labels are interpreted properly' An edge labeled with a singlc symbol a is interpreted a^san edge labeled vritlr the expression a, while atr edge labeled with mrrltiple syrnbols a,b,... is interpreted as an edge labeled with the expr()ssioll a f b * .,.. Flom this observation, it follows that for everv regula'r latrguage, there exists a

Figure 3.8


Chopter 3 Rncur,en LANGUAGESaNo R.pctll,.cR GRAMMARS

e

Figure 3.9

)

t

,T

a, L

j

"

\

J

(a)

generalized transition graph that accepts it. Conversely, every language of iccepted by a generalizecltransitiolr graph.is,reqular.. Siriffiiabel Tvery w-at[-In a genetralized transition graph is a regular expressi6n, this appears to be an immediate consequenceof Theorenr 3.1. However, there are some subtleties in the argument; we will not pursue them here, but refer the reader itrstead to Exercise 16, Section 4.3 for details.

tai

a,.1,aJ 1

want to create an eqiiiVEittint'f6neralized transitiorr graph with one less state by removing g. We can d0 this if we do not change the language denoted by the set of labels that can be generated as we go from q0 to q.f. The corntructiorr that achieves this is illustrated in Figure 3.9, where the state q is to be removed a"rrdthe edge labels &,b, ...stand for general expressions'

th@ ntosfi'e"eiili" tlie sense' rhe casiae------fiiEteilETfie to all three Yg$!c9! qiL+j,(l: Irr ca^reswhere an edge is missitrg in (a), wtr ornit the corresponding edge in (b). The construction in Figure 3.9 shows which edges have to be introduced so that the language of the generalized transition graph does not change when we remove q and all its incoming and outgoiilg edges. The complete process requires that this be done for all pairs (qi,Si) in I - {q} before removing q. Although we will not formally prove this, it can be shown that the construction yields an equivalent generalized transition graph. Accepting this, we are ready to show how any nfa can be associated with a regular expression. Let tr be a regula,r language. Then there exists a regular expreti$ion r such that .L: L(r). Proof: Let M be arr nfa that accepts .L. We can assullle without any loss of generality that M has only one flnal state and that q0 f F'. We interpret the graph of.il4 as a generalized transition graph and apply the above construction to it. To remove a vertex labeled q, we use the scheme in Figure 3.9 for all pairs (qt,qi).After all the new edgeshave been added,


3.2 CoruuncrroNBnrwporu Rnculnn ExpRpsslot-rsAND REGULAnL.c.Ncuncps

Figure 3.1{)

83

r4

n

rl

rr

\

/

-------},f\{ / lt-,,r' *\,1-l.xcr )) .-------'='-/ \+-J/

(rn (t-(rq +r-r r,rr '*)

q with all its incident edge$ can be removed. We continue this process, removing one vertex after the other, until we reach the situation shown in Figure 3.10. A regular expression that denotes the language accepted by this graph is

r : rlr2 (r+ + rsrirz)-

(31)

Since the $equenceof generalizedtransition gra.phsa^reall equivalent to the initial one, lve can prove by an induction on the number of states in the generalizedtransition graph that the regular expressionin (3.1) denotes L . r HxsmplC fl"S

generalized transition Considerthe nfa in Figure3.11(a).The corresponding graph after removalof state q1 is shown in Figure 3.11(b). Making the identificatio[ ln1: b + a,b"a, rz : o,bb,TB* H, T4- af b, we arrive at the regular expression r : (b a ab*a)* ab*b(a * b)* for the original automaton. The construction involved in Theorem 3.2 is tedious and tends to give very lengthy answers,but it is completelyroutine and alwavsworks. I

Figure 3.11

b

n

n

b

){'--L-!,

t

a,h

A

a9'--__-J-9-"--\?

l

a+b

b+abta

0 F.

l"t ab*b A

-{ r' }----11 r';1 \---'l \+J/

o

,",

(b)


Chopter 3 Rnculan Laucuaces amo R.nculaR GRAMMARS

trnq'tmplut.t n

Firrd a regular expressionfor the language L-- {* e {rr.,b}. : ru,,(ur) is even and n6(ttr) is odd}. An attempt ttt construct a regular expression directly from this description lcacls to all kinds of difficulties. On the other hand, finding a nfa for it is easy as long ir,swe rr$evertex labeling effectively. We label the vertices with EE to dcnotc zrrl e,ven number of rr,'s and b'E with O;Lto detrote an odd this we eaJilil * ott]With ,rFumberof a's anr-an ,:.re.n*t*ne"Iffia*nd get the solFion in Figffe.3'fZ' We can now apply the conversion to a regular expression in a nre<:hanictrl way. First, v/e rcmove the state labeled oE, giving the generalized trarrsition graph in Figure 3.13. Next, we rerllove tlrc vertex labeled O0' This gives Figure 3.14' Finally, we applv (3.1) with

rr: aa+ ab(bb).ba, 'rz:b+ab(bb)-a, \ rc: I

\ r+:

b*a(bb)*ba,

!---+---'

Figure 3.12

Figure S.13

/"\+

d\Itl)

0,. .-


: i . 2 colnpcjrrohr Bp'rwpEN RnculaR Expn,ESSroNS eur R.ncur.anLeNcuecns

Figure 3.14

85

+ ab(bb)rba

b + a(blt).ba

e9

EE b + ab(bh)*a

The final expression is lorrg irrxl corrrplicated, but the way to get it is rela. tivcly straiglrtfbrward. I

RegulorExpressions for Describing SimplePofierns Irr Exarnple 1.15 and in Exercise ls, Section 2.J.,we explored the t:orrncx:ti6rr bt'twenn finite accepters and sorne of the simpler constituents rif prr.rgratrrrtirrg larrgrrages,such as identiflers, or integers and real numher$. Tlxl rclation bctweerr finite autorna[a and regular expressions mean$ thilt wc can also use regular expressions irfJrr wiry of describirrg these features. 'Ihis is eas.yto see; firr examplc, thc sct of all acceptable Pascal integers is clefined by the regular expression

f')r'q' stl,d*,

wherc.s staruls fcrr tlte sign, with possible values from {t,-,A}, ir,rxl r/ stands frlr the dicits 0 to L Ptrsr:al irrtcgers are a sirnple case of what is sometimes clallccl a "pat$crtt," a terrn that refers to a set of objects having $ome common propcrtics,

g{t-T,,tf a1q5brrg ref'ersttr gories. Often, thc kcy to successfulpattern nratching Lsfincling an effective way to describe the pattertis. This is a cornplicir.tccl ancl exterrsive area of computer scienut to which we calt only briefly allude, The example below is a simplifitrd, but rrcverthelessinstructive, dernonstration of how tho irluas we ha,veta.lkNl attout stl far have been found useful in pattern mtr,tchirrg.

rrir"tt:lrirrgoccurSintextecliting.Alltexteclitors allow files to ht: st:arrrrcdfor the occurrerlce of a given strin65; mclst rNlitors c:xtt:rrd this to penrtit searching for patterns. For extr,mple,tlrrl eclitor ed in

the UNIX opera-tingsystenrreqg$ll?fl jlq jgl[t4;I*

iry" Ll"t'x "t)


86

nNu llecuI,an GRAuneRs Chopter3 Rr;cur.enLANGUA,GI;s

as eln instruction to sea,rchthe file for the flrst occurrence of the string ab, followed by arr arbitritry ntrmbtlr tlf tz'tl, f'ollowed hy a c. We see from this example tltat the UNIX editor can recogtrize regular expressions (although it uses a sornewhat different convention for specifying regular expressions than the one used here)' A challenging task in suclt ar} applicatioil is to writc an effici<rrrtprogram for rec:ognizingstring patterns, Searching rr file for occurrellces of a givcn string is a very simple progranming exercise, but lterc tlie situ.rtion is more complica,tecl. We have to deal with an unlitnited rrurrrber of rrrbitrtr.rilv rrlmplicatefl pa,tterns; furthertnore, the pattertrs are not fixed beforeharrd, but created at mrr tirnc:. Tlxr pa,ttern description is part of the input, so the recognition pt:ocessmust be flexible. To solvc this probltrm, ideils from arttoma'ta'theorv a,reoften used. If the pa,ttern is specified by a regular expression, the patterrr recxrgrrition prqgram can take tltis description and cotlert it into arr cc|rivalerrt nfh, lsing the constructiorr in Theorem 3,1, I'heoretn 2'2 rrray therr be ust:tl ttl recluce this to a dfa. This clfa, in the fomr of a trtrnsition tatrltt, is eft'ectively tltr pl,ttern-ma,tching a,lgorithrn. AII the progranllllel has to do is to proviclt) a rlrive,r that gives the general framework for using the table' Irr this way we carl autotnatically }tarrtlltl il lilrge number of patterns that are defined at run titne. Tlxl lffit:iunr:y of the progra,m must be considered also. Tlte constructiorr of finite autorrrataI'rornrtlgrrlar tlxpression$using Theorems 2.1 and 3,1 tettds to yielcl autorrrirta with rntrny states. If memory space is a problem, the state reductiorr rrrethotl rlt:st:ribedin Section 2.4 is helpful. /

T

l

llse the conrrtrrrction in L (ub'aa, I bln,*a,h).

'l'heorem

3,1.to find an nfa that arcepts the language

2. Fiflcl aII rrfa t]rat accepts the complement of the language in Exercise 1, @ [1)

Ciu* an nfa that accepts the larrguageL((a+

b)- b(o + bb)")' ffi

Find clfa's that accept the Ibllowing larrguages. (a) r(aa."+-aba.ti.) ffi (b) t (ab (a * ab). (a * aa)) (c) I ((obab)" * (aaa. f b).) @) r (((aa.). b)-)


3.2 cor'rNncrtor'rBnrwnnrq R.nculnn Expnnssror-rs ehrr R.ncuLaRLANGUAGES 87

5. !'ind dfa's that accept the following languages,

(a) L :

L (.ab*a*) U I ((ab)- ba),

(b) L : L (ab*a*) n r ((ab)- tu).

6 . Find an nfa for Exercise 15(f), section 8.1. use this to crerive a regular etpression for that language, t .

Give explicit rules for the cons-truction suggested in Figure 3,g when various edges in 3.9(a) are missing, ffi

-., f$,/Consider the following generalized transition graph.

(a) Find an "^equivalerrt gerreralized transition states. ffi

graph with only two

(b) What is the languageacceptedby this graph? W

".'--* qLlwhat

langr.rageis accepted by the following generalized transition graph?

a+b

a+b'

aab+t

Q0r)Find regular expressions for the languages accepted by the following automata,


88

Chopter 3 Rlculan

AND RECULAR Gnauuil,Rs LANGLTAGES

(")

0

11, Rework Exarrrple i3.10,this tirrre elinrinating the state OO first' (@

f itta a regular expressiorrfor the following languages on {a'b}' (u) I : {rrt : no (ru) and rr,6(rrL) are trot}r everr} (b) , : {tt : (n,,,$u) - nt, ('u)) mod it : 1} (.) ,: { t n : ( n o ( t n ) n ' 1(,T a )r)n o d 3 l 0 } (d) 1, : {ur : 2n* (u) + :3n6(u.') is even}

1 3 . !'ind a regular expression that generates the set of all strings of triplets defining correct binary addition as in Exercise 23, Section 2'l'

r4.

Prove that the constructions suggested by Figrrre 3.9 generate equivalent generalized transition graphs.

1 5 . Write a regrrlar expression for the tret of all Pa^scalreal rrumbers' 1 6 . Find a regular expression for Pa^.rcallrets whose elementsrare integer mrmbers' ileed t T . In sorrre applicatiOns, srrdr a-sPrograrrfr that cheCk spelliilg, we Inay rtot notion the Once otre. pattern, a1 approxirnate only t[e of arr exact rrratch


3,3 Rrcur.nrr,Gir,alrlr,q,ns

89

of an approximate matt:h has been madc precise, autornata theory can be applied to corrstruct approximate patterrr rnatchers. As arr i llustratiorr of this, consider patterrrs derivcd from the original orres by inscrtion of orre syrnbol. Let 1, bc a, regular larrguage on }J and define 'i,rtsert(L) *

{uau: a E X, utt â‚Ź L} .

ln effer:t, i'rtsg'L(L) contains all the words crcated from tr by inserting; a spuriorrn^syrrrbol anywhere in a lvrrrtl. * (a,) CJiverran nfa for 1,, shr:w how orre can construct an nfa frrr rirr.serl (.L). ffi **(b)

LJiscusshow you might use this to write a pattern-recognitiorr program for i,n,ser,t(tr), using a,sinprrt a regular expression for .L.

* 18. Aner'logousto the previous cxcrcise, corrsi<lerall words that can be forrnetl frorn 1, by dropping a single syurtrol tif the string, lbnnally rlefirre this operation rlrop for lang;uagcs. Ctustnr<:t arr nla lbr droTt(L), given an nIa for .L. Usc the conrrtnt:tion in Theorem 3.1 to find nfa's for L(aa) arxl -L(O*). Is the result corrsistent with the defirrition of thesc la,nguages?

RegulorGrqmmqrs tffiffiffi A third wa.y of describing regrrla,rlariguagesis hy means clf certain simple grer,mrniirs.Grarnmars a,reof'tr:il ail alternative way of specifying languages. Whcrrrlver we define a, ltrrrguage family through irrr automaton or in sorncl other way, we err()irrterested in knowing what kind of grammar we can associatewith tlrrl farnily, First, we look at grarrrmarsthat gurxlrate regular Iangua,ges.

Right-qnd Left-LineorGrqmmqrs

A grirrrtrnar G : are of the fbnn

(y,T,S,P)

is said to be right-linear

if a,ll prorluctions

A -- rB, A --+ it:, wlrere A, B E V, a,nd l; E 7'+ A gra,rnrnilris said Lo be left-linear prloduct,ionsa,reof tlu: lirlrl A -

I1t:,

if all


90

Chopter 3 Reculln

GRAMMARS LeNcuacrs ann R,rc;trl,q,R

or tt,

AA regular

gramfirar

is tine tha,t is either righl-liilear

or lcft-lirreirr.

Note that in a regular Hrirrnma,r, a,t nlost one varia[le appt)ilfs tl1 the tllsistent'lv right sicle of any productiorr. f'\rrtheilnore, thet Erlablgjlxsl

bir eitherthe rightrnostor leftmostsy"tbg]gi$Effi tion.

..| .$N$iN$Ns[liEihi.-"egramma,rGr:({s},{o,l,},S,Pr),withP1givenas S-

rtbSla

is right-linear. The granunar G2 : ({S,5r, Sz} , {o, b} , S, fz), with producticlrrs S - Stab, Sr -r Srnbl5zo '52 +

ft''

is lcfb-linear. Both Gr and G2 are regula.rgramma,rs. The stltpttntle: S+a,h9+ahab9+rtbaba is a clerivatiorr with Gr. Flom this single instance it is ea"syto conjecture that .L (G1) is thc lirrrguageclenotedby the regular exprtrssionr =. (ab)- a. In a sinilar way, we tlirn $eethat L (G2) is the regular language I' (aab (ob).)'

I

,i.lti$[$

, BI , {o,,t} , S',P) with producti''s '5-,4. A - u.Bl.)., B-Ah. is not rcgrrlrr,r.Although everyproduction is either in right-linearor leftIinear fbrrn, the gra,mmaritself is neitlrt:r right-linear nor left-linear, and


il.lJ Fl,ncul.q,n GnAr,IrueRs

9l

tlrt:refore is not regular, The gramnlirr is *rr .x*r'ple of a linear A litrear granrmar is a grammilr irr which at most, one iEliiElEorrr (xr(:rrr on tne rlght $Idetrtf arry ]1lp,4tletjo-Lwithout restriction orr tht: llositiorr of *+-"-#. +':**-::-*:* ttris varia,bte. cleir.rly, a,fcsular qratunalh_alwavslilear, tlrt rrot all linear gl'aloqrar$ arreregqtal_ I (Jrrr rrcxt goal will be to show tha,t regulrrr grarrrrnars are associa,tecl with rcgulirr larrguagesand that fbr every regulirr language there is a,regula,r grirnrrrrirr. Thus, regular grammars a,rea,nothcr wrry of talking about regu)a,r Iirrrgua,ges.

Right-Lineor GrommorsGeneroteRegutqrlonguoges First, we show tlmi a language generir,tcdliy ri right-linear gra,rnmar is tr,lwir,ys rtlgulil,r. To do so, we coustruct tr.nnfir that ruirnics the clerivations of a,riglrtlirrear grammar. Note that ther scntcrrtial forrns of a right-linea,r grirrnrrriir lul'e the special fonn in whic:h thcrc is exactly one variable a,nclit rir:<:1rsas l,he rightmost symbol. Srrl4xlsenow Lhat we have a,step in tr rlr:r.ivation a b - . . c D+ a . b ' . ' r d E , a,rriveri tr.t try usirrg a production D + rIE. The corresponding nftr <:rr.n imitatc this step lry going from statc D to state E w]:err a, svmbol d is r:nrxlurrtered.In this scheme,the stir.trlof the automaton corre$ponclstcl t|e variable in the sententia,lftrrm, whilc the part of the string a,lreaclypr'cessecl is identical to the terminal prclix of the sentential form. This sirnple icleais the basis for the fbllewirrs theorem.

LeL G : (V,T,S, P) ht: * riglrt-lirrear grarnmar. Then tr (CJ) is a regular ranguage. Proof: We assumethat V = {V0,y1,...}, that s: Vo, a,nclt}ra,twr: have of the form Vs -i r)tvi1V - uzVj,,,,or V, -1 ,t)t,t... If .rois Jrrodrrctiorrs a stritrg in tr (G), then btx:inrstrof the lorm of the proclu:tiorrs in G, t]re derivation must have thc fbrrrr V o1 u t V , 1

*

tt'ttJ,tV.;

__

4I)11)2-..UpV71 4

1 1 1 1 ) 2 '"' u k l J l = ' I t J .

(3.?)

rhe automaton to he t:orrstnrcted will reprocluce the rklrivation by ,,consuming" ea,chof thtlstl u's irt turn. 'Ihe initial stir,tc of the autornaton will


92

Chopter 3 Rncuren Lencuacns .lrun RpcuLRR GRAMMARS

btl labeled Vs, and for each variable Va there will be a nonfinal state labeled Vl. For ea<'hproduction f i - - +a 1 a 2 " ' a * V i , the automaton will have transitions to connect V.i and Vi that is, d will be defined so that , r a z '' ' a * ) : V i ' d . ( Y o& For each prodrrction V'ataz"'ftrntt the corresponding transition of the automaton will be 6 * ( U , & t & 2" ' g , r ) : V y , where V7 is a final statc. The intermediate states that are needed to do this are of no concern and can be given arbitrary labels. The general scheme is shown in Figure 3.15. Thc mrmplete automaton is a,ssetnbledfrom such individual parts. Supposenow that w e L (G) so that (3.2) is satisfied' In the nfa there is, by construction, a path from V6 to V; labeled u1, e pa,th from V; to Vi labeled u2, innd $o on, so that clearly V1 â‚Ź d* (H6,'u.r) , and tl is ar:cepted by M. Corrversely, ir*uuumethat trr is accepted by M. Because of the way in which M wa,sconstructed, to accept ur the automaton has to pass through a seqqenceof states Vo,Vr,.,. to y/, lr$ing paths labeled u1,a2,'... Therefore, rr rmrst have the form 'ut = ltt\1'2"'LlnUt

(\-_-/ F A n , ) \ -./

ReprcscntsV|+ arar... a^V.

fl=-6 \J \7 RepresentsV;+

arar... d*


3.3 RnculnR GnauueRs

93

Figure3.16

and the derivation V " + u t V 4 u 1 u 2 V i3 u p 2 . ' - u n V n4 u 1 u 2 .. . . t ) 7 y , t ) 7 is possible. Hence ur is in I (G), irrrd the theorem is proved.

to'that

I

accepts the languagr: generated by the

grammar V(t -

aVt,

V -

o,bVolb.

We start the transition gra,ph,with vertices Vo,Vt, and VJ. The first pro* drrctiott rule creates an edge la,l,releda betwecrr tr/sand trzr. Fbr the se<;ond rule, we need to introdurxl att additionnl vertex so tha,t there is a path Iabeled ob between lt and Fo. Finally, we neecl to rrdd an edge labeled b between Vr and Vy, giving the autornaton shown irr l'igure 3.16. The language gerrerated bv the gramma,r ir,nd accepted hy the automaton is the regular la;rguage L ((aab). ab). I

Right-l-ineqr Grqmmqr$for Regulorlonguoges To show that every regular langua,get:irrr be generattxl by some riglrt-linea,r grammar, we start fronr the dfa fbr thc language ilnd reverse thc construction showrr in Theorem 3.3. The sta,tcsof the dfa now become the variahles of the grarnmar, and the symbols causing the tra,rrsitions hetxlrne the terminals in tlte productionu.

If .L is a regular language on the a,lphabetE, then there exists rr right-linear g r a m r n a rG : ( V , t , 5 , P ) s r r c ht h a t L : L ( G ) .


g4

Chopter 3 Recul.+n LANculcEs ANDREGULAnGnaltltaRs

(Q,t,d,qo,F) bc a,dfa that ac:r:epts,L'We assumethat Proof: Let M: : Q { q o , q r 1 . . . t q n }a n d E : { a t , a z 1 . . . 1 a m } . C o n u t r u c t t h e r i g l r t - l i n e a r grarunar G : (V, E, S, P) with v = {qo,Qt,"',Q,r} and ,5 : gu. For each tra,nsition il (qt,oi) : qn of M, we put irr P the produr:tion qi + a.jqh.

(3.3)

In addition, if 96 is in F, wc add to P thc productiott {r

--+ .\.

(3'4)

We first sltow that G dcfined in this wir,v can gelrerate every strirtg in ,1,. Consider u â‚Ź L of the fortrt w :

(].i,aj"'anilt,

For &/ to acccpt this string it must make moves via d(qo'at): qo, 6 (q, ai) : e,,

d ( q " ,a n ) : q t , 6(qr,at):qy e F' By txrnstructiorr, the gralnllrar will have orttt productiort ftrr each of tlxrse dts. Therefore we cirn make tltc derivatiotr "'(LnQt. QoI ut,ep4 aiai4r I aiai 1 u , ; , a.,'.'ia k a r q l 1 a i a i , ' ' ( I k & r , ,

(3.5)

with the granrlrlar (J, and u E L(G). Conversely,1f u e L (G), tht:n its derivation must havc the form (3'5). But this implies tliat ii* (qo,ataj ' ' ' apal) : Ql t complcting the proof.

r


3.3 It.[]cuLAnGRnnrnraRs

I'igure 3,17

o)={sJ s(so,

lo -d(t

E(qt,,) = {92}

{t-Ett

E(q, b) = {qt}

9z-bg,

E(qyo) = {s}

lt

?.rEF

95

--aQJ

9t -L

For therpurpose of constructing a grammar, it is usefhl to note that the restrictiorr that M be a. tlfa is uot essential to the proof of Theorem 3.4. With minor modification, the sarneconstnr(:tion can be used if M is an nf'a. H$nuwtplo$',I $

Construct a right-linear gr&mmirr fot L (aab*a,). The transition firnction for att ttfa', togcther with thc corresponding gramrrrar prodrrctions, is given in Figure 3.17. The result was obtained by sirnply following the construction irr Thtxrrern 3.4. The string aabcr, can bc clerivedwith the constructed grarrrrnar bv Qo 1

(rh )

0,aQ2I

aahq2 =+ aabaqs + aaba.

I

Equivolence BetweenRegulorLonguoges ond RegulorGrommors The previous two thrlclrerns esta,bliuli the connectit:rr between regular languagcs and right-lincirr grammars. One can make a simila,r connection between regular langrragesand left-linttar gramma,rs, thereby showirrg the complete txlrivalence of rcgular gr&mmirrs and regular languages. A lanSirragetr is regrrlar if and only if there exists a left-linear ggammar G s u c ht h a t L : L ( G ) . Proof; We orrly outline the main idea. Given any left-Iirrear grammar with productiorrs of the fbrm

A --+Bu,

A+rj,


Chopter 3

Recrrlan

Lnuclul,cus ANtl RH(IuLAR GnRunlns

we r:onstruct from it a right-linear grammar G try repla,cilg every suc:h prodrrction of G with A -

uRB,

1{ _+ t.t&,

respectively. /

/^\\R

(, ("))

A few examples will ma,ke it clear quicklv that ,L (G) :

. Next, we useExercise12,Section2.3,whichtells us that the rc-

verseof any regular larrgrageis also regular. Sinced is right-lirrcrrt,/, (d) is regular.But thenso are.L((d))R,'nd,L(G)'

r

Prrtting Theorems 3.4 and 3.5 together, wc a.uive a,t the equivalence of regular larrguages and regular grarnmars. A Ianguage -L is regular if and orrly if there exists a regrrlar grammar G suc:h that L: L(G).

We now have several ways of describirrg regular languages: dfa's, nfa's, regular expressions, and regqlar gra,mmars' While in sorne irrstarxle one or the other of these may be most suitable, they are all cqually powerful' They all give a complete and unambiguous definition of a regular langnage' TIle connection between all tlrcse concepts is established by the four theorems il this c:hapter, as sltowrr in Figure 3.18.

Figure 3.18

Regularer<pressions

Theorem3.2

Theorem3 . 1

dfa or nfa

Theorem3 . 3

Theorem3.4

Regular Brammars


ll.3 Rncure,n.GRerr,rrunRs 97

1. Construct a dfa that accepts the language generated by the grammar S -

abA,

A + baB, fi + u.Albb. 2. Find a regular Blarnmar that generates the language L(aa* (ab + a)-). S. Constnrct a left-linear grarnrnar for the larrguage in Exercise 1, t

Lj,)

Corrstruct right- antl left-lincar grarrlrrrars for the larrguagc L : {atth"' :rr,} 2,rn > B}. ffi

5, Construct a riglrt-linear grailrrrrar lbr the langrrage L((aah.ab).). ', '

(O

f itta a regular grarrrrnar that generates the language on E : {r.r,,b} consisting of all strings wittr no more than three a's.

7. In Theorem3,5,provethat l, (e) : (f (G))8. W 1

Suggesta construction by whicJra left-linear grammar can be obtairreclfrom

G

an nfa directly. 9. Find a left-lirrear grammar f<rr thc language in Exereise 5. 1

l-inrl a regular grailrrrrar for the larrguagc L - {a"lt"'

@

tn I

mis even},

S

11. F'ind a regular gramrnar that generates the language L = {w â‚Ź {rr.,b}- t n" (w) * 3rr.6(ur) is even} . (t

Find regular grammals for the following languages on {a, b}.

G

(e) L : {ut : no (u) and nr, (zrr) arc both even} ffi '

(b) , : {w : (no (w) - no (to)) mod 3 : 1} (c) I - {w : (n* (ur) - ru6(,rrr))mod:l I 0} (d) ,:

\\

{tn: lno (ur)

- 1 1 6( t r ) l i s o d d } .

(l$y'Show that for every regular language not containing.\ there exists a rightv linear grammar whose productions are restricted to the forms

A-aB

A+tL,

w h e r e A ,B e V a r d a e T


98

Chopter 3 Rnculen Leucuac+psaFIIrRnculan Gn.q'r'tM.cn,s

{r+} \--'l

Sho* that any regular grarrrlrrar G for which L (G) + g must havc at least one procluctiorr of the form 4+r. whereA e V and r ET*.

15.

Find a regrrlar gramlnar that generates the set of all Pa"scal real numbers.

( r c . ) f e t G 1 = ( V r , E , 5 r , P r ) b e r i g h t - l i n e a ra n d G r = ( V r , X , S r , P ? ) b e a l e f t V lin*u. grammar, atttl a^ssumethat Vr and l/z are disjoint. Consitler the linear ({S}UyrUyz,t'.9,P), where.9is not in VtJVz and P: g r a r n m a rG : * -L (G) that is regular. $l# Pr Show u Pz, SrlSz} u {5


P r o p e r t i e so f R e g u l q r[ q n g u q g e s

e have defincd rergularlanguages,studied $om(]wirys irr which they c:arrtre represented, and have $eenir f'cw rlxirrnples of their usefillness. We now raise the qrrcstirlrr of how general regula,r la,ngrralgcs arc. Could it be thnt cvt:rv forrral language is regula,r'/ Rrrhaps irrry set we can specify c:a,nllct acx:epteclby some, albeit velry txlrnplex, finile automaton. As we wiII scc shortl.y,l,heatrswerto this r:rirrjrx:tureis cleliniLely no, But to understarxl why this is so, we rnust inqrrirc rnrlrtl cleeplyinto the nature of rcgulirr larrguagesancl see what propt:rtics the whole fa,rnilv ha,s. Thc fir'st que.stion we raise is wha,t ha,Jrpr:nswhen we perforrn operations on regular latrguages.The opera,tiririswcrcorrsiderare sirnple set opuratiorrs, such as concatena,tion,ns wc:ll as operations in whir:h etr.chstring of a language is chrrngrxl,a,sfrrr inslance in ltlxercise22, St:c:tiorr2.1. Is the resulting langua,gestill rtrgular? We refer to this a,sir,closure question, Closure propertios, tlltlxrtrgh tnostly of theoretica,lintorcst, ireltrlus in discrirnila,ting hetween thc various language fa,milieswe will crrt:ounter. A srx:orrtlset of questions a,bout ltrngrra,gc: frrrnilies deals with our ability to tltx:ideon certain properties. Frrr r:xirnple, cari we lell whether a lirrguage

99


Chopter 4 Pnorenrrns on RrculAR LANGUAGES

is finitc or not? As we will see,such qttestions are easily tr,nuweredfbr regular Ianguages, brrt are not as easily answered for other larrgua,gefamilies. Finallv we consider the important question: How can we tell whether given language is regular or not? If the language is irr fa<:t regrrlar, we a can always show it by giving some dfa, regular expressittn, or regular grammar for it. But if it is not, we need another line of attack. Ont: waY to show a language is not rcgultr,r is to study the geueral properties of rcgrrlirr Ianguagtrs, that is, characteristics that are sharcd bv a,Il regular languages. If we know of some such property, attd if we can show that the candidate languagc rloes not have it, then we catr tell that thc larlguage is not regular. Irr this chapter, we look at a variety of propt:rtieu of regular languages' These properticu tell rrs a,great deal about what regular languages can and canrrot do. Later, when we look at the same quttstions fbr other language families, similarities and differences in these propcrtics will allow us to contrast the various language families.

C l o s u r eP r o p e r t i e so f R e g u l o rL o n g u o g e s Oonsider the following questiorr: Given two regular languages -L1 atrd -L2, is their union also regular? In spec:ifit:instances, the auswer may be obviotrs, but here we want to addrcss thc problem in general. Is it true for all rt:gular trr and Lz7 It turns out that the arlswur is yes, a fact we express by saying that the family of regular langtragcs is closed under union. We catr ask similar qucstion$ about other types of operationu on languages; this leads us to tlte study of the closure properties of larrguagesin general. Closure properties of variorrs language farnilies utrder differcrrt operations are of considerable theoretical interrest. At first sight, it may rrot bc clear what practica,lsignificancethese propcrtieu have. Admittedly, sotne of Lhenr havr: vt:ry little, but many results are useful. By giving us insight iulo the general nirture of language families, clostrrc properties help us answer other, more prac:tic;illqrrestions.We will see ittstancesof this (Theorem 4.7 and Example 4.f3) lir,ter in this chapter.

Closureunder$impleSetOperotions We begin by looking at ther closrtre of regular lattguages urrder thc (:ommon set operations, such as ttnion and irrtersection. If L1 and.Lz are regular liulgtrages,then so are -L1U L2, L1 -tL2, L1L2,fi arrd.Li. We say that Lhe farrfly of regular languagesis closed ullder ulri()rl, irrtersection, concatenatioll, cornplcmentation, a'ttd star-closure. Proof: If .Lr and L2 tne regular, then tltere exist regula,r expressions rr a,nd ?'2'fl'Ild 12 such that,Ll : L(rt) and tr2 : L(rz)'By defilititlrl, rl *?'2' ?"1


,1.1 CLosuREPRoPIIR:I'IEs orr Rncur,nRLANGUAcEs

101

ri are regular exprtlssiottsdenoting thc latrguagestr1 U L2, L1L2, and -Lit and star-clttsrrreis respectively. Thrrs, closure unrler trrrioll, conca,tena,tiotr, inmedia,te. To show c:losureunder rcrnlllerneutation, k:t, M : (8, X,,)' 96,i-) be a 'Ihen ther tlfrr, dfa that accepts tr1. M : ( Q , X , d ,( 0 , Q - F ) accepls Zr. fnis is rather strariglrtforwa,r'd;we havc already slggcstcrd Lhe result irr Exercise 4 in Scrtion 2.1. Noter tltat in the defirritiotr of a' dftr., d* to be n total function, so that d* (qu,rrr) is deliled fbr all we a,ssrrrnecl tr € E*. Cotrsequentlyeitlrt:r d* (qo,u) is a final state, in which caseru € tr, o r t ) * ( q 6 , ' u )e Q - F a n d w e L . Dcrnonstrating cllosureunder intel$t:ction takes a little tnore work. Let (Q,x',Ir,{o,Fr) and Mz : : Lt L ( M 1 ) a n c lL z : L ( M 2 ) , w h t r r cM t : from Ml a,nrlM2 a combintltlau(P,X,dz,Po,.F2)are clfir,'s'We constnrr:t w h o s es t a t e s e t 8 : t o m a t r r r rL 4 : ( A-,'f , f , ( g o , p n ) , F ) , e x P consists / \ of lrairs (q.i,:lrj),atrd whosc trarrsiliou functiorr d is such thnt M is itr state (q,,f,.) whenever M1 is in sta'te qi a,rrdM2 is iu stir,tcPi' Ihis is:rt:hieved by taking

f ( ( q o , p r, )a ) : ( t r * , , p t ) , whencvcr

d1(q1,a) : {r a,nd dz (Pl n,): P' F is clefitreclas the set of nll (,Jo,pti),such thirt Qr e Ft a,rrdpi € Fz. Thcrr it is-r, sitrple matter to show Lha,t ur E tr1 f'l ,1,2if a,nd orrly if it is arx:cptecl by M. Consequently,trr fl 12 is regrrlirr. I

The proof of t:losureunder irrtt:rsectionis a gtlod exampltl tlf a coustruc:tive proof. Not orrly does it t:stablish the dcsircd result, but it also shows explicitly how to construct ir,finile acceptt,'rfor lhe intelrscctiouof two rc:gula,r la,rrguages.ConsJtructivepr:oofstlt:cur throughout this book; thcy are irnpurtirtrt because tlx:y give 1rs iniriglrt into the rrlsrtlts and ofttlrt serve a"s thtr strrrting point fix practica,l tr,lgtlrithms. Hert:, as itr tnany c:ases,there 4rr: sfiorter but nttrrt:ottstructive(tlr al least nrtt so obviously r:onstructive) argurrtettts, For tlosure uuder irrtrlrsection,we start witlr DeMorgan's la,w, Erpation (1.3), ta,king the complcrnetrt of hoth sicles.Then L1) L2:

Ltl

Lz


Chopier 4 Pnopnnrrns ol R.Bc(rr-AR Larqcuecps

for any languages -Lr and tr2. Now, if tr1 and L2 are regular, then by closure under complementation, so are Z1 and -L2. Using closure under union, we next get that Z1 UL is regular. Using closure under complementation once more, we see that

ffit:

Lttt Lz

is regular. The following example is a variation orr the same idea.

WowthatthefamilyofreguIarlanguagesisclosedunderdifference'In other words, we wallt to show that if .L1 and .L2 are regular, then L1 - L2 is necessarily regular also. The needed set,identity is immediately obvious from the definition of a set difference. namely Lr-Lz:Lrf\Lz. The fact that ,Lz is regular implies that Ia is also regular. Then, because of the closure of ragular languages rrnder intersection, we know that tr1 rf 12 is regular, and the argument is complete. I

A variety of other closure properties can be derived directly by elemen* tary arguments.

The family of regula.r languages is closed under reversal. Proof: The proof of this theorem wa^ssuggested as an exercise in Section 2.3. Here are the details. Suppose that .L is a regular language. We then construct arr nfa with a single final state for it. By Exercise 7, Section 2.3, this is always possible. In the transition graph for this nfa we make tha initial vertex a final vertex, the final vertex the initial vertex, and reverse the directiorr on all the edges. It is a fairly straightforward matter to show that the modified nfa accepts ruE if and only if the original nfa accepts u. Therefore, the modified nfa accepts .Lft, proving closure under reversaJ. t


4.1 CI,osuRr.: Pn,opnnrrnsor R,tcuu,R LAN(+rrAGEs 103

ClosureunderOtherOperotions In acldition to the standarcl operations on la,nguarges,orre ca,rl dclirre other operatiorrs atrd investigatc closure propertics for thenr. Thr:re are ma,ny srrr:h results; we select only two typical ones. Others are explored in the exrlrc:ises at tht: crrd of this sectitxr.

' 1i',, lil$finirt.lt'l't:|"ll#1 Supposc X and I a,realphabels. Then a,function h:}j-f* In words, a hornomrlrphism is a, substitution is tr,llcld a homomorphism. in which a single lettcr is repla,cedwith a string. Tlrc dourain of tlrtr function h is c:xtendedto strings in a,nobvious fashion; if 1r.):uLaz'..Qro,

thcrr h,(ut) : h,(a1) h (az) .. . h (a.) . lf tr is a, la,nguageon X, tlrcrt its homomorphic

image is defined a,u

h(L): {h(r) twEL}.

Sxsnple 4,1

Let X:

{o,,b}and f : {a,b,c} arrddefineh by h'(u):1,6, h,(h): bbc.

Tlren h,(n,ba): abbhca,h.Tlte homomorpltic ima,geof L : {aa, aba} is the Ianguagc h (L) : laha,b,abbbcabl. I If we have a regular exprt:ssion r for n latrguage I, tlten a, regular exprrrssion lctr h(L) carr Lreobtnirrcd by simply applyirrg the homomrirphism to each E syrrrbol of r.


104

Chopter 4

Pnornnuns op R,uctrt.ARLAN6SAGES

E: {a,b} ',rrrtt:

W,ke

{b,c,d}. Defirxrh hv h(a):46""' h (b) : 6,1,,'.

If f. is the regular latrgttagedenotedby r = (e * b.) (aa). , then 11: (tlbtc + (bdc)-) (d,bcctlbt:c)* rlenotes the regulirr lernguageh (L).

I

The generrrl result on the closure of regular languirges under any horrxlmorphistn folklws f'rom this exarnplc in an obvious lrlanner. Let /r,be a,holrorrtorllhism. If -L is a regular la,nguage,tlten its homomorphic irna,gcft,(I,) is also rtlgular. The fanrily ttf regula,r latrguagesis therefore closed rrrrrlt:ratbitrar.y }tottrorrurrphisms. Proof: Let ,1,trt: a. regular languag<l denoted by sonre regular expressioll r. We fincl h(r) by substituting h,(a) for each sytnbol o € E of r. It r:an be shown directly by an appeal to the definition of a regula'r expressiorr that the result is rr regular expression. It is equally ea^syto see that thc rcsrilting expressiorrdenotes h (r). All we need to do is to show that for ()very ru e L (r'), thc txrrresponding h (rr') is in L (h (r)) arrd converselythat for every u in L (h,(")) tlterc is it ttt itt .L, such that u : h(w).Leaving the details as an exert:ise, we claitn that h (tr) is regular. I

l|iM Let L1a,nd L2 be latrgurrgcson the sarnealphir,bet. Then the right quotient of tr1 with trz is cleflled rrs

Lt/Lz-

{r:ry

€ . L 1f o r s o m e y € t r 2 }

(41)

To form thr: right quotietrl of -L1with /12,w€ t,akcir,llthe strings in.L1 tha,t have a sufrx bclonging to tr2. .Evcry such striug, af'tcr removal of this sullix, trelongsto L1f L2.


4.1 Clostlns

Fn,opnRTrns ol' R,l:cjulan LANCUAGFT$

105

-

s'4 Hr+r{rtrplo

If L1 : {a"b"' : n } 1,rn > 0} U {ba,} and L2 :

{l|!I : m, } 1} ,

thcln l,t I L2 :

{att6"' ; rz }

I, ?n,> 0} .

The strings itt .L2 consist of orre or rrlore b's. lllhrlrt:fore, we arrive at the answer try rcrnoving one or rtrore b's from those strings in .L1 tha,t tr:rrrrinate with at least one l) as a sulfix. Note tha,t hcre L1, Lp, and Lt/Lz ale all reE;ular. This suggesto-that the right quotient of any two regular la,nguagesis also regular. We will prove this itr lhe next theorem by a, constnr:tion that takes the dfa's for L1 ir.rrrl.L2 altd constnrcts from them a, tlfa for Lt/Lz.Bt:ftrre we describc the r:onstnrction in firll, lct us see how it a.pplies to this exer,rnple.We start with a, dfir for l1i sa/ thr: automaton Mt : (Q,X,6,qu,F) irr l'igure 4.1. Since iln inrtotttaton for L1f L2 urust accept ir.nyprefix of strings in,L1, we will trv to rrtodify M1 s{r that it accepts z if there is any y satisfying (4,1).

Figure 4.1


Lnrlcue,cps Chopter 4 PRopeRuEsor.'R.Bcur,nn,

Figure 4.2

The dififrculty comes in finding whettrur there is some g such that ty E L1 nnd 37€ L2. To solvc it, we determille, f<rrca<:hq € 8, whether there is a, walk to a final sta,te labeled u sur:h tha,t u € Lz, If this is $o, anY r such that d(q6,2) : q will be in L1f L2. We modify the automaton a,ccordingly to make t7 a finnl sta,te. To apply this to our present ca^se,we check each state qq, Q1-,Qz, Qt, Q+, qb to see whether there is a walk labeled bh* to any of the q1, 92, or 94. We see that otrly 91 and 92 qualify; eo, Q3,94 do not. The resulting automaton for L1f L2 is shown in Figure 4.2. Check it to see that the construction works. The idea is gerreralized in the next theorem I

If L1 and L2 axeregular lauguages, tfu:n L1f L2 is also regular. We say that the family of regular languages is closed under right quotient with a regular la,nguage. Proof: Let Lt : L (M),where M : (8,8,,5,q0,F.) is a dfa. We constnrct u. follows. For each et. € Q, determine if ir,nother rfia fr : (A,X,li,qo,F) there exists a y €. L2 such that d* (qo,y) : q1 € F. This carr be done bv looking at dfa's M; = (Q,X,d,{i,F). The arrtomaton Mi is M with the irritial state q0 replaced by gi. We now determirre whether


4,1 Closunr: Pnopr:t{rlr:sor RucuLA.RL.c.I-tcuacns

107

there exists a E in L(Mt) that is a,irtoin -L2. Frrr this, we can use the construction fbr the interscctiorr of tvro regular languages given in Theorem 4.1, finding the transition graph for L2-tL (M1). lf there is any path between its initirrl verttrx arrrl any final vertex, then .L2 a L (M) is not empty. In tlrat case, acld qi to F. Repea,ting this f'or every et, E Q, we determitre F and thereby construct M. let r be arry elemetrt of L1f L2. Then To lrrove that "L (fr) : Lr/Lz, '' ' \ / there must be a, g € L2 such that rg e trr. This implies that d* (qo,ry) e F, so that there rnust be some q € Q such that d. (gu,r) : q alnd

d* (g,g) e F'. Therefore,by construction,q € F, and fraccepts u becaused* (go,z) is in F. Converscly,for any r acceptedby M, we have d * ( q u , r:)q e F . Rut again by rxrnstructiotr,this implies that there exists a 37€ tr2 suclt tha,td* (q,'a)e F. Thereforer3t is in L1, and r is in LtlLz.We therefore corrcludethat

,(fr): and fiom this that LtlLz

is regular.

LttLz, t

Lt:

L(a*batt*),

Lz:

L ( a b * )'

We first find a dfa tha,t trt:cepts tr1. This is easy, and a solution is given in Figure 4.3. The examplc is sitttple etrough so that we can skip the formalities of the r:onstrrrction. Flom the graph in Figure 4.3 it is quite evident that

L(Mo))L2:s, L (Mt). Lz : {o},1 s, L (Mz). Lz: {o} # a, L ( M : t )i L 2 : s .


Chopter 4

PRopnnrrEs oF REGuLAR LANGUAGES

Figure 4.3

Figule 4.4

\s'

Therefore, the automaton acceptin1 LtlLz is determined. The result iq shown in Figure 4.4. It accepts the language denoted by the regular exL1f L2: pression of a*b* a,*batt*,which can be simplified to a*ba*. Thh:us L(a"ba"). I


4.1 Cr,osuREPRopERTTEs or Recur,eR LANcuAcEs

l0g

1. Fill in the details of the constructive proof of closure under intersection in Theorem4.1, 2. Use the construction in Theorem 4.1 to find nfa's that accept (a) r ((a + b) a.) n.L (baa.), ffi (b) L (ab.a.) n L (a*b"a). s. In Example 4.1 we showed closure under difference for regular languages, but the proof was nonconstructive. Provide a constructive argument for this result. 4. In the proof of Theorem 4.3, show that h (r) is a regula.rexpression. Then show that h (r) denotesh (I). show that the family of regular languagesis closed under finite union and intersection,that is, if L1,L2,...,.L, a,reregular,then

Lu=

U

Lt

n

Le

; : 1 1 ,e , . . . , )n and

Lt:

d:{ 1,2,...,n}

are a"lsoregular. 6. The symmetric

difference

of two sets Sr and Sz is defined a,$

Sr $,Sz = {n t n E ,Sr or fr â‚Ź 52, but,r is not in both .5r and Se}, show that the family of regular languages is closed under symmetdc difference, 7, The nor of two languages is nor (L7, L2) : {w : w {, L1 and w f L2} . show that the family of regula,r languages is closed under the nor operation.

ffi

8' Denne rhe compremenrffjTl,T':;1."::.-", J",, ,.

-b

/-l

r

Show that the family of regular languages is closed under the cor operation.

(g/Whi"tt

of the following are true for all regular languages and all homomor-

phisms? (a) h(hu

r , . z :) h ( L r ) u h ( L {

*fv-*

(b) h (21 r-r.L2)= h(Lr) nh(Lz)

. F,r ls.

(c) h(LrL2): h(L1)h(L)

lr.-.-*r


oF REouLaRLat'Icuacns Chopter4 PRopuwrttrs

110

./.\ fto)t"t

\

FirxlL1fL2. and.Lr: L(abo'-). Lt: L(o,*haa*)

,'7.--'R

: LrLzlL:r is not true ftrr all languagesI't and Lz' \O "., fojjEhow that L t SuDDosewe krrow that .Lr U .Lz is regula,rand that -L' is finite. Can wc " {fi\ \\ U "on"tuae frotn this that Lz is regula.r'l ffi 13. If I/ is a regrrlar language' prove that L1 : {ua : u, E L,lul : Z} is also regular' 14. If tr is a regular language, prove that the langrtage {uu : u E L,u E Zr} is also regular. W

4il lI

t\

L,,,n//,(rq..

F)

The left quotient of a language -Lr with respect to 1,2 is definecl as LzlLl

: {g : r € L7,'r11F- L1],'

show that the farnily of regular languages is closed under the left quoticut with a regular language. "If 'Lt is regular n'rrtl trr utrz is also regular' then 1 6 . Show that, if the statement ^Lr must be regular" were true for all .Lr and .L2, thefl all languages would be

regular, Wffi L 7 . Theto,il of a laflgrrage is defined as thc set of all suftixes of its strings, that is tu,i,t(L) : {y : ny € -L for some r € t- } . Show that if .L is regula,r, so is toal (l). 1 8 . The head, of a language is the set of all prcfixes of its strirrgs, that is, h.eo.tl(L) - {n : ny e L for solne gr€ t.}

.

Show that the family of regula,r languages is closed under this operation.

ffi 1 9 , Defirte an operation thi,rd on strings arrtl languages as t h i r d ( a n z o " t a 4 a ! , t t ( ; ' ')' : a s a a ' ' ' with the appropriate extension of this defiflitiorr to languages. closure of the family of regular larrguages under this operatiotr.

20.

Prove the

Forasrring(rrtrz...^,:::,;nt,:,,:TT:Trutr,:, Flom this, we can define the operation on a language as shift(L):

{u:'u : sli,ft(ru) for some trr € .L}.

Show that regrrlaritv is preserved utder the.shilt

\\ c90-0"'

e n c h a n , gc ( . a t a z ' " t t n - 1 I t ' ' ) :

operation.

Q r y t , t 7 z ' 'f'tn - r t r 7 ,

aI]cI 'u exch,ange(f) : {u : a : exchange (Tr.,)for some e -L} ' Sbow that the family of regrrlar langualges is clcisetl under euchan'ge,


4,2 Elnvnrvrnrr,v QrrusrroNsABour Rncur.eR LANLITJA#E$ 111

* ??. The shulfle of tu'o languages ,r

an(l Zz is rlefinetl as

shuffie(L1, L2) : {w1u1utzuz,. .lr}6?}n1 i wtwz...wm € Lt, u1u2.,,1)m € Lz, for all ru.l,ul € X*) show that operation.

the farnily of regular languages is crosed under the sh'ffie

* 28' Define an operati.rlmi'nuss on a language .L as the set of all strings of tr with the fifth symbol frorn the left rcmoved (strings of length less than fivc are left unr:harrged). Show that the family of regrrlar larrguages is closed under the rruinusSoperation, * 24. Definc the operation le-ftsi,tteon .L bv Ieftsi,de(t):

{?,,

,ruruq e L} .

Is the family of regular languages closed under this oper,ation? 25. The m,in of a language .L is defined as rni.n(L):

{tu e l:

t } r e r ei s n o u E L , u E E + , s u c h t h a t u r : . r o } .

Show thal, the family of regular larrguages is closed under the rnriruoperation.

,/_-\ 2.+\39Let -

Gr and Gz be two regulargrammars.Showhow()rrecanderiveregular

glrammars tbr the larrguages \a.l

L(Gt)r L(Gz) ffi

(b)

L @r) r, (Ga) ffi

((rJ L (Gr)-

ffi

E l e m e n t o rQ y uestiono sbout R e g u l o r[ o n g u o g e s wc rrow conre to a very fundamenttrl issue: Given a language ,L and a string to, can we determine whethel ol not ,ar is an elemcnt of .L? This is the membership q*estiorr and a method fbr anuwrlrirrg it is called a membership algoritlmr. very little can be done with languages for which we carrrrot {ind elficierrt niembers}rip algorithms. The question of the existence arrclnature of rrrerrtbership algorithms will be of great concern in later cliscussiorrsl it is a;r issue that is often difficult. trtrr regular languages, thorrgh, it is an easy lrratter, We first consider whir,t cxactly we rnearl when wet say ,,given a language...." In ma,ny a,rg1r'rr:rrts,it is irnportant that this be unambiguous. we have used several ways of describing regular languages: informal verhal


tt2

Chopter 4 Pnoprnuns or Rpcul,nR Lerucu.qc+l:s

rlcsr:riJltiqns,set nqtation, finite altomata, regular expressions,and regrrlar grarrrrrrars. Orrly tlxr la,st three are sufficiently well defined for use in theorerns. wc thcrcfcrrc sa,y tha,t a regnlar language is given in a standard representation if arrd orrly if it iiJ dQst:rihedby a flnite automaton, a regular expressiotr, or a rt:gtrlir'r grilrnfiIirr. Given a, standard representation of any regular larrguagc tr on x and any u â‚Ź X*, there exists an algorithm for detertninittg wltether or nttt ru is in I,. 'ur ttl stNrif it is Proof: We represent the lalguage by some dfa, tltetr test accepted by this automtr,ton. I Other irnportant questiorrsarc whether a language is finite or itrfinite' whether two languages are the same, attd whether orrc langrttr,geis a subset of another. For rtlgular langrtagJesat least, these questions are easily arrswcrtld. given there exists an a,lgorithm fbr determining whether a reglrlar lirrrgrtir,ge, representa'tion,is etnpty, Iitrite, or infirritc. irr sta,ndir,rcl Proof: Tlrg itrrswt:r is a,ppa,rentif we represent t,he latrguage as ir trirrr,qititln graph of a dta. lf there is ir,sinrpkr pnth fiom the initial vertex to any firLal is ttot enrpty. velrtex, then the la,ngr-tage To dt:termine whethel or ttol a,Ianguage is infinitc, find aII the vertices that are the base of sorlrecyclg. If irlY of these a,reon a path from au initial to a final vertex, the lattgtttr,gcis irrfirritc. Otherwise, it is finite. I -The question of the eqrrality of two languages is also an importailt practica,l issue. Often several defilitions of a programming language exist, and we need to know whether, in spite of their dificrent irppearances, they specity the same language. This is gelerally a difficult problertt; cvtrn ftrr regula,r larrguages thel ilrgument is not obvious. It is 1ot possible to arguc on il serrtence-by-sentencecomparison, sirtce thiS works only fbr finite languages. Nor iu it ea-$yto see the answer by looking at tlte regular expression$' gramma.rs, or clfa's. An elegant solution uses tfue alreatly estahlished closure propr)rties. Given standard representatiols of two regular lalgtrirgils -L1 iilIrl tr2, there t:xists an algoritlrm to determiue whether or not Lt: Lz. Proof: Usirig .L1 irntl -L2werdefine the la'ngua,ge t4:

( L 1 . 1 6 ) | ( t ' t n t ' z ).


_____-----

4.2 Er,nttnltrnnv Qun$.r.lr)Ns AIlotTTR.BculeR LnFrL+r.J.Ac.;r{s 118

Bv closure, -t3 is rcgular, arlcl we ca,nfincl a dfa M tha,t accepts 1,3. Once wtl httvc: M we can theu use the a"lgoritlrrn in I'heorern 4.6 to cletsrrnirre if tr;j is ettrpty. But from Ex(n:is(: 8, Section 1,1 we see thirt Lt : fr if ancl trnl.yif Lt : Lz. I

These resrrlts are fundamental, in spite of br:irg obvious ancl unsurprising. For tlgular languages,the questions raisrld by Theorems 4.b to 4.2 r:an bc zrnswcredeasily, but this is not ir,lwir,ysthe case when we deal with larger farnilies of langrrages.we will encounter cluestionslikc these on several occasions la,ter orr. Anticipating a, Iittle, wc will see that the answers ber:gr1(: irrcrcasirrgly rnore difficult, tr.nrlrlvcrrtually irnpossible to fincJ.

For all the exercises in this strtion, itr stanclard representtr.tion.

assume tha,t regrrlar lir,rrguages are given

1. Show that there exists irn algoritlrnr to determine whether or not ur E Lr - Lz, for arry given to and any regr.rlar larrguages.Lr arlcl /.1. ffi 2. Show that there exists ari algorithm for determirrirrg if tr1 C 1,2, for any regular languages Z1 arrd Zl. ffi 3, show that there exists an algorithm for cleteruining if A e ,1,,for any regular language tr. 4. Show that for any regular Zr and.L2, there is an algoritlrru {,o dctermine wlrcther or not ^L1: LtlL,t. 5, A larrguageis said lobe apa,l,irtd',rne languagc if ,1,: tr/i. Fi'd an algorithm for dctermining if a given regular language is a Jralildrorne language. ffi 6. Exhibit an algorithm for tleterrnining whcther or not a regular language .L contains any string rrr such that tuF € 1.. 7. Exhibit an algorithtr that, given any three regular languages, L,L1,Ls, termines whether or not L : I,tLz.

de-

g. Exhibit an algorithrn that, giverr any regular language .L, tletenninerr whcther or rrot .L : .L*, g, Let tr be a regular langrrage orr X arrd,fi bc any string irr E*. Find an algorithrn to dctermine if "L coltains any tu such that fi is a substring of it, tlrat is, such that w : ufru, with u, ? € X*. 10. Show that there is an algorithrn to detemrine if 1, : shuf f Ie (tr,,L) fbr any regtilar L. 11. The operation tnil(L)

is defincd as

tail (L) : {u : uu E L,u,,'r.' € X" } .


114

Chopter 4 PRoenRrrus on Rncttl,n'n tn'ucu.q.cns

: show that there is an algorithm for determining whcther or uot .L tail (L) .L. for any regular 12. Let .L be any regrrlar larrguage on E : {a, b}, show that an algorithm exists for determining if tr contairrs arry strings of even length, ffiS 1 S . t'ind an algorithm for cletennining whether a regular language tr contains an infinite numher of even-length strings' Describe arr algorithrn which, when given a regular gramrnar G, can tell us whelher or not L (G) : E-.

r onguoges l d e n t i f y i n gN o n r e g u l o L R.cgular languages ca,p be inflnite, as rlost of our examples have demonstrated. Thc ftrct that regular languages are associated with autt>mata tha,t have finite nrcm()ryl however, impOses Some limits on the structure of a regular language. Sotne ttarrow rcstritltions mlst be obeyed if regularity is to holcl. Intuition tells ns that a language is rtlgulir"r only if, in processirrg arry strirrg, the infbrmation that has to be rernernbertxl zrt a.ny stage is strictly lirnitecl. This is true, but has to be shown prtx:iuely to be used in arty rntlarringful way. There are several ways in which this precision can be achieved.

PrinciPle Usingthe Pigeonhole The term "pigeonhole principk:" is rrsed by mathematicians to refer to the fbllowing simple observation. If we put z objects into rn boxes (pigeonholes), ancl if TL> rrll then at least one box must have rnore tltarr one item in it. This is such an obvious far:t that it is surprising how many deep result$ can be obtained frotn it.

..fi'lll

Efufffnfm 4rS

Is the larrgrragetr - {anbn : rz > 0} regular? T}rc answer is no, as we show using a Proof bY contradiction' Suppose.L is regular. Therr some dla M : (8' {a, b} , d, {0, F) exists for I , 2 , 3 , . . . . S i n c et h e r e a r e a n u n l i m i t e d i t . N o w l o o k a t d * ( Q o , o t )f o r i : of states in M, the pigeonhole nunber brrt a finite zj's, only of rrurrrber say q, such that state, sonre be thtlt must us there tells urinciple d* (qu,o") : q

d* (q,i, e*') : q,


4.3 InururrnyrNcNor-rRpcuLAR Lnlrcuacos

lfE

with n t' m. But sirrce M accepts a"bn we must have d. (g,b') : q7 â‚Ź. F. Flom this we can conclude that d* (go,a*bn):

d. (d. (qo, an'),b") : 6 * ( q , b "t : Af.

This contradicts the original assumption that M accepts a^bn only if n : rn, and leads us to conclude that tr cannot be regular. I In this argument, the pigeonhole principle is just a way of stating precisely what we mean when we say that a finite automaton has a limited merlory. To accept all anbn, an automaton would have to differentiate between all prefixes a* and a-. But since there are only a finite number of internal states with which to do this, there are some rr and m fbr which the distinction cannot be made. In order to use this type of argurnent in a variety of situations, it is convenient to codify it as a general theorem. There are several ways to do ttris; the one we give here is perhaps the most famous one.

A Pumping Lemmo The following result, known as the pumping lemma for regular larrguages, uses the pigeonhole principle in another form. The proof is based on the observation that in a transition graph with n vertices, any walk of length z or longer must repeat some vertex, that is, contain a cycle. Let L be an infinite regular language. Then there exists some positive integer ?rl such that any w e L with l.ul ) m. can be decomposed as w

-

egk,

with

l r u l1 m , and

l s l> 1 , such that ^ ' , -wi

i s a l s oi n . L f o r a l l i : 0 , 1 , 2 , . . . .

* . , ' ' z* &g e)

(4 2)


116

Chopter 4

PttopsRrlns or RnculAR LANC;UAGES

this, everyuuFicientlylong strjqg in 4 cql_b9*b[aken To paraplrrase p.rT an--arbitrarynumberof repetitionsof iT . isg_!L:e.: iy"h y thc rnidd-le part yields another string in .L. Wtr say that the rtriddlc strittg is "pumped,tt ltence the term puntping lemma for this result.

Proof: If ,L is regular, there exists a dfa that rexrogrfzesit. Let such a dfn,havtl states labeled qo,qr,Q2t...tqrr.Now take a,string trr in,L such that Since tr is assumed to be infinittr, tfuis can always he dqne. l",ffiEl] Cornider the set of statcs tlte automaton gocs through as it proces$cs ?rJ' say Q oQ , t 'Q i' , , , ,Q fSince this sequencehas exactly l,r.'l+ I entries, at lea"stone state must be repeated, arrd such a repetition rrnrst start tro later than the nth move' Thus the sequcrrcernust Iook like Q o tQ l . ,Q i , . . . , ( l r ' , - . - , Q r t ' . ' t Q f t

indicatirrg there tnust be substrings 'x,r Ai z of ru such that d * ( q u ,r ) : e , , d* (g", u) : q,', d* (q', z) : q.r, with lrgl { rz * 1 : m, and lyl } t. F}om this it immediately follows that ,5*(qn, rz) : Qf, as well as rf. (eo, nyzz) : qr, d* (80' rY3z) : q7, and so on, completittg the proof of the theorem.

r

We have givcrr the pumping lemma only f'or irrfinite languages. Finitt: Ianglra,gr:s, although always tegttla,r,r:annot bc putnped since prrmpirrg automrrtir:a,llv an infinite set. Tlrl tlrtlorerrtdoes hold for finitt: lirrrguages, " creales ,{^Jl.ry 'l'he ^, but it is Vac.r.roliE/ nl irr the putnpitrg lemtna is to bt: taken larger tha'tr tlte lotrgest string, so thir.t rro strilg can be purnptld. The pumpirrg lcrnrna, Iike the pigeonhole argtrtnettt in Example 4.6, is rrsed to strow that cert,aitrla,ngrtagesa,rttnot regular, The demonstration is ir,lwaysby cotrtracliction. Therc is rrothirtg in lhe pumping lerrrrnit,as we hir,vc'statedit here, which ca,nllc rrsc:tlftrr proving that a languir,gcis regular'


4.3 Irnrurrr'ylNc NoNREcLTLAR LaNcuacns

IIT

Even if we coulcl show (ancl this is rmrrnir.llyqrrite difficult) that any pumped stringJrnust be in lhe original larrguir.gc,there is nothing in the stalernent of Theoretn 4.8 that irllows us to r:orrrhrrleflrrm this that the language is rcgrrlar.

[ltulnflg

Usirrgt]rc prrmping lemma to show LhaL.L : {unltn : n > 0} is not regula,r, Asstrrnttthrr,t L is regular, so thaL Lhe puutping lt:rnrnil must hold. We do trot krrow the va,lue of nt,, but whatever it i$, wc ctr,nalways .hodeE t&] Tlxtreftrre, the substring y musl consist c'rrtircly of a's. Suppose lyl : ft.> t Thcrr thel string obtained by using i : 0 irr Erpatiorr (a.2) is

[;t','

'Ll)O :

t1711-lrhm'

rr.rrdis clearly not in .L. I'his conLradicts thc pumping lemma and thereby indit:irtcs thnt the a,ssumptionthat .L is regular rmrst bc fir,lse. I In applying the purnlrirrg l{:rnrna,we must keep in mind what the theorem says. We are guirrarrttxxl tlrc existence of an nz as well as the decomposition rEz, but wc:rkl rrot krrow what they are. We cannot clain that we have reached a contradiction just bcr:auscthe pumping lenrrna,is violated firr some speciflc values of rn, rsr 'ryz. Orr tlxr otherr hand, the punrping leurua holds fbr cvery ?1,e tr and every 2,,'I'herefore, if the purnpirrg lernrnil is violatecl everrfor orro 'iu or i, thcn the language cannot be regular. The correct argument can be visualiztxl ir,iirr game we play aga,inst an opponent. Our goal is to witr Lhe gatrte by tlsttr,blishingir, contradiction of the pr.rmping lernma, wirile the opponcrrt triels to foil us. There are fbur moves in the garne. it+

1.

'Ihe

opporrerrt pir:ks rru.

2, Givt:n ?7?,, we pick a, string ,urin ,1,of length cquerlor grcatcr tharn rn. We are free to r:]rorisr:irrry ?1r,srrb.jer:tto 'roâ‚Ź .L ancl lrul > rn,,

) \

T IL\tvdlF**

, t | /1 vdr'e

3. l'he opponent c:Iroosrls tlrr: rkx:ornllositionrgra, subject to lzyl f nz, l3rl> 1. We have to assurnctirir"tthr: opllonent ma,kesthe;choicethat will make it hardest for us to wirr tlrt: girrntl. 4. Wc try to pick i in such a way that Lhe purrrpeclstrirrg ti.r,,,rk:firrcrl irr Erlrir,tion (4.2), is not in.L. If we can do so, we win the garrrt:. A strategy that rrllows us to wirr whiltcver the opponent's r:hoicesis ta,nta,mountto a, proof thal the language is rxrt rcgulir,r. Irr this, Step 2 is t:rtrt:ia,l. Wrile we canuot forc:eLhe opponerrt to pick il pa,rticulilr tltx:ompr-rsitirlrrof 'rl, wc mrly be rr,bleto choose tl so that the opponent is very


Lewcua.cns Chopter 4 Pnornnlus or.'Rr,:(+lrl,.q,n

118

Figure 4.5 u...ah...bh..,ha...a

F-+-t--------------E J

Z

restricttxl irr Step 3, forcing a choice of r, g, ir.rrcl.ethat allows us to produce a,virilrrtion of puttrping letnma on our next rrrtlve.

Nnl(

th at'

he o f lle , ,

t -

Jr) t

L0

/t

ili

/ a r n . 6 1 * t.

(qna6t

/

sltorla/

.

tts I/g &#tra/ t

for.(e :

/ a 5

o r h , s ) < , n ,' flo,l

j r',t{

t

{o

k

o{ '

l

/ht ;f u,'n u); hor{ lro'/.n,+rtL. y, It,.o I I

7,:{wwfl:rueX*} is trot regrtltr.r. Wha,tcvclr rrl the opponent picks on Step 1, wc carr always choose a34 aE-shetynirr Figure 4.5. Because of this choice, trrrtl the requirement that

ffisrestrictedinSteplJtoc}roosirrgaythatconsists *cirtirblflbf a's. In Step 4, we use i : 0. The strirrg obtained in this fashion has fewer o,'$ on thc lcft tharr orr the right and so cannot be tlf the fbrrrt u,tuft. Tho,rt:ftrrc-L is trot regular. F Note that if we irad chosen ru too short, then the opponent could have t:hosen a g with a,rl even numher of bts. In that case, we could not have reached a violation of the pumpirrg lerrrrna orr the last step, We would also fir"ilif we were to choosea string cxrnsistirrgof all ats, say, 'til :

ttrZ'o,

which is in L. To dcfca,tus, the opponetrt need only pick U:ua' Now zr,. is irr -L for all i, atrd we lose. Ttr ir,1l1llythe purnping letntna we cannot ir.Jsurncthat the opponent will rnirk(r ir wrong rnove, If, ,in the ca-sewhere wu pit:k 'u) : [tr?"', the opponent wc:rt:to pick a:Q, then t 0 is a strirrg of odd length ilrrd therefore trot iu .L. But any trrgument that assurne$thtr.t thc oJrportettt is so accommodating is autornatically incorrer:t. I


4.lJ Itrurq'r'rr,'ylNcNoNREcuLARLexcuacns

119

Let X : {a, b}. The languirge 1, = lut E X* : rzo(,ru)< n6 (u)] is not regular. Supposewe are given rn,. Since we have cornplete freedorn irr choosirrg rrr, wc pick ut * tt"'b""t r. Now, beca,uselrgtl cannot be greater than rrz, the opllonent r:a,nnotdo a,nything br.rt pick a g with all a's, that is y:ak,

r!h<rn'

We now pump up, using zi: 2. The resultirrg strirrg ?il2:

5*trl-|E6m+L

is not in .L. Therefbre, the purnping lerrrrna,is violated, ancl -L is not regular

The larrguagtl

7, : {(ab)" ak : r-t> h, fo > 0} is not regular. Civen rrz,we pick a,sour string ut:

-

(ab\"'t7 a"'

which is in -L. Because of tire constraint l*El < ?r-4,. bo!l;g-g+cl*U-rlrJ:! bqilltp_pg1! g{_tbg p_tlirrgnade up qf."g!lS, The choice of r does rrot affect tht,'+r,rgument,so let us seewha,t ca,nbe done with y, If our opponent picks (gL-:-6,,wc t:hoosei = 0 and get ir.string not in .L ((a,b)-a-). If the opponent piclks;r7:,g.b,ywcca,rrchoost)rl :0 rr,gir.irr. Now we get the string (ab)tt'a"', rvltich is _Illl_tiqLl. Irr thc sarrrc wily, wc can dea,l with ir.ny possiblu r:hoir:el by the opponent, thereby proving our clairn. I

Slmw tha,t L:

{ e r r l; r z > 0 }

is not regular. Ciiven Lhe opponentts choice for rrz,we pick as 'u the string a"'l (unless the opponent picks nt 13, in which case we can use aill as trr). The various


120

Chopter 4 Pnorenuns or RnGtJtan, Latqcuecns

decotnpositions of tu obviously differ orrly in the lengths of the uubstrirtgs. Suppose the opponent picks y srrth thtrt

lul: h < m' We then Iook at z.s which has length m,l - k. This string is irr l, only if there exists a j su<:hthat mt-h:jl But this is impossible, since for rn ) 2 and k { m we havc

h!:.rnl-tu>(m,-1)! Therefore, the lartguage is not regular. I In some cases,closure propcrties can be used to relate a given problem to one we have a,lready cla"rsified. This ntay be much simpler than a direct application of the putrpirrg lemma.

Show that the language L : { a " b k c ' ' t k I h > 0 .h > 0 } is not reeular. It is rrot clifficult to apply the pumping lemma, clircctly, but it is even easier to use closure uuder homomorphisrn. Ttrktl h ( a ) : u , h , ( b ): a , h ( c ) : c

lhF" h ( L ) : { a n t h t " n + k: n * k > 0 } :{aici;i>U}, but we krrow this language is not regular; thereforc .L carrrrot be regular either.

I

Show that the ltr,ngrrtr,gt: l:{ar,.bt.:nlI} is not regula,r.


4.il Innur'lpyrNGNoNRECLILARLar'rcuA,cns121

I'Iere we need a trit of irrgemrity to apply the purnping lernrrra dirrx;tly. Clhoosinga string with ri, : l+ 1 or n, E l*2 will not do, since our opporrerrrt citrr illwalys choose a decomposition that will ma,ke it impossible to pump thc string orrt of the language (that is, prrurp it so that it has an equal trutrrlrerof a's arrd b's). We rnrrst be rnore inverrtive, Let us ta,kl n.: m,l and I : (rn * 1)!. If thc ollporrcnt now choosesa g (by necessitycorrsistirrg of all a's) of length h < rt,, we purrrp r. tirrx:s to generate a string with rn! + (i,* I)k a's. We can geL a corrtradir:tionof the pumping lemma if we r:ilrr pick i suclt that nz!f(i-1)A:(rrr,*1)! I'his is rrlwir,yspossible since nt, ntl . . r- rT fr arxl A { rn,. The riglrt side is tirerefore an inttlgur, irnd we have srrcceedecl irr violtr,tingthe conditions of the purnping lernrnir. Howcvrlr, there is a, rnuch more elegarrt way of solvirrg this prriblem. firrllpose L were regular. Then, by I']rxrrclrn 4.1, Z ancl tire lilnguage I't:T

r - t . L( o * b * )

woulcl also be regular'. tsut -L1 : fntlhtt: rz > 0], whicli we have already classifit:das rurrrrcgrrlilr.C)onsequently,tr cannot be regular. I The purnlling lcmmil is difficult fbr several reasons. Its staterrrerrt is complicated, irrrtl it is oir.syto go a,stray in applying it. But everr if wr: tnaster the teclurique, it rniry still be ha,rd to seeexactly how to use it. The pumping lernrna is likc a girrnc with r:omplica,tedrules. Knowledge of the ntlqrs is esseutial, br.rt ihat alotre is rrot errough to plir,y rr.good game. Yon alsrl rrrxrd er,good strategy to win. If yoLr can apply the purnpirrg lcmma, correctly to sotne of the rnorc rliffir:ult r:asesin this book, vou a,re to be congratula,tecl.

(t

Prove the fbllowing version of the pumping lemnra. [f L is rcgular, thcn there is an nz such that, every .u â‚Ź Z of lerrgth greater than nr, ca,n be decomposed as 'tD:

frUE1


r22

Chopler 4 Pnopnnrtns oH ltr:ctrlan, Lar'{cuacns

with lyzl < rn,

l s l> 1 ' such that nyoz it irr .L fbr all i. 2

Prove the frrlkrwirrg generalization of the prttrrping lemma, which irrclutles Thcorcm 4.8 as well us Exercisc 1.as special cases. If .L is regular, therr there exists an m,, such that the following holds for every sufficiently lurg ur € .L and cvery one of its decornpositions llr : I.t,tltl.tzj with u,r,rr,z€ X*,lul ) tn, il'he middle string u can bc written asu:'rAz1 < rn, lgl ) 1.,such t,hat,ulrgi z'uz E L tor all i : 0, 1, 2, ... m with le;'grl

3 . S l r o w t h a t t h e l a r r g u a g et ' :

{w:n*(ut):"r6(tr')}

is trot rcgular. Is.L.

regular? that the following larrguages are not regular.

(i)e.o"*

( u ) f : { a , " I t l t t k: I + } n + l }

ffi

(b) ,I:

{ o " 6 l o a: A t I n - l I )

(c) I :

: I or l f k} {anblak : n

(,1).L:{o."bt:n{11 (c) Z, : {ut : n,,(tu) t' nt (,ru)} ffi -(f) t - {unn :'ur€ {4,b}-} (g) L:

(fl

4 .t)l

if the following larrguagesou X : {a} are regular

(a) I : {an : n } 2, n' is a prirne nutrber}

*-;D

, Fa.,t

\J'',

n"t*.*inc

{ t u r r l , r r , r , Rt l: e { a , b } " }

.gt*.,

f . ' i ; '-. l_

\_/

Lrro,*+-(b)

L : {a"' : tt' is rxrt a prime number}

(c) l,:

tv t-

-(.t)

ffi

{ttn : n:

A 2 f o r s o m eh > 0 }

.L - {a" : n. :2k

1 -oqK

l ' , .^

for some h > 0}

(e) I : {an : n. is the product of two pritre numbers} (f) , = {an : n is either prirne or the product of two or more prirne numbcrs} 6. Apply the pumping lcmrna directly to show the result in Example 4.12. .

j

L-'\fl

nr la

t-

(r]

rno* that thefollowinglarrguageis trot regular.

n>i1-r\,

t5*'i

o l"-gxn _ . ": , "nl_ L:{*n'-h n=

\

n>

lL-

\

(\( * \ . r t _ . t ,"L;;r,} " ,,t'-;" r}_1,{""r,-

.-,,,/t} ,(i)

ordisprove the ftrllowing statement.""If trr arrtl .Lz are nonregular lan@erov* guages, therr -Lr U -Lr is also nonregulat. ffi


4,3 IuuvrrnyrxcNoun,ncuI,An,L.tNc;ulcrs

123

Consider the latrguagesbelow. For each, make a corriecture whether or not it is rcgular. Then prove .your conjecture. ( a ) Z , : { a n b t u , k: r r l l l k

> 5} m

(b) fI : {a"blak:rr,}5,1>3,h Sl}

ffi

(c) .L -- {a"bl : nf I is an integer} (d) /. : {a*b{ : rz "1-I is a prirne nurnber} (e)

{a."ht : n, < t. < 2nI

( I ) , I : { a " b t : n > 1 0 0 , 15 1 0 0 }

( e )r : { a " b r : l n - l l : 2 } \fOJ l,r the following langua,geregular? [, : {w1c'u2 i Lurl,rrzâ‚Ź {o,,b}. ju)r + :iuz}

( il{ -J |2.

(

l"t -Lr and .L2bc regular lt-r.nguages. Is the languags 7 : necessarily rcgular'l ffi

Apply the pigeonhole arguuent rliret:tly to the larrguage irr Exarnple 4.8.

1_t:)A*

the followirrg larrguagesregular?

(a) I :

{u'*rr,''r

i'rtrj'Lt1'u e) { a , b } + }

W

* (b) /, : jrLte {a, 6}+ , lrl } lul} dS {tnuwRui u,1r)

| ",. i ' r ' i

G4

,utâ‚Ź L1,utr e LJ} {tu :

/ $

tr thc following languagcrcgular'l ' "'l l , : { r , * t ' r , r , u,wE la,bl L J

'/5-,

rl+(L5. , P E P A '--- )fet f be a,n infinite Lrut courrtable set, arrd associate with ear;h languagc .Lo. Thc smallcst sct containing every .Lu is the union over the infinite set P; it will be c.lenotedby Upep1,p. Show by cxample that tirc fatrily of regular larrguages is rrot r:losed urrtler irrfirrite urriorr. ffi * 16, Consider the argurnerrt irr Set:tiorr iJ.2 that the langrrage associated with any getreralized trartsition graph is regular. The larrguage assor:iaterlwith sut:h a graph is

r-, : |L J| L- \/, r ' p ) , DeP

whcrc P is thc sct of all walks through the graph and ru is the expression associated with a walk p. I'he r+etof walks is genera,llyinfinite, so tha,t in light of Exerr:ise 15, it tloes rrot irnrnetliately follow that -L is regular. Show that in this case, beca,useof the special nat,ure of P, the infinite uniorr is regula,r.


Chopter 4 Pnopnnrlrs or RncutAR LANGUAGES

the family of regular languagesclosedunder infinite intersection? ffi Supposethat we know that Lt I Lz and trr are regular' Can we conclude \L /\.r- dE-J .#

tr /+

*fipfr

trom thls that .Lz is regular?

19. In the chain code language in Exercise 22, Section 3.1, Iet .L be the set of that describe rectangles, show that .L is not a regular all u e {u,r,ld}* language.


Context-Free Longuoges

n thc la,st chapter, we rliscoverud that rrot all latrgrta'gesartl rcgular. Whilc rcgular langua,ges are ttfft:r:tive in describing t:elrtilirr sitttple llatterns, one does not rrt:c:tlto look very fa,r fbr exir,rnplesof nonregular languages. The rclcvartceof these limitations to programming larrguages becomes evirlt:rrt if we reinterpret somt: of tlte exatnples. If in L: {q*\rn : rz > 0} we sutrstitute a left pa,renthesisI'or a atrd a right parellthesis for b, then parenthesesstrings such as (0) and ((0)) are in -L' tlrt (0 is not. The la,nguagetherefore clescribesa sirnple kincl of nestud stmcindicating that somtr llrollerties of ture fbund irr programmittg la,ngutr,gos, programmirrg lattguages reqrtirer sorncthitrg beyond regrrlar lirrrguages. In rlrrlclr to cover this and otlrt:r rnore complicated fuaturcs we tttLtst enla,rge This leads us to considt:r context-free langrrir'gcs the farnily of langJrragt:s. ancl grammars. We begin this r:ha,1lterby clefining context-f'rtxr gralrunars a'nclltr,nguirgcs, illustrating the dqfirritions with some simplc: cxarnples. Next, wc txrrrsider the importa,nt nx.'rrtbtlrship problem; in prrrticular we ask how wt: t:irn tell if a given strirrg is clerivablefiorrr a givtlti graurtnar. flxpla,irririg ir setrletrce through its grilrnrnirtical deriva,tion is fir,rriiliar l,o tnost of rrs f'roru ir stucl.y

125


126

Chopter 5 Cournxr-FRno Lancuacns

of natural languagesand is callexl parsing. Parsing is a way of describing sentence structure, It is irrrportirrrt wh(lnevrtr we need to understand the meaning of a sentence,as we do frrr irrstirrrcein tra,nslating from one language to a,nother, In computer science,this is rclt:vilnt in interpreters, compilers, a,nclother translating prograrrrs. The topic of context-free languagesis perhirys the most important aspect of firrmal la,nguagetheory as it applies to llrugrilmming la,ngua,ges. Actual progritrnrning la,nguageshave many fealures that c:arrbo clescribed elegarrtly try means of context-free languages. What frrrrnal lar.nguagetheory tells us irtrout rxrntext-fiee languages has irnportant applic:rrtiorrsin the design of prograrnrnirrg ltr,ngua,gesas well as in the constructiclrr clf clfficient conpilers. We touch rrpon this briefly in Section 5,3.

Context-Free Grommors ffi;;mffim The procluctions itr it rtlgrrlir,rgrarnmar are restricled in two ways: the left side musl be a sirrglr:variilblr-',while tlre right sicle has a spcc:ialforrn. To crea,tegl'alrtlrrars thirt irlu rntire powerfirl, we rnust relax sorne of tlrrlsc:rostrictions. By retaitritrg the rcstrit:tion on the left side, but perrnittirrg arrythirrg grarnma,rs. on the right, we get corrtc:xt-fi'rxr

M A gramrnar G: (V,T,S,P) is said to be context-free if rrll prodrrctions itr P have Lherltrrrn A+il,

where,4 e V and 'r t (V u ?).. A lir.rrgutr,ge I, is said to be corrtext-frtxrif and only if there is a contextfreegrarnmarG srrchtha,t tr : L(G).

Every regular grarrurrar is rxrntext-free, so a regular langrragr: is trlso a, rxrntext-free one. But, as we krrtlw f'rom simple examples such as {u"h"l, thclrc are nonregular languages. Wt: ha,ve already shown in Exarnple 1.11 thrrt this language can be generirtctl tly tr.crrntext-free grarnmar, so wc s(xl tltat tlxr fer,milyof regular Iatrguirgesis a proper subset of the fanily of cotrtext-frrt larrgrrirgesgrirmmar$ derive their trattrt:frorn thcl fa,ct tha,t the subCouLext-1'rtx) stituliotr of thc: variable on the left ol a produr:tiori t:ilrr be rnade any time sttch a, variable appears irr il st:rrtential form. lt, does rrot dt:pcnd on the


5.1 Conrnxr-li*Rps GnalrunRs

127

'I'his f'eaturtr is symtrols in the rest of the senterrtiir,lfrrrtrr (lhe contcxt). of rrllowittg only a, sirtgle va,ria,bk:ori Lhe left sidc of the the consecluenr.:et procluctiorr.

Longuoges Exomples of Context-Free $*sqtrsfH $;l

The gramma,r(J : ({S} , {o, b} , S, r), with productions S --+a5a', .5 --+ b$b, ,9-4, is t:orrtext-free. A typical clerivation irr this gramma,r is S + a.Su + aaSe,a + aubSbaa ) 'I'his

n,a,bbu'tt'.

makes it clear that L(G):

'rrr {u,ruB, e {a,b}.}.

The languagc is context-f'rrlc, but as shown in F,xample 4.8, it is uot regrrlar

S - abR, A - uaBb, R + bbAa, A-4, it to the rea.tlt:rto showthnt is context-free,We letr,vt: L(G):

{ab(bbaa)" hha,(ba)'': n, > 0}

T Both of the above erxanples involvcl gral]lmars thirt ate not only c:ontextfree, blt lirrc:ar, Regular ir,rrdliuear grafiIIIIarS are clearly croutext-fitlt:, but a context-free gramnlirr is uot neces$rrrilylinear.


128

Chopter 5 Conrnxr-Fn,pp Lerrc+u.q,cns

WWWW

rhe language 7,:{a"b*:nlm,} is context-frrxl. lb show this, we need to protlurxl a, context-free grarnmar fbr the language. The castl rtf n : ?7?was solved in Exa,mple 1.11 and we c:a,nbuild on that solutiorr. Take the case ??) rn. We first generate a string with an equal number of ats and b's, then add extril fl,'s on the left. This is done with ^9-

ASr,

5r - a5rblA, A --+aAla. We can use sirnilar rea^soningfor the casc n { m,, and we get thu answer 5 - ASrl,SrB, ,5r -i a$rbltr, 71-+ aAla, B --+bBlb. The resulting gra,mmar is contuxt-free, hence tr is a context-free languager. However, the grammar is rrot lirrerirr. The particular ftrrm of the grarrmrar given here was choserrfrlr the purpose of illustration; there a,remany other eqrrivalent context-frrle grammars. In fact, there are some sirrrple linear ones for this language. In Exert:ise ZE at thc end of this section yolr are asked to firrtl one of them. I

:IIWiliWWilWWMW Consider the grammar with

productions .9 -

a.5'bl55l,\.

This is another gramma,r that is context-fiee, but not lirrrrar. Some strings in .t (G) are abaabh, aababb, and ababab. It is not difficult to urnjecture and prove that * L : {, e {4,,b} : no (w) : nr, (ut) and no (u) } 26 (u) , where u is any prefix of 'u,').

(b.l)

We can sec the connection with programming larrguir,gesclearly if we rt:place c, and b with lc:ft a,nd right parerrthtrses,respectively. Thc language .L


5.1 Cournxr-FRnn GR,ttutrr.c'Rs 129

incluclessur:hstrirrgs as (0) ancl 0 () () and is itr fa'ct tht: sct of all lrroperlv nestr:rlparetrthesisstructtrrt:s lbr the colrllnorl prtlgratntnitrg la,ngtrilgt:s' Here again therel ilnr rriany other eqrriva,lcrrtgralrtlrla,rs. Brrt, irr contrast to Example 5.3, it is rrot so easy to sexlif there are any lint:irr oiles' We will have to wait r.rntil ()hrrpter I befbre w(t (tirrr auswer this qrltlstiott'

T

Derivqtions Leftmostond Rightmost In context-freegrirfirilrarsthat a,renot lirrcar, a derivation rnay involve sententia,lfrrrmswith more thaln tlrrc variable. In srrch(iases,we have a chtlice in the order in which variiltrlesare repla,ced.Trlkc for exampletht: grarrrrnar G : ({A,8,5} , Io,bl ,5, P) with produr:tions t. S -+ AB. 2. A --+aaA. 3.,4-4, 4.8 - tsb. 5.8*+A' getrerates the languageL (G) = {aznb*' , trt is ea,syto sr:cthat this gra,mrnar rz ) 0, rn,> 0]. Conuidcrrrow the two dt:rivatiotrs s 1,ta

4 aaAB4 naB S uaBb4 aab

i.r,ntl

\ aab. s 4 .+n 1 e'at'4 aaABb4 q,a,Ab In order to show which producliott is a,ppliexl,we have numbcred the productiotrs and written the appropriate mrrnber on the + syrttbol. F\'om this we see that the two deriva,tiorrs rtot only yield tht: sarne sentent:tl but use exactly thc sarne procluctiqrrs. The clifferentre is etrl,irely in tlrc order in which t}r: productiols arc aplllied. To rem6vt: suclt ilrelevant factors, we often reqrrirc that the va,riabltrsbe replaced in a specific order.

M A rlcrivatiou is sa,id to be leftmost if irr each step tlrt: leftmost varitr,ble in the sententia,l forrn is replaced. If in each step the rigltttnost va,ritr.bleis replaced, we call thc derivatiou rightmost.


Chopter 5 Cor'rlrjx'l'F nr:r: Lnrvc:uaclns

Figure 5.1

),-\

\_1/

the gI'aIIuIIar wit} prochrctions

WnsicJer

S --+a,AB, A --+hRh, Al.\, fi __+ Then S + uAB + a,hBhB+ abAbB+ abbBhhB+ abbbbB+ o'bbbb is a lefturostrlcrivationof the string abltbb.A rightmost derivationof thc samestriug is + ahhhh S + aAB + aA + u.hBh+ abAh+ abbtsbb I

DerivotionTrees A second way of showing derivat,iotrs,indt:pcndent of the order irr whitlh prodttctiotrs arc usud, is by a derivation tree. A derivation tree is irrr orclered tree itr which rxrdes are la,beled with the lcft sides of productiotrs arrcl irr which the children of a node rcpresent its corresporrdirrg right sides. For example, Figrrre 5. 1 shows part of a tlcrivation tree representirrgthc prodnction A -

o.bABc.

In a derivation tree, a uode labeled with a variable occurring on the left side of a production ha,schildren consistiug of the symbols ou the right side of that productiorr. Bcginning with the root, latrtrlerdwith the start syntbol and ending in leaves that ir.re tertninals, a derivatiorr tree shows how each variable is replaced in thc durivation. The followirtg tlcfinition makes this trotiorr precise.

[Rlnfii.fii 'f,tf,ffi,n,,.Nil$,lil, Let G = (V,7,5,P) be a c:orrtr:xt-fieegramlnar. Atr ordcred tree is a derivation tree for G if ancl orrly if it has the following propcrtics.


Gnnurutnns 5,1 CoNTEXT-Fnnn

131

l . The root is latreled ,9.

2. Every leaf has a la,belfrom T U {I}. 3 . Every inttrrior vettex (a vertex whic:h is not a leaf) ha,sir.la,trcl frotn V. 4. If a vertex has labc:l A â‚Ź V, and its chiklrt:rr are Iabeled (from ltrft to right) o,1,a2,...,e,n, then P must conta,intr,llrclductiou of the ftrrrn A - + u 1 u 2 ,' , a n , 5. A leaf lahtllcxl\ Itas no siblings, thrr.t is, a vertex with a t:hiltl labeled A ca,nhave no other children. A tree that has properties 13,4 and 5, but itr which I docs rrot rrecessarily holrl and in which property 2 is replacecl by: 2a. Every lcaf has a label fiom V U 7'U {I} is said to be a, partial derivation tree. The string of syrnbols obtained by reading the leaves of the trct: frotn Ieft to right, omittirrg itrry ,\'s encoutrtered,is sirid to be the yield of thc tree. Tlre descriptive term Icft to right ca,nhe givt:rrit precisemeaning. Tht: yield is the string of trlrrninals in the ordt:r they are ettcoutrteredwhetr the tree is traversed in a depth,first rnarlrrer, always ta,king thc lefttrtost, tttrexplorttd branch.

,$u\WWNNWW,W$I thc grarnmarG, with procluctions Consicler S --+aAB, A -+ bBb,

fr --+rtlA. Thc tree in Figure 5.2 is a partial derivatiqn trce for G, while tlrc tree in Figure 5.3 is a deriva,tiotttree. The string abBbB, which is tlrc vield of the first tree, is a sentential form of G. The yielld of the second trtrtr, abbbbis a sentenceof I (G).

I


Chopter 5 Cournxr-Fnnn L.lNc.tt)AcHs

Figure 5.2

Figrre 5.3

RelqtionBetweenSententiolFormsqnd DerivotionTrees Derivation treeu give a very explicit and easily comprehended tlescription of a derivation. Like transition graphs ftrr {inite automata, this explicitness is a great helJr in making argurnetrts. First, thongh, we must establish the connection between derivntions and derivatiott trctts.

't .Let G -- (V,T,S, P) be a context-free gralilnar. Then fbr every â‚Ź I (G), tlNrre exists a derivation tr*,' of G whose yield is ir.'. Conversely,the yield of trny derivation tree is irr -L (G). Also, if t6; is atry partirr.l derrivationtree for CJwlxrsc root is labelecl5, tlrcrr thc yield of fs is a senterrtial fbrm of G. Proof: First we show that frrr every sentential fortn of I (G) there is a corrcsponding partial derivatittn tree. We do this by indrrction on the number of stcps in the derivation. As a basis, we note that the clainred result is true for every scntential form clerivable irr one step. Since S + u implics that there is a production .9 -r u, this follows imrntldiately from Definition 5.3. Assr.rmethat for every sentential form derivablc in n, steps, there is a corrcsponding partial derivation tree. Now any ?rrderivable in n * I steps


5.1 CoNTEXT-I"IIIJEGnaultarr$

133

must be such that S I r A y , n , aE ( v u 7 ' ) * , A â‚Ź V , in rr,steps,arrcl r A y + ' I A 1 A 2 " ' Q , , , . ! J :W r A LE V t l T , Since try tlte iucluctivc assunrption there is a partial derivatiorr tree with yield :rA'g, aud sint:t:the gramrnar tnust have llroduction A + a1a2' ' '(trrrLr we see that bv expanding thc leaf labeklrl A, we get rr,partial derivation tree rlr. By itrdrrctiorr,we therefilre claim that the with yield r&rfl|"'amA: forms. for all sententirrl rcsult is true r:arr show that t:very partia,l derivation trr:c reprevein, we In a similar loave this a"stlrr exercise. We will fbrrn. sents some scrrtential partial derivatiorr tree whosc leaves a,re is also a, clcrivatiorr trr:c Since a (G) is the yield of some derivain I c:very sententlc it follows that terninals, ()f tr:eeis irr l/ (G). I yielcl every derivatio[ the of G anrl tha,t tion trcrr:

Dcrivation tretls show whic*r productitlrrs are userl irr o[ta'ining a sentclrx:e,but do rrot give tlx: order of tlx:ir applica,titlrr. l)erivtltitlrr trees are ablt: to represent atry derivation, reflet:ting the fa,ct that this tlrtler is irrelqvattL, au gfus(:rvation whidr allows lltJ to close tr. gap itl the preceding discussion. By cle{initionr any u E L(G) has a dcrivation, }rut we havet rrot cla,imefl that it a,lsohad a leftmost or rightrrrost derivtr,tiotr. flowevcr' once wo have a, derivatiotl tree, wc catt alwaYs get a leftmost clerivatiott by thinkirrg of the trce as having been brrilt irr such a, way that thc leftmost variable in thc tree was rr,lwaysexpantled lirst, Fillirrg iu a, firw details, wc are Iecl to thr: rrot surprisirrg result that any ?trâ‚Ź I (G) has a,leftmost and a rightmost rlerivation (fbr cleta,ils,sce Exercise 24 at the erl(l of this scction).

1 . Conrplete tire argumerrts in F)xample 5,2, showirrg that the latrguage giverr is gerrerated bv the gra,mtnar.

,

l)raw the dcrivatiorr Lrcc corresponcling to the dcrivatiorr ilr Example 5.1,

3.

for the grammar in Example 5.2. Give a derivation tree for w : ahhltu,u,bbaba Use the derivation ttee to find a leftmost derivation' show that the grarnrnar irr Example 5.4 does in far:t generate the language describerl irr Equation 5.1. W

5.

Is the language irr Exatnple 5.2 regular?

s.

Cornplete the proof in Theorcrn 5.1 by showirrg that the yieltl of cvery partial c{erivation tree with root ,9 is a serrtetrtial form of G'


134

Chopter 5 Conrrnxr-FRnr L,q,mcuacps

b

Find corrtext-freegrammars for the following languages(with rr.> 0, rn ) 0)

f!

(u) I : {anb* : rt.17n + 3} ffi (b),I:{u,"h"',nInz-L} (c)I:

{a^b*:nl2rn}

( d ) . L: { a ' b * : 2 n . 1 r r " < 3 n } S (*) I : {tl e {a, b}* : n^ (w) t' nu (w)} ( f ) . L - { T re r { a , b \ * : n o ( u ) > n 6 ( u ) , w h e r er . 'i s a n y p r e f i x o fu r } (e) I : {w e {o.,bl" : no (,w): 2nt,(ur)+ t}. Find context-free grarnrrrarsfor the following languages(with n, ) 0, rn )

o,h>o). (a) I:

{a^b ch :rl:m

(b),I :

{a"6*o"n: n : rn or m I k}

(c) tr:

{anb*tk:h:nIm)

( d ) . L:

{a'ob*ck:n+2m:t+)

(e) L:

{a"b*ck : k : ln *l}

or rn !,h}

ffi

W

(f) f : {w e [a,b,c]* : nn (tu) + nr, (ut) I n. (w)]

9' /\\ (10)

\-/

*11. *--,

(e) I :

{a"h"'ck,h I n + rn)

(h) f:

{a"b*ch:k > 3}.

Find a context-free grammar for head(tr), where .L is the language in Exercise 7(a) above. For the definition of head,see Exercise 18, Section 4.1. {o,b} forthelanguage p=

Tndacontext-freegrammarforE:

x-,n > 1].

{anu^u?b : w E

Given a context-free grammax G for a language tr, show how one can create from G a grammar I so that l, fe) = head.\L). \ ./

/'\ | ( L2/ Lct L: {a"hn: n > 0}.

\J

(a) Show tl::,;l Lz is context-free.

S

(b) Show that .Lh is corrtext-frec for a"trygiven A ) 1. (c) Show that Z and -L* are context-free. 13. Let .Lr be the language in Exercise B(a) antt .Lz the Ianguage in Exercise g(d). Show that Lt J Lz is a context-free language. , 14.

Show that the following; language is context-free. L :

*r5'

{uorur,,,

i rt,rr,,tt) e {a, b}+ , lul :

l..ul: z}

show that the complement of the language in Example 8.1 is context-free. W


5.1 Cot'lrnxr-FRer:ClR.nnrlunRs

135

16. Show that the cor[plefirent of t]re latrguage in Exercisc 8(h) is cotrtcxt*free. ' fu-f) Sir.,* that thc langrrageL : {wicll)z i'u.l,.tt)2e {o. tr} ,*, t urit}, with }j : {n,,b, c}, is context-free. 18', Show a tlerivation tree for the string aabbbbwit'h the grarntnar g +,481.\, A + ttB, B-Sb. Civc a verhal tlescription of the language gerreratcd by this gralrlmar' 4ti9)orrsicter

\_--.,"

the grarnrnar with prorluctions S *

aaB,

A * bBiiltr, fi+Aa. show that thc striilg gau,rrrrrt"a. w

aabbabba is rrot in the larrguaE;c generatecl by this

2O. Consider the tlerivation tree below.

I.'inrl a simplc graurrrrar G for which this is the clerivation tree of the strirtg Thcn find two ntore serrtettccsof I (G)' -rr:1ntrb' what one rrright mean bv properly rrested parenthesis stnrctures in/ Uu)n"ntte volving two kincls of pareflthescs, say 0 and []. Irrtuil,ivcly, properly nestetl strings irr this situatiort are ([ ]), (tt ll) t()l' but not (Dl o' ((ll Using vour clefirrition, give a t:ontcxt-free gramrnar for ge[erating all properly nested parelrtnescs. Fincl a rrrrrtext-free glalnlnar alphabet {a,b}. ffi

for the set of all regular exJrressions on the

Find a context-frec grallllllar that carr generate all thc production rules for context-freegrammars with T: {a,b} and V: {A,B,C}' ( 24. hrouo that if G is a context-fi'ee grammar, then every u E L (G) has a lcftmost V ancl riglrtnost clerivation, Give arr algorithm for finding sudr derivations from

,fa*"""^"

a derivatiotr tree.


136

Chopter 5 Cournxr-FRr:n Larqcu.q.c;ns

2.5. F'ind a lirtear grammar for the larrgua,gein Example 5.1J. 2 6 . Let G : (V,T,S,P) bc a context-free gramrnar such that every one of its productiotrs is of the form ,4 * u, with lul : h ) 1. Show that the dcrivation tree for anv ?rl â‚Ź I (G) has a height h such that

I o g i ' l5uh. '<l q +

{ffiffi Porsingond Ambiguity We have scl firr c:oncentrated orr the generative aspects of grammars. Given ergr&rnmar G, we studied the sc:t of strings that c:anbe derived usirrg (J. In t:itscsof practical applications, we are also concerned with the analytical sidrl of thtr grammar: giverr rr.string tu of tclrmina,ls, we warrt to know whether ot rrot ru is in L(G). If so, we may want to find a deriva,tion of ru. An algorithm that can tell rrs whether'ru is irr z (G) is a nrernbershipalgoritlrrrr. Ihe tcrm parsing describes finding a se(luenceof productions by which a w ( L (G) is derived.

Porsingond Membership civen a string ru in r (G), we can parse it in a rather obvious fashion; wtr svstematically construct all possible (say, leftrnost) derivations arrd see wlrcther any of thern rntr,tchru. Specifically, we start at round one by looking at all productions of the fbrm S+JDr

finding all r tha,t can be derived ftom ,5 irr one step. If norrc of these rcuult in a rrratch with tu, we go to the next round, in which we irpply all applicable prod'ctions to the leftmost variable of every r, This gives us a sct of sentential forms, some of thcm possibly leading to ru. On each subsequent rr)und, we again take all leftmost variables and apply all possible productions. It rnay be that sorrrc of these senterrtial fbrms can be rtrjected orr the grounds that ur c&n never bc derived from them, but in general, we will have on each round a set of prnsibJe sentential fbrms. After the first rt)und, we have serrtential forms tlmt can be derivcd by applying a single production, after the second round wtl ha,vethe sentential fbrms that carr be derivecl in two steps, arrd so on. If u â‚Ź L (G), then it rrlrst have a leftmost derivation of flnite lerrgth. Thus, the nrethod will eventually give a leftmost derivation of tt. For referrlnr:ebelow, we will ca,ll this the exhaustive search parsing rnethod. It is a ffrrnr of top-down parsing, which we can viuw ir,sthe c:tirrstnrctionof a dcriva,tion Lr.ec: f'rom the root down.


5,2 Pnnstllc .q,no Atr,lrrIctlIrv

tBT

S - - +5 5 l o 5 b l b 5 a , l I tr,ndtlre string ru : aabb.Round tlrregivesus 1. 5 +,5,9, 2. S + aSb, 3. ,9+ bSa, 4. ,9 + .\. ftrr obvious The last two of theser:irnbe removexlfrom further <:orrsideration reason$.R.oundtwo thetr yields selntctttialforms S =+ 55'+ S,95, S+55+aSb,9, S+$,9+bSaS, 5+55+5, which are otrtainedbv re:placingthe lcftrnost S in serrtentialform l with all applicable substitutes. Sirnilarly, fiom senteltial fqrm 2 we get the additional sentcrrtialfonns S+aSb+tr,95b, S+aSb+aaSbb, S+ab'b+u'bSab, 5+aSb+ab. Again, severalof thesectr.rrbe removetlfrorn contentirlrr.On the rrcxt round, we fiud thc actual targct string from the sequenc(] S+aSb+aaSbb+aabb. Tlrerefore e,ahl)is in the larrguage gencrated by the gratnlnar rrnder consideration.

I

Exhaustive search ltarsiug ha,s serious fla,ws. The most obvioqs orrtl is its terliolsless; it is rrot to be rrsed where erflicietrtpa,rsirrgis requirtxl. But even wfien effic:itlrrclvis a setxxrdary issue, there is tl rrtore pertirrrlrrt objection. While the rrrelhod alwir.ysparses au â‚Ź L (G), it is possible tirat it never tt:rmiuates for strings not in L(G). This is certaitrly the cilse in


Chopter 5 CoFrrnxr-Fnnn LRwcjunoes

the previous cxir,mple;witlt m : a,hb,the method will go on producing tria,l sentential forrns indefinitely urrlcsswe br,rildinto it $ome way of sloppirrg. pa,rsingis relativrly TIte probklrn of nonterminatiorr of exhaustive $c:irrt:h elasvto overcornc if we restrict thc: forrn that the grirrrlma,r ca,n have. If wtl examine Exarnple 5.7, we see that the difticulty cotn(ts tiom the productions ^5'-* tr; this prorhrction can be ustld to clecreasethc length of successivc scntcntia,l fornts, so that we canttot tell *r,sily when to stop. If we do Irot have arry srrr',hproducbiorrs,then we have rttarry f'ewer dilTiculties. Irr fh,ct, tirere are two types of produc;tion$we wanl l,o nrlc out, those of tlxl firrnt ,4 --+ ,\ as well as those of the ftrrrn ,4 --+ B, As wc will see in the rrcxt chtr.pter,this restriction does not irffcr:t the power of the resulting gratnrna,rs irr ir.ny signilicarrt wav.

The gr*tnrnirr S' *+ 55 laSbl bSu la,hlba satisfies thc given requirerncnts. It generates the language irt Exnmple 5.7 without thc empty string. Given any Iil â‚Ź latb]+, the exhaustive seitrch parsing metlxrd will althc length ways termirrate in no more than lrul rounds. This is clear becauser of the senteutial ftrrm grows by at lea*stone syrnbol in each rountl. After lruj rounds we have either produ<:eda,parsing or we know that w { L(G).

The idea in this cxample catr be gtrneralized and rnade into a theoreur for context-fiee languages in genera,l. Supptrse that G : (V,7,,9, P) is a context-free gralrrilrirr which does not have any rules of the fornr

O-*tr,

A*

R.

where A, B e V. Thetr tltt+ cxhillrstive seart:hlla,rsingmet,Itorlr:ilrr be made into an algorithm whiclt, ftll any u â‚Ź X*, cither pr:oducesa parsing of trt, ot tells us that nrr parsing is possible. Proof; For each sentential forrrt, consider both its length and the number of terminal synrbols. Each step in the derivation increases at lcir"st one of these. Since neither the lerreth of a senterrtial form nor the rrtrmber of


5.2 Pansruc ANDAMBrcurrtr

1SS

tertninal symbols can excced lrul, a derivation carrnot involve rlore thtr.n 2 lul rorruds, at wlfch ti're we either havera succerssfulparsing or .u cannot bc genertr.tedby the gra,rrrrrar. I While the tlxhaustive search tnethod gives a theoretical guarantee that parsing can rr,lwaysbe done, its practical usefirlness is limitecl becausc the number of sentential ftrrrns generated try it may be excr:ssivelvlarge. Exactly how many sr:ntentia,lforms are generated diffcrs from case to case; no precise gt:rreral resrrlt ca,n be ester,blished,tlut we can put nome rough upper bounrls cln it. If we restrict our$clves to leftmost derivations, w() can have no more thnn jPl scntential forms after one rountl, no ilrore than lpl? sentential fclrurs aJter the set:ond round, a,nrJso on. In the proof of Theortlrn s.t, we obsclved that parsing ca,nnot involve more than ? lul rounds; therefore. the total numtrer of sententiir,l forms cirnnot exceed

M = l F l +l P l ' + . . . +l e f t * t

( 5 , 2)

This indic.r.tes that the work for t:xhaustive seatrh parsirrg may grow cxponentially with the le.ngth of the string, makirrg the cost of the mcthocl prohibitive. of rxrurse,Equation (5,2) is only a, bound, and often the rrumber of sentcntial forms is rnuch snraller. Neverthcless,practical observation shrws that exha,rrstivesearch parsing is vcry inefficient irr most cases. The const.l:tion .f more eflicient parsing methods for context-free gralllmar$ is a crlnrplicated mattcr that belongs to a course rlrr compilers. We wjll rrot pursue it here excr:pt for sorne isolated results.

Frlr every context-free grarrrrnar there exists a,n algorithrn that parses any ut E L(G) itr a nrrurber of steps liroportiorral to lrirls.

There arer severtrl known rnethods to at:hieve tliis, brrt all of them are sufficiently conrplicatcd that we ca,nrrot everr descritre thern wittrout developing some additiorral results. In section 6.8 we will take this question up agnirr briefly. More detnils ca,n be fountl in Harrison 1gz8 and Hopcrofb rrnd Ullrnan 1979. onc reasorr for not pursuirrg this in cletail is that evcn these algolithms are unsatisfactory. A rnethod in which the work rises with the third power of the length tlf the string, while better than an e,xponential algorithttt, is still cluitc iuefficient, antl a corlpiler hased on it woultl neetl an excessive amount of time to parse (lverr & nroderately long program. What we would Jikc to havc is a,pirrsing rnethod which takes tiue proportional to the krrgth of the strirrg. wc refer to such a,rnethod as a linear time parsing


140

Chopter5 Corurpxr-FRuuLerucu.q,cns

algorithm. we do not know any lindar time parsing methods for contextfree languagesin general,but such algorithms can be found for restricted, but important, specialcases'

],,tifii ltiffiffifiF|il,fi.H

ir

A context-free gra,mmar G : (V,7, S , P) is said to be a simple grammar or $-grammar if all its productions are of the form A --+ar, where A â‚Ź V, a E T, * â‚Ź V*, and any pair (A, a) occurs at most once in P'

The grammar g -+ c,,9lbS$lc is an s-grammar. The gra.mmar g + a,9lb55la$Slc is not an $-grarnmar becarrse the pair (^9,a) occurs in thc two prodnctiolE ,S __+aS ald S + a,SS.

I

While s-gramma,r$ are quite restrictive, they are of some interest. As we will see in the next section, Illaily featlres of comtnon programming languages can be described by s-grarrrmars' If G is arr $-grammar, then any string trr in f, (G) can be parsed with an effort proportional to ltul To see this, look at the exharrstive sea,rchmethod ancl the string rl : tt1Q,2.,.a",. since there can be at rnost one rule with S on the left, and starting with a1 olr the right, the derivntion rmrst begin with S + a1AyA2.-'A*. Next, we substitlte for the variable 41 , but sinr:e again there is at most one r:hoice, we rnust have S 1a4azBtBz"'Az"'A^, We see frorn this that eaclt step produces one tertrinal syrnbol arrtl hence the whole process must be t;omplettld in tro more that ltul steps'


5.2 PARSINcaNn AMeIcuIrY

141

Ambiguity in Grommqrsond Longuoges On the basis of our argument we can claim that given any w e L(G), "4" exhaustive search parsing will produce a clerivation tree for tu. we s&y "the" derivation tree because of the possibility derivtrtion tree rather than trees may exist. This situation is derivation of different a number that to as ambiguity. referretl

A context-free grammar G is said to be ambiguous if there exists some ut E L(G) that has at least two distinct derivatiou trees' Alternatively, ambiglitv implies the t:xistence of two or more leftmost or rightmost derivations.

The graurmar in Example 5.4, with productions ,9 - aSblS^91I,is ambigrrous. The setttence aaltb has the trvo derivation lrees shown in Figure 5.4.

_

l

Ambiguity is a common feature of natrual languages, where it is tolerated and dealt with in a varicty of ways. In programming languages, rvhere there should be urly one interpretation of each statement, ambiguitv rnust be rerrtoved when possible. Often w(] can achieve this by rewriting the grammar in an equiva,lent, unamlliguous fbrtn.

Figure 5.4


n

Chopter 5 Conrnxr-FRnu l,,nncuacns

Figure 5.5 Two dcrivatirxr tre.esfora*b*c.

(b)

Wonsider

the grammarG:

(y,T'-E,P) with

v = {E,I}, T : { a , b , c , ** ,, ( , ) }, and productions

E-1, E-ElE, ff + E,rE, E - (E), 1-alblc. Tlre strings (a+b)+c and q,rb + c arc in L(G)' It is easy to see that this grarnmar getrera.tesa, restri(:ted subset of arithrnetic expressions lbr C and, Puq(rill-like programming Innguages. The EI&IluIIar is arnbigutnts. For instatrce, the string u, I b*c ha^stwo differtlrrt derivatitln tree5, as shown itr Figure 5.5. I One way tq resolve ttre ambiguity is, as is done in programmilg marlrells, to associatelrrecedencenrles with tlttl operalors * and *' Since + norrrra'llY has higher preccdencetharr *, we would ta,ke Figrrre 5,5(a) a-sthe Correct parsing as it indicates thal b*r: is a subexpressionto be evaluated before perfortring the addition. I-Iowevr:r,this resohrtion is cornpletely outside the grammar. It is better to rewrite the gramntar so that only one parsing is ptlsuihle.


5.? PansrucArun$.lrercrrr'rv

143

Figure 5.6

$,*qfiplo S, t t

Ttr rewritethc;grarnmtlrin Exa,mple5.11we introducencw varia,blcs, taking V as {ll,7,tr',1} and repla,ce thc productiorrswith E-+7, T ' - - +F , F --+I,

E*:E*7, T-+T*F, F -+ (E), I -

o,lblr:.

A derivation tree of tlrc sentenrxrtt lb + c is shown in Figure S.6. No other derivation tree is possible for this string: the grammar is urrarnbiguous. It also is cquivalent to the grrrrnrnar in Exarnple 5.11. It is not too hard to justify these claims in this speciflc instarrce,but, irr general, the questiorrsof whether a, givt:rr context-free granrrrrar is amtliguous or whether two given txrntext-free grarnrnars are equivalent are very difficult to answer. In fact, we will latr:r show thrrt there are no general algorithms by which thesc questions r:iut always be resolverd.

r

In the foregoing examplrr the amhiguity came from ttre grammar in the sensc that it crruld be removed by finding a,n equivalent unambiguous grammiir. In somr: instanrxls, howevcr, this is rrot possitrle becausc the a,mhiguity is in tire langrrage.


L44

Chopter 5 Corirnxr-FR.r:r:Luqcuaclli

ni.rig1v611 ifliif 11Spfi If L is a cont(rxt-hee langurrgtl tbr which thcre exists an rtn1mbiguotls Brtt'mrnilr, tltett -L is snicl to be una'rnbiguotts' If t:verY Branlrlrtrr that geueratt:s L is aurlriguous, thcn the languirgtl is called inherently ambigrrous'

It is a, somewha,t clifficult tniltter even to exhi|it

art inherently aIII-

rwithsor'': 1lrff;.,:flll::lil llH,llffl'''r' il,lfl.l ;,,I'"I;ff i;:]"";;"'If, Exomplo 5.13

The languagt: L -- {a"bn t:"'} u {atbtttctt'}

t

with rz alcl rn non-negativr,., is an irrhcrently anrbigrrous conttlxt-fiee la[gllage.

That tr is cxrntext-freeis ea,syto sitow. Notice tha,t L:

LtU Lz,

-L1 is generir,terJby

51- 51clA, A -- a,Abl.\, and .L2 is rlur.:tious

by arr a,na,logousgrermtnar with start syrrrbtll ,52 atrd llrtr

Sz- aSzlB, B - bEr:ll. Thcrr -L is gette*r,teclby the t:ombiuatiorr of these two grarntnars with the additional produt:tiorr ii -

51lS2.

The grarrrrrra,ris ambiguorrrisiltce tlrtl stritrg A''bncn has two distinct clerivirtions,o1e strrrting with s =* 5r, tht: pther with,5'+ sz. lt tloes of course not follow I'rom this that -L is inhererrtlv a'mbiguous a's there rnight other tronarnbiguoltsgraIIIIna,rSfor it. Brrt in sotne:w,ty ['1attd -L2 exist Sorrr() hi1vecolflicting requirerncnts, tire first puttitrg a rtlstrictiotr orr the uutnber of a,'sa,trclbts,while the scctlnr.ldoes thcl seunefor bts il,nd c's. A ftlw tries will


5.? Fansrrvc+ arunAuucurly

l4E

tluickly cortvince you of the imllossibility of combining thgsc requiremelts itr a single set of rules that c:over the case n : m, uniq'ely. A rigorous argulnent) though, is quitrl teclurical. one proof cilrr be founcl in Flarrison 1978.

r

Gl

"u,.t arr s-Eirarnmarfor tr (a,a,u,rtt a b).

2. Firrl arr s-gramrnarfor L : {.a,"lt" : ru > 1}, ffi B, l'ind an s-grarrilnar fbr l, : {a'',n,rr :.ru} 2}. L9

show thal sno.t s-Br'alnlnaris unambigrrous. Let G: (y,il,S,P) be an s-srarnmar. Clive an expression for the maximurrr sizc of P in term.r of ll,/l ancl l7l.

6.

Sltow that thc following grarrrrnar is ambiguous. S *

ABlaaB,

A -

alAa,

E*b,

{ffi

7, Construct an urrarrrbiguous grammar equivalent to the grammar in Exercise 6. E. Give the dcrivation tree for (((a "f b) + c)) * o f t), using the grammar irr Exarnple 5.12. L 10.

Show that a regular languagc cannot be inherentlv arnbiguorrs. ffi Give an unambiguous grammar that gerrerates thc sct of all regular express i o n s6 1 X : { a , D } .

11, Is it possihle for a regular grammar to be arnbiguous? tt

that the larrguage7:

{uu,Il:tu

E {a,b}-}

is not inhererrtty ambigu-

|li:* 13.

Show that the folkrwing gramlnar is amhiguous. S + aSbS lbSasl A

L4. Show that thc grammar in Example 5,4 is ambiguous, but that the lnnguage denoted bv it is not. W 15,

Slxrw that the grammar in Exarrrple 1,13 is ambiguous.

16, Show that the grammar in Exarnple 5.5 is unambiguous. 17,

Use the exhaustive scarch parsirrg rrrethod to pa,rse ttre string abbbbhbwith the gramrnar in Exanrple l-r.l'r. In general, how rnarry routcls wil.l be neederl to parse any string ru in this language?


146

Chopter 5 Corurnxr-Fnnn LeNcuacns

1 8 . Slxrw that the grammar in llxar,rrplc 1.14 is unambiguous' 1 9 . Prove the following tesult, Let CJ: (y,T, s, P) lre a cofltext-free Btailrrrlar iil wlrich every A e V oc(ilrft on the left side of at urost o1c productiol' (J is urrarrrloiguorrs.

Thert

2 0 . Fistl a grammar equivalent to that in Exa'rurple5.5 which satisfies the trrrxlitiorts of Theorem 5.2. m

Grummorsdnd Context-Free $ffiffilii P r o g r o m m i nL go n g u o g e s One of tlte rnost imporlant uses of tlrtl tlxxrry of fbrmal languagesis in the clefinition of programmitrg languagcsa,rxlin the construction of itlterpreters nrrclcompilers fbr them. The basic problclll }rtlrttis to define a,programming lar,ngrr:lgeprecisely arld to use this clefinitiorr a.sthe starting point for the writirrg of c:ffir:icrit a,ncl reliable trauslation prograllls. Both rcgrtltr,r and cotrtext-frec lirrrguirgt,'str.re important in achieving this. As wtl hirvtt $een, regular lalguagels irrc u$ed in the recognition of certain sitrtpltl ltirttl'rns which occur in prograrrrrning la,ngrrages,but as we arguecl in the irrttoduction to this chapter, we rrefil c:ontext-fiee langnages to model more cortrplic:attxl turpects. As with most other latrguages, wc (:irrl define a programming latrguage by a grir,mmar. It is traditiorral irr writing orr pr()gfammirrg languages lo usc ir rxrnvention f'or specifyilg gIaIIIInars calltlrl the Bachus-Naur form or BNF. This forrn is in essencethe sarne as the notatiorr wc havtt used here, are enclosedin triarrgular bqt the appearallceis diffcrr:rrt. In BNF, va,ria,bles special marking. BNF withorrt any brir.r:kets.Termilal symbols are writtr:rr in we have done, Thus, rmrt:h way the illso uses subsicliary syrnbols such as l, BNF irr a*s appear tlre grirrnmtr,rin Example 5.12 miglrt (erpressi,on) ::: (terrnl | (t:nytre.ssion)I (term) , (terrn) ;;: (/ar:tnr) |(term,) + (/octor) , aqcl so o1. Thg symbols * ancl + are tetgrirtals. The symbol I is used -+' RNF as an alterrrator iL'Jin our rrotatiotl, but ::: is usetl instearl clf va,riable t:xltlit;it more use to lir,ngua,ges tend clescriptious of programrning there othcrwise But production explicit. of icleltifiers to make the irrtcrrt the two notations' bctwtlen the irre no significant dilleretrccs Ma,ny parts of a Pascal-likc progra,mtning language are susceptiblc to rlclinition lry restricted forrns of txlntext'fiee glammars. For exarrrplel, ir, Itascirl if-then-else statetnent can be tlclfirrtld as (i,f -state.me.nt)::: i,f (erprt:ssi,on) (the'n-clause)\else-clause) '


Laucuacns GnalrlreRs ANDPROGRAMMING 5,3 Cox'r'Erxr-Fn,nn

L47

Here the keyword z/ is a terminal syrrrbol. AII other term$ are variables which still have to be defirrcd. If we check this against Definititlrr 5.4, we see that this looku liktr arr s-granunar production. The variable (if -statement) on the left is always associated with the terrnirral z/ on the right. For this reason such a statement is easily arrd efficietrtly parsed. We see here a reasorr why we use keywords in prograrrntitrg languages. Keywrlrd$ not otily provide some visual structurqr that can guide the reader of a prclgrarn, but also make the work of a, rxlrnpiler rnuch easier. Unftrrtunatelg not all features of;r typical programtning language (jaII be trxpressedby an s-gramrna,r. Thtl rules for (erpress'ion) abovtl irre rtot of this type, so that parsing hecomeslt:ssobvious. The question then arisesfor what grammatical rrrles we calr pertnit and still parse cflicierrtly' In compilers, extensive use has bcen tnade of what are called LL arrd LR grammars, whic:h have the ability to express tlrt: less obvious features of a prograrrrrning larrguage, yet allow us to pilrst: irr linear time, This is not a,sitrple rnatter, ancl much of it is beyond thc scope of our discussion. Wtr will briefly touch on this topic in Chapter 6, tnrt for our purposes it sufficersto realize that $rrch granulrars exist and have brlcrr widely studied, Irr connection with this, the issue of ambiguity take$ orr added signifl(jarr(ic. The specification of tr prograrrlllring language must bcl urrarnbiguous, otherwise a program mav yield very different results wherr processed by different compilers or rurr on different systems. As Exarrtple 5'11 shows, a naive approach catr easily introdrrce a,mbiguity in the gra,mmar. Ttl avoid such rrristakes we must he a,blc to recogtrize attd remove arnbiguities. A related question is whether a language is or is not irrherently ambiguous. What we need for this purpose are algorithms for cletecting attd removing ambiguities in cx)rrtr:xt-freegrammars and ftrr tleciding whether or not a context-free ltr,nguirgeis inherently ambiguous. Unfortunately, theutr irre very difficult tasks, irnptlssible in tlte most general scIISe'as we will see later' of a programming language which can be modelul by a Those a^spert:ts However, it context-fieel grallrlrlar are usuallv rerfilrrcd to as its syntax' is rrortnally the case that rrot all programs which tr,rt:syrrtactically correct in this serlseare in fact acceptatrleprogram$. For Itascal, the usual BNF definition alkrws constructs such a,s 1)Qrr111 : r'eal; r, z '.'irtteqer

u&T 'r i ?n.[egel";

n ::3.2. to a Pascal compilt:r, sitrcethey Neither of tht:stltwo cotrstructs is a,crccptirble "etrnirrtcger variable cannrlt bc assigtred violate otherr constraints, such iLT


Chopter 5 Corurnxr-FRnn Lamcuecns

a real value," this kind of rule is part of progra,mming latrgua'gesenlalltics, sinrxr it ha,s to clo with how we ittterpret thtr rntrir,ningof a, pa,rticular corNtruct. sema,ltics are a courplicatctl rnatter. Nothing Prograrrrrrrirrglrr,ngrra,ge gramilrars exists ftlr tIrc Specifica'tion context-free cxlrrt:ise a,s as elegattt atrtl Colsequerrtly some semantic feaaud semantics, lnngrrage of prograrnrrring It is atr orrgging concern both ambiguotts. poqrly or definecl tures ilrav bc thexrry to find effective latrguage formal itr and larrgrtages i1 prograrnrrrirrg Several methods scrtrilntics. lauguage prograrnming tle{ining rnetho,Is ftrr accepted as uttivcrsnlly }ee1 have none them btrt of proposedr have beel bt''en havtl languages context-free tlcfinition as sernarrtir: for and as successful for svntax.

1' Give a complete dcfinition of \ex'pressiora)for Pascal, 2, Give a BNH dcfinition fbr the Pa^rcalwhile statement (leavilg t]re general concept (sfc,te,nent) undefinccl)' 3. Give a BNF Hril,r[rrrar t]rat shows the rela,tion between a Pascal program ancl its srrbprogrants, 4 . Give a BNF definititxr of a FOttiI'RAN do statement. 5 . Give a definition of ttre correct form of the if-else stateuent in C' 6 . Find examples of fcatures of c that <:annot be described by context-free gtamIIIATS.


S i m p l i f i c q t i oonf Context-Free G r q m m q r sq n d

N o r m oI F o r m s

efore we carr study context*free languagesin grealer clepth, we must atttlrxl to st)rnt:te:t:lrrrit:irlrnirtters. Thc tlcfirrition of ir contexl-free grirrrrrrriil irnpOses ntr rtlstrit:tiorr whatstltlver oil the right side of a freedonris rrot necessary,ancl in facl, llrorlur:tirirr. However, corrrplelter is a detrirnent il sorne argumenls, In'I'heoretr b.l, we saw the corlelielce of certairt restrictiorts olr gralrulratical forlm; elirninating rules of the forn A - A and ,4 - B made the arguments easier, In rnany irrsta,nces,it is desirablettl place evell lror(l stringcnt rustrir:tirlnsrlrr tht: grirrrrrrriir.Bccausc of this, we Ileed to look irt rnethoclsfor' lrirnsfonning an arlritrary corltextfree gta,mrna,rinto a,neqrriva,lentone tha,t sa,tisfiescerta,inrestrir:titins on its f'rrrrn. [n this r:ha,pterwel strrdy $evrtrill trrr,nsftlrmir,tiorrs arrtl sutlstituticxrs thu,t will bc rrscfirl irr srrbsr:tlrrclrrt rlisrlrssiorrs. \A/tl alstl irrvt:stigtrtc rrorrnal forrrrs f'rlr cxirrtr:xt-fr(xi grarrrrrrars, A 1orrnal filrrn is orrt: tlurt, a,lthough rr:strir:tc:rl,is llroad errough so that, a1y glirrrlrrlirr hlm itrr tx|rivirlertt trorrrtirl-forrrr versiorr. We itrtroduce two of the tnost useful of these, the chomsky norrnal form ancl the Greibach normal form. .Both ltave tnatr.ypractical and theoretical uses. An immecliate applicalion of the chonrsky normal forrn to parsing is giveri in ser:tion 6.3.

149


I 150

FoRMS ANDNORMAL or..Cournxr-FnEEGRAMMARS Chopter6 Slr,rpl,t.lcnrroN

The somewhattedious na,tureof the material in this chapter lies in the fact that many of the argumentsare manipulative and give little intuitive insiglrt. For our purposes,this technical aspect is relatively unimportant and can be read casually.The various conclusionsare significant;they will be usedmany times in later discussions'

ffiffi Methodsfor TronsformingGrommors we first raise an issue that is somewhat of a nuisance with grammars and Ianguages in general: the presence of the empty string. The empty string plays a rather singular role iil many theorems and proofs, and it is often necessary to give it special attention. We prefer to remove it from consideration altogether, looking only at languages that do not contain ,\- In doing so, we do not lose generality, as we see from the following considerations' Let .L be any context-free language, and let G: (V,T,S,P) be a contextfree grammar for -L - {I}. Then the grarnmar we obtain by adding to V the new variable ,96, tnaking Ss the start variable, and adding to P the productions

5o- ,91tr, * genera,tes.L. Therefore any nontrivial conclusion we c8,n make for , {I} grammar will almost cortainly transi'er to .L. AIso, given any context-free {I} G, there is a method for obtaining d such that -L (e) : , (G) (see Exercise 13 at the end of this section). Consequently, for all practical purposes, there is no difference between context-free languages that include .\ and those that do not. For the rest of this chapter, unless otherwise stated, we will restrict our discussion to .\-free languages.

Rule A UsefulSubstitution Many rules goverrr gerrerating equivalent grarnmars by means of substitutions. Here we give one that is very useful for simplifying grammars in various way. we will not define the tenn s'implffication preci$ely, but we will use it nevertheless. What we meaJl by it is tlte removal of certain types of undesirable procluctions; ttre process does not necessarily result in an actual reduction of the number of rules. Let G : (V,T,S, P) be a context-free grammar. Suppose that P contains a production of the forrn A --+r1Br2. Assume that ,4 and B are different variables and that

B+atlazl"'la"


6.1 MerHoos poR TRnl,rsFoRMrNGGnelrlreRs

LEt

is the set of all productions irr P vrhich have B as the left side. Let E : (V,T,S,F) U- the grammarin which F is constructedby deleting A - ntBrz

(6.1)

from P, and adding to it A --+n rarrz lnraznzl. . . lrtanrz. Then

,(*) =r(G) Proof:

Supposethat ,r.rr e L(G), so that S 4c tr.

The subscript on the derivation sign * is used here to distinguish between derivations with different grammars. If this derivation does not involve the production (6.1), then obviously

S36-. If it does, then look at the derivation the first time (6.1) is used. The B so introduced eventually has to be replaced; we lose nothing by assuming that this is done immediately (seeExercise 17 at the end of this section). Thus S 4c urAuz 46 u1r1Bfrzuz +G utixr,gjr:zuz. But with granrmar d we can get S 16 u1Au246 u1fr1!in2u2. Thus we can reach the sarne sentential form with G and d. tr 1e.r; is *secJ again later, we can repeat the argument. It follows thrrrr, lry induction on the number of times the production is applied, that

S46u. (E\. \ ,/ By similar reasoning,we can strowthat if w â‚Ź L(t), completingthe proof. I T h e r e f o r1ef,w â‚Ź L ( G )',,t h e n , e t

tnu" w e L(G),

Theorem6.1 is a simpleand quite intuitive substitutionrrrle; A production A - ntBrz can be eliminated from a f{rammar if we put in its prace


152

Fc;R.MS aNl NOR\,IRL Chopter6 SrurunrC.lrlgN OFCOlrrnxr-li'Run GrtRMH,laRfi

the set of procluctions in which B is replaced by all strings it derives in one step. In this result, it is necessary that A and B be different variables. The case when A : B is partially addressed in Exercises 22 and 23 at the end of this section.

o.t nxqmpli

Consider G : ({A, B} ,Io,b, c} , A, P) with productions A --+a,luu.AlabBc, B -+ ahbAlb. using the suggestedsubstitution for the variable B, we get the grammar d with productions A --+alaaAl aba,bltAclabbc, B + ahbAlb. The new gramma,r d is equivalent to G. The string aaabbchasthe derivation A+

aa,A4 aaabBc4 aaabbc

in G, atrd the correspotrdittg derivatiotr A+aaA+aaabbc ln

(j.

Notice that, irr this ca*se,the variable B and its a^lsociatedprgductions are still in the grammar even though they can no longer play a part in any clerivatiorr. We will see shortly how ulch unnece$$ary productions can be removed from a gralrfnar. I _

Productions Removing Useless Onn invariably wants to remove prodlctions from a glammar that can ne,ver take llart itt any derivatitln. FOr example, in the grAmrn1l wlxlse errtire production set is S-

a ^ 9 bl . \ l A ,

A-aA, the protluctiorr ,5 -+ A Clearly playS nrl l()ler a$ -4 CAnnqt be trarrsfortrtxl into a terminal string. While A can occur itr a string derived frorn 5, this can never lead to a sentence. Retnoving this production leaves the Ialguage unaffected and is a simplification by any definition'


6,'l

Mr:'r'sor$ FoR TRANSFoRMTNGGn.a,rr,rlrnns

lEB

Let G : (V,7,5,P) be a ctlntext-free granrmar. A variable A e V is saicl to be useful if and only if thenr is at least one ?rrâ‚Ź tr (G) such that

S# rAy 1w,

(62)

witlr r, y in (V U7').. In words,a variableis rrsefulif and only if it occurs irt at least one derivation. A variablethat is rrot usefulis called useless. A production is uselessif it involvesarrv uselessvariable.

[xornpl+ s'2

A variable may be u$e]essbecause there is no way of getting a terminal string from it. The casejust mentioned is of this kind. Another reason a varia,ble may be rneless is shown in the next grarnmar. In a grammar with start symbol S arrd productions 5-iA, A -

aAlA.

B-bA, the variable B is uselessand so is the production B - bA. Although B ciln derive a terminal string, thert: is rro way we can achieve S # nBy. I This example illustrates the two reason$why a variable is useless: either because it cannot be reached fiom the start symbol or because it canrrot derive a terminal string. A procedure for removing useless va-riables and productions is based on reco6pizing these two situations. Before we present the general case and the corresponding theorem, let us look at another example.

iinx.d$$frqirbisli*Eliminate uselesssvmbols atrd productionsfrom G

* (V,T,S,P), where v : {s, A, B,cI and ? : {4, b}, with P consistingof S - aSlAlC, A+at B -+ aa,

C - aCb.


Chopter 6 SlvtplIllcRtloN

oF Courpxr-FREE

GRAMMARSAND NoRMAr, Fonnrs

Figure 6.1

@ First, we identiff the set of variablesthat carrlead to a terminal r*ring. BecarrseA -+ a and B -r oG,,the variables A and B belong to this set. So cloes,9, becarrseS + A *# a. However,this argument cannot be made for c, thus iderrtifying it as useless, Removing c and its corresponding productions,we are led to the grammar Gr with variablesyl : {,S'A, B}, terminalsT : {a}, and productions S - aSlA, A+0,, B-aa, Next we want to eliminate the variables that cannot be reached from the start variable. For this, we can draw a dependency graph for the variables. Dependency graphs are a way of visualizing complex relationships arrd are found in many applications. For context-free grammars' a dependency graph ha-sits vertices labeled with variables, with an edge between vertices C and D if and ontv if there is a productiotr form

C * rDy. A dependerrcy graph for Vr is shown in Figure 6.1. A variable is useful only if there is a path from the vertex labeled s to the vertex labeled with that variable. In our case, Figgre 6.1 shows that B is useless. Removing it and the affected productions and terminals, we are led to the fi.nal answer

e : (t ,i, s, F) with fr : {S,A} , i : {o}, andproductions S --+aSlA, A --+a. The formalization of this process Ieads to a general construction and the corresponding theorem.

I

L e t G : ( V , 7 , 5 , P ) b e a context*free grammar. Then there exists an tnutdo*,not contain any useless equivalent gra*rrrur I : (t,.fr.s.F) \ / variatrlesor productions.


6.1 Merlrops non TRarqsFonMrNc GBenruens

lEF

Proof: The grammar I can be gr,,ncrateclfrom G by an algoritirur consisting; of twtr parts. In the fir'st paxt we rxrnstlLrct an intt:r.mediate graqrrrrar G : 1 (V,T2,5,P1) such that V1 contirirrsonly va,riablcsA for whir*r A4.Loâ‚ŹT* is possible. The steps irr the algoritirm irre; 1. Set V1 t,tt fr. 2. R.epeatthe ftrllowing step rrrrtil no rnore variables are arlcleclto v,. lbr every A e V for which l, has a, prorluction of thc forrn A -

rtnz' . .'rr,t with all irr in yr U ?,

adrl A to I/1. 3. Ta.k* 1l as all thr: productions irr -P whose syrrrbols are rr,I1in (yl u "). ck*r.r'ly this prornclure termi'*tes. It is eq'ally crear that if A e vr, tht:rr A 4 ru e 7* is a possitrle cleriva,tionv/ith Gr. Thc remaining issue is wlrether every,4 f'or which,4 $ rr : ah.'. is acrcledto v1 before the procedure tclrrninates. Ttr see this, cxlrrsiderany such ,4 ancl look at the partia,l rltlrivation trcc correspondirrg to that dcrivation (Figure 6.2). At level A, tht:re are only terrnina,ls,so cvery variablc Aa at level A - 1 will bel aclded to l/1 on the first pa"ssthrough step 2 of the algorithtr. Any va,riirlile at level a - 2 will thr+n be added t, v1 on the sex;ondpass through step 2. The third Lirne throrrgh step 2, all variables at level a * :l will be rr,rldecl, and so on. The algorithnr cilnnot terminir.terwhile therc a.reva,riable.sin the trcc that are nrt yet in I,l . Hence A will cventually bt: adcledto [. Irr thn seconclpart of the constructiorr, we get the lina,l answer t fro* G1. wc draw the va.riable dept:rrdencygraph ftrr Gl and from it find all

Figure 6.2

\,{ )

- T.evelr -2

- Level,&-1

-------Levcl,4


nnn Non,lral Fottr,t$ Chopter6 SrN{1'1,lprc.cr,roN OFCoNttlx'trli'ttEECIRAMMARS

r56

va,riables tha,t c:aunot be rtler,chedfrotrt 5'. These artl removed frorn the va,riable Set, it$ a,rethe proth.rctions involving them. Wtl t:an also elirnina,te tr,ny ternrirurl that d,oesrrot cicclnrin sotrrc rrseful prothrction. The rcsult is

'

A

t*.f.s.F\.

ttleglarlrrllirru:\v

/ ^ Becaust:6f the colstnrr:tion, G does not contain any trselesssyilrtlols or procluctions. Also, for each tu E L (fJ) wtr have a clerivation SltAylu;' proclrrc:tions'we $iuce the c:orrstructionof d retains A a,rrr.lall assoclia,ted Irave everyttring ueeded ttl rnake the tlt:rivation S t--6 nAy 11u ru.

1,tr.1 t , r ? + $

The grirrnma. d ir t:orrstructeclfrorn G by tlrt: rtlrnoval of prcitlucltious, Putting the two rt:srtlts sr tha,t F g p.()o.seq*cntly 1,(e) e LG). ^ \ / togethr:r, we see that G a,rrclG arc ttqrtivaleut' I

i.fi

.

{t f"

* : i

{ .Jf7

u r s f

RemovingA'Productions One kincl of llroc]uction tha,t is sornetirrrcsunclesiratrleis otre in which the right sicleis the ernpty string.

o

iF fll'ili!,,ic,nr"1f;i?,riilr"rr i1t,,,, Arry producliorr of a,cotrtext-fi'eegralrllrlirr of the folrn A-A is calledir,.\-production, Auy variir,bleA for which the deriver,tiotr

A+A

(ri3)

is uossible is crallerinullable.

A. graurrrrilr tnay gt:ritlrate a lauguir'ge not clonta,irringtr, yut have sornt: .\-llrocJucliorisor: nullir,ltlt:vtlriables. Irr suclt cascls,the .\-prorluctious cillr [rt: renroved.


6.1 Ml:'rHops l.'oR TRAN$FoRMTNC; Gnelrrverr,s

167

S n aStb, 51 -+ a51blA. This grammir.rgenera.testhe A-freelanguage {a,nb" : n, > i}. T}re .\-production 51 --+ ), ca,n be removed after addirrg new prtiductions obtairred by srrbstitutittg A for 51 wherc it occurs on thc right. Doing this we get the gr:Irrimar $'-

aS'rblab,

5r -

a.$rblab'

We can ea"silyshow that this n()w gramrrrar genr:rates the samr: language:as the clrigina,l otre. Irr rnore grlrreralsituations, substitutions for .\-protluctions can ht: rnade in a, similar, althougtr rnore r:ornplicil,ted, marrner. I

Let G be arrv r:otttext-fiee gr:r.rnmar with .\ not in L ((;). 'fhen there cxists an eqrlivalcnt gramrnar I having uo .\-productions. Proof: We lirst find tire set U1s of all mrlla.ble variables of G, usirrg the fbllowing strllls. 1. Fbr all productions A --+.\, prrt A into H1y. 2. Repeat l,he following step until no firrther vrrriables ir.readrlt:d to V,n,. For all nroductions fi _- 4142. . . A,,, where .41,Az,...,An art:in !fu, put B irtto l/r,'. Onc:c the stlt V7lr hirs been founcl, we are ready to construct F. we look at all productions in P of the fbrrrr

'Io

.1,, ro,

u l + a D r f f t " , f r m , r r r2l 1 , wherc each nt,( V U ?. For t:ach srrr:hprocluction of P, wc put into F tn,nt procluction ir.swell ar,sall th<lsegenerated lry repla,cing nullable variables witlr A in all pousible corrrbina,tions.Fcrl exa,mple,if z1 and r:,i are botlt nullable, tirere will he oue prorluction in F with fir r()placeclv/ith .\, one irr which ri is rcplaced with A, and onc iu which both ri and 11 a,rerepltr,t:edwith .\.


158

ANDNoRMAI Fon,vs GRAMMAR,sI Chopter6 Slrrrurlc.qltoN r)r' CoNrnx'r'-lr-nEE

Tht:re is onc exception: if all r,; fl,rernrlla,ble,the prodrrctiotr A -+ .\ is rrot Put into P. The argumerrt that this grauunnr d is cqrrivalerrt to G is straiglrtftrrwarcl atrtl will bo left to the rearlor. I

grarrrma,rwithout .\-productionsequiverleutto the gratrFirrcla t:rintext-f'ree ruilr defitrtldby S - ABaC, '1 - BC, B - blA, C --+Dl.\, D*d,. Fhomthc first stcp of tlrc coustntctiorrin Theorern6.3, we lirrrl that the rurllablevariablesare A, E, C, Therr, folkrwing the secondstep of the construt;tion,wc get

S - ABaClEn,ClAaClAB alaC lAalE o,la, A - BlClBC, B-b, ('--+ D, D -+ tl,

I

RemovingUnit-Produciions '-fheorent As we see from 6.2, prrlrfilctious in which both sides irre a single variable art: at tiurcs uncltlsirahle.

,i$fi'f,initien-'0,,i iiil,fii Arry production of a, contcxt-free grilmrrrirr of lhc fbrm A -->B, where,4, B â‚ŹV is called a unit-production.


i.1 Msrnons

noR TRehrsFoRMrNGGRerr,rr\l,rRs

159

To removeunit-productions, we use the substitution rule discussedin Theorem6.1. As the constructionin the next theoremshows,this carrbe done if we proceedwith somecare. Let G : (V,7,5, P) be any context-foee grammarwithout .\-productions. Then there existsa context-freegrammar e: (t,f ,S,F\ that doesnot / have any unit-procluctionsa"urlthat is equivalent\toG. Proofr Obviously,any unit-productiorrof the form A -- A can be removed from the grammar without effect,and we needonly considerA -+ B, where ,4 and B are different variablerl.At first sight, it may seemthat we carr use Theorem6.1 directly with rr : nz = .\ to replace A- B witlt

A-utlazl"'1a". But this will not alwayswork; in the specialcase A- B, U-4, the utrit-productionsa-renot rentoved.To get arourrdthis, we first Iind, for each,4, all variablesB such that

t4 n.

(6 4)

We can do this by dtawing a dependency graph with an edge (C, D) whenever the gtammar has a unit-prodtrction C -+ D; then (6.4) holds whenever there is a walk between .4 anci B. The new grammur d ir generated by first putting into F ail non-unit productions of P. Next, for all A and B satisfying (6.4), we add to P

A - y t l a t l" ' 1 a " , whereB + A1lyzl.. .U* is the set of all rulesin F with B on the left. Note that sinceB - !r,,lgl| ' . ly, is takerrfrom F, rroneof the yacan be a single variable, so that no unit-productions are createdby the last step. To showthat the resulting grarnmaris equivalentto the original one we can follow the sameline of reasoningas in Theorenr6.1. I


160

Chopter 6 SturlrnceuoN

Figure 6.3

---------.

o

oF Conrexr-FnnE

GRAMMARSAND NoRMAL Fonns

@[]

-:J

G x o m p l6i . 6

Remrtveall unit-productions from S * AtrlB, B _ Albb, A --+alhclB. The dependencygraph lbr the utrit-productionr is given in Figure 6.3; we seefrom it that ,9 + ,4,S + E, B + A, and ,4 4 B. Hence,we add to the origJinal non-unit prodrrctiotrs ff+Aa, A + albc, fi+bb, the new rules S - alhclbh, n --+bb, B * albc' to obtain.theequivalentgrammar 5 - a lbclbblAa, A --+albblbc, B ---+ albblbc. Note that the removalof the unit-productions has made ,B and the a*ssociated productions useless I We carr put all tltese results together to show that gramma"rsfor contextfree language$ can be made free of useless productions, A-procluctions, and unit-productions.

Let .L be a context-free language that does not coutain A. Then there exists a context-free grarnmal that generates .L and that does not have any useless productions, .\-productiorr$, or unit-productions.


Gn.q,rtlra.ns 161 6.1. Mnrsons roR TR,aNsFoRMING

given in Theorems6.2, 6.3, and 6.4 removethese Proof: The proceclures kinds of productiortsin trrrn. The only point that needsconsiderationis that the removal of one tyJre of production may introduce productions of another type; for example,the procedurefor removing .\-productions can createnew unit-productions. Also, Theorem6.4 requiresthat the grammar ha,veno A-productions.But note that the removalof unit-productions (Exercise15 at the end of this section),and doesrrot createtr-produc:tions thc removal of uselessproductions does not create A-productionsor urritproductions(Exercise16 at thr: end of this sectiorr). Therefore,we can renroveall undesirableproductions using the following sequenceof steps: 1. RemoveA-productions 2. Removeunit-productions 3. Removeu$elessproductions The result will therr have none of these productions, and the theorem is uroved. I -

1" Cornplete the proof of Theorem 6.1 by showing that ,5 :+d

rlj

implies D ;+G: tlr.

,

In Example 6.1, show a derivation tree for the string ababbbac,using both the original and the modified grammar. Show that the two Erammars

S + abABlba, A +

o..an,l

B - aAlbh and

S + ahAaAlabAbblba, A +

are equivalent. ffi

aaa,


L62

Chopter6 Slupr,rnce,rtoNoF CoNroxr-FREE GRAMMARS ANDNoRMALFoRMS

4, In Theorem 6.1, why is it necessaryto assumethat A and B are difierent variables? 5. Eliminate all uselessproductions from the grammax S - aSlAB, 4-bA, B-AAWhat Ianguagedoes this grammar generate? 6. Eliminate uselessproductions from g + alaAlBlC, A * aBlA, B*Aa, C - cCD, D - ddd. 7. Eliminate all .\-productions from S-

AaBlaaB,

A*I, B + b6,41.\. L

Remove all unit-productions, from the grammar

all uselessproductions, and all A-productions

S - aAlaBB, A + aaAl,\, B + bBlbbC, C-B, What languagedoes this grammar generate? ffi L

Eliminate all unit productions from the grammar in Exercise7.

10. Completethe proof of Theorem6.3. 11. Completethe proof of Theorem6.4. 12. Use the construction in Theorem 6.3 to removeA*productionsfrom the grammar in Example 5,4. What languagedoes the resulting grammar generate'l


6.1 Mer'sons FoR TRANSFoRMTNGGnertlreRs

163

1 3 . Suppose that G is a context-free grammar for which I â‚Ź .L (G). Show that if we apply the construction in Theorem 6,3, we obtain a new grammar d such

r h a r z ( S ): r ( G ) - { . U . 1 4 . Give an example of a situation in which the removal of .\-productions irrtroduces previously norrexistent rrrrit-productions. ffi 1 5 . Let G be a grammar without A-productions, but possibly with somc unitproductions. Show that the construction of Theorem 6.4 does not then introduce any,\-productions. 1 6 . Show that if a grammar has no .\-productions and no urrit-produrtions, then the removal of useless prodrrctions by the construction of Theorem 6.2 does not introduce any such productions. ffi

1 7 . Justify the claim made in the proof of Theorem 6.1 that the variable B can be replaced as soon as it appears.

r8,

Suppose that a context-free gramrnar G: the form

(V,T,S,P)

has a production of

A+fry,

where tr, l/ â‚Ź (y U ")t.

Prove that if this rule is replaccd by A+

By,

B+x,

where "B { V, then the resulting grammar is equivalent to the original one. 19;, Oonsider the procedure suggested irr Theorerrr 6,2 for the removal of useless 'l:pfoductions. Reverse the order of the two parts, first eliminating va"riables that cannot be reached frorrt S, therr renroving those that do not yield a terminal string. Does the new procedure still work correctly? If so, prt;ve it. If not, give a counterexample. 2O. It is possible to define concept of complexity of them is through the For example, wc might

the tcnn sirnplificatiorr precisely bv introducing the 'Ihis of a grarlimar. can be done in rrrany ways; one length of all the strings giving the productiorr rules. use

(G) : compt,e.nity

E

{1 + lul} .

A+uEP

Show that the rernoval of useless productions always reduces the complexity in this sense. What can you say about the removal of .\-productions and unit-productions?


164

GRenuaRs nNo Noti[4A.L FoRMs Chopter 6 SrllplrrrcATroN oF C<.rr'r'ruxr-FR,tlp:

?1,

A contcxt-free gramrrrar G is said to be rrrinimal for a giveu languagc .L if r:omplexi.t11(G) < torn7tlerZtlv (d) for arry d gcnetating f,. Show by examplc that the rernoval of useless produt:tions does not necessarily protluce a minimal gramnrar. ffi

*ZZ,

Prove the following resrrlt. Let G : (y,T,S,P) be a context-free grammar. Divide the set ef productions whose left sides arc soilre given va,riablc (say, /), into two disjoint subsets A+

AnllA*rl','lA*",

and

A *,!!rlazl ...1,!t,", where fii,;r7i are in (Ill?')-,

but A is not a prefix of any gl. Consider the

where Z f V and F is obtained try rc(u J{Z},T,s,F), placing all productit;rrs that have A on the left by gramrnar d:

A + yil11iZ, z: Z + xil'tiZ,

1,2,...,n1.,

t . : 1 , ? , . . . ,r t , .

rhentr(c):r(d). 23. Use the result of the precedingexerciseto rewrite the grarrrma,r A + Au,laEcl.\ B - Bblbc so that it no longerhas prtxluctionsof the forur A + A:r or IJ * Br, *24.

Prove the following courrterpartof Exercise22. Lct the set of productiorrs involving the variahle,4 on the lcft l-redivided into two disjoint subsets A+ntAlr,tAl .,,l*"A and

A + , s t l a z.l. . 1 u . , wher'e ,4 is not a suIfix of any :qi. Show that the grarnmar ohtained try rcplating these productions with A * 'ltrlZth, 'i : 7,2,...,n't, fr '* rr.lZnt, i :1,2,...,n' is etlrivalent to the original grarnmar,


6.2 Two lr,tponrenl Nonna,l Fonus

165

iffiffi Two lmportont Normol Forms Tht:re are nrany kinrls of norrnal forrns we cir.rrestahlish fot coutext-fi'ee grirrrrrrlars.Sone of these, br:c:ause of their widt: usefhlrress,havtl beetr studitxl extentively. Wc considr:r'lwo of thettr bricfly.

ChomskyNormqlForm One kirxl of nrlrrrral forrn we cir.rrlook ftlr is one in whit:ir Lhe mrrnber of s.ymholson thc riglrt of a prodrrc:tiotla,r()strictly liurited. ltr particular, wc can ask that thc: string rln the right of a productiotr consisLof no trtore tirirtr two symbols. One instance of tltis is the Chomsky normal form.

uR'funllfi ,ftftilrltiuw A contcxt-free grarrlrrlirr is in Chornskv rtorrna,l form if all production$ at'e of thc form

A _- BC, OI

A _-+ ill

whcre l, B, Ci are irr I,', a,ptl a is in 7'.

g'- ,45'la, A - $AIb iu iti Chornsky norrnal forrn. fhe gralrllnilr

fi - ASIAAS, A is rrotl both proclucticxts5 Dt:{irtition 6.4,

S Ala'tt'

AA5 attcl zt -

u,a violate the conrlitions of I


166

Chopfer 6 Stlrpllrtc,q,rroN oF CourFxt-FnEE

GRAMMARSAND NoRMAL FoRMS

Any context-freegrammar G : (V,T,S,P) with ^ # L(G) has an equivalent grammare : (?,i, s, F) in Chomskynormal form. Proof: Becauseof Theorem 6.5, we can assumewithout loss of generality tlrat G has no A-productionsand no unit-productions. The construction of G will be done in two steps. Step1: Construct a grammar G1 : (V1, T, S, .Pl) from G by consideringall productionsin P in the form

A+ntrz"'frn,

(6 I'r)

where each 14 is a symbol either in V or T. If z : I then trr must be a, terminal since we have no unit-productions. In this case, put the production into P1. If n > 2, introduce new variables Bo for each a â‚Ź ?. For each production of P in the form (6.5) we put into Pt the production

A - CrCz...Cn, where Ct = rt if z,' is in V, ancl Ca: Bo if ra : s. For every Bo we also put into Pr the production Ba --+ a'

This part of the algorithm removesall terminaJsfrom productions whose right sidehaslength grea,terthan one,replacingthem with newly introduced variables. At the end of this step we have a grammer G1 all of whose productionshave the form A--a,

(6 ri)

A - - +C t C z ' ' ' C * ,

(6 7)

where Cl â‚Ź Vr. It is an ea,syconsequenceof Theorem 6.1 that

l'(Gr) : L(G).


6.? Two lMponrerur'Nonrranr, Fon,Ms

167

Step 2: In the secondstep, we introduce additional variablesto reducethe length of the right sidesof the productions where necessary.First we put all productions of the form (6.6) as well as all the productiorrsof the form (6.7)with n:-2 into F. For n j 2, we introducenew variablesD1,D2,... and put into P the productions A -->CtDr, Dt * CzDz, : Dn-z + Cn-tCn..

Obviously,the resulting grammar d ir itr Chomskynormal form. Repeatecl applicatiorrsof Theorem6.1 will showthat f (Gr) : , (,d), so that \ ./'

, (") : LG). This somewhat informal argument can easily be made more precise. We will leave this to the reader. r

ffi

Sxnttpl+, fi;fi

Convertthe gramrnarwith productions S -+ ABa, A -+ aab, B -+ Ac, to Chomsky normal form. As requiredby the constructionof Theorem 6.6, the gralnmar doesnot have any tr-production$or any unit-productions. In Step 1, we introduce new variablesBo, Ba,B. and usethe algorithm to get S - ABB,, A - BoBoBa, B -+ AB., Bo ''

a,

Bn*b, Ba+c'


ANDNoRMALF onlrs Chopter6 Srrururlc,q'rroNoF'Colrnxr-Irrtnn GR,,+vrr,tnns

In thc secondstep, we irrtroduce additional variables to get tlrc first twrr producLiorrsinto normal forrn and we get the fina,l result S-

ADr,

Dt - BRo, A --+ BoD2, D2 -+ RnB6, p --+ABt, Bu. a

Q,

Bt, --+b, Bc+c'

I

GreibqchNormolForm Another usefrrl grammatical fotrn is the Greibach normal form. Here we put restrictions not on the lcngth of the riglrt sides of a prodrrction, but on the positions in which tcrminals ir,nd variirhles carl appear. Arguments justifying Greibach normal ftrrm are a,Iittle complicated and not very tra,nspartrnt. Sirrrilarly, constructirrg a grarnmar irr Clreibach normal form ertlrivalerrt to a given context-free grammar is tedious. We therefore deal with this rnatter very briefly. Nevertheless, Greibach trormal forrn has marry theorctical arrd practica,l conseqrrences.

,nelf'l;ri,!,ii1qniy.ffi tN A txrntext-free grarnmar is sa,id to be in Grcibach rrormal forrn if all productions havtl the forrrr !

--+ o,fr,

whercaâ‚Ź.Iandrâ‚ŹV*

If we compare this with D<:fi,nition 5.4, we $ee that the form A * o,r is comrtron to both Greiba,ch normal forrn and s-grammar$i but Greibach normal form does not carry the restriction that the pair (-4, a,)occur at most once. This additional fret:dom gives Greitra,ch norrnal forur a generality not lrossessedby s-grammars.


6.2 Two IturoRrnrqr NoRl"ler.FoRrrs

169

If a gramrnar is not in Greibach normal form, we may be able to rewrite it in this form with sorne of the techniques encrluntered above. Here are two simple examples.

S+AB, A -- aAlbBlb, B ---'b is not in Greibach rrormal form. However, using the substitution given by Theorenr 6.1, we immediately get the equivalent grammar

S + aABlbBBlbB, A --+aAlbBlb, B-b, which is in Greibach normal form. I

Exoilrple 6,10

Convert the grammar S-

abSblaa.

into Greibach normal form. Here we carl use a device similar to the one introduced in the construction of Chomsky normal form. We introduce new variables ,4 and B that are essentially synonyms for a and b, respectively. Substituting for the terminals with their a"ssociatedvariables leads to the equivalent grammar S --+aBSBlaA, A+d, B+b, which is in Greibach normal form. I In general, tltough, rreither the conversion of a given grammar to Greibach normal form nor the proof that this can always be done are $imple matter$. We introduce Greibach normal form here because it will simplify the technical discussion of an important result in the next chapter. However, from


Chopter6 SrupllnrcauoNoF Colrruxr-FnpnGn-q,uve,ns ANDNoRMAlFoRtvrs

a conceptualviewpoirrt, Greil-rachnormal form plays no further role in our discussion,so we only quote the following generalresult without proof. For everycontext-freegrammar G with ^ # L (G), there existsan equivalent Bramrrar d in Greibachnormal form.

1 . Provide the details of the proof of Theorem 6.6.

,

Convert the grammar S + ttSblab into Chomsky normal form.

3.

Ttansform the gramma.r S form,

aSaAlA, A -

abAlb into Chomsky normal

4. Transform the grammar with productions

S * abAB, A + bABl,tr, B - BAalAlA, into Chomsky normal form, 5. Convert the gramma,r

S - ABIaB A - aabll B-bbA into Chomsky normal form. ffi 6 . Let G : (y, ?, S, P) be any context-freegrammar without any tr-productions or unit*productions. Let k be the maximum number of symbolson the right of any production in P, Show that there is an equivalent grammar in Chomsky normal form with no more than (k - 1) lPl * lTl production rules.

, . Draw the dependencygraph for the grammar in Exercise4. L A linear languageis one for which there exists a linear grammar (for a definition, see Example 3.13). Let Z be any linea^rlanguage not containing ,\. Show that there exists a grammax6 : (V,T,,S,P) all of whoseproductions have one of the forrns A+aB, !+Ba, A-e,, w h e r eo E T , A , B E V , s u c ht h a t . I ,= r ( G ) ,

m


6.2 lI'Wo Ir,rpoRrAru.rNoRMAL Fon,prs

171

9, Showthatfrrrevervcontext-frcegrarrrtrarG:(V,T,S,P)thereisaneqrrivalent one irr which all pr'64u"tr1on.rhave the forrn A + a,BC,

or

,,1* tr, ffi

w l r c r ea € X U { I } , A , B , C € V . 10, Corrvert tl:regraurnar S -

aSblbSola,lb

irrto Greibar:h rrorma,l forrn. 11, Convert the following grarrrrilar into Greibach norrnal form. S + rt,Sblab. 12, Convert the grarnrnar g + ab loSlaa5 into Greibach norrnal form.

W

13. Convert the 6grammar S + ABblo,, A+

uaAlB,

B -

ItAb

into Greilrach norrrral form. 14. Can every lincar grar[mar be convcrted to a fbrm irr which all productiorrs Iook like A - an,where a e ? and tr € yU {,U? 15. A context-free grammar is said to be in two-standard forur if a,ll production mles satisfy the following pa,ttern ,l + o,BC, A-aB, A_+e1 wher'eA,B,CeUandae?. Convertthe grammar G:

({S,A,B,C},{n,,b},S,P)

witli P givena,s

S + a,9-4, A + bABC, B -b, e + aBCt irrto two-starrdard forrn. *16,'I'wo-$tandard

ffi

f<rmr is generall for any context-lrce grarrrmar G with A { ,L(G), there exists an ecluivalenf grammar irr t,wo-sta,rrdard form. Prove this,


Chopter 6 Stltrurlc;a'r'loN oF CON:rEXT-FRljtl GRnr,tuaRS At'ro Non.lvlel, F ORl{s

iiffiffiffitil A MembershipAlgorithm for ContextF r e eG r o m m q r s * In Ssrtion 5.2, wt: cla,itned,without any elaboration, that mcmbership and parsing erlqorithttm f'or cotrtext-fiee grarnmars exist that retpire approximately ltol'' steps to parse a string trr. We are lr()w in a position to justify this clrrirn. The algorithm wt: will describe here is t;tr,lledthe CYK algorithm, a,fter its originators J. Cocke, D. II. Vlrnger, arxl T. Kasarni. The algorithrn works only if the grarnrrrar is itr Chomsky rrormal forrn aucl sutr:eeds by breaking one pr()hlem into a $equell(xl of stnallt:r ones irl the following wir,y. Assume that we trave a grarnmar G : (V,T', S, P) in Chornsky llorlllill forrn and a string 'til:eraz..'en,

We definesubstrings wij:(ri"'ajt

and sulnetsoI U

vi : {e Ev :A5 u,,,}. Clearly, u E lt (G) if and orrly if S e Vr",. ancl only if G contains a, proTo cornpute Vl.i, r.rbservethtr,t A â‚ŹViif drrction A -- ai. Therefortt, Vil call llt+ computed for all I < z < rz by inspection of u arxl the prtlrluctions tlf the grarrrmar. Ttl r:ontiurrc, notice tlrnt for j > i,, A derives ruii if a,nclonly if tltere is a, produt:tion A--+ BC, with-B 4r1* v,ij:

a . r r dC 1 ' , r , 1 " a 1 f, oi r s o m e f u w i t h i < h , h < r . I n t r t h e r w o r d s ,

U

{A: A-

BC,wit}rBEHtr,câ‚ŹVr+r,r}.

(6.S)

LE {r;,rt l,.,.,j - i I

An instrrrx:tiono[ t]rc indices in (ti.8) shows that it catr bc used to t:ompute rr.llthe F1r.if we prot:ced in tirtl secluetrctl l . C t r m p u t eV r , V z z , . . . , V n , 2. Ctimpute Vrz, V2r, ..., tr/r*1,r, , ,-2,,, 3 . C t r m p u L eV t : t , V 2 t , , , ,W attd srt ott,


0.3 A MnlrnnnsHlr Ar,colr,rrrtl,rFon {loFl'r,r.xr-FnnnClHe,l,rn.tns

Ultqftn$f$ld;'Il',

Dcttlrmine whetlrrrr the string In: tlul H-ratnutar

L7B

(rabbbis irr the langua.gt:generatecJlw

5-AR, A _ BBla, R -+ ABlb, First rxltrl Lhi'rl'ur1I ,: .r,iso V11 is thc scl of all vir,riablesthat irurric:rliately rk:rive a, that i*, l'lr: {,4}. Sirice ur22: e,,wc irlso havt: yz:: {,4} at:ci, similarly, E r : { , a } , V t t : { A } , V r , : { B } , V 4 a= ={ E } , Y r r : { B } Now we use (6.8) to get I4z : {A : ,! --+B(.i, B E Vtt, O t V22} , Since V11= {A} ancl V22: {A}, the sct c:onsistsof a,ll variables t}iiit occnr ott Lhe lefl ..ridcof a prorluction whosu righl sicle is A,4. Since tlrt:re ale nrltre, V12 is crrtllt,y,Next,

V4 : {4., } _-tB(i, B â‚Ź Vn, (_,'e t/rr} , strtlre reclrtircrl righl sirleis AB, tinrlwc haveH23: {S',8}, A straiglitft-rrward argrrrnt:ltalongtlu:srlIinesthrlu gives Vn: Y r 3: Vr+: V ' n:

fr,I/rr : {5, RI ,Vt+: {l} ,l/+o: {A} , { 5 , B l , l / z r : { , 4 } ,V r r : { 5 , 8 } , {A} , V:s : {S, R} , { 5 ,R } ,

s ot h a t u t e L ( G ) . I

Thc CYI( algorithru, as clrrst:r'illccl her:e,dctt:rrrlinesrncrnbershipfor ir,riy langrra,gcgr:leraterl lr; a grarnrnal iu flhonr,sky lolrng,J frirrn. With sorlc a,dditiorrsto keep trac:k of how tlrc elertrentsof l/1, alc r1r:r.ivecl, it r:arr l_re qrnvcrted iulo ir,irarsirtg rnetlxltl. Iir see thrr.t the CYK rnemlrershill algoritlrrrr lecluiles ()n,:r sl,e1-rs, ntitic:c Lhat exa,ctly rr,(rz -1.1) /2 sets of V;.; have to lltl cotrtl:rtted. Eir,chinvtilvr:s 1,heeva,lrrntiorrof a,t rriost ll, terms irr (ti.$), so tlrc cla,inrt-xl rt:srrlLli-rilows.


774

Chopter 6 Srrvtt,r,rprcetloNol Corvrrxr-Fn,Ilri GRAMMARSAND NoRMAI Fonlrs

1

Use tlre CYK algorithm to dcterrrrirte whethcr ihe strings nabb, aabba,,and obtibb a,re in the language generated by the gtarnmar irr Examplc 6'11'

2 . Usc the CYK algorithm to find a parsing of the strirrg aab, using tlre gratnrrrar of Examplc 6,11. ffi

3 . Usc the approach crnployed in llxcrcise 2 to show how the CYK mcmbership algorithm can be trrade itrto a parsing metlxrrl.

* * 4 . IIse the rcsult in Exercise ll to write a cotnputer prograln firr Jrarsing with any contcxt-free gra.mfllat in Chomsky norural form,


P ush d o w n Automoto

he description of cnntext-free languages by means of context-free gramrrrars is convenicrrt, as illustrated by the use of BNF in programming language dcfinition. The next question is whether there is a class of automata that can be associated with context-free langua,ge.s.As we have seern,finite automata cannot recognize all context-fiee languages. Intuitively, we understand that this is because finite autotrara have strictly finite rrrernories,whereas the recognitiorr of a context-free language may reqrrire storing an unbounded amount of information. For example, when scanning a string from the language 7 : {a"b,, I n } 0}, we mu,st not only check that all a's precede the first b, we rnust also corrrrt the nrrmber of a's. Sinc:en is unboundrld, this counting cannot be done with a finite memory. We want a machinc that can count without limit. But as we see from other examples, su()has {wwR}, we need more tharr unlimited cxrurrting ability: wc rreed the ahility to store and match a sequence of symbols in reverse ortler. This suggests that we might try a stack as a storage mechanism, allowing unbounded storage that is restricted tci operating likr a stack. This gives us a (lass of machines called pushdown automata (pda)'

175


Chopter 7 PusHooww Aurorr,r.tre

Input file

Figure 7.1

In this chapter, we explore the connection between pushdown autornata and context-free languages. We first show that if we allow pushdown arttomata to act nondeterministically, we get a class of automata that accepts exactly the family of context-free languages, But we will also see that here there is no longer an equivalence between the deterministic and nondeter* ministic versions. The class of deterministic pushdowrr automata defines a new family of languages, the deterministic context-free languages, forming a proper subset of the context-free languages. Since this is an important family for the treatment of programming language$, we conclude ttre chapter with a brief introduction to the grammars associated with deterntittistit: context-free languages.

M

Automoto N o n d e l e r m i n i s tP i cu s h d o w n

A schematic representation of a pu$hdown automaton is given in Figure 7.1. Each move of the control unit reads a symbol from the input file, while at the same time changing the contents of the stack through the usual stack operations. Each move of the control unit is determitred by the current input symbol as well as by the symbol currently on top of the stack. The result of the;move is a new state of the control unit and a change in the top of the stack. In our discussion, we will restrict ourselves to pushdown automata acting a.seccepters.

Definitionof q PushdownAufomoton Formalizing this intuitive notion gives us a precise definition of a pushdown automaton.


7,1 NowlnrER.MrNrsrIcPrJsurrowruAuroltnrA 177

,irRR.$lr|"q.,l..iirqnri,,fil,lll lr A nondeterministic pushdown accepter (npda) is defineldtry the septrrple M : ( Q , 8 , 1 , , t , q o 1 z 1 F|) whcre Q is a finite set of internal sta,tesof the control unit, X is the input alphrrbet, I' is a flnite set of syrnbols ca,llerlthe stack alphabet, d ; Q x (E U {A}) x l' - finite suhscts of Q x f* is thc transilion functirin, Qoâ‚Ź Q is the initirrl state of the corrtrol unit, r â‚Ź | is the stack start symbol, F e Q is the sct of linal states.

The ulrnplicated formnl appearance of the drlrrrairr and range of d rnerits a closer o'xamitration. The irrgutnentsof d a,rethc current state of the control unit, tht: current input syrnl-rol,and the currrlrrt symbol on toJ) of the stac.k. The re:sult is a set of ptlirs (q,z), where q is the next state tlf the control unit trnd r is a string whiclt is pnt on tcill of the stack irr place of the singlc syttrbol there before. Note that thc second argument of d may be .\, irrdicating that a move that does not collsurne an inprrt symbol is possible. We will call such ir rrlove a .\-tra,nsitirirr. Note also thnt d is defined so that it needs a, sta,t:ksymbol; no m()v()is possible if thrr stack is emptv. Finallv, the requirernetrt that the rarrge of d be a finitrl subset is ner:essarybeca,use I x f* is atr infinite set ir,rrdtherefore ha"r irrfirrite subsets. While an npda may have several choit:cs for its move$, tiris choice must be restricted to a firrite set of possibilitics. ________-____'Emunrpl+ f - | Suppose the set of trir.rrsitionrules of arr npda crrntains d ( , t r ,a , b ) : { ( , 1 r ,c d ) , q 3 , , \ } . If at any time the control unit is in statc q1, the input svurbol read is a, irnd the symbol on top of the stack is b; then one of two thirrgs can ha,ppen: (1) the t:otrtrol unit gJoesirrto state {2 ancl t}re string cd replirces b on top of t}re stack, or (2) the u)ntlol unit goes irito state qs with the symbol b rcuroved frotn the top of tht' stack. In our rrltation we &ssurnethat the instrrtion of a string into a,stack is done symbol by symbol, stir,rting at the riglrt end of the string. I


178

ffi

$xomple 7.2

Chopter 7 PusHnowN AuroMlt'rA

Consideran npda with Q : { q o , Q t , Q z , q,t } p : { a ,b } ,

r : { 0 , 1 ,} z:0,

r': {qr}, and /-,

..-di : {(q',10),(q3,.\)}, d (sor.$;o) d ( q 0 , . t r ,:0 ){ ( q 3 , I ) },

6 ( q r , a , 1 ): { ( q 1 , 1 1 ) ,] d ( q t , l r ,1 ) : { ( q 2 ,A ) } , 6 ( q r , 1 , , 1 :) { ( q 2 , , \ ) } , d(qz,A,0):{(q3,I)}. What can we say about the action of this automaton? First, notice that transitions are not specified for all possible combinations of input and stack symbols. For instance, there is no entry given for d(qo,b,0). The interpretation of this is the same that we used for nondeterministic finite autoutata: an unspecified transition is to the null set arrd represents qdead configuratiqllfor the npda. The crucial transitions are {I--=....'_

6(h,a,1) = {(q1,11)}, which adds a 1 to the stack when an o, is read, and 6 ( q r , b , l ) : { ( q 2 ' . \ ) }, which removes a 1 when a b is encourrtered. These two steps count the mrmber of a's and match that count against the number of b's. The control unit is in state (1 until the first b is encountered at which time it goes into state q2. This assures that no b precedes the last a. After analyzing the remaining transitions, we see that the npda will end in the final state q3 if and only if the input string is itt the language 7,-{anb":n>0}u{"} As an analogy with finite automata, we rnight say that the rrpda accepts the above language. Of course, before making $uch a. claim, we must define what we mean by an npda accepting a language. I


7.I NoNDETERMrNrsrrc PusHoowruAurouarn

l7S

To simplify the discussion, we introduce a convenient notation for de* scribing the successiveconfigurations of an npda during the processing of a string. The relevant factors at any time are the current state of the control unit, the unread part of the input string, and the current contents of the stack. Together these completely determine all the possible ways in which the npda can proceed. The triplet ,n/ld Jfn"J

(5 {r, lc )

t,;,r'ol u^-rt

L t' '+-

j,,t,70/

JTdeF

a

where q is the state oJ_tlp_gqntrol uaiL tr.' is the unread part of the input string, and u is the stack contents (with the leftmost symbol indicating the top of the stack) is called an instantaneoua description of a pushdown automaton. A move from one instantaneous description to another will be denoted by ihe symbol F; thus (qt,aw,b!) | (qz,y,U!) is possible if and only if i t ( q z , a )â‚Ź d ( q 1 ,a , b ) . Moves involving an arbitrary number of steps will be denoted Uy [. On occasions rvhere several automata are under consideration we will use Far to emphasize that the move is made by the particular automaton M.

TheLonguogeAcceptedby o PushdownAutomolon

lllfim$,tm$ilnl ril, irtri (8,E,f,d,80,a,F) be a nondeterministicpushdown automaton. Let M: The language accepted by M is the set

L(M):

{,

e r . i ( q o , w , r ) i u @ , 4 , u ) ,ep F , u e f . } .

In words, the language accepted by M is the set of all strings that can put M into a Iinal state at the end of the string. The final stack content z is irrelevant to this definition of acceutance.

; , , i 1 ,t , i , , t i ; , i '

ii,l i

Ht(dftfpls y,S

Construct an npda for the language L : {* e {a,b}* : no(w) : nu(ru)}. As in Example 7.2, the solution to this problem involves counting the number of a's and b's, which is easily done with a stack. Here we need not even


Chopter 7 PusHnowu Auronere

w6rry a|out tlte order of the a's atrd b's, We ca1 iusert a counter symbol' say 0, into lhe stack whenever an a is read, then pop olle counter symbol from the stack when a, b is fbund. The only difficulty with this is that if wtl will not find a,0 to usr:. But there is a prefix of ur with more b's thiln r?,'it, this is easy to fix; We can ll$e iI negiltivtl cotrntttr symbol, sa,y 1, ftlr t:ourrting thel b's that a,rc to trc rnirtchtxl irgairrst ats later. Tlte cotttplete solution is r1o,z, {r1l}), with ri given as ir,nnpdrr,A'[: ({rlo,q.fI , {&,b} , {0, 1, z} , 15,

.I (qo,,tr,z) : {(qt, z)} , d ( q o ,a , z ) : { ( q 1 1 , 0 4 )' } 1z)} , d (rJs,tr,r) : {(q11, d ( q o ,a , 0 ) : { ( r t 6 , 0 0 ) }' d ( q o ,b ,O ) : { ( q n ,I ) } , d ( q o ,a , f ) : { ( q n , ' \ ) } ' d ( , j n ,b , 1 ) : { ( q 6 ,1 1 ) } . In processing the string baotr, the nprla, mtr,ke$the move,s ( q s , b a a b , z )f - ( g o ,a a b , ' I z )F ( q 0 , a b ' a ) l- (qo,b,0z) F (qg,.\, z) F (qr, tr, t) alrd hcrrt:t:thcl strirrg is ircx:epted.

. 'W WMWWiN*t:tlrrstrtrtltilnrrpclilfilrtr,cx:tlptirrgt}rtllirrrgrri'r,ge -

I, :

(

p

\unun

r

, - r \

: ur â‚Ź {ct,,b}-} ,

we use the fact that the syrnbols are retrieved frorn a stack itr the reverse or<lerof their insertion. Wlren reacling the first part of the string, we push rxlrrsgr:rrtivr:svrnbols rirr tiru sta,r:k. For thc stx:tlnd pa,rt, we r:ompi.r,rethe current input symbol with Lhe top of lhe stack, cotrtinuitrg as lotrg as the

?iil: ix,lT*il:,:lT:'"'*ni[::r:]:. #,.i;i:i1,1]:f ''An

apparent clifliculty with this suggcstiorr is th.r,t wtt rltt rrttt know tfie rniddle of the strirrg, tirirt is, w]rttrc tl cnds ilrrd rirE sttlrts. Brrt the lolcleterrlinistic rrature of the irutornatr)rr hclllls us with this; thtr npda correctly guesseswhere the rrriddle is artd switt:hes stirtes at that ptlirrt. A s o l u t i o nt o t h e p r o b l e m i s g i v e u b y M : ( 8 , 8 , I ' , 6 , q o , 2 , i ' ) , w h e r e Q : {,1(t,Qt,4z}, y:{a,b}, | : {a,b,z}, F : {arl


7.1 NolrnnrERMrNrsrrcPusHlowl Arlrouara

181

The transition firnction can be visualized as having several parts: a set to push zr orr the stir.r:k,

\ / )

J""sh

I

{ )

d (so,o, a) : 6 (qo,b,a): d (so,o, b) : (t(,10,1r, b) :

{(so,aa)} , {((t(r,b(z)}, {(qs,ab)} , {(qo,bb)},

d ( Q o , a ,z ) :

{ ( q g ,a e ) } .

d ( q o , b ,z ) :

{ ( q o ,b r ) } ,

a set to guesstlre middle of the string, where the npcla switches frorn state 9r1ttl qt '*-d ( q o ,A , a ) : { ( q t , a ) } , d ( q o ,t r ,b ) : { ( q t , b ) } , a set to uatch zrrRagainst the contents of the stack, 6(h,a,0) : {(q1,I)}, : , ){ ( q t , I ) } , d (,1r,1r,1 and fina,lly d ( q t , A , , ) : { ( q z ,r ) } , to recognizea successfulmatch. The sequence of moves in a,ccepting abba is (qo,abba,z) t (qo,Itaa, az) | (Qo,ba, baz) F ( q 1 , b a , b u z ) l( t 1 1 , o . , n F 2 )( g r , A , . z )F ( q 2 , . e ) . The nondeterministic alternative fbr loca,ting the miclclle of the string is taken at the third move. At that stage, the pcla ha-r the instir,ntaneo's descriptions (qo,ba,baa) a'rl ha,stwo choices fbr its next rnt)ve. c)rre is to uue d(gs,b,b) = {(qo,bb)i ancl make the move (q6,bu,rt, baz) | (qo,a, bbaz), the secondis the one used above, namely d(q0,,\,b) : {(qr,b)}. Only the latter leads to acceptanceof the irrput.

I


Chopter 7 PusHnowt-rAurol\lArA

1 . Find a pda with fcwer than four states that accepts the same language as the pda in Exarnple 7.2. D,

Prove that the Jrda in Exarnple ?.4 rloes not accept any string not in {tutrtF}'

ffi C.,nrt.lrct nptla's that atxrept the followirig regular languages'

fD

(a) Lr : L(aaa.b) (b) IIr : L(aa,h"aba-) ({

the urrion of trr an<l Lz

(d) Lt - Lz

languageson X : {4, b,c}' fD Construct npda's that acceptthe following (a)L={**bt*,rz>0} W (b) .L : {tnctu'r : ru e {n,,b}* } (") f.L: {.a^h*c'+*: n, } 0,- > 0} ( . 1 f) : { o " ^ h + * c * : n } Q , m } 1 } (*) t:

r z: > 0 } {a3b'oc'o

(f)r:{a*b*:nlnz/-3n}

W

( S ) r : { u : n o ( * ) : r r u( t , ) + t } ( I t ) . L : { u t : n o ( u t ): t r , 1 - 1 y (i)

{*

'. n,, (to) + nr, (u) : Tr..(tu)}

(j) .L : {'u:2no (trr) { n;, 1tr,)< 3n" (ta)} (k) .L = fw : no (tr,) < na (tr')] )<rrrstrtrt:tarr nlrda that ac:t:eptstfue language 7 : {anb* : n' } Q,r1 f v,ltt. Firrrl arr nprla on 5 : {n,,b, c} t}rat accepts t}re language L :

i t t r tl ' t r t â‚Ź z { r r ,b } * , r u , I * { } {zntcutz

'

?, Filtt arr rrpda for'the cotrcatetratiotrof tr(a*) and the language irr Exercise 6. @

fitta an trpda lbr the language 7 = la'b(ab)" b (ba)" : n > 0].

9. Is it possiblc to lincl a dfa that acccpts the sarne language as the pda M : (Iqo, qr]1, {a,,1,}, {t} , qo,{qr }) ,


7-1 NoNnerERMrNIsTrc PusHpownrAurouArR

1BB

with d ( q o ,a , z ) = { ( q r , e ) } , d (qo,tr,e) : {(So,e)} , 6(qr,tt,z): {(q',r)}, d ( q r , b , " ) = { ( q o ,z ) } ? f f i 1O. What language is accepted by the pda M :

UrIr, qt i qil qli 4+, qsj , {a, b} , {0, r, a} , qo,{qsl) ,

with , ,z ) : { ( q 1 ,1 z ) } , d ( q ob 6 ( q t , b , 1 ): { ( q ' , r r ) } , d ( q r ,o , 1 ) : { ( s . , } ) } , d ( q . ra, , 1 ) : { ( q n ,f ) } , 6 (q+,u, z) : {(g+,z) , (gs,z)l'l /+\

1.,

S1,/Wtrat languageis acceptedtry the npda M = ({qo,$,gzl ,{a,b},{o,b,a}, 6,Qo,z, {qz}) with transitions 6 ( q o , a ,z ) : { ( S r , n ) , ( g 2 , . \ ) } , d ( q r , b ,a ) : { ( q t , b ) } , d(qr,b,b): {(q,,0)}, 6(qr,a,b): {(Sr,A)} ffi 12. What language is accepted by the npda in Example 7.8 if we use F : {qo,q'}? 13. What language is accepted by the npda in Exercise 11 above if we use F : {qrt,$,qz)? 14. Find an npda-lvith no n:rorethan two internal states that accepts tlre language

tr(aa*ba*).ffi

15. suppose that in Example 7.2 we replace the given vaJue of d (g:1,A,0) with

,5(sz,tr,0): {(qn,A)} . _ (rel

What is the languageacceptedby this new pda? we can define a restricted npda as one that can increa,$ethe length of the stack by at most one symbol in each move, changing Definition ?.1 so that d : Q x (Xu {I}) x | + ?Qx(rrufu{r}). The irrterpretation of t,his is that the range of d consists of sets of pairs of the ftrrnr (q6,ab) , (qt, a,), or (Q,,A). Show that for every npda M there exists such

a restrictetlnpda fr suchthat L (M) : L (fr)

ffi


Chopter7 PusunowNAuroMArA

184

dD -

O" alternative to Definition ?.2 for languageacceptanceis to require the stack to be empty when the end of the input string is reached. Formally, an npda M is said to accept the language N (M) by empty stack if

N ( M ) : { t u . E . , ( q o*,, r ) i * 1 p ; , r ) } , L

J

where p is any element in Q, Show that this notion is effectively equivalent to Definition ?,?, in the sense that for any npda M there exists an npda fr such that L(M) = t (fr), atrd vice versa'

M

P u s h d o w nA u t o m o t oo n d C o n t e x t F r e eL o n g u o g e s

In the examples of the previous section, we saw that pushdown automata exist for $ome of the familiar context-free languages. This is no accident. There iS a general relation between context-free languages and nondeterministic pushdown accepters that is established in the next two major results. We will show that for every context-free language there is an npda that accepts it, and conversely, that the language accepted by any npda is context-free.

[onguoges PushdownAutomolqfor Conlext'Free

-1 "1d "* 7{t-tilrtl+-jh ")t)F

.P*gnrnlS.,fir,5|

We first show that for every context-free language there is an npda that accepts it. The underlying idea is to corEtruct arr npda that can, in some way, cerry out a leftmost derivation of any string in the language' To simplifii the argument a little, we assume that the language is generated by a grammar in Greibach normal form. The pda we are about to construct will represent the derivation by keeping the va.riables in the right part of the sentential form on its stack, while the left part, consisting entirely of terminals, is identical with the input read. We begin by putting the start symbol on the stack. Aft'er that, to simulate the application of a production A + afrt we must have the variable A orr top of the stack and the terminal o as the input symbol. The variable on the stack is rerrroved and replaced by the variable string r. What d should be to arhieve this is easy to see. Before we present the general argument, let us look at a simple example.

Construct a pda that accepts the languagegeneratedby grammar with productions 5 - oSbblo,.


7.2 PusHoowr-r Aurordaraeuu CoNrnxr-FRnrLANGuAcEs IBE

We first transform the frammar into Greibach normal fbrm, changingthe productionsto 5'-+ a1Ala, n ---+ bB, B --+b. The corresponding automatonwill havethreestates{go,qr, gz}, with initial state gs and final state q2. First, the start symbol ,9 is put on the stack by d (qu,A,") : {(qr,S")} . The production 5 --+ aSA will be si*rlate4 tr lbq]p4gby_fg*oViqg g to*

th* tto,"k o.tt.l "o['lo."it g it o.ith-

Similarly, the rule ,5 - a should cause the pda to read an a w.h:ilesimply removing S. Thus, the two prodrrctions are represented in the pda by d ( q t , a , S ) : { ( s r , 5 , 4 ) , ( q 1 ,A ) } . In an analogous manner, the other productions give

6 ( q r , b , A:) { ( s r ,B ) } , 6 (h,b,B) : {(qt,f)} . The appearance of the stack start symbol on top of the stack signals the completion of the derivation and the pda is put into its final state by d ( g r , A , z ) : { ( q 2 ,I ) } . The construttion of this example can be adapted to other cases,leading to a general result. I

For any context-free language Z, there exists an npda II such that L:

L(M).

Proof: If tr is a A-free context-free language, there exists a context-free grammar in Greibach normal forrn for it. Let G : (V,T,S,P) be such a grarnmar. We then construct an npda which simulates leftmost derivations in this grarnmar. As suggested, the simulation will be done so that the unprocessed part of the sentential fornr is in the stack, while the terminal prefi.x of any sentential form matches the corresponding prefix of the input string.


Chopter 7 PussoowN Au'roMA'I'A

Spar:ificallg the npda, will be M : ({qo,{r,{l},T,V l{z},d,rlo, z,{gl}), whetrez I V. Note that the input alphabet oI M is identical with the set of terrninals of G and that the stack alphabet <ffitains the set of variables of the grammar. The transition functiotr will include

( 7l )

d ( q o ,A , a ) : { ( q 1 , 5 a ) },

so that after the first move of M, the stack contains the start symbol ,9 of the derivation. (The stack start symbol a is a marker to allow us to detect the end of the derivation,) In addition, the set of transition rules is such that

(7.2)

( r n , u )€ 6 ( q 1 , aA, ) , wllenever

A-

au

is in P. This reads itrput a and removesthe variable A from the stack, replacingit with u. In this way it generatestlte transitiorrsthat allow the pda to simulate all derivations. Finally, we trave

( 73 )

d ( q r ,A , z ) : { q t , z ) } ,

to get M into a final state. To show that M accepts any ru € L(G), consider the partial leftmost derivation

'**.::::

.";i:o';rf

A*

If M is to simulate this derivation, then a"fterreadirtg ar&z" '4r,, the stack must contain AtAz' ' ' A^. To take the next step in the derivation, G rnust have a production A1 + ttBl , .' Bp. But the construction is such that then M has a transition rrile in which ( q r , B r . ' . E r ) € d ( q r ,b , A r ) , so that the stack now contains Br ' ' ' BnAz'" An after having r€fld a1a2' ' 'anbA simple induction a,rgument on the number of steps in the derivatiorr then shows that if

E+u,


7.2 Pusnnown Auroua.rA ANDCor-{rnxr-FnEELANCUAoE$

fgz

then ( c 1 1 , u t , s at ) ( q 1 , . \ z, ) . Using (7.1) and (7.3) we have (qp,'rr.,,F . z )( g r ,w , s z ) i ( n r , ^ , a ) F ( q 1 ,. \ , . e ), so that L(C) e L(M). To prove that .t (IUI) e LG),let

u â‚Ź L(M).Thern by delinition

( q o , r uz, ) f k t y , A , u 1 . Brrt thcrc is orrly orle wfly to get fiom qg to tlt and orily one way from q1 to {1. I'herefore, we must }rirvt: ( q r , * , 5 e ) i ( q 1 ,A , z ) . Now let us write ut = eje,zil;i. " a* Then the first step irr ( q 7 , a L a z a 1 " . . . a n , S z1) qi 1 , , 1 , 2 ;

(2.4)

must be a, rule of thc forrn (7.2) to get ( q y ,q a 2 a 3 . , . e n , S a ) F ( q t , a , z f l :. r' . a , , 7r ., 4 2 ). Rr.rtthen thc grarnrnar has a rule of the fllrrn S + aturt so tha,t 5 + a 1 t i ,.1 Repeating this, writirrg 'LtL: Auz, we have ( e 1 , u , 2 t't,;,1a , r , A u 2 z ) | ( q 1, u , 1. .1. e , , , , , u 3 I t 2,l ) irnplying that, A + u2rn is in the grailrnrar rrncl tha,t, S # a,1u2'u3'u2. This mir.kr:sit quite clear a,t any poirrt the slack contents (cxcludirrg z) are idt:nticirl with the unnra,tcheNl liirrt of the sentential forn, so that (7.4) implics 5 # a142.',ar,. In conseqrren<:c, L(M) g , (G), completirrg the proof if the larrgrra,gcr cloes not txrntairr .\. If A e -L, we add to the constmcted npda the transitiorr d ( q o ,) , z ) : { k t , z ) } so thnt the empty string is a,lsoac:r:cpted. r


Chopter 7 Pussoowt'r AuroN4ArA

-Jfi

rr.rfu,"il

Exonple 7.6

Consider the grammar S-aA,

A+ aABClbBla, B-b, C-c. Since the grammar is already irr Greibach normal form, we can use the constrrrr:tion in the previous theorerll irnrneditltely. In addition to rules d ( q o ,t r ,s ) : { ( q r , , g " ) } and d ( q . , t r ,a ) : { ( q r , z ) } ' the pda will also have tratrsition rules d ( s r , o ,S ) : { ( s ) , A ) \ , d(qr,a,A) : {(q', Mg),(qt,tr)}, d ( q r , b , A ): { ( q t , g ) } ' d ( q t , b , B ): { ( , 1 1 , t r ) } , 6 (qr,",d) : {(q1A , )}. The sequence of moves mrrde by M in processing aaabc is (go, q,afl,bc,z) F (qr , aao,bc,S z) | (q1,aabc,Az) F ( g 1 , a b cA , BCz) | ( q 1 , b cB , Cz) F ( q 1 ,c , C z ) F (91''\' z) F ( q 7 ,A ,a ) . Tltis rxlrresponds to the derivrrtion S + aA+

aaABC + aaaBC + aaabC 4 ua'ahc.

t Irr order to simplify the argurnents, the proof in Theorerr 7.1 ir,ssumed that the grirmmar was iu Greibach rrorma,l form. It is not necessary to do this; we carr rrrake a similar and only slightly more complicated constru<:tion


7.2 PusrrDowmAurolrnrn nulr Clt)rurr:xtFn.EELaNcuacns

189

from a genelral t:ontext*free gramma,r. For example, for productions of the f'orm A --+Bx, we remove :,1frrlrl the stack ancl replace it with Br, but coruume no input symbol. Frrr procluctions of the form A + abCr, we must first rnatch t,he ab in the inprrt agrrinst a similar string in the sta,r:k anrl then replace A with Cz. We leirvc thc details of the constnx:tiorr arrd tlrc associatedproof as an exerr:isr:.

Coniext-Free Grqmmorsfor PushdownAufomolo The converseof Thtlclretn 7.1 is also true. The rrlrrstruction involved readily suggests itsulf: reverse the proce$$ in Theorctr 7.1 so that the grammar sirnulates the moves of the prJa. This rrreans that the contq:nt of the stack should be reflected in thc variable part of the sententirr,l ftrrrrr, while the processed input is the trrrrnirrirl prefix of the sentential ftrrrn. Quite a few details are needed to mir.kt:this work, To keep the discussiorras sirnple as possible, we will assurnethat the npda in question meets the following recluirements: 1. It has a single fina,l sta,te rly that is entereclif and only if the stack is cmpty; 2. All transitions rmrst have the form 6(qt,,o,,A): {c1,c?,...,cr}, whelrc

r:a: (qi, tr) ,

(7 lit)

r:i : (qi, BC) .

.(76)

Tha,t is, eaq:htrove either iucrease$or decrcascsthe stack content by n sirrgle symtxrl. Tltese restrictions rniry appeer to tre very sev()rc)but they are not. It can be shown thrrt for any npda there elxii*s an equivalent one ha,ving properties I rr,nd2. This equivalence wiL,rexpkrrtxl partiallv in fixercises 16 and 17 in Srxrtiorr7.1. Ilere we need to cxplore it further, hrrt a,gairrwe will Jeavetht: a.rgurirents as an exercisel(st:c Exelcise 1tj at the elnd of this section). Takirrg tltis as given, we now rxrrrstruct a context-free grarnrrrar ftrr tlx: lirrrguage accepted by tlxl rrllda. As statecl, we want the senturrtial forur to represent thc content of the sttrck. But the configura,titirrof the npda also involves arr irrternal state, and


Chopter 7 PusnrrowN AUToMATA

this has to be remembered in tht: sontential form as well. It is hard to see how this can be done, and thc rxrn$trrrctirtn we give here is a little tricky. Suppose for the motnent that wc ctr.nfind a grammar whose variablt:s irrrr of the form (qiAqi) and whose prothtt:tions are such that ( q i A q 1 )1 u , if and only if the nptla ora-sesA from the stack whilc rtltr.tling u a,nd going fiom state qi to state 4i. 'oErelsing"here mealls that A atttl its cffttts (i.e., a,ll the successivestritrgs bv which it is replaced) are retnoverdfrom the stack, trringing the symbol originally bckrw ,4. to the top. If we calr firrtl srrt:h a, gra,mmar, and if we choose (qorqt) a"rits start symbol, therr (qozq.1)3 u, if and only if the npda nlmoves ,z (creating an empty stat'Jr) while reading rr and going from (lo to qf. Brrt this is exactly how the rrpdtr,a.ccepts'tu' Therefore, the lattguagt: genera,ted by the gralrrlrlar will be identica'l to the Iangua,geaccepted by thc npda. To construct a grarrllrrirr thnt satisfies these conditiorrs, wâ‚Źlexamine the cliffererrt types of transitions that c*n trc ma,clehy the npda. Since (7.5) involves arr immediate erasure of A, tlrtr grtrmmar will have a cotresporrtling production (t1iA(t.i) --+ (t'.

Productionu

of type (7,ti) generate thc set of rules

( q i A q n ) - - +a ( q 1 B q 1 )( q C q n ) , whcrr: 96 nnd g1ta,ke on all possiblc va,lrresin Q. This is due to thc fir.ct that to erase A wc {irst replac.e it with BC,', w}rile reading a,Il o' elrd going frorrr state 91 to qr. Srrbsequently, we go frorn t1i t'tt q1, erasing B, then frorrr q; ttr gA, erasing C. In the last step, it may seem that we havc arlrltrd too much' as there may be solrle stat()it ql that cannot be reaclttxl lrom qi while erasing B. This is true, but this docs not aff'ect the gratnmar. Tlxr resulting variables h.i Bqi are uselessvariirbkrs ilnd do not affect Lhe larrgtra,getr,cceptedby the grarrrrrrar. Fina.lly, as start variable wt: tirke (qozqt), where qy is thc single final state of the nuda.


7.2 PussoowN AureMArA aruo Cortrux'r'-FRpeLANcLTAcEs

191

Consider the lpda with transitions d ( q o ,* , a ) : { ( q 6 ,A z ) } , 6 ( , k , , a , , 4 ): { ( q u , , 4 ) } , 6 ( q o , b , 1 ): { ( q r ,A ) } ,

d (gr,tr,") : {(sz,A)} Usirrgqo 4qinitial stato aryl ga qqth* fi the npda satisfiestxrndition 1 above,but not 2. To satisfy the latter, we irrtroducea,new state 43 and an intermediate step in which we first rcmove the A from thc stack, then replaceit in thc ,r**t *ou*. ilhe new set of traffi d (qo,a, z) : {(qs, Az)} , d ( q r , A , z ) : { ( q s ,A e ) } , 5 ( q o ,a , . 4 ) : { ( q l , ) ) } , 6 ( q , o , b , A:) { ( S t , A ) } , d ( q r , A , a ) = { ( q 2 ,I ) } . Thc last three transitions are of the form (7.5) so that they yield the correspotrding productions (goAqs)-

a,

(gu,4gr)* lr,

(g1ag2)-i .\.

From the first two transitiorrs we get the set of productions (sgaqo)-

a (qsAqg)(spaSo)lo (qo, qt) (Sr",Zo)l

il (qoAsz) (qr, qo) k, (qn Aq=) (qt, qo), --+ (sozqr) a (qsAqs)(s,raqt)l" (SoAqt)(Sraqr)l (-r(qoAs.t) krzrlrtt)lo (qoAqr) (qtzqt) , (qo"qz) -

a {qs Aq1) (qozqz)lo (qrAr1r) (qr rqz)l

a (qoAqz)(qzzqz)lo (qoAS,r)(qtrqz) , --+ (So"qr) a (q0,4qr))(qo"qr) lo(,toArtr)(qraqr)l a(qoAsz) (qrrs.r) lo (qnAqr)(sszsc), (q;tzqo)- (s0Aq0)(so"qo)l(qnAqt) (qrzqo)l(qoAqz)(qrrsu) I (qoAq.) (q:aqo), (ss"sr) -. (s0Aso)(so"qr) l(qoAsr) (qraqr)l (qoAqz)(qr",tt) I (soAqr) (qtzqt) , (qszqz) + (s0Aq0)(qorqr) l(qoAqt) (qnzsz)l(qrArrr) (,nrqr) | (soAs,])(qtrqr) , (qzzqz) -' (qoAqo)(qorqt) l(qnAst) (sfiqz)l(srArt ) ktzzqs)| (qoAqs)(qrrq*) .


Chopter 7 PusHpown Autouara

The start variable will be (qorqz).The string aab is accepted by the pda, with successiveconfigurations (qq,aah,z) | (qo,ab, Az) F (q3,b,z) | ( q o , h ,A z ) F (g1,A,z) F (q2"\' '\) ' The corresponding derivation with G is (qorqr) =+ a (q6Aq3)(qtrsz) + aa(62q2) + aa(qsAq1)(qzq21 + &eb(qrzqz) * aab. The steps in the proof of the following theorem will be ea^sierto understand if you notice the corresponderrce between the successiveinstantaneous descriptions of the pda and the sentential forms in the derivation. The first gi in the leftmost variable of every sentential form is the current state of the: pda, while the sequenceof middle symbols is the same as the stack content. Although the construction yields a rather complicated grammar, it can be applied to any pda whose transition rules satisfu the given conditions. This forms the basis for the proof of the general result. I

Iffi

language' u h: r(:r,r)lforsoqenpdaM, then.Lis a context-free Proof: Asslrme that M : (8,X,f,d,S,1,2,{gl}) satisfies conditions 1 and 2 above. We use the suggested construction to get the grammar G : (V,T,S,P), with ? : E and V consisting of elements of the form (acq.i). We will show that the grammar so obtained is such that for all e t , Q i , EQ , A e f , X â‚Ź l * , u , ? â‚Ź E * ,

( q o , r r , A Xi) ( t i , u , x )

(7.7)

implies that (qiAq1) 3 u,

and vice versa. The first part is to show that, wheneverthe npda is such that the symbol A and its effectscan be removedfrom the stack while readirrgz and


7.2 PusHoowu Autorrrara .q.ul Conrnxr-F'Rpp L.cr-rcuacps

f03

going from state qa to qi, then the variable (StASi) can derive z. This is not hard to see since the grammar was explicitly constructed to do this. We only need an induction on the number of moves to make this precise. For the converse, consider a $ingle step in the derivation such as (rlrAqx)+ a(q1Bq1)(q,Cq*) . Using the corresponding transition for the npda

(78) we see that the A ca,n be removed from the stack, Bd put on, reading a, with the control unit going from state qt\o qj.Similarly, if (qiAqi) + a,

( 7,e)

then there must be a corresponding transition

6(qr,o,A): {(si,A)}

(7,1r.))

whereby the A can be popped off the stack. We see from this that the sentential forms derived from (qAqi) define a sequenceof possible configurations of the npda by which (7.7) can be achieved. Notice that (qiAq1) + a(qiBq1)(qfiqn) might be possible for some (qiBq) (qfiqn) for which there is no corresponding transition of the form (7.8) or (7.10). But, in that case, at least one of the variables on the right will be useless. For all sentential forms Ieading to a terminal string, the argument given holds. If we now apply the conclusion to ( q o , r , a ) i ( q . 1A, , A ) , we $ee that this can be so if and onlv if ( q s z q y )1 w . ConsequentlyL(M)

: I(G).

I

1 . Show that the pda constructed in Example 7,5 accepts the string aaabbbb that is in the language generated by the given grammar. Prove that the pda in Example 7.5 accepts the language 7 :

{an+7b2^ : rz Z 0}.


Chopter 7 PusHoown AuroNrar.c,

/

e. )C"tt*ttrrct

an npda that accepts the language generated by the gramma^t S + asbblaab. W

Construct

an rrpda that accepts the language generated by the grammar

^9- aS,9Slab,ffi 5. Construct an npda corresponding to the Brammar

S + aABBlaAA, A u aBBla, B + bBBlA. 6 . Construct art rrpda that will accept the languagegeneratedby the gramrnar s + AAla,A - SAIb. G : ( { S , A } , t o , l r } , S , P ) , w i t h p r o d u c t i o nS 7 . Showthat Theorems7.1 and 7,2 imply the following. For every npda M, there existsan npda frwith at urost three states,suchthat.L (M): l, (i|f) . ffi 8 . Show how the number of states of ff in the aboveexercisecan be reducedto

two. (E

tr"O an npda with two statesfor the languageL:

L0. Find an npda with two states that arcepts 7:

{a'ob*+r: n, > 0}' ffi

{a*bz* I n I 1}.

11. Show that the npda in Example 7.7 accepts I (aa-b). ffi 12. Show that the grammar in Example ?.7 generates the language L (oa-b). 13. In Example 7.7, show that the variables (+o.Bqn;and (q6agr) are useless' 14. Use the construction in Theorem 7.1 to find an npda for the language Example 7.5, Section 7.1. \ \ / ff/

nina a context-free grammar that generates the language accepted by the npda M : ({qo,Qr},{a,b}, {A,z},d,{0,s, {qr}), with transitions 6 (qo,o, e) = {(ge, Aa)} , d (so, b, A) = {(qo, AA)} , d ( s o ,o , A ) = { ( q ' , A ) } .

1 6 . Show that for every npda there exists an equivalent one satisfying conditions L and ? in the preamble to Theorem 7,2.

L 7 , Give full details of the proof of Theorem 7.2, by which an arbitra"ry context-free grammax can be used in the proof of Theorem 7.1.

1 8 . Give a construction


7.3 Dprnn,vINISTrc Pusunowrq Aurclnrnra, Ar'lrr DptpRL{rNISTrc Corurnx:r-FRHH L.q,Nc;uacns

ffiffi DeterministicPushdownAutomqtoond D e t e r m i n i s t iC r o n t e x t - F r eLeo n g u o g e s A deterministic pushdown accepter (dpda) is a pushdownautomatorr that neverha^rrr,choicein its move. This can be ir.chieved bv a modification of Definition 7.1.

lffim,finrnqffi"!,ffif,il,il i A puslrdown automaton M : (Q,I, l, d, qo)zt F) is said to be deterministic if it is a-nautomaton as defineldirr Definition 7.1, subject to the restrictions ttrat, for every g E Q,e E X U {I} and b â‚Ź f, 1. d (g, a, b) conttrins at rnost one eleinent, tust be empty for evcry c â‚Ź E. The lirst of these conditions simplv requires that for any givcn input symbol and any stack top, at most one mov{} carr be made. The second condition is that when a A-move is possible for some configuration, no inJlut-consuming alternative is available.

#,fl \1

#

It is interestirrg to note the differt:rrce between this definitiorr and the corresponding definition of a detcrrrrinistic finite automaton" The do4ra,ig

gf_!h!_Eg+q4g4 funqlio" iu still asin Definition7.1rathertharyb E;11

be,caPrtc"wewant to {gtain )-transitions. Since the top of the stack playsE role in determining the ncxt move. the presenCbof .\-transitions does not automatically imply nondeterminism, AIso, $ornc:transitions of a,clptla may be to the empty set, that is, undefined, so there may be dead rxrrrfigurations. This does not affect the definition; the orrly criterion for determirrism is that at all times at rnost one possible rnove exists.

Dcfinition

7-4

A language ,L is said to be a deterministic context-free only if there exists a dpda "lb"sn'ch"tlrat t; =l;(MJ.

language if and


Chopter 7 PusIItowt't Aurtltrlare

196

The language

Exomple 7.8

7:{a"b":n>0} language.I'he ptlir,M : ({qo,Qt,Qz},{o, lr}, is a deteruritristict:orrtcxt-fiee with {0,/},d,40;ro,{qs}) b .,./o,r,,,\ \ ,u

1 (J"\

r'',|l

't

,l

r { ; t - 1 , tf ; n " l 1 "

l

f

i

\

;- L + ' l '

: {(sr,10} , 6 kto,u,o)

. l ( q ' , o ,t ) : { ( q 1t.i ) } , , l ( 4 r , b , 1: ) { ( q r , I ) } , 6 ( q r , b , 1:) { ( q r , , r ) } , d (qz,,tr,o) : {(q,r,i) } ,

accept,stlrtl givcn lir.ngrta,ge,It satisfics the conditions of llelirritiorr 7.4 anrl is therefore deterrninistic.

I

Look now elt lixample 7.4. Thc npder,there is ttot dett:rrninistic because d (116, a, a) = {(qe, aa)}

and d (qo, ,\, o,) :

{ (s' , a)}

violate cotrditiorr 2 of Definition 7,3. This, of course, does rxrt irnplv tha't the language {trlurR} itsolf is nondetertrtinistir:,since there is tlte possi}rility of irrr cqrriva,lent dpda. Brrt it is knowu that tlxr lir,ngrtage is indeed rxrt detrlrrninistic. Fl'otn tiris a,rrtlthe next exalrrplc wo see tha,t, in cout,rastttr deterministir: irrrrlnondeterrninistit:llrulidowtr automata art: finite irutorrra,ta,, nol ecluivak:nt. There are context-f'reela,nguagesthat ilrc not deterministic.

E x o m p l e7 . 9

Lt:t L1=la"b";zl0) and

:n,>0]. A,1 obvirlrs moclification of tlrc argrrment that -L1is a r:ontext-freelatrguirgrr shows that ,L2 is also context-frtx:. The language L = I'tl) Lz


7.:l DETr.lRMrNrsTrcPuSHDowN Aurorvt.q,la nnp DprnnulNls'l'I(i Coltrnxr-FRnn

LAI'rGu.q.c;ns

197

is context-freetu'Jwell. This will follow from a ge,ncral!+g{:-t}H be prcsenteclin the ncxt chaptcr, but ca,ncasily be -"d* plultffiHt-THis point. Lct G1 :

(Vr,T,S1,P1) attd G2 :

(Vz,T,52,P2) be context-freegrarn-

rrrarssuchthat -Lr : L(G) and 1,2: L(Gil' If we assumethat I and the two, grarrlnlar V2 are disioirrt and tha,t S # U U V2, then, cxrrrrlrinitrg G : (Yr U VzU {5} ,7, ,9,P), w}rtrre P:h

U P z U { , 5- . 9 1 1 5 2 } ,

shorrld be fairly clear a,tthis point, brrt the details of generatesLl)L2.'Ihis the a,rgumentwill be clefirrrecluntil Chapter 8. Accepting this, we sec that' 1, is context,frce. But .L is not a derterministic context-frce langua.gtl' This seerns rea$onable, sint:c the pdn" has either to match tlrre b or two aga'inst each a, rrnd so has to rnake ir.rrinitial choice whetlter the irrput is in -L1 };eginning of the strirrg or in .Lz. There is rro informirtion availa.bleat the t_rywhich the choice,,ut. b* marle deterministically. Of courstl, this sort of argumelt is basecl9rr a partit:ular algorithrn we havtl in mind; it rnay letr.tlus to the r:tlrrect conitlcture, hrrt cloesnot prove anythirrg. Therrl is always tlte possibility of a completely clifferent a,pproach that avoids nrr itritial crhoice' 'lb but it turns orrt that therc is not, ir.rrd.L is indeed nondetertninistic' see tlis we first establisli the folkrwing claim. If .L wcre a dett:rrninistic rxlrrtext-freelangua'ge,then L:

t ' l ) { a " h " c n : r z> 0 }

wouldle a c.ontcxt-free language. we show the la'tter try constructitrg a,n npda M fbr tr, given a,tlPda M for L. The icfurabehincl the constnrction is to add to the control rtttit of M a sirrrilar part irr whiclt tratrsitions c:ausedbv the iuput symbol b are replacxxl with similrrr ones fttr input c. This new part of the control utrit mar,ybe enterecl rrfter M has reacl atb"'. sint:tl the second part rtlsponds to c'iu the stlrrre wav as the flrst part cloesto b"', the llrocess thir,t recognizes Q,'"bZ" now also accepts (trrbrrcn.Figure 7.2 describesthe construction graphically; rr formal argurnent ftrllows. Let M : (8, X, f , d, 40,z, F) with Q : t q o 'Q t , ' . . , Q n-I Then consider f r :

( d , E , r , d u F , . zwith ,F)

, ar,...,8,,} 8 : a u {ao, F':Pu{fi'eeEFlt, frornd by irrcluding attdfr constructed F ( 0 rA , , , . s:) { ( f l ys, )} ,


Chopter 7 PusHnowruAtJ'rouara

Figure 7.!

C)

Addition

Control unit of,44

i.

I It :r{ q

f

r+''

sl

for all ql e 4 s_€ Il, and

: {(fr,u)}, 8(ir,r,s) for all 6 ( q o , b , s :) { ( q i , u ) } , et EQ,s € 1,ru € l*. l'or M to accept anb" wE must have (qoa , n b n r, ) i * ( g oA , ,r ) , with q4 € F. BecauseM is cleterministic, it must a,lsobe true that ( e o ,a " b r n , , ) t * ( q i , h " ,u ) , so that for it to accept unbz" we nrust further have ( q r , b n, r l f * , ( q j , A , t l r ) , for some qj E F. But then, by r:onstruction

(ti,c",r)im (fr,tr,zr;, ,,

i-:

so that M wille,ccept a,nb"cn.It re,mainsto be shownthat no strings ottrer than those irr .L are acceptedby M; this is considereclin severalexercises at the end of this secbion.The conclusionis that L : t (fr), .o that i is context-free.But we will show in the next chapter (Uxirnpte 8.1) that i is not context-free. Therefore,our assumptionthat L is a deterministic ctlntext-freelanguagemust be false. I


Z.B DnrnnvrNrsTlc

Pussoowu

AurouRrn

I . Show that 7:

,

.q.NoDnTTRMINISTICCoNr:sxt-FnEE

LANGUAGES

199

language, {a*bz": rz > 0} is a deterministic context-free

Show that 7 : {6nlt't" : rn I n*

2} is deterrninistic.

3 . Is the languag. 7 : {a"bn : n } 1} U {b} deterministic? 4 . Is the languag. 7 : {a*b D.

: n > 1} U {a} in Exarnple 7.2 deterministic? ffi

Show that the pushdown automaton in Example 7,3 is not deterministic, but that the language in the example is nevertheless determirristic,

6 . For the Ianguage -L irr Exercise 1, show that.L* is a deterministic context-free language. Give reasons why one might conjecture that the following language is not deterministic. L:

8 . I s t h e l a n g u a g e7 :

G.

in=rno.*:k) {a,,b,,,cr,

{anh* in:n7

or n:

m}2}

deterministic?

* the language{ucwT : w E {a,b}- } deterministic? ffi

1O. while the language in Exercise I is deterrninistic, the closely related language 7 : {wwR : w E {a,b}- } is known to be nondeterministic, Give a,rguments that rnake thi.B statement plausible'

fi-J.

Srro*that .L : {u e {a, b}. Ianguage.ffi

n,,(u) f n6 (ur)) is a deterrninistic context-free

1 2 . Show that -[4-itt E*u*ple

7.9 does not accept anb ck fot k I n.

1 3 . Show that fif i., E***ple

7.9 does not accept any string not in .L (a"b-c-)'

7.9 does not accept a"bznch with fr > 0, Show alstr that it does not accept a"b*ca unless 'trl: n or m:2n.

1 4 . Show that fr} in E***ple

show that every regular language is a deterministic

context-free language.

ffi "-\ (ro.l stto* that if .Lr is deterministiccontext-freeand lz is regular,then the * context-free.ffi Iu.rg.,*g"LtJ Lz is deterministic irz/ 'J

show that under the conditionsof Exercise16, .Lr f-l-Ls is a deterministic context-freelanguage.

ri) V

ci-r" an example of a deterministic context-free language whose reverse is not deterministic.


Chopter7 PusrruowNAurou.ue

ffiffi Grommorsfor DeterministicContext-Free Longuoges* The importarr(:c of deterministit: <:ontext-free larrgrrageslies in thc fir.ct that they catt be parsed efficiently. We can see this intuitively by viewing the pushdown autorrraton as a parsing device. Since there is no backtracking involved, we can easily write a computer progralrr frrr it, and we rrray expect thtrt it will work t:fliciently. Sittce there may be .\-transitions involved, we cannot immediatcly claim that this will yield a linear-time parser, but it J.rutsus on lhe right track nevertheless. To purrrue this, let us see what grarnmars might be srritable for the description of deterministic contextfrec languages, Herc we enter a topic important irr the study of contpilers, but sorrrtlwhat peripheral to our interests. We will provide only a brief introduction to some irrrportant results, ref'erring the reader to books ort r:ompilers for a more thorouglt trea,tment. Supposr:we are parsing top-down, attempting to find the leftmost derivation of a particrrla,r sentence. For the sake of discussion, we use the approach illustrated in Figure 7.3. We scan the input ru from lefl to right, while developing a senterrtiir,lfbrm whose terminal prefix matches the prefix of ur up to thc currently scanne.dsynrbol. To proceed in matehing consecutive uymboftl, we would likc to know exactly which produr:tion rule is to be applied at each step. This would avoid backtracking and give us an efficient parser. The question then is wltether there are grilmmars that allow us to do this. For a general r:ontext-free grammar, this is not the case, but if the form of the grammar is restricted, we can achieve our goal. As first casc, ta,ke the s-grammars introdtrtnd in Definition 5.4. From the discussion thcre, it is clear that at every stage in the parsing we know exactly which protlrrction has to be applied. Suppose that to : rul?{rzand tlnt we have developed the sentential form tulAu. To get the next symbol of thc sentential forrn matched against the next symhol in u, we simply look at the leftmost symbol of ur2, sfl,$a. If there is no rule ,4 * ay in the grammar, the string ru does not belong to the language. If there is such a rule, the parsing can proceed. But in this case there is only one surh rule, so thcre is no choice to be made.

Figure 7.3

tt1

d2

d1

t t .

d3

|

oo. . . r,

I " *+* a 1

Matched part

|

4 . . .

Input ro

Sentential form

Yet to be matched


7.4 Gnelru,q,RsFoR DorenMlr.u$uc Cor-rrext-Fnnn L,q,rucuacns

201

Although s-grammar$ are useful, thcy are too restrictive to capture all aspects of the syntax of programming languages. Wc tieed to generalize the idea so that it becomes ntore powerfirl without losirrg its essential property for parsing. One type of grammar is called a,n LL grammar. In an .L-L gramrrlax we still have the property that we carr, by looking at a limited part of the input (consisting of the scannod symbol plus a finite number of symbols following it), predict exactly which production rule must be used. The term LL is standard usage irr books on c:ompilers; the first .t stands for the fact that the input is scarrned from left to right; the second tr indicates that leftmost derivations are constructed. Every s-grarnmar is an LL grammar, but the concept is rrrore general.

Exomplo 7.10

The grammar S --+ oSblab is not an s-grarnma,r, but it is an LL gralnmar. In order to determine which production is to be applied, we Iook at two consecutivesymbols of the input string. If the first is an a, and the second a b, we mu$t apply the production S - ab. Otherwise, the rule S --+aSb must be used. I We say that a grirmlrra,ris an -L,L(ft) grammar if we can uniquely identifu the r:orrect production, given the currently scanned symbol and a "lookahe.:ad"of the next ft - I syrnbols. The above is an oxarnple of an LL (2) grarrrmar.

Exornpla 7.1,1

The grammar S -, ,9,9laSbl ab generates the positive closure of the language in Example 7.10. As remarked in Example 5.4, this is the language of properly nested parerrthesis structures. The grarnma,r is rrot an LL (k) grammar for any h. To see why this is so, Iook at the derivation of strings of length greater than two. To start, we have available two possible productions 5 -r ,S,9 and S --+ aSb. The scanned symbol does rrot tell us which is the right one. Suppose we n()w use a look-ahead and consider the first two symbols, finding tlnt they arc aa. Does this allow us to make the right decision? The &nrJweris still no, since what we have seen could be a prefix of a number of strings, including both aabb or ao,bbab.In the fi.rst ca*se,we must start with S - aSb, while in the secorrd it is necessary to use 5 * 55. The grammar is thereforc not an LL(z) grammar. In a similar fashion, we can see that


Chopter 7 Pusnnowr*lAuroN,IalA

no matter how many look-ahead symbols we have, there are always some situations that cannot be resolved. This observation about the grammar does not imply that thc language is not deterministic or that no ,L-L grammar for it exists. We can find an .L,L grammar for the language if we alalyze the reason for the failure of the original grammar. The difliculty lies in the fact that we c&nnot predict how many repetitions of the trasic pattern a"b' there are until we get to tltc end of the string, yet the gralnmar requires an immediate decision. Rewriting the gramrrrar avoids this dilliculty. The grammar g -

a$b^91.\

is an L,L-grarrtrnar nearly equivtrlent to the original grammar. To see this, consider the leftmost derivation of ru : aba6. Then S + aSbS + ab.9 + q,baSbS+ a,hq,bS+ abab. We see that we never have any choice. When the input uymbol examined is r:r,we must u$e S - aSbS, when the syrnbol is b or if we are at the end of the string, we must use S --+ A. But the problern is not yet completely nolved because the new grammar can generate the empty strirrg. We fix this by introducirrg a new start variable 50 and a production to en$urethat some nonempty string is generated. The final result 5'6 S-

aSbS aSb,SlA

is then arr .L^L-grammar equivaletrt to the original grammar. f While this itrfcrrmal description of trtr grammars is adequate for understanding simple exarnples, we need a more precise definition if any rigorous results are to be develoned. We corrclude our discrrssion with such a definition.

IliMl Let G : (V,7, S, P) be a context-free grammar. If for every pair of left-most derivations Sl

w t A r t 4 1 i l 1 ! ! 1 n 11 w 1 ' * 2 ,

Sl

wtAxz 1trt1y2r2 l

wrtts,

witlr zr1,?t)z,'u)B â‚Ź 7*, the equality of the ft leftmost symbols of ur2 and tl3 implies lJt -* Uzt then G is said to be an LL(h) grammar. (If lu.'21or lur3l is less than fr, then h is replaced by the smaller of these.)


7.4 GRAMMARSl,'oR Dnrnnlrrt-rrsuc CON.I'EXT-FREELANcuAcES

203

The definitiorr rnakes precise what has already becn irrdicated. If at nrry sta'gein thr: leftrnost derivatiorr (w1Ar) we know the next h symbols of the input, the next step in the cleriva,tion is uniquely determined (as expressed bY gr : l/z).

The topir: of .L-L gramma,rs is an importa,nt t)rre it the study of courpilers. A nrrrntrcr of progranuning larrguagescan bu defined by LL griilrrfilars, and rnany rrrrrrpilers have het:n written using .L.Lpilrsers. But -L-Lgranrmars are not srrflicietrtly genertr,lto deal with &ll dertt:rrninisticcontcxt-free languages. Ctlrrsequently,therc is itrterest in other, rrroregenera,ldrltcrrninistic grammirrs. Particula,rlv irnportant a,rether so-called -LR grir,rnrnars,wlrich also allow efficierntparsing, trut can bc viewed as constructing the derivn"tiorr tree fiorn the bottom up. There is a great dur,l of material orr this subject that can bt: found in books orr compilers (e.g., Hunter 1981) or trooks specifically devott:d to parsing mt:thods for fbrmal larrguages (such ns Aho and Ullman 1972).

1 . Show ttrat the strcond gramffrar in Exa,mple 7.11 is an Z/: grailrnr&r and that it is equivalent to the original gramrrra"r.

2 . Sirow that the gramrnar frrr -L : lw : n* (r) : rrr,(trr)) given in Examplc 1.13 is not arr tr.L gramrnar.

ffi

3 . li'irrd ari 1,.L gramnrar f<rr the language in Exercise 2, 4 . Construct an -L-Lgraurnar for the languagc L (a*hu.)t-t.L (abbb"). ffi i),

Show that arry LL gramrrrar is unambiguorrs.

6 . Show that if G is an /,L (k) grarnuar, then I (G) is a dcterministic contcxtfree languagc. Show that a dctcrnrinistic contcxt-free larrguageis never irrherently amhiguous.

L Let fJ be a corrtext-frcc granrrrar in Grcibach rrorrual form.

Describc an algorithm whi<;h, Ibr any given ft, tletermines whether or not G is an Z.[ (A) Srammar.

9 . Give.L.L gramrnar$ for the lbllowing larrguages,assumirrg 5: (a) Z: (b),

{ a t t l 1 r r t , , r t t " r : n Z 0 , m ,0} } W tzlt*7'n+"n: n, : 70,rrr, } 0} {a,"

(") fI :

{a'"tn+2rn':rr,}fl,trz>1}

(d) Z, = {u : n,,,(u) < n6 (ur)} (e) f, : {tu : n,,"(w) + nu (ut) t' n. (u)l

{a,b,c}.

ffi



P r o p e r t i e so f Context-Free Longuqges

he family of cotrtext-free Ianguages occupies a central position in a hierarclty of formal languages. On the one hand, (:orrtext-free languages include important but restricted language farnilies such as regular and deterntinistic context-fiee languages. On the other hand, there are broader language families of which context-frce larrguages are a special case. To study the relationship between langtage families and to exhibit their similarities arrd differences, we inveutigate characteristic properties of the various fatrilies. As in Chapter 4, wc look at closure urrder a variety of operatiorts, algorithms for determirring properties of members of the family, and structrrral results such as pumpirrg letnmas' These all provide us with a mear$ of understanding relations between the different families as well ir,sfor classifying specifit: languages iu an appropriate category.

205


Chopter 8 Pnopnnuns oF CoNTlJxr-Fnnn LeNcuacns

ffiM Two PumpingLemmos The pumping lemtna girrcnitt lThe,oEm?37is an effectivctool for showing Similar pumpitrg lrtmmas are knowu art: not n that we will discuss twtl strt:tr resrrlts, one for gonertr,l, other for a restrir:ted type of contextirr the cgntext-free languages language. free

A PumpingLemmofor Conlexl'Free[onguoges Let .L be an infinite r:ontext-fiee language. Then there exists some positive integer rn such that atty ut F L with lrul > ?n calr be clccotrposed as

(8.1) with

(8.2) er,nd

lr vl> 1,

(83)

; ; 'uu"x:'!'z E L1

/R d.I

sur:h that

fbr all i : 0, 1,2, .... This is known a-sthe pumpirrg lemma for context-free languages. Proof: Consider the language L - {,\}, ard assurne that we h*ve for it rr grarnmar G without utrit-productions. or A-productions. lJince the Iength of thc string on the right sidc of any production is bounded, say by k, the Iengtlr of tltc clerivation of any'ur â‚Ź tr must be at least lwl lh- Therefbre, since -L is infinite, there exist artritrarily long derivatiorrs and corresponding derivation trecs of arbitrary height. Consider ttow sut'h a high derivation tre.c a,ndsome sufficienbly long path from the root to a lca,f. Since the lrumber of va,ria,hlesin G is finite, thcrtr rmrst be some va,Iiable that repeats on this pathr as shtrwn sclrematically irr Figure 8.1. Correspotrdittgto the derivation tree irr Figrrre 8.1, we have the derivation Sl

uAz l

uuAgz A'u,unrtrz,

wlrert.' r,r,,u I t I AI and z art: tr.ll strings of termitrals. Flom the above we see t l r a t A 3 u A y a n d A 5 , " , * . , a l l t h e s t r i n g s u ' u i . ' g i z ,r l : 0 , 1 , 2 , . , . , c a t t be gerrera,tedby the graInIIIaI arrcl tlre therefore in -L. Frrrthermorer in the


8,1 Two PuuPtllc LetuN,Ias

207

Figure 8.1 Derivation trec lbr a long string.

+ \A)

X

/

\

derivations A I uAy and A 4 ,r, *e calr assuilte that no variable repeats (otherwise, we just use the repeatitrg variable as A). Therefcrre,the lengths of the strings u, fi, and y depend orrly on the productions of the granrmar anrl carr be bounded independently of ur so that (8.2) holds. Finally, since there are no unit productions and no.\ productions, u and y cannot both be ernpty strings, giving (8.3). This completes the argument that (8,1) to (8.a) hold. I

This pumping lemtra is useful in showing that ar,larrguage does not belong to the family of context-free languages. Its application is typical of pumping lemrpas irr generall they are used negatively to show that a given Ianguage does not belong to some family. As irr Theorem 4.8, the correct But argument can be visualized as a game against an]Urlelligâ‚Ź@!. now the rulos make it a little more difficrrlt for us. For regular languages, the substring-zE-\hose length is bounded by rn starts at the left end of zu. Therefore*tliEiubstring gr that can be pumped is within rn syrrrbols of the beginning of ur. For context-free language$, we otily have a bound on luryl. The substring u that precedesura carrbe arbitrarily long. This gives additional freedom to the adver$ary, rnaking arguments involving Theorem 8.1 a little more complicated.

Show that the language [,:{u,nbncn:rz>0} is not context-free.


Chopter I

, | .r r q rr Lr{ a E I

PRopnnrtns on Conrnxr-FREE Lancuacns

which is Once the adversary lrqs choserrrn,,we.pick jl: ttli"efq3id! in tr. The adversary now has-EETffil-cltoices. If he choosel urq to contain only a,'s,then the pumped string will obviously not be in ,L. If he chooses 4 slriqg rnntaining a,n equal number of a's and b's, then the pumped string k #;.,ot L* generated, and again we have generaterl a string [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[fJ*ith not i1 .L. In fact, the only way the adversary could stop us froilI wirrrring is to pick ufra so that u3rhas the same number of a's, bts, and cts. But this is not possible becauseof rrstriction (8.?). Therefore, -L is not conte4!-free. If we try the sarne argutrrerrt on the language 7 - {a"b"}, we fail, as we must, since the language is context-free. If we pick any string in tr, such as u i : e , " ' b " ' , t h e a d v e r s a r y c a t r p i c k r , : t t r ka n d g r - b A , N o w , n o m a t t e r w } r a t

lHi-*: . -Dc-.tlffiJlil,':,,';-*#il1-iJ * Y - , L :,!rt,r;rH,H"l-;-:T""Tl:'i:-,ffiT:i umpinglemma.Thattris ,

-

-lT*.,/ ),g,,,,' ,".'dir' i''--r*lir'lr+r

context-freemust conrefrom someother argurnent,suchasthe con$truction of a context-free granrrrra,r. The argument also justifies a claim made in Example 7'9 and allows us

to closea gap in that examPle.Tlq L:

{a"b'}|

{a^bz^} | {anb"cn}

is not ctrntext-free. The string artlbrnctrl is irr Z, but the pumped result is not.

T

irixUd$lle,raigi

Consider the language

L - {ww : tu

:glt]

Although this language appears to be very similar to the context-free language of Example 5.1, it is not context-free. Consider the string a*lfna*b*, There are many ways in which the adversary can now pick uzg, but for all of them we have a winning countermove. For example, for the choice in Figure 8.2, we call use i : 0 to get a string of the form akYa*b*,h < rn or i < m, which is rrot in .L. For other choices by the adversary similar arguments can be made. We conclude that ,L is not context-free'


8,1 Two Pulrerrvc Lnrvruns

209

Figure 8.2 a...ab...ha...ab...h

F_i_]___+ ilv

fr!

E

,--

{a"t:n>0}'/

l_p

nul

r'eilr rt-r

c.t

n_st

is not context-free, In Example 4.11 we showed that this language rg11g!_tgg!g:. However, for a language over an alphahet with a single symbol, t[eEEfrttle clifference between Theorem S.l a,nd the prrmping lemtra for regular languages. In either case, the strings to bt' 1'r'rped corrsist entirely of ots, and whatever new string can he gcncrated by Theorern 8.1 can also be generated by Theorem 4.8. Therafore, we can use essentially the same arguments as in Example 4.11 to show that .L is not context-free.

r

,

r ,,,1*.\ ---*.lo*

t

fr,#sthplqS,,4

<J

Showthat the language

I '1

t

r^ bb 'trra lrrrcroao6a

is not context-free. Given rn in Theorem 8.1, wc pick as our strirrgitr,r"'6n,)The adversary now has several choices. The only one that requiris mffh thought is tho

q4ovrn_illFigur* S.?-..$q!11g11!jry will yield glrâ‚Źw-striue;v/ith ,on=q lg + \i - I)kr a'sJandW + U -I)hz b's.lIf the adversarytakesk, I 0, E;FO@TIffi i:0-.-Sinre--r (* - hr)'g (* - 1)' = n 7 , -22 m * r < rrf - lq, the result is not in tr. If the opponerrtpicks A1 : 0, frz I 0 or h * 0, hz : 0, then again with i : 0, the purnpedstring is not in ,L. We can concludefiom this that .L is rrot a context-freelanguage. I

, 5lf.


PRopeRrtrs or- Corutnxl'-FREELRwcu.+cps

210

Chopter I

Figure 8.3

# n -- j_----E-

]

d . . .d . . . a . . . a h ' . ' h ." b . . ,b

#

u

i

l

l

r

l

z

A PumpingLemmofor Lineqrlonguoges We previouslymade a distinction betweenlinear and qog}llrj3rg4text-freg grammaxs.We now maftea similar distinction betweenlanguages.

IIM A context-free language .L is said to be linear if there exists a linear contextfree grammar G such that .L : L(G).

. ,--1, C'l

tl ,.''ri.',r -)),,.*

Clearly, every linear languageis context-free,but we have not yet established whethet or not the converseis true-

I

r?> 0) is a linear language' A linear Brarnmar for it is given in Example 1.10. The grarnmargiven in Example 1.12for the languageL : {.; no (tu) : nr,(tr)} is not linear, so the secondlanguageis not necessarilYlinear' I _ Of course, just because a specific gramma,r is not linear does not imply that the language generated by it is not linea,r. If we want to prove that a language is not linear, we mu$t show that there exists no equivalent linear grammar. We approach this in the usual wa5 establishing structural properties for linear languages, then showing that some context-free languages do not have a required propertY. Let -L be an infinite linear language. Then there exists some positive integer ?n, such that any w â‚Ź. L) with lurl ! m can be decomposedas w:uafrllz with

luuyzl 1n-t',

lual>r,

(8 5) (8 6)


8.1 Two Purr.rprr{c Lnlrues

ZLt

suchthat uu'ny'z â‚Ź L,

(,$7)

f o r a l l z : 0 , 1 ,2 . . . . . Note that the conclusions of this theorem differ from those of rheorem 8.1, since (8.2) is replaced by (8 s). This implies that trie strings u and 3r to be pumped must now be located within rn synrbols of the leff and right ends of ru, respectively. The middlc strirrg z can be of arbitrary length. Proof: our reasoning follows the proof of rheorem g.l. since the language is linear, there exists soure linear grammar G for it. To use the argumerrt in Theorern 8.1, we also need to claim that G contains nei urrit-productions arld no .\-productions. An examination of the proofs of Theorem 6.8 and Theorern 6.4 wilt show that removing A-productions ancl unit-productions does not destroy the linearity of the gra,mmar. we ca.rrtherefore assurre that G has the required property. consider now the derivation tree as shown in Figure g.l. Because the grammar is linear, variables can appear only on the path from s to the first A, on the path from the first A to the second one, and on the path from the second .4 to sorre leaf of the tree. Since there are only a finite number of va"riablerson the path frorn 5 to the first -4, and sirrce each of these gerreratesa finite number of terminals, u a'd z rnust be bounded. By a similar argurnent, u attd y are bounded, so (g.b) follows. The rest of the argument is as in Thcorern 8.1. I

llH$$$si$.Niiil rhuru,.g,,us*

is not linear. To show this, assume that the langua6Jeis linrrar and apply Theorem 8.2 to the string ,uJ :

arnbzrrl(f,ttr,

Inequality (8.7) shows that in this cas. thc strirrgs ut Dt a must a,llconsist at entirely of a's. If we pump this str'ing, we get orrt,lk62nzont+t, with either h > 1 or I ) t, a result that is not in tr. This contracliction of Theorern g.Z proves that the languageris rrot linear. I


F 2r2

Chopter I

Pnt)pr:nrrr:soR Cot't:rsxr-FRenLeNcuecns

This exatnple answers the general qtrestion ra,iscd on thtt relation betweel the f"tr,miliesof context-free and linear languages' Tlttl farnilv of lineflr Iangrrelgesis a, proper subset of the family of context-free larrguagtrs.

1. Use reasonirrg similar to that in Exarnple 4.1I to give a complete proof that the language in Examplc 8.3 is not cotrtext-Iree. that the langrrage L : {a* I rz is a prime trutnber} is not t:orrtext-free. @Stto* 3, Show that -L : {trr.utrttt : w E {a,b}- } is not a cotrtext-fi'ee larrguage W 4. Show that .L : {ur e {tl, b, c}" r n2"(*) + "3 (*) : til| (to)} is not context-frcc, 5. Is the language 7 - {anb* : n * 2*} corrtext-free? 6. Show that the language I: q

{o"'

, " } 0} is not coutext-free'

Show that the following languagcs on X = {a, b, c} u,re not context-free.

(3)

( a )I :

{a*ll ; " <.i'} m

(b) I : {alhi :" > (j - 1)'} (c)L:{anbich:k:jn\ (d) /,:

{a'obich:k>n,f >i}

( e ) L : { a ; " h ic k : n I j , n ! h S . i } (f) .L : {ru : n,,,(tu) < nu (tu) < n.. (tr)} (S) -

(e) f - {w: no(w) lrr't,('u,): n" ('u')} (h) Z, :

,/

-)

,{\ !dr)

: {u â‚Ź {4, b,t:}* : n,,.(zr)* nr (tu) 2n" (tu)}'

Deternrinc whether ol not the following languagcs are cotrtext-free ( a ) Z , : { a ' f " u u H a ' t :' n , } 0 , u e { a ' b } - } (h) L:,{a"'bia^H (c) 1, :

: ? r} 0 , i l 0 i

W

> 0} {a'ltirtib" : tr.> 0,j

(tt) L : {a"H akht I n + i t- k + I} (c,) L = {anbi ahlsl,r, a- k, J < I} j} (f) r: {u"'b"t! ,," < g. I1 Theprenr 8.I, find a bound for rrz in tertns of the properties of the grammar G.

'{^ , m

nctcrmi*e

whetrrer .r *.t

t6e followi.g

language is cnntext-free.

[, = {.,w1ttn2i ,u]r1,ttze {4, b}- ,wt I wzl

ffi


8.2 CLosuRePRoreRrlns ewo Dncrsrol Ar,conrrHMs FoR Cournxr-Fnpn LaNcuecns

b

drb V q

QH

Sfro* that the language7 : {a'ob"a"ob'*; n 2 0,m > 0} is context-freebut not linear. Show that the following languageis not lirrear. t,:

1 3 . S h o w t h a t t h e l a n g u a g eL : context-free, but not linear. \' D

213

{w : n*(w) > nu(.)}

W

{w â‚Ź {a,b,c}*:rro (w)+na(*):rr"(u,)}

is

- 1} islirrear. {a"bt :i 1,n<2i 15. Determine whether or not the language in Example 5.1? is linear. ffi

(4.

O e t e r m i n e w h e t h e r o r n o t t h e l a n g u a g e7 =

16. In Theorem 8.2, find a bound on m in tenns of the properties of the glarrrrrrar G. 17, Justify the claim made in Theorem 8.2 that for any linear langrrage (not rxrntaining A) there exists a lirtear grarrtrna,r without }-productiorrs arrd unitproductions. L8,

Consider the set of all strings af h, wherc a and b are positive decimal integers such that u I b. The set of strings therr represerrts all possible tlecirnal fractions. Determine whctircr or not this is a context-frec languagc.

* 19, Show that the complement of the language in Exercise 6 is not context-free.

\t

@

t* the followinglanguagecontext-free? L:

{o""' : rz and m, are prime numbers}

W

C l o s u r eP r o p e r t i e so n d D e c i s i o nA l g o r i t h m s f o r C o n f e x t - F r e eL o n g u o g e r In Chapter 4 we looked at closure under certain operations and algorithms to decide on the properties of the farnily of regular languages. On the whole, the questions raised there had easy arrswers. When we a,sk the same questions about context-free la,nguages,we encounter more difficulties. First, closure properties that hold fbr regular,rlalnguilges do not always hold fbr context-free larrguages.When they do, the arguments needed to prove them are often quite cornplicated. Seconcl,many intuitively simple and important questions about srntext-f'ree larrgrragesc:artrrottrtt arrswered. This statement rrray seerrrat first surprising and will need to be elaborated as we proceed. In this section, we provide only a sample of some of the most important results.

longuoges Closureof Context-Free The familv of context-fierela,ngrragesis closed und(lr union, ctltrcaterratiorr, arrd star-closure,


2L4

Chopter 8 Pnornnuss ol' CoNrexr-FRen LANcuAGEs

Proof: Let trr and -L2 be two context-free languages generated by the c o n t e x t - f ' r eger a m m r r n JG r : ( V r , ? r , S r , P 1 ) a n d G z : ( V z , T z , S 2 , P 2 ) ,r e spcctivcly. Wt: c:arra^$$urnewithout loss of generality that the sets V1 and Vz <rredisioint. Cotsider rrow the larrguage I (Gr), gerrerated by the grammar Gr : (H u V2 U {S.} , r, u 72,,93,P3) , wlrere S'3 is ir variirtrlc rrot in V U Vz. The productions of Gs are all the prochrt:ti{rrrrtof G1 a,rrclG2, together with an alternative starting production that allows us tr) us(l orro or the other grammars. More precisely, P:t: Pr U Pz U {5r * 5rl5z}. Obviously, 63 is a context-free grammar, so that L (Gz) is a context-free languagc. But it is ca^syto see that L(Gs) : Lru Lz.

(8,8)

Suppose for instance that u â‚Ź trr. Then S:t+Sr4ru is ir possiblc derivation in grammar G:J. A similar argument can be made for u â‚Ź -L2. AIso, 1f w e L (G3) therr either (8.9)

53 + ^91 or 5h + 5'z

(8.I0)

rmrst trr: the first step of the derivation. Suppose(8.9) is used. Since sentential ftrrrns durivcd fiom Sr have variables in I{, ancl V1 and V2 are disjoint, the derivation SrSru can involve productions in P1 only. Hence trr must be in -L1. Alternatively, if (8.10) is used first, then tu must be in.L2 and it follows that .L (G3) is the rrnion of L1 a,nd 1,2. Nr:xt, rxlnsider .,

G+ : (Yr u Yzu {50},il', u'I'2,Sa,Pa).

Here a,gain Sa is a new variable and P+: Pt U Pz U {5i -

SrSz} .


8.2 ClosunE PRopERTTES ANDDucrsroN AlcoRrrHlrs

FoR CoNTEXT-Fn.en LANGUAcES

215

Then L ( G + ) : L ( G r )L ( G r ) follows easily. Finally, considerI (Gs) with Gs : (VLU {Ss} ,7.1,55, P5), where 5'5is a new variable and Ps: Pt U {S5- 5155l.U. Then I ( G s ): L ( G t ) * . Thus we have shownthat the farnily of context-freelanguagesis closed under union, concatenation, and star-closure. r

The family of context-free languages is not closed under intersection and complementation. Proof:

Corrsider the two languages L 1 : { a n b " c n ': n Z 0 , r n } 0 }

and L2:

{ a " b * c * i n , } 0 ,r n > 0 } .

There are several ways one can show that .L1 and "t2 are context-fiee. For instance, a grammar for -L1 is S * Sr,5'2, 5r --i a5rblA, 5z --+ c5zlA' Alternativelg we note that tr1 is the concatenation of two context-free languages, so it is context-free by Theorem 8.3. But L1 -t L2:

{ u n b n c n: n } 0 } ,

which we have already shown not to be context-free. Thus, the family of context-free languages is not closcd under intcrsection.


Chopler I

PnornRuns on Cowrnxr-l'Ree LnxcuRcps

The second part of the theorem follows from Theorem 8.3 and the set identitv

::=..: L 1 t 1 L 2 :L t U L z . If the family of context-free languages were closed under complementation, then the right side of the above expression would be a context-free language for any context-free L1 and L2. But this contradicts what we have just Shown, that the intersection of two context-free languages is not necessarily context-free. Consequently, the family of context-free languages is not closed under complementation. r

While the intersection of two context-ftee Ianguages may produce a Ianguage that is not context-free, the closure property holds if one of the languages is regular.

Let .L1 be a context-fiee language and tr2 be a regula,r language' Then L1 tt L2 is context-free. Proof: Let M1 = (8, X,I, dr, q0,s, .F"l) be an npda which accepts -Lr and (P,E,dz,po,Fz) be a dfa that accepts.L2' We construct a pushMz: qo,r,F) which simulates the parallel down automaton ffi : (0,8,f,t, action of M1 and M2: whenever a symbol is read from the input string, -ffi simultaneously executes the moves of M1 and M2. To this end we let

e:exp, 0o: (go'po), * t':

I\x

Ezt

and define d such that

( ( q r , p r )r, ) e 6 ( ( q a , p r )o, , b, ) , if and only if

(gn,r) e d1(qa,a, b) , and 6 z( p i , a ) : p t . In this, we also require that if a : tr, then py : p7. In other words, the states of M are labeled with pairs (U,pi), representing the respectivestates


8.2 Closunn

PRoRnnrlns ,q,FID Dnctslol

AlconrrHMs

FoR Colrrexr-FREE

LANGUAGES

217

in which Mr and M2 (:arr be after reading a certain inprrt string. It is a straiglrtforward inductiorr argument to show that ( ( q t " P t ' )' u ' "z ) i ; ( ( q ' , P " ) , n ) , dtlr q, â‚Ź .Fr and p" â‚Ź Fz if and only if ( q o , w , " ) i - , ( g , ,z ) , nd d* (Po,w) : p". Therefore, a string is acceptecl Uv fr if and only if it is accepted by Mr and M2, that is, if it is in tr (Mr) ) L (Mr) : Ltfi Lz. I The property addressed by this theorem is called closure under regular intersection. Becauseof the result of the theorem, we say that the family of context-free languages is closed under regular intersection. This closure property is sometime$ useful for simplifying arguments in connection with speciflc languages.

Exomplo 8.7

Show that the language

tr-la"h"rz)0,nlI00j is (:tlntr:xt-free. It is possible to prove this claim by constructing a pda or a context-free granrrnar for the langrrage, but the process is tedious. We can get a muclr treater argument with Theorem 8.5. Let L t : { a l o o b l o o }. Then, because trr is finite, it is regular. Also, it is easy to see that, 7,:{a"bn:n>0}1121. Therefore, by the closure of regular languages under complementation and the closure of context-free languages under regular intersection, the desired result fbllows. I


Chopter I

W[*Blii$l'tt,\N

PnornRTrnson Corcrnxr-FREEL,q,rucuRcns

showrhat the languase L :

{*

â‚Ź {cl, b, c}* : no (*) : n, (tr) : t,, (",)}

is not context-free. The pumping lemma can be used for this, but agaiu we can get a tnuclt shorter argurnerrt urting closure under regular intorsection. Suppose that .L were context-free. Then L f t L ( a , * b * c * ): f a ' b " c ' : r z l 0 ] would also be context-free. But we alreadv know that this is not so. Wtr conclude that tr is not context-free.

I

Closure properties of langrtages play an important rolc in the theory of fclrrnal languages and many more clt$ure properties for context-ftas languages can be established. Some additional results are explored in the exercisesat the end of this section.

Longuoges SomeDecidqblePropertiesof Context-Free By putting together Theorems 5.2 and 6.6, we have already established the existence of a rnembership algorithm for context-free languages' This is of course an essertial feature of any language family useful in practice' Other simple properties of corrtext-fiee languages can also be determined. For the purpose of this discussion, we a$$ume that the Ianguage is described by its grarllma,r. (V,T,^9,P), there exists an algotithm Giverr a context-free grammar 6: for deciding whether or not .L (G) is empty. changes have to be Proofr For simplicity, a^$sume that.\ # L(G).Slight made in the argumerrt if this is not so. We use the algorithm for removing uselesssymbols and productions. If '5 is found to be useless,then tr (G) is ernpty; if not, then ,L (G) contains at least one element. I

(VrTr^S,P)' there exists an algorithm Given a context-free grammar 5: fbr determining whether or not I (G) is infinite. Proof: We assume that G contains no .\-productions, no unit-productions, and no useless symbols. Suppose the grammar has a repeating variable in the sense that there exists some A e V fbr which there is a derivation

Al

rAs,


8.2 Cl,osuRrilPnopnnrrns eNr Dncrsror,iAlcoRt't'HMSFoR,(,'oltrnxr-FnnE La,NcuncF:s

zfg

to have no A-productions ir.rrtltto unit-productions, r Since (J i$ ir,$srrrned and .t7cilnnot bc sitttultatreously emptv. Sintxl A is tteither nullable nor ir, rrsclesssvmbol. we have

SluAuSut and

A1z, where, u, ?, and a trre irt 7'*. But then S 1u,Au 3 ux'''Ay"u 3 ur"zynu is possible for all n,, so thrr.t I (fJ) is infinite. If no varia,ble r:an cver repeat, then the length of irny derivation is boundecl by lI/1, In that t:itsc,.L (G) is finite, I'hus, to get a,n tllgoritltrn for detertniuing wlxrtlrcr I (G) is finite, we rreed only to cletermine whclthcr the gramtnar ha,sstturt: rcpeating varia'bles. I'his can be done sirnply by drawing a depenrlerrc:vgraph for the va,ritlblos in sucir a way tha,t therc is att edge (A, R) whenever thcre is a corresponding production A --+:nB,!t. 'I'hen

any varia,trle that is rrt' Lhe base of a, cycle is ir repeatiug one. Corr$e quently, the gra,mmir.rhirs ir repeating va,riahlcrif rrrrd otily if the clepelncltlncy graph ha,sa,cyt:lo. Since we now have an algorithm fbr rlu:itlittg whether tt, grilrnrnitr ltas a repeating va,rierblc,wc have an algorithm frrr rlctertnining whetlrcr tlr ttot I (C) is infirritc. I Sornewhatsurprisitrgly,other sirllllt: properLiesof context-f'rtlclanguages rrr(: rrot so easily clealt with. As itr I'heoretn 4.7, wtl rniglrt look for a,n a,lgtr ritlrrrr to deternrine whcthcr two context-free grilrrrrnarsgenera,tethe sarntl Iarrguage,But it trrrrrs rnrt that there is rrclsuc:trrr.lgorilhn. For the mtlrntllrt, we do uot have the ttx:hrrical tnachiner:vfirr ltrtlllerly clefinirrgtiro rnt:arritrg of "there is no ir,lgoritlrtn," but its intrritivt: rncatrilg is clea,r. This is an important poirrt to which we will retrrrn lirter.

1 . Is thc conrplernent of the languagc in llxamJrle 8.8 contcxt-free? ffi 'lheorem B.4. Show that this larrguage is litrcar Consider the language .L1 in


220

Chopter I

Pnopnnuns

ol CoultLxr-!'nnn

LANGUAGES

3. Show that the family of t:ontext-free languages is closerl under homornorphism. 4'

Show that the family of linear lartgrrages is closed under hornorrrorphism.

5. Slxrw that the family of context-free languages is closecl under reversal. ffi 6. Which of the languagc faurilies we have discussed are not closed under reversal? 7. Show that the farnily of context-free languages is not closed untler difierence in gerteral, brrt is closed under regular difference, that is, if .Lr is rrrntext*free and .Ls is regula,r, then /,r - .Lz is context-free. I'

Show that the farnily of deterministic context-free languages is closed urrder regular difference.

9. Show that the family of litreat languages is closed under union, but not closcd unrler concatenation. ffi 1O. Show that the family of linear larrgrragesis not closed under intersection, 11. Show that the family of deterministic contcxt-free larrguages is not closed utrtler union and intersectiorr. 12. *13.

Cive an example of a cotrtext-free language whose complernent is not contextfree. Show that if .Lr is lirrear and 1,2 is regular, therr .L1.L2is a linear language.

W 14.

Show that the family of urtarrrbiguons context-free languages is not closed under urrion.

15'

Show that the farnily of unarnbigrrous contcxt-free languages is not closecl under interset:tion. ffi

16, Let .L be a rleterministic context-free language and defirre a new languagc L1 : lw : aw â‚Ź L,a â‚Ź X], .[s it necessarily true that .Lr is a tleterrninistic corrtext-free language'l 1 7 , S h o w t h a t t h e l a r r g u a g e7 : free.

{anb": n } 0,n is not a multiple of 5}iscontext-

18. Slxrw that the following language is context-free.

19'

7 : {tu e {4, b}- : n." ('u) : rr,t (ur),tu does not cttrrtain a substring aab} Is the farnily of deterministic corrtext-free languages closed under homornorphisrn'i

20. Givc the cletails of the inductive argurnent in Thcorcm 8.5. 21. Give an algorithrn which, for any giverr trntext-free grarnmar G, r:arr determine whethcr or not A e I (G). {S 22. show that therc cxists an algorithm to dctermine whether the language gerrcrated by some context-free gratrrrnar contains any words of length less than some givetr llumtrer rr. 23. Let -Lr be a context-free language arrd .L.: be regular, Show that there exists an algorithm to deterrrrine whether or not .Lr arrd .L2 have a cornrnon elerrrent.


Turing M o chi n e s

n the foregoing discussion, we have encountered $Qmefundamental ideas, in particular the concepts of regular and context-free languages and their association with finite autontata and pushdown accepters. Our study has revealed that the regular languages form a proper subset of the context-free languages, and therefore, that pushdown automa,ta are more powerful than finite automata. We also saw that context-free languages, while firndamental to tha study of progranmfng languages, a,re limited in scope' This was made clear in the last chapter, where our results showeclthat some simple languages, such as {atb"c"t} and are not context-frtle. This prompts rrs to look beyond context-free {**}, Ianguages and investigJatehow one might define new Ianguage families that inclrrde these examples. To do so, we return to the SEeneralpicture of an automaton. If we compare flnite arrtornata with pushdown automata, we see that the nature of the temporary storage creates the difference between them. If there is no storage, we have a finite automaton; if the storage is a stack, we have the more powerful pushdown automaton. Extrapolating from thiS observation, we can expect to disCover even m()re powerful language families if we give the automaton more flexible storage. For example,

22r


222

ChopterI

Tunlnc MACHTNES

what wtlrrld happen if, irr the general scheme of Figure 1.3, we rrsrxl two stacks, thrne stacks, a queu(],or $orneother storage device? Does each utorage device tlcfine a new kind of ir,rrtomaton and tlrrough it a new langrrirge family? Tlis approach raises a lirrge number of qucstions, rnost of whit:h turn out to bc rrrrinteresting. It is more instructivcl to ask a more arntritirlrs question and rxlnsider how far the concept of arr arrtomaton can btr ptrshed. what carr we $ay about the rnost powerful of arrtomata and thc limits of computatiorr? This leads to the funda,mentalcorrr:r:ptof a Turing machine and, in turn, to a precise elefinition of the idca of a nrechanicalor algoritlunir: computation. We bcgin our study with a fbrmal defirrition of a, Ttrring rnar:hine, thel develop sorne feeling fbr whtrt is involved by tloing some simplc programs. Next wc errgrrethat, whilu the mechanisrn of a, T[ring rrrachine is quite rudimentary, the concept is tlroad enough to c:oververy complcx processes. The discussibnr:rrlminatesin the Turing thesis, which maintains thirt any urmputational llrocess,such as those carried out by present-day cornprrters, t:elnbe done on a T\rring machine.

T h eS t o n d o r dT u r i n gM o c h i n e Although we carr cnvision a variety of automata with complex and sophisticaterdstorage deviuru, a,Thring rrrachine'sstorage is actually quite sirnple. It cart lle visualized as a single, one-dirrrensionalarray of cells, each of which can holtl tr single syrnbol. This array extcnds indefinitely in both directiorrs and is thercfirre capable tlf holdirrg an urilirnited amourrt of infbrmation. The infornmtion can be read irrrd changed irr any order. we will ca,lJsuch a storage device a tape hecauseit is analogousto the magnetic t,apesused in actual curnputers.

Definitionof q TuringMochine A Trrring rnat:hineis an automaton whose ternporary storage is a tape. rhis telpe is divided into cells, eac:hof which is caJrrrbleof holding one symbol. Associated with the tape is a read-write head tha,t can traval right or left orr the tape and tha,t ca,n read arxl write a single symhol on each nrove. To deviate slightly frorn the general schcme of chaptcr 1, the autornatorr that we use as a Thring machirre will have neither an input filc nor any special output rnu:hanism. whatcver input and output is necessarvwill be done orr the machirrc's tape. we will see later that this modificatiorr of our general model in sec;tion 1.2 is of little r:onsequence.we could retairr the input file arrd a specilic outprrt mechanisrn without affecting any of the corx:lusions we artl rrbout to draw, but we leave them out becausrl the resulting automaton is a little easier to rkrscribe.


9.1 THn Srer"rrrARlrTuR,Inc MacntNn

223

Figurc 9.1

Reed-write head

A dirrgrani giving an intrritivtr visualization of a T\rrirrg rnachiue is shown irr Figure 9.1. I)efinition 9.1 rnakes the notion prccise'

A T\rrirrg tnachine M is derfirrtxlby M : ( Q , X , f , ( t ,q 6 ,i l , F ) , where Q is the set of intcrnal states, X is the input nlphabet, I is a finite set of symbols called the tape alphabet, d is the transition function, ! e f is a special symbol called the blank, qo E I is the initial stntc, F C Q is the set of firral states.

In the rlclinition of a Ttuing rnachine,we assumethat E ! f * {n}, that is, that the input alphabet is a,srrtrsetof the tape alphabct, not including the blank. Blanks are ruled out a,r irrput for reasons that will become apparelrt shortly. The transition funt:tiorr d is defined as d:Qxf-Qxfx{,1,.R}. Irr gerreral, d is a pa,rtiirl futrction on Q x f; its interpretatiott gives the prirrciple by which a, Thritrg tnachine opcrates. The arguments of d are the current state of the control unit and the current tape symbol being read. The result is a new stattl of tlte control unit, a new tape sYmbol,


Chopter 9 TunIuc

Figure [|.2 The situation (a) before the move and (b) after the In()ve.

flnternal

MACIIINE$

state7o

l'l'l'l

f

Internal smte fI

I

l'l'Fl(b)

which replaces the old orre, and a move symbol, L rtr R- The move syurbttl indicates whether the read-write head moves left or right one cell after the new symbol ha^sbeen written ott the tape.

r,lll;f

ii*tnele

+,1

Figrrre 9.2 showsthe situation before and after ttNr move causedby tlx:

.i (qo,o) : (91,rJ,It) . We can think tlf a Thring machilre as a rather simple computer. It has a processing unit, which has a finite llrelnorYr and in its tape, it has a secondary storage of rrnlimited capacity. Tlrg instructiols that sut:h a c6mputer ca,n cs,rry ttut ttre very limited: it can $ensea symbol on its tape and use the result to decide what to do next' Tltc onlv actions the machine can perform are to rewrite the current symbol, to ctrange the state of thc control, and to move the read-writc head. This small instflrction set may seem inadequate lor doing complicated things, but this is not so. T\rring machines are quite powerfirl in principle. Tlte transition function d defines ttprogram" of the machine. how this computer acts, and wc often call it the As always, the automaton stai-t$ in the given initial state with sorne informatiorr on the tape. It then gotls through a sequenceof steps controlled by the transition lunction d. During this process, the conterrts of any cell on the tape may be cxamined and changeclmany times' Eventuallv, the whole process may ternrinate, which we achieve in a Tlrring machine bv putting it into a halt state. A Ttrring maclfne is said to halt whenever it reaches a configuration for which d is not defiiled; this is possible because d is a partiat function. In fact, we will assume that rro transitions are defined for any final state, so the T[rring machine will ha]t whenever it enters a final state.

I


TunInG MlcnIbrn 9.1 Tnn Sr.q,r.tnRRo

226

Figurc 9..3 A seqrrerrr:eof moves.

h't(HtilFld f -t

Consider the Tirring machine delined by Q : {qo,qr} , 5 : { a ,b } ,

f:{a,b,tr}, r : { q r }' ano d (qo,o) : (qo,b, ft) , d (qo,b) : (go,b, ft) , d (qo,!) : (sr, tr, tr) . If this T\rring machine is started in state qs with the symbol a under the read-write head, tlte applicable transition rule is d (go,o) : (go,b' ft)' Therefore the rea,rl-write head will replace the a with a b, then move right on the tape. The machine will remairr in state q0. Any subsequent a will also be replaced with a b, but b's will not be modified. s4ren the machine errcounters the first blank, it v/ill move left one cell, then halt in final state q1' Figure 9.3 shows several stages of the process for a simple initial configuration.

. H,fi,ffin#Ili.Hi$"

I

Take Q, E, I as definedin the previousexample,but let F be empty. Define dbv d (qo,o) - (q1,a, E) , d ( q o ,b ) - ( q 1 , b ,R ) , d (qo,n) : (qr, E, ft) , tI (qr, a) : (qo,a,L) , d ( s r , b ): ( q o , b , L ) ,

d (sr,il) : (qs,n,.L).


TuRrNc Macmbru$ Chopter Chopter 9 Tunrr,rc

To see what happens here, we cirn trace a typical ctr,se. Suppose that the tape initially corrtainsc,b..,,with the read-write head on the a. The machine then reads the a, brrt does not change it. Its next strrte is {1 and the readwrite head moves right, so thai it is now over the b. This symbol is also rt:ird and Ieft unchanged. The machine goes back ittto state {q and t}tc readwrite head moves left. We are rlow track exactly in tlre original state, trnd the sequence of moves stilrts again. It is dear from this that the maclfrrc, whatever the initial information on its tape, will run forevet, with the readwrite head moving alternately right then left, but making no rnodifications to the tape. This is an instalxjc of a Ttrring rlachine that does rrot halt. As an analogy with programmirtg tcrminology, we say tha,t the T\rring machine is in an infinite loop. I Since omr can make several different definitions of a Tirring machine, it is worthwhile to nummarize thc main features of our model, vrhich we will call a standard T\ring machine; 1. The T\rring machine hils a tape that is unbounded in both directions, allowing any number of left and rigltt moves. 2. The Turirrg machine is detcrministic in thc sensethat d defines at most one move for ea,chconfiguration. 3.

'fhere

is no special irrput file. We a"ssumethat at the initial time the tape has some specified r:ontent. Some of this rnay be considered input. Similarly, there is no special output device. Whenever tlte machine halts, some or all of the contents of the tape may be viewed as output.

These corrvnntions were cltosen primarily for the convenience of subse* qrrent discussiorr. In Chapter 10, we will look at other version$ of Ttrring nratrhines and discrtss their relation to our standard model. To exhibit the configurations of a Ttrring machine, we uso the idea of an irutantaneous description. Any configuration is completely determined by the crrrrent state of the control unit, the conterrts of the tape, and the positiorr tlf the read-writc head. We will use the notation in which frtIrz

or a t f l z ' ' ' a k - r q 0 , h o , k + 1 o' ' "' ' is the instantaneou$ description of a machine in sta,te q with the tape depictetl in Figure 9.4. The symbols et,...tiln show the tape contents, while q defines the state of the corrtrol rrnit. This convention is chosen so that


9.1 TnE Srlrun.qRnTuntuc MncHttqn

227

Figure 9.4

the positiorr of the read-write head is over the ccll contaitting thc sytnbol immediately followirrg q. I'he instantnrrcous description givcs otily a finite amtlrrrt of informatitlrr to the right a,nrllt:ft of the read-writc head. The unspet:ified part of the ttrpe is assumed to <xlrrtain all bla,nks; rxlrrrrally such blanks are irreleva'nt ir,ntl are not shown elxplicitly in the irrstiurtaneousdescriptiotr. If the positiorr of blanks is rc:levant to the clisr:rrssion,however, the blank syrubol rtray appear in the irrstantaneous desr:ription. For examplc, the insta'nttrrrtxrus description qE.ru indicates that the read-write herr,tlis on the cell to the immecliate krft of the first symbol of to and that this cell coutains a blank,

$fisrhpls S,il

The picrtrrrc:sdrawn in Figrrrc:9.3 correspond to the sequenceof instantaneous dt:scriptions Qoaa,hqya, bbqsl, bqlbA move fron one conliguration to a,nothtrr will tre denoted by F. Thus, II

d (s1, r:) : (q2,e, R) , then the rnove abqlctl I abe.q2d, is ntacle whenever the internal sta,ttl is q1, the tape t:orrtainsabcd, arxl the reacl-write head is on tlte c. The symtrol F has the u$rral neaning of irrr arbitrary nutnber of mtles. Subscripts, such as F14, are used iu argumtlnts to rlistinguish betweort several ma'crhirrcs. I

ll.i-r,

llfi

i:i-l-

.Et{frr$Sl'$9;,$

The action of thc T[ring machinc irr Figure 9.3 can be representul by rl11{r0.1bq(raI bbqptrF bqlb

qoilill bqLb,

T


Chopter9 Tunltrtc MncHrlrns Chopter

For frrrther discussion,it is convenient to summarize the: vir,rious observations jrrst macle in a forrna,lway,

'fhrirrg ma,chine.Tlrtlrr ir.nystring a1 . . . Let M : (8, X,l, d, Qo,n, F) be a a,k-Lqruht&t1+t.-.ant with n,; g I and h â‚Ź Q, is ir,ninstantalttx)lliJdescriptiott of M. A move { 1 1' " f i , A , - r q I A h A h + l

is possible if and only if d ( , 1 ro, r ) : ( q z , b ,H ) . A move A1 "'

" 't t L n | & 1 " ' tl,1r-1Q10 ri0.tt+

qTuk-rbAp'r1 "'

A,

is possible if ancl only if

d (qr,*n) = (qz,b,tr) , M is sa,icito halt sta,rtingfronr sorncinitia,lconfigura,tion rfli:x2 iI it:,rQiIzI tltqiauz

firr any qj arr(l a, fbr which 6(qi,o) is unrlcfined. The sequc:rrr:e of couligtrra,tionsleaclirrgto a halt statc will be called a computation.

Iixample 0.lJshorvsthe possibility that a T\rring machitrewill never hall, proceeclingin an crrdlessloop I'rrlm which it, carrnot escape. This situatior plays a, firnclarnentirl role in the dist:ussion of 'I\rrirrg ma,chines,so w() u$e a, special notation for it. We will reprt:sent it by x:lqf,?F co, indicating tha,t, starting frorn the initial l:ontiguratior :t,1qfr!1the tnar:hirrel never halts.


9.1 THn Sr^r'lDnn,l TuRIlrcj MacnIt-ln

229

TuringMochinesos LqnguogeAccePters Ttrring machintls can be vitrwed as accepttlrs in the fblklwing sense. A string tu is written orr the ta,pe, with bla,nksfillirrg out ther urtused portions. The rnachine is started in the irritial state q1 with the retr.tl-writehead positioned on the leftrrtost symhol of tu. If, after a sequellce of tnovels, the Ttrring rnachine entets a fina,l state atld halts, tlten tl is ctlrrsideredto be accepted.

Let M : (Q,t,f,d,qo,E,F) accepted by M is

L (M):

b e a , T \ r r i t r g m a c h i r r c Then thc language

( * I n1qyn2for sorneqy t r -F , t t 1 , r , â‚Ź f . ) ' i z,'â‚Ź X+ ; qptt., t

'l'his rlerfirritionirrdir:atesLha,tthr: irrput ,ruis written on the tape with blalks on eitfier sicle. The reason for excluding trlanks fiom t]re iuput rrow becomesqlcar: it assurosus that a,llthe itrput is rcstricted to a well-defined region of thc tape, bra.cketeclby hlanks ou the riglrt and left. Without this convention, the rna,chirrecould not lirnit the relgionin which it must look for the inprrt; rro matter how ma,ny hlrrtrks it str,w,it coukl ntlver be strrtr that thelrt:was not sorntl nonbla,nkirrput somcwltere else orr the ta,pt:. It savs Defirrition 9.3 tells us what tnust happetr when ut e l'(M)' rrotfiing abqrrt t]re outcorrrefor any ot]rer iuput. W]reu u is trol itr L(M), 6rre of two thirrgs ca,nhappen: the rrrachinecan halt in a norilitral sta,tcror it can errter an infinite loop a,ntl trever ha,lt. Any string for whit:h M does not halt is by clefinitiott trot in 1, (M)'

accepts tht: lauguage dcnoted ,.f,ffit*,fitg6l,Sr,$;#,:,iFbr E : {0, 1}, design a T[ring rnar:hinethat by the regular exPressiou00*. This is a,n easy exercise irr Turing rnin:hiile prograrrttrting. starting nt the lcft end of theri1put, wr: reacl each syrnbol anrl tlheck thnt it is a 0. If it is, we contimre by movirrg right. If we reaclr a blank withtlut eucorrntering alything hrrt 0, we terrrrinate anrl irccept the string. If the input txrntains a i any*fiere, t|e strirrg is not irr r (0i].), arrclwe halt itt a tronlirral state. Io keep track of thc r:omputtrtiotr, t,wo interrral sta'tesQ: {qo'{t} and one: final stirtc F: {qr} are sufticicrrt. As trir,nsition functiotr we carr take

d ( s , r , 0:) ( q 0 , 0 , I t,) d (qo,tr) : (q1,n, fi) .


230

Chopterg Trrurrc MACHTNEs

As long irs a 0 appears rrnder the re^d-write head, the head will move to ther right. If at a,nytime a 1 is read, the rnachine will halt in the nonlinal state q()' $ince d (,nr,1) is undefined. Note thirt the 'rhrirrg ma,chineaJsohalts in a Iina,l state if utarted in stnte {s on a blank. we could interllret this as ilt:ceptance of A, but for ttxlhnical rea^s()nsthe ernpty string is not includecl irr Definition 9.3.

T The recogrrition of morr: complicated languagesis more difficult. since illrring mac]rincs have a primitive instrrrr:tion set, t]re computations that we carl pregram ea"sily in a higlxlr level lailguaEle ere oftr.:n cumbersome on a Turing machine. still, it is possible, and thc concept is ea^syto understand, as the next exarnrrlcu illustratc.

q.i Exorrrple

For E : {a, b}, design a Thring machine that ar:r:eprs 7:{a,rh":n}1}. Intuitively, we solve the problern in the following fashion. starting at the Ieftmost a, we check it off by replacing it with some symhol, say z. We then Iet tlxl read-writc head travcl right to fincl the leftrnost b, which in turrr is checked off by replacing it with another symbol, say y. After that, we go left again to the leftrnost a, replrr,ceit with a,n r, then move to thc leftmost b and replrrce it with .r7,and so on. T}avelirrg back and forth this way' we ma,tch each .r,with a correspondirrg b. If after some time no c,'s or b's remain, thc:n the strirrg must be irr .L. working orrt the cletails, we arrive a,t a conrplete solution for which Q:

{qo,qt1qz1qBtq4} 1

F:{qn}' g:{a,b}, | : { a , b , . r :y, , E } . The tra,nsitions can be broken into several parts. The set , i ( q o ,a ) : ( q t , r , R ) , d (gr, a) : (qr, a, ft) , d ( q r ,y ) : ( g r ,u , f t ) , d ( q r , b ): ( q z , y , L ) , rerplacestlrc leftmost a with an lr, then causcrJthe read-write heatl to travel right to the first b, rcpla,cingit with a y. when the y is written, the machine crrtersa state qz, indicating that an a has beru successfullypaired with a b.


9.1 Tsn SrannaRn TuntNc MaoHItIp

231

The next set of transitions reverses the direction utttil an z is encountered, repositions thqt read-write head over the leftmost a, and returns control to the initial state. d ( s z ,g ) : ( q z , a ,L ) , d ( q z ,a ) : ( q z , a ,L ) , 6 ( q z , n ) : ( q o ,r , E ) . We are now back in thc initial state q0, ready to dcal with the next c, and b. After one pass tlrrough this part of the computation, the machine will have carried out the paftial computation q o e ,"a. , r t b b... b i * q o o . ' . a y h" ' b , so that a single a has been mtr.tclted with a single b. After two passes' we will have completed the partial cornputation qoao,''a ' h h ' ' ' t i * * q s " ' ( r A Y" ' b , and so on, indicating that the matching processis being carried out properly. When the ir4lut is a string a'b', the rewriting continues this way, stop ping only when there are no more o'$ to be erased. When looking for the Ieftmost a, the read-write head travels left with the machine in state q2. When an r is encountered, the direction is reversed to get the a' But now, instead of finding an a it will find a 37.To termirrate, a final check is made to see if all a's and b's have been replaced (to detect input where an a follows a b). This can be done by 6 ( q o , a )= ( Q : t , a , R,) 5 ( q z , a ) : ( q : t , a ,R ) , d (Sr,!) : (ga,tr, E) . If we input a string not in the language, the computation will halt in a nonfinal state. For examplc, if we give the machine a string a'b*, wittr rL 7 ffit the machine will trventually encounter a blank in state q1. It will halt becauseno transition is specified for this case. Other input not in the language will also lead to a nonfinal halting state (see Exercise 3 at the end of this section). The particular input aabb gives the following successiveinstantaneous descriptions

araabbl rqlabbl raqlbbl nq2ayb tsq2raybl rqsaybl nrtnyb I myqlbl rnq2yyl ng2ryy I r:xqoyuI rryqy I myqsJ I rnyyZqal.


282

Chopter 9 Tunrr'rc;MACHrFres

At this point the T\rring machine halts in a final state, uo the string aabb is accepted. You are urged to traccl this program with several more strirrgs in -L, a"s well a-rwith some not in -L. I

i

Design a T\rrirrg macrhinethat accepts 7,:{a"bncn:n > 1}. The ideas userl to Exrr,mple9.7 are easily carrierl over to this case. We matclr each a, b, and c by replacing them in order lty r, u, a, respectively. At the end, we check that all original symbols have been rewritten. Although conceptua,lly a simple exterrsion of the previous exarnple, writing the actual progralrr is tedious. We leave it as a somewhat lengthy, but straightforward exercise. Notice that even though {o"bn} is a txrntext-free language and {anb"c"} is not, t}tey r:an he acceptedby I'uring rnachineswith very similar structufesI Otte ctlnt:hrsion we can draw frorn this example is that a Thring machine can recognize sorrre langrrages that are not contrlxt-free, a first indication th.r,t Tirring machines are rn()re powerful than pushdown arrtomata.

TuringMochines os Tronsducers We have had little reason so far to study transducersl in language theory, accepters are quite adequate. But its we will shortly see, T\rring machines ate tttlt only interesting as language accepters, they provide us with a simple abstract model for digital comlrrrter$in general. Since the primary purpose of a conrputrlr is to transform input into orrtput, it acts as a transducer. If we want to rnodel comprrters using T\rring rrrachines,we have to look at this er,spectmore closely. The inprrt for a computation will be all the nonblank symbols on the tape at the initial time. At the conclusiorrof the computation, the output will be whatever is then on the tape. Thus, we can view a Tirring machine tra,nsducer M as arr implementation of a furrction / defined by

fr=f(w), provided that QywI L.rQyfr, for sornc finrr,lstate q7.


!|l

THn SrAFtrrA,ItoTunINc M,tcsItqn

233

inn.1,$.i'l$fimmir.Plil, or just A function / with clomain D is uaid to be Turing-computable : qo, E, F) (Q,t, f d, tnachine M computable if there exists some Ttuirrg , such that

QowluWf@t),

eyeF,

for all z.' â‚Ź l).

As we will shortly claim, all the common mitthematical firrrctions, no matter how cofirplicated, are T\rring-computable. We sta,rt tty looking at some simple operations, such as addition and arithmetic comparison.

Given two pgsitive integers l' and y, design a T\rri1g machirre that computes nlA. we first have to choclsesome convention for represerrting positive integers. For sirnplicity, we will use unary notatiol in which aly positivc irrteger " i, ,*pr"."nted by ur (r) e {1}+, srrch that

lur(r)l : z. we rnust also cler:idehow r and 3r are placed orr the tape irritially and how their sum is to appear at thc end of the ctlrttputation. We will assume that ro (r) and tu (37)are on the tape in unary rrotation' separated by a single 0, with the read-write head on the leftmost symbol of trr(z). After the computation, ru (r + U) will be on the tape followed by a single 0, and the reacl-write head will be positioned at the left end of the result. we thcrefore want to design a Ttrring machine for performing the compltation

qour (r) 0t, (s)i qp (*+ s) 0, where q1 is a final state. constructing a prograrn for this is relatively simple' All we need to rlo is to move the separating 0 to the right end of u (g), so that the addition amounts to nothing more than the coalescing of the two


234

Chopter 9 'Itltrrruc MncnrNns

strirrgs. To achieve this, we construct M : (Q,X, l, d, qs, E, F), with Q : { s o ,q t , q z t q J ) q 4 } , F : {qn}, ,l (s,j,1) : (qo,1, ft) , ,i (,?0,0) : (qr, 1, ft) , d ( S r , 1 ): ( r 1 r , 1R, ) , d ( q r ,t r ) : ( q z , Z , L ) , , 5( q z ,1 ) : k t t , 0 , L ) , d ( q r ,1 ) : ( r l l ,1 , [ ) ,

d (Ss,tr) : (s+,tr, ft) , Note that in moving the 0 right we temporrrrily create an extra l, a fact that is remembered by putting the machirrc into state q1. The transition d (qr, 1) : (r1:1,0, iB) is needed to remove this at the end of the r:omputation. This can be seen from the sequen(reof instantaneous descriptions for adding 1 1 1t o 1 1 ; q n 1 1 1 0 1F1 1 q 6 1 1 0 1l 1* 1 l q s 1 0 1 1F 1 1 1 q 0 0 1 1 F 1 1 1 l q 1 1 F 1 1 1 1 1 q 1F1 1 1 1 1 1 1 9 1 t r F 1 1 1 1 1 q 2F1 1 1 1 1 q 3 1 0 i q s t r l 1 1 1 1 0F q a 1 1 1 1 1 0 , utrary notation, although cumbersorrrc for practical computations, is very convenierrt f'or programrrring T\rring machines. The resulting programs ara much shorter and simpler than if we had used another representation, such as binary or decinnl.

I Adding numbers is one of the furrdamental operations of any comprrter, one that plays a part in the synthesis of more complicatecl instructions. other basic operations are copying strings and simple comparisons. These can also be done easily on a Ttuing machirre.

Exotnple fl"10

Designa Thring ma,chinethat copiesstrings of I's. More preciselv,find a machinethat perftlrms the corrrputation tyl'..,i q.1ww, for any ?u e {1}*


9.1 THn Sremoann TuRrr'rcMrc;Hrrue

235

To solve the problem, we irnplemcnt thc fblklwirtg itttuitive process: 1 . Rtllllaclc:c:very 1 lry an r. a. Firrd thc riglrtrnost r and replace it with l.

3 . Tlavel to l,he right end of the current nonblank regiorr arrd create a 1 there, 4. R,c1lr:irtStcps 2 and 3 until there are no more tr's. A T\rrirrg rnachine version of this is d (qu,l) = ({0, r, Il) ,

d (so,n) : (gr,!, r) , . l ( q r , z ): ( q 2 , 1 , f t ) , ,5(Sr, 1) : (gz,1, ft) , d ( S z ,n ) : ( r 1 r1, , t r ) , , )( S r ,1 ) : ( q 1 ,1 , t r ) , ,l (,1r,tr) : (q3,D, ft) , where q3 is the only final stirtc. Tlfs trtay be a little hard to see a,t Iirst, so let us tra,ce the pr()grailr with Lhe sirnple string 11. Tlrc c:orrrputation perfbrmeldin this case is qsll F rrgeI F tru(en I rqln F r l g 2 n F r q 1 1 1F q 1 r l 1 F 1 q 2 1 1F 1 1 s 2 1F 1 1 1 q 2 t r F 1 1 9 i 1 1F l q 1 1 1 F s 11 1 1 1F q 1 t r 1 1 1 1 F q31111.

T

I,et r a,rrd.r7trt: two positive integers represented in utrary notation, Clorrstruct a TLrring urachine tha,t will halt irr ar,Iirral state q, if r j 91,ir.nd that will lralt in a nonlinal sta,teqn iI n { g. More specilically, the mat:hint: is to perform the computation

(z)0, (E) t qrut(r) 0, (y) , ryprr'

if n Z A,

+

qn,rr(r)0, (E) | q.w (r) Oru(y) ,

if r < A.


236

Chopter 9 Tunrug MecnrNns

To solve this problem, we carr rrrrethe idea,in Example 9.7 with sott.e minor rnodiflcations. Insteacl of rnatr:hing tz's and b's, we match each l on the left of the dividing 0 with tlrr: I on the right. At the end of the matching, we will have on the taoe eithc:r tyr-..lllnn...rn

rn...m]m...21ltr. deperrdirrgorr whether u > E or A > fi. In thc first case,when we attempt to match arrotlrr:r 1, we encounter the blank rrt the right of the working space. Ihis can bc usecla,sa, signal to enter the state gr. Irr the second case, we still find a 1 on the right when all I's on the lcft her,veheerr replaced. We use this to get into thc other state qr,,. The cornplt:tc progra,m fbr this is straightforward and is left as arr exercisc. Tltis t:xample ma,kesthe irnporlant point tha.t a,Thring machine carr bt: progralrurlc(l to tra,ke decisions based otr aritlrrntltic cnmparisons. This kind of simplc dtxlision is comrnon in the rrrachirxrlangua,geof computers, whrlrc alternate instrut:tion strea,msare entered, deprlrrdirrgon the outcome of an arithmetic operation.

t

**

L, Write a Turirrg rnachine simulator in sorne higher-level programming larrgrrage. Such a simulator should accept as input the tlescription of any Ttrring tttacltirre, together with a,n inititl,l corrfiguration, and shoulcl produce ar^s outprrt thc result of the rrrmprrtatiorr. ?. Design a Turing rnachirre with no morc than three states that acccpts the language L(a(a-l b)-). Assrrrnethat X: {a,b}. Is it possi}rleto do this with a twogta,te machine'l ffi 3, Deterrrrirre wha,t the 'I.\rring rnachitre irt Example 9.7 does when preserrted witlr the inputs aba arcJ.aaabbbb, 4. Is thcrc any input for which the Thring trachirre irr Example 9.7 goes into at irrfinite loop?


THr,i S'rAlrr^,nrr TuRrr.rc MaclHrlrn

9.I

5. What language is accepted by the rnachine M { a , ,b , ! } , d , Q o , l , { q s } ) w i t h

:

237

({qo,qt,qz,qt},{.r,b},

ii (qo,a) : (qr , a, ,t) , d (r70,b) : (q2,b,R) , d (qr , b) : (sr, b, Il) , ,i (Sr, n) : (gs,E, ft) , ,i (qz,b) : (sr, b,,C) , ,i (qz,a) : (qs,a, R) . 6. What happens it Exarnple !1.I0 if the string u conta,ins any syrnbol other than 1? 7, Construct l\ring

machines that will accept the fbllowing lirnguagestxr {o, b}.

(*) r:Z(oba-b)

ffi

(b) f : {rr.,: lrrrl is even} ffi (") ,:

{ t u : l t o l i s a r n u l t i p l eo f 3 }

( . 1 )f I : { e | ' 6 " ' ' n } I , n f m,} : : ( e ) / , { u : n , "( u ) n 6 ( t a ) } (f) I,:

{ a n b " " s " + ' n ;n } 0 , r r r>, 1 }

( g ) f l l :l l { a ' " I ) " t L " b " : r r . } 0 } (h) I:

{flitbzn:n> I}

For each problerrr, write out d irr rrrrnplete detail, then check your answers by tracing several test exarrrples. 8. Design a T\rring machine that a,cceptsthe ianguage L:{tntr.t:rue{a,b}+}. L

Construct a Ttrring uachinc to compute t,ire lunction ./ (r) : ,*, whereur â‚Ź {0, 1}+.

10. Design a Turing machinc that fincls thc micldle of a string of cvcn lcngth. 'I\lring machine Sper:ifically,if ru : e,tez...e,*e,u;_t...oz',, with o1 â‚Ź X, the -X. {ffi slrorrltlprorlrrc:efi:ataz...a^cetu+t...e,zrrj where c g f 'I\rring 11. Design machines to corlpute the following functions for ;ir and 3r positive irrtegers represerrtetl irr rrnary. (a) /(u):3a

(u) / (t, 'a): x -'!, --0,

'r > a x<A


238

Chopier 9 Tunnrc MAcHTNEs

(") f ( n , u ): 2 n - l 3 a ( d ) f (") : E,

r-tl

if t is evert r r f f 1 $ (rdd

(e) / ( t ) : r m o d 5 (f)

t2.

the /(*) : Lfrl, whereLfJ denotes cqualto f.

largest integer less tltarr or

Design a T\rring machine with f : {0, 1, n} that, when startcd on any cell containing a blank or a 1, will halt if and only if its tape has a (l sornewhere

on it. ffi 1 3 . Write out a complcte solution for Example 9,8. L4, Give the sequence of instantaneous descriptions that the Turirrg rnachirre in Example 9.10 goes through when presented with the it4lut 111. What happens when this machine is started with 1lt) on its tape? 1 5 . Give convincing arguments that the T\rring mar:hirreirr Exarnple 9.10 does in fact carry out the indicated t:ornptrtatiorr.

1"6. Cornplete all the details irt Exatnple 9.11. 1 7 . Supposc that in Exarnple 9.9 we had decided to represent r and gr in birrary. Write a T[rrirrg rnac]rirre progranr fbr doing the indicated cornputation in this representation,

1 8 . Sketch how Exarnple 9.9 trlrltl be solvetl if r and g/ were represented iu decirnal.

1 9 . You may have noticed that all the examples in this scction had only one final state, Is it generally true that for any Thring rnachine, there exists another one with only one final state that acccpts the same language? ffi

2 0 . Definition 9.? excludes the empty string frorn arry language ac:ceptedby a Ttrrirrg rrrachine. Modify the definition so that languages that contain A mav be accepted.

ffi

C o m b i n i n gT u r i n gM o c h i n e sf o r Complicoted Tosks

We have shown explicitly how some important operations fbund in all computers can be done otr a Thrirrg rna,chine.Since, in digital complrters, strt:h primitive operations rrre the brriklirrg blot:ks for mclrc c:orrtplcxirrstructions, lct us seerhow thcstr basit: opura,tiorrs carr also bc put together on a Thring machine. To demonstrate how'I\rring machines can be cornbined, we follow a practice common in prograrnming. We start with a high-level description, then refine it successivelyr.rntil the program is in the a,ctua,lla,nguagewith which we are working. We can describe T\rring ma,chines$evcral wirys irt ir, high level; block diagrams or pserrdocodeare the two alpproir.r:hr.'s wc will ust:


9.2 Cor"lerNlr'rcTURIr'tcMacnrNns pon, Corr,rpl,IcArnn Tasxs

239

Figure 9.5

J'@,t)

rnost frt:qrrently in subsequent discussions. In a block diagram, we encapsule computatiorrs in boxes whose function is described, but whose interior details are not slrown. By using such boxes, we implicitly claim thai they r:nn actually be constructed. As a first example, we cornbine the machines in Exarnples9.9 and 9.11.

Design a Turing machirre that cxrmputes the function f(n,y):nt'!t - 0,

ifnlt!, If n < 31.

For the sake of discussiorr, asrrume that r and E are positive integers in unary representation. The value zero will be representedby 0, with the rest of the tapc blank. The cornputation of / (r, gr)can be visualized at a high level hy means of the diagrarn in Figure 9.5. The diagram shows that we first use a comparing nrachirre,like that in Exa,mple 9.11, to determine whether or not r> y. If so, the cornparer serndsrr,start signal to the adder, which therr computes n*y. If not, arr erasirrgprogram is started that changesevery 1 to a blank. In subsequent discussion$, we will often use such high-level, blackdiagram representations of Thrirrg rnachines. It is certainly quicker and clearer than the corresponding extensive set of d's. Before we accept this high-level view, wc muut justiSr it. What, for example is meant by saying that the conrparer serrds a start signal to the adder? There is nothing irr Definition 9.1 that offers that possibilitv. Nevertheless, it can be done in a straightforward way. The program for the cornparer d is written a; suggested in Example 9.11, using a Ttrring machine having states irrdexed with C. For the adder, we u$e the idea in Example 9.9, with states indexed with ,4. For the eraser E, we corr$tnrct a Ttrring machine having states indexed with E. The computations to be done by C are

qa;,gur (r) \ut (y) I nop*(r)Oru(s) ,

if r'P a,

arrd +

qc,ow(r) 0, (E) | qa,ow(z)Oto (g),

if x < s.


240

Chopter 9 Tun,rwcMAcHrNos

If we take {,1,0 and qp,g as t}re initial states of ,4 and E, respectively, we see that C starts eithcr A or .8. The computations performed by the adder will bt: ( . n ) O r (ug ) i q o , y r u ( r * u ) 0 , tt.s,ow and that of the erraserE will be q n , o t (un ) 0 , ( y ) i q r , r o . The result is a single Thring machine that comhines tlrc actiotr of C, A, and E a^sindicated in Figure 9.5.

T

Anotlrt:r useful, highJevel view of T[uing1rnac]rirresis otre itrvolving pseudocode. In computer programming, pseudocodeis a way of outlining a computation usirrg descriptive phrases whose meaning we claim to understand. While this description is not usahle on the cornputer, we assume that we can translate it into the appropriate language when needed. One simple kind of pseudocode is exentplified by the idea of a macroinstmction, which is a single-statement shorthand for a sequence ()f lov/er level statements. We first defi.ne the macroinstnrction in terms of tlte lower level language. We then use the macroinstructiorr irr a progralrr with the assumption tha,t the relevant low-level code is substituted for each occurrence of the ma,t:rtr instruction. This idea is very useful irr Ttrring machine programming.

Consider the macroinstruction i/ a then Qi else q*, with the following interpretation. If the Thring maclflle reads an a, then regardless of its current state, it is to go into state {i without changing the tir,pc content or moving the read-write head. If tha syrnbol read is not an a, the machine is to go into state 96 without changitrg anything. To implement this macroinstnrction requires several relatively obviorrs 'f\ring steps of a machine. 6(qt,a): (qio,a,R) d(qr,b) : (qftg,b,ft) d (gio,c) : (qi,c, L) d (qon,c) : (qp,c,L)

for all qt€Q, f o r a l l q a€ Q a n d a l l b € f - t o ] , for all c € f, for all c e l.

Thc states qio and {6s are new states, introduced to take care of complications arising from the fact that in a standard T\rrirrg machine the read-write


9,2 Corr.rerlrrNc Tunrr'rc MAcrrrNrilsFoR CoMpllcAruu T.q.sxs

241

head changesposition in each rnove. Irr the ma,croinstruction, we want to change the state, but leave Lheread-writt: hcir,dwhere it is. We let the head mrrve right, but put the machine into a stirtc (Ii1 $ qko. Tliis indicates that a lefb move mrrst be ma,debefore entering the drlsirul statc qi or q4. I

Goirrg a stt:p firrthtlr, we (:an replace macroinstructions with sutrprtr grarIIS. Norrnillly, rr, milcroinstruction is replaced by actual corlc a,t ctr,ch occurrerrcci wlrcreiul er,subprogram is a single piece of code that is irrvokcxl repcirtt:dly whenever needed. Subprograms are fundarnerrtal to high-lcvcl prograrrrrning languages, but they can also be used with TLrring rnac:hirr:s. To rnake this plrrusible, let rrs outline briefly how a 'I\rring rnachine carr bc used as a subprograrrr that t:arr bt: irrvoked repeatedly by another Thring machine. Ihis requires a new featurc: ther ability to store infornration on tlrt: r:a,llingprogram's configuration so the configuratiorr carr br: rtx:rcilterd on returtr frorn tlN,' subprogram. For example, say rnachine A in state q,; invokes rrtircltint: B. Wlxlrr B is finished, we would like to resurne prograrrr A itt sta"teq,;,with the read-write head (which may have rnoved during B's operatiorr) irr its originrrl pla,ce. At other times, A may call B frorn state {i, itt w}tich c:itscc:rlrrtrol shor.rldreturn to this state. To solve the control transfer problern, w(l rmurt be ir.bleto pa,ssinforrnation from A to B ancl vice versa, be able to r(l(:r{lntcA'u cxrnfigurationwhen it recovers control from B, aucl to assurethat the tetttporarily suspcndcd comprrtations of ,4 are not afftx:ted by the execrrtion of /1. To solve this, we carr divide the tape irrto several regiotrsas showrr in Figurtl 9.6. Before A calls B, it writes the irrfrrrurirtionneededby B (e.g., ,4's current state, the argunents for B) on thc: tirpr: in some regicln 7. .rl then pa*sses control to B by making a trarrsitiorr to the start sta,teof B. After transfer, B will use 7 to find its input. The workspaur ftrr B is seprlrtlte fiom 7 and from the workspace for A, so no irrterfertlrr(lc (jirn or:<:rrr.When B is linished, it will return relevant results to region T, where A will expect to firxl it. Irr this wiry, the two programs can interact in the required fashion. Note that this is very sittrilar to what actuirlly htr,ppensin a rerr,lrxrmputer when a subprograrn is called.

Region separator

Figtue 9.6

Workspace forr{

Workspace for,B


Chopter 9 Tururuc MnctIIwns

Figure 9,7

v

We can now program T\uirtg rrrachinesin pseudocode, provided that we know (in theory at least) how to translate this pseudocode into 8,n actual Ttrring machine program.

Design a Ttrring machine that multiplies two positive integers irr unary notation. A rrrultiplication machine can be constructed bv combining the ideas we encourttered in adding and copying. Let u,'Ja$rJlrrnethat the initial and final tape contetrts are to be a,sindicated in Figure 9.7. The process of multiplicatiorr can then be visualized as a repeated copying of the rnultiplicand g for each 1 in the multiplier r, whereby the string g is added the appropriate number of times to the partiallv cornputed product' The following pseudocodeshows the main steps of tlte process. 1. Repeat the following steps until r corrtairrs tto more I's. Find a 1 in r and replace it with atrother symbol a. Replace the leftmost 0 by 0y. 2. Replace all a's with I's. Although this pseudocode is sketchy, the idea is simple enough that there should be no doubt that it can be dorre. I In spite of the descriptive nature of these examples' it is not too farfetched to conjecture that Ttrring machines, while rather primitive in principle, can be combined irr rnatry ways to make them quite powerful. Our examples were not general a"rrd detailed enough for us to claim that we have proved anything, brrt it should be plausible at this point that Tirring machines can do some quite complicated things.

1. Write out the complete solution to Example 9.14.


9.2 CortelNlr.rc Tunnrc Mecumns FoR, CoMplrcArun

Tasxs

Z4B

2 . Establish a conventiorr for representing positive an{ legative integers in unary rrotation. with your conventiorr, rrketch the construction of a subtracter for t:omputing a - 3y,

3 . using adders, sutrtracters, comparers, copiers, or rnultipliers, draw block diagrams for Thring nrarhines that computc the functions

( a )/ ( " ) : n ( n a 1 ) ,

ffi

(b) /(n):r'1', (c)/(rz):2", (d) /(n):,11, (e) /(n) : p'!, for all positive integers z.

4 . use a bhck diagram to sketch the implementation of a function / defined for all w1, w2, 'rrrrg {1}+ by

f (wr,wz,ws) : i., where i is such that ltui j = ma-x(lturl,luzf , fu3l) if no two tu have the sarrre length, and i:0 othcrwise. "high-level" 5- Provide a rlescription for T\rring rnachines that ar:cept the forIowing la'guages orr {a, b}. For each problern, define a set .f appropriate macroinstructions that you feel are reasonably easy to implement, Therr use them for the solution, (u) Z:

{,ruturt}

(b) t:

{*r.,

t wr # wz: lurl :

l,rrrrl}

(c) The complement of the language in part (a) (d),

: {u"bn' : m,:

ffi

n?,r, > 1}

(e) f : {an : n. is a prime nurrrtrer} 6, suggest a method frrr represerrting rational numbers on a Ttrring rrrar:hine, then sketch a method for addirrg and subtracting such numbcrs. 7, Sketch the constructiotr of a Ttrrirrg machine that can pcrform the adclition antl multiplication of positivc intcgers x ard g given in the rrsual decimal notation. 8. Give an implementation of the macroinstruction searchright (.a,qt,,qi), which indicates that the rnachine is to sea.r'r:hits tape to the right of the current position for the first occurrcnce of the syrnbol a. If an a is encountered befrrre a blank, the machine is to go into state ql, otherwise it is to go into state 4i. W


244

Chopter 9 TunIwc MAcHINEs

Use the macroinstruction in the previous exercise to design a Ttrring rnacltirte on E: {a,b} that a,cceptsthe language L(ab-ab"a). 1 0 . Use the rlacroinstructiotr searchright irr Exercise I to creatc a Thring machine program that replaces the syrrrbol immediatcly to the left of the leftrrxrst a by a blank. If the irrput cofltains no a, repla,ce the rightmost rrorrblaflk syrnbol byab.

Turing'sThesis 1flffiffitr Thc prtrcxling discussiou not only shown how a T\rrirrg rnachitre can be constnrr:ted from sinrpler parts, but also illrrutrat()sa llegative aspect of working with such low-level a,utoma,ta.whikr it takes very little imaginatiott or ingenr.rity to tra,nslattt a block diagram or pseudocode into the t:orrtlsporrdirtg Tlrring rnachirreprogram, actually doing it is timtl t:orr$urnirrg,error prone' irntl irdds Iittle to our understa,nding. Tht: irrstnrction set of a Turing ma<:hirrt:is so restrictecl that any argument, soltrtitlrr, or proof for a nontrivial probletn is quite tedious. Wt: rrow face a dilenuna: we want to t:la,im thilt T\rrirrg lllacllines can pcrfrrrrrt ttot only the sirnple operations fbr which wc have provided explicit prograllls, but nrore complex procert$esils wtlll, clescribableby block diagrarrls or pseudocode. To defilrrd such claitns against challenge, we sholld show the releva,nt progrirm$ cxplicitly. But doing so is unpleasant and diutracting, a,rrclouglrt to be avoided if possible. Somehow,we would like to firrd a way of carrying out a reasonably rigorous rlist:rrssiorrof T\rring machines without havilg to write lengthv, low-ltlvtll code. There is unfortutlately no completely satisfactqry way qf gt:ttirrg out of the predicament; the best we can do is to rea,cha rea,sona,bklt:tlrnprotnise' To see how we might achieve such a comprclmise,wo ttrrrr to a sornewhat,philosophical issue' We can rlrirw sornrlsirnple conclusionsfrom the examplesin the prtrvitlrs 'furing macllines appeal to be more pqwtrrful thir,n section. Ttrc first is that ,:omment on this, stlt: Ext:rcise 2 at the etrd of pusltclowtr aul,o[tata (for a, the construction of a T[ring mawt: skctchcd this section). In Example 9.8, ancl for which, r:orrscqucrrtly, uot coutext-free which is t:hinc fbr a larrguage 9.9, 9.10, ailtl 9.11 s]row thal Iixamples puslrdown exists. au[omaton lro perform string optlrittittns, simple arithmetic can do sorne Tlrrilg machines also ilThc discussion simple compilrisons. tnake sotne rrarripulations, attcl cotnplex to solve nore t:orntrirred lxl prirnitive cilrr opera,tionu Iustrates how problems, how sever.r,IT\rring rnttclltirtescan lre cotnposed, and how olle program can act a,sa, subprogrrur filr arxlthrlr. Since very complex operations ca,n he brrilt this wiry, w(.'rliglrt srrspect that a'ftrring tnachine begins to ilpproar:h ir,tyllir:irl t:ornputcr itr power. Suppose we were to make tlle conjecture that, in some $enst:, T\rriug nachitres are equal itr power lo a typical digital computer? Htlw trtiulrl wtl


9.3 TuRrNc'sTnnsrs

245

def'endor refute such a hypotlxrsis'l To def'end it, we could take a seqrrerrcc6f increasingly more diflicult problerns arrd show how they are solvecl by sorne Tlrring machine. we might also takc the ma,chinelanguage instruction set of a specrific computer and design a Tlrring machine that can perform all the instrr.rctionsin the set. This would rrndoubtedly tax our patience, but it ought to be possible in prirrciplc if orrr hypothesis is correct, Still, while (lvery ,5llcce.$s irr this directiott would strcngthcn our convictjon of the truth rlf tIrc hypothesis, it would not lead to a proof. Tlxr dirlir:rrlty lies in the fact thir,t we don't know exactly what is meant by "a typictr,l cligita,lcornputer" arrrl tha,t we have no lneans for rnaking a precistl rlcfirrition. we crrrr *lsr approa.r:hthe problem from the other side. wrr r'ight try lo find sorne proctxhrru filr which we ca,nwrite a computer prograrrr, but frrr which we can show thirt rxr T\rrirrg ma,chinecan exist. If this were possible, we would hirvt: ir lrtr,sisf'rrr rejecting the hypothesis. But rlo one has yct lrr:crr able to prorluctl a r:ountt,rrexa,mple; the fact that all such tries have beerr utrsuccessfulrnust btt takttn as circunrstantial evidence that it cannot be done. Every irrrlic;ation is tha,t Ttrring machines are in principle as powerfirl a"sarry corrrputer, Arguments of this type lul A. M. Ttrring and others in the mid-lg30's to the celebrated coqiecture called the T\rring thesis. This hypoihetis states tha,t any cornputation lhat can be carricrl orrt lly nxrchanical mea,n$can be perfbrtrr:d by some T\rring rna,chine, Thin iu rr.sweeping sta,tement,so it is important to keep in rnirrrl what Thrirrg's thesis is. It is not something that can be proved. Tb do so, we would hrrvr:trl define precisely the term ,,mechanicalrneans." This would requirc sorne clthcr irllstrar:t model tr,ndleave us no further aheacl than before. Ihe 'rhring thesis is rrrorc protrxrrlyvierweda,sa,definition of what constitutes a mechanical cornputirtiorr: ir cxlrnllrtatiorr is mechanical if and only if it can be performed by sorne Turirrg rnachirrc. If ue take this attitude attd regat'd the TLrlirrg thcsis sitrply as ir dclinitiotr, we raise the clucstiorras tti whcther this definition is sufficiently broad. Is it far-reaching enough to covcr' cvcrythirrg wr) now do (a,nd r:onceivablv tright do in the future) with computers? An unequivocill "yes" is rxrt possible, but the evidence in its firvor is vr:ry strrlrrg. Some rlrgrrments for a,cceptingthe Ttrring thesis as the definition of a rnechanical comDrrtation'

1 . Arrythi'g tha.t r:,,n be done or any existing digital computer can also be dotre by a T\rrirrg rnir:hirrt:.

2 . lrToorre,'ha,syet been a,bleto suggest a problern, solvable by what we irrtuitivtly rxrnsidcr ar.na,lgorithm, for which a Thring rnachine program canrtot llr: writtcrr.

3 . Alternative moclels have beel proposed for nrechanical ctrrnputation, brrt none of them are tnor:epowerful than the T\rring rnachine model,


246

Chopter 9 Turrtuc; MacHhrns

Tlx,rse irrguulellts ale circurnstantial, and Thring's thesis ctr.nnot trtl provrxl by therrr. Irr spite of its plausibility, Thring's thesis is still itn alJsrrmptiorr. But viewitrg Turing's tllesis simply a"san arbitrary defirritigrr rlis$gs an impottant poinL. In solne sense) Tbring's thesis pla,ys tlrt) rltrrrrt: role in cxrrnputer science as do the basis laws of physics and chemistry. Classical physics, for example, is based largelv on Newton'$ Ia'wsof r'cttitttr' Although we call them laws, they do not htr,vekrgical n(xx)ltitity; rather, tltey are plausible models tha,t explain mur:h of tlrt: physical world. We accept them because the conclusions we dra,w fiom thcrn irgrtlc with our experietrce ancl our ohserva,tions. Slch litws t:arrrrot be proved to be true, although they can possihly bc invalidatecl. If an exPerimellta,l result contradicts a conclusion based qn thc laws, we might begin to question their validity. On the other hancl, repea,terl fir,ihrrc to invalidate a Iaw stretrgthens our confideuce in it. This is the sitrra,tirlrrfclr TUringts thesis, so we have some reason for c:onsiclcrirrgit a basic law of computer scieuce, The conclusions we draw f'rom it agrt:e with wltat we ktrow alrout real computers' and so far' all irtttrrrrpts to invalidate it have failed. There is always the possibility that sorrreorrewill come up with a,nother definition that will accotrnt fttr somtl subtle situations not covered by T\rrirrg rnitchirrcs lnt wlfch still fall within the range of our intuitivc notiorr of rrrechatrical cotrtputatiotr. Itr such atr eventu6lity, sorne Of t)trr $trllsctlttent discussions would have to be modifled signifir:ir,ntly. Howc)vcr, thc likelihood of this happenilg seems to be very srnnll. Hirvirrg accepted T\rritrg's thesis, we are in a positiou to give a precise rlcfirritirlrr of art algorithttt.

lM" Arr algorithm for a function f : D --+ ft is a Ttrring machine M, which given as input any d â‚Ź D on its tape, eventually halts with the correct answer / (d) e .R on its tape. Specifically,we can require thtr,t

qodl nr ,UI (,1),q1 e F, frlr all rJ â‚Ź D.

Iclentifying an algorithm with a, Ttring ma,chine progrirm tilIows us ttr "there exists a,n algorithm ." or "thcr(l prove rigOrously sUch claims as ." However, t0 trurrstrtt<:ttlxlllit:itly irrr algoritlrrrr for is rro algorithm. even relativelV simple problems is el vcrv lcrrgthy urrdertakirtg, To avoid Sur:hunpletr,str,rrtprosprx:t$, wr: t:ilrr irllltc:itl to T\rrirrgts tltesis atrd clain that ir,rrythirrgw() (titrr tl6 orr arrv txrrnputer carr also be dotre otr a Tirring tnachine,


9,3 Turumc'sTunsrs

Conseqrrtlntly, we could substitute ttPascal program" for ,,Ttring machinet' in Defirdti'rr 9.5. This would easethe burderr of cxhibiting a,lgorithmsconsiderably. Actrrtr,lly, a,swe have already done, we will go t)ne step further and accept vurtrirl descriptions or block diagrams as algorithms on the a"ssumptiotr that we could write a T[rring machine prograrrr frrr them if we were challenged to do so. This greatly simplifies the discussion, but it obviously Ieaves us open to criticisrn. While "Pir,st;rr.l program" is well defined, ,,clear verbal description" is not, and we are irr rlrrnger of claiming the existence of nonexistent algorithms. But this da,nger is more than offset by the fact that we can keep the discussion simple arrd intuitively clear, and that we can give cclnclisedescriptions for sorne rather c{lmplex processes. The reader who hirs any doubt of the validity of these clirims can dispel them by writing a suihrble program in some prograurrrrirrg l;rnguage.

**

1' Conrrider the set of trar:hine langua,ge instructions for a corrrputer of your choice. Skctch how the various instnrr:tions in this set could be carried out by a Ttrring machine. 2. In the above discussion, we stated at one point that Tlrring machines appear to he more powerful than pushdown autorrrata. Since the tape of a Turing rnachine can always be made to bchave like a stack, it woukl seem that we can actually clairn that a Tlrring machine is more powetful. What irnlrortant factor is not takerr into account in this argurnent'l W

**

3. There are a number of enjoyable articles on Turirrg machines in the popular Iiterature. A good orre is a paper in Sci,entifi,cAmerican, May 1984, by J. E, Hopcroft, titled t(T\rring Machines." This paper talks about the irleas we have introduced here and also gives some of the historical context in which the work of Turing and others was done. Get a copy of this article and read it, then write a brief review of it.



O t h e rM o d e l so f T u r i n gM o chi n e s

ur definition of a standard Tirring rnachine is not the only possible one; there are alternative definitions that could serve equally well. The conclusions we can draw about the power of a lfuring machine are largJely independent of the specific structure chosen for it. In this chapl,er we look at several variations, showing that the standarcl T\_rring machine is equivalent, in a sensewe will define, to other, more complicated models. If wc a<:r:erpt Tirring's thcsis, we exper:t that cromplicating the standard Thring rnachine try giving it a morc txrrnplex storage device will not have any cffrtct on the power of the automaton. Any computation that can be perforrrred orr surlh a n{}w arrarrgement vrill still fhll under the category of a rnechanical corrrputation arrd, therefore, can be dorre by a sta,ndard model. It is nevertheless instructive to study rrrore complex models, if for no other teasoll than that arr explicit derrrorrstratiorr of the expected result will demonstrate the power of the T\rring machine and thereby increase our confidence in Ttrring's thesis. Many variations on the basic model of Derfinition 9.1 are possible. For example?we can consider T\rring machines with more than one tape or with tapes that extend in several dimensions.

249


250

Chopter l0

O'rsun Molnls or Turut'rc MecHtlrns

We will c:ortsidervariants that will be useful in srtbscqucrrt discussious. we also look at nondeterministir: T\rrirrg rnachiiles and show that they are no more powerf|rl tharr dclterrlilistic ones. This is unexpected, since T\rring's thesis covrlruonly rnechanical computatious and does not tlddress the clever guessing implicit in nondctt:rrnirrisrrt. Atrother issue that is not irnmediately resolved by Ttrring's thesis is tha,t of one uachine executing "reprodifferent programs a,t different tirnt:u. This leads to the idea of a "universal" T\tring rnilt:hirre. grarnmable" or Finally, in prepa,ration firr lrrtcr cltapters' we look at linear bounded automata,. These ar,reT\rring rnacltirres tltat have atr infinite tape, but that can rnake use of the tape only in a restricted way.

ffiffiffi Minor Voriotionsorl the Turing Theme Mochine We first consider some relatively mirxlr {:}rilnge:sirr Defitrition 9.1 and investigate whether these changes make a,ny dilltrrttntr: irr the general cotrcept. Whenever we chlnge a dtrfinition, we irrtroduce a trew type of autotnata and raise the qgestion wlrt:thcr thcse new automata are in any real sensedifferent frorn those we have alreacly encounterecl. Wlrat do we mean bY ir.ntlsst:ntiir.l clifferencebetweetr one cla,ssof automa,tir,tr,ndalntlthcr'l Although there rnay be clear differences in their clefinitions, these difference$ maY rrot have atry interestiug consequences. We have seen tln exampltl Of this in tlte case of cleterministic and nondeterministic Iinite arrtornata. These have quite different definitions, but they a,re elqrrivakrrrtirr the serrsethat they both are iclentifled exactly with the fa,mily of rcgular larrguages. Extrapolating from this, we carr defirre equivaletrce or nonecluivalence for cla-ssesof automata in gerleral.

of Aulomoto Equivolence of Closses Whenever we define equivalence for two automata or classes Of atrtornata, we must carefully state what is to be nnderstood by this equivalcrrcc. For the rest of this chapter, we follovr the precedence established for nfa's ald dfa's and define equivalence with respect to the ability to at)rxrpt larrguages.

ilM Two automata are equivalent if they accept the same language. Corrsider two cla,ssesof automata C1 and dr. If f'trr everv tttrtomatort M1 in C1 there


10.1 Mn'ron. VARTATTONS oN .r.Hr,tTuRrrqc Mec;Hluo THnlrn

25L

is an autorrraton M2 in C2 srrch that

L (Mt) : L (Artz), we say that Cz is at lur^st as powerful as C1. If the converse also holds anrl fbr every M2 in C2 tlrcre is arr M in (J1 such that -L (Mr) : L (Mr), wc siry that C1 and Cz are equivalerrt.

There are ma,ny ways to establish thc equivalence of automata. The r:onstruction of Theorem 2.2 does this for clfir,'salndnfa's. For demonstrating ttrc equivalence in connection with T[rrirrgts machines, we often use the inrportant ter:hnique of simulation. Let M btl an rrrrtomir,ton. We sav that another autorrraton -ffi r:a,nsimulate a corrrputrr.tionof M, if ffi .un mimic the computatiorr of M irr the following rrrarutlr. Lt+t ico,d1, ... he the sequence of instantarreous dr:scriptions of the comprrtatiorr of M, tha,t is d,0 | Aadl

Therr M siurula,testhis computation if it carries out a r:orrrllutir,tion analogous to that of M,

wlxrre d0,d1,... &râ‚Ź instantaneous descriptiorrs, srr{;}rt}rir,t err,chqf them is associa,tedwith a unique configuratiorr of M. Irr otht:r wrlrd$, if we know the cotnputatiorr r;a,rriedorrt hy M, we ca,n determine from it exactly wlnt computations ,\,4 woukl havr: done, given the corresponding starting configura,tion. Note that t,he simulatiorr of ir singk: move rJi ts*d,,+, of M may involve scvt:ral move$or ff. the intermecliateconfigurations in 4 io 4*r rnay not corresporrd to arry r:onfigurrrtiong! M, but tltis does not affect anything if we can tell whitlh configuraltions of M a,rereleva,nt. As long as we can deterrnine from the cornputatiorr of ,fr what M w.rrlrl have clo-ne,the simulation is proper. If M can sirnulate erely cotrprrtation of M, we say that rt ,o, sirnrrlate M. rt should be clear that if M can sirnulate M. thcrr uratters catr bc arrarrgrld so tha.t M ancl fr a,ccept the same language, ancl the two automata arc rx|rivirlent. To demonstrate the equivalenceof two classesof autotnata, we show that frlr t:vury mar:hine in one cla,ss,there is a machine iu the second class capirtllc:of sirnrilatirrg it.

TuringMqchines with q Stoy-Option Itr our clcfirrition of er,sta,ndard rhring machine, the read-write heacl rnust tnove either to thc riglrt or to tlrc lefl. Sometimesit is convenient to provide


262

Chopier l0

O'r'ntn Montls oF TURINGMacntnns

a third option, to have the read-write hcad stay itr place after rewriting the cell content. Thus, we ca,n define a Tiuing rrtachitre with stay-option by replacing ci in De{initittn 9.1 by

d:Qxf-8xfx{tr,,R,5} with the interpretation that 5 signilies rlo IIIoveIIIeItt of the read-write head. This option does not extcrxl the power of the automaton. The class of Ttrring rnachines with stay-option is eqqivtrlent to the class of standard I\rring rnachines. Prooft Since a Ttrring machine with stay-option is <:ltlarly iln extension of the standard model, it is obvious that arry starrtlard T\rring machine can be simulated hy one with a stay-ttptiorr. be a TlrringtnaTo show thtt txtrrverse,IetM : (8,X,f,d,{o,n,F) chine with stay-option to be sirmrlated by a statrdard Ttrring machine M : ^\ //^ f , f , d, fio,tr, F ). For each rntlvc of M, t}te simulating machine M does {O, ' \ / the following. If the move of M does not irrvolve the stay-option, the simllating machine perfbrms one nl()vc, esscntially identical to the move to be simulated. If S is involvtxl in tlter rnove of M, lhen M will make two moves: the first rewrites the syrnbol artd rttoves the read-write heacl right; the secorrtl rrrovesthe read-write head left, Ieaving the tape t:ontents urtaltc:rcd. The sinulating machine can be constructed by M hv defining li as fcrllows: For ea,chtransition d ( q r ,a ) : ( qj , h , L o r R ) ,

we put into f

E(A;,,) = (Qj,b,tr or rt) . Frrr cach S-transitiotr 6 ( q r , u ): ( q i , b ,S ) , tve put into E the corresponding transitions

E ( a ; , o: ) ( i j , , b , R ) , a,nd F,^

0 ((.1.,, , cJ :

\Qi , c, r,) ,

fbrallcel. It is reastlnallly otrvious that every cornputaLion of M llas a corresponding computation of M, so that M can simr.rlateM, t


10,1 MrNoR V.tRIarIoNs oN THt TLIRINGMncHIun THntrae

25S

Figure 10.1

b t

Track1 Track2 Tiack 3

Simulation is a standard techniqut: for showing the equivalence of automata, artd the forma,lismwt: have described makes it possible,as shown in thc above theorem, to talk about the proces$precisely and prove thetlrtlrns about equivalenr:c.In our suhsequerntdiscussiolr,we use tht: rrotiotr of sitnulatiol freqrrently,but we generally mahe tto attempt to dcscribe everything in a rigortltrs and detailed wtr,y.Complete simulatiorrs with'I\ring ma,chirttls are often crrrlllersonte, To avttid this, we keep ortr disc:rrssioudescriptive, rather thtr,nirr theorem-proof forrn. The simulations artt givetr only iu brotld outlirre, bgt it should not he trirld to see how they r;a.rrbe made rigorous. The reader will find it instnrctive to sketch each simulatiotr in some highcr level la,rrguageor in pserrdtltrrtle, Befrtrcl irrtroducing other rntldels, we make one rcrrtark otr the sttr,nda,rd Ttrrirrg rnachine. It is implicit in Definition 9.1 that each ta,pe svrlbol can be a compositc of characters rtltlrt:r than just a singlt: tlrre. This can be macle morc explicit by drnwirrg att expanded versiort of Figr.rre9.1 (Figure 10.1), in whic:h tlte tape symbols are triplets fiom sorrtesimpler alphabet. In thc picture, we hervt:divided each cell of the tape into threN:parts, calltxl tracks, each crontainitrgotte memhtlr of the triplet. Ba.sedon this visualization, such an arrtomaton is sometirrrescalled a T[rring tttachine with multiple tracks, hrrt such a view in no way extends Definitiott 9.1, sitrce all we need to do is rnake I au a'lphir'bt:tirr which ea,chuynrbol is composed of several parts. 'fhring machitre urodels involvc a change of defirrition, so Howcvcr, otlter ther txluivaletrce with thc stattdard machint: ltas to be demonstrated. I{ere wc look at two srrctr trtodels. which are sometime$ us(xl as the standartl defilition. Some variarrts tha,t a,reklss cottttnotr are explclred in the exerc:iscs at the end of this section.

Tope with Semi'lnfinite TuringMqchines Many authors do not consider thtr rnodel in Figlre 9.1 as standa,rd,hrrt use o11ewith a ta,pc that is unbortndtxl only in one dircctiotr, We can visualize this a,sa tapc that has a left bourtdary (Figure 10.2). t'his Tlrring nrachine is otherwist: identical to otrr statrdarcl nrodel, cxcept that no left move is permitttxl when the retrrl-write head is at thc boundary. It is ttot difficult to see that this restriction does not affect the power of tlrtl tnachine, To sirrtulale a sta,ndir.rtlThring machirrc M by a machirre fr with a semi-infinitr: tape, we ufJetht: arrangenrent showtr in Figrrre 10.3.


254

Chopter l0 Otnnn Monnls or TuHrNc Macnrruns

Figure 10.?

Figure 10.3

-

Treck 1 for risht nart of standard tepe" Tiack 2 for left part of stefldfid tape

The simulating rnrr,chineM has a tape with two tracks. On the upper onli, we keep the irrfrlrmation to the right of some refercrrr:e point on M's ttrpe. The refererrccpoint could be, for example, the position of the readwrite head at the start of the computation. The lower trirc:k conta,ins the left part of M's tape in r{}verseorder. M is programrned so that it will use information on the upper track only a,slong as M's read-write head is to the right of the refererrr:epoint, and work on the lower track as M moves into thc left part of its tape. The distinction carr be made by partitioning the state set of M into two parts, say Qu and Q;: the flrst to be usr,'dwhen workirrg otr tho upper track, tlrc second to be usud on the lower one. special end markers ff a,re put on thc left boundary of the tape to facilitrrte switching frorn one trrlck to the othcr. For example, a$$rrmethat the rntr.chineto be simulated arrd the simulatirrg machine are irr the respective corrfigrrrations shown in Figure 10.4 and that the move lo be simrrla,tedis gerrcratedbv d(qr,a):(qj,c,L). The simulatirrg machine will first rnove via tlic transition

E ( A , @ , r: , )( )f r ,( c , b ) , L ) , wlrerefi â‚Ź Qu. Bcca,use fi belongsto Qu, only irrfrrrmationin the rrpper trat:k is consideredat this point. Now, the simulatirrgmachinesees(ff, ff) Figure 10.4 (a) Mat:hine to bc simulaterJ. (b) Simulating machinc.

Ref'ercncepoint

/'' h

a

(a)

)q, {


10,1 Mrr.lon V.qn.Inuolts oN THtl TURIr'rc MnclHttrs Tsplrp

Figure 10.5 Setltretceol conligurations in simulating d ( s r ,a ) :

255

)0t

,l

( q i ,c , L ) .

in state ti e Qu. It rrcxt uses& transitiorr

t ( t , , ( # , # ):) ( F i , ( # , # ) , R , ) with fr â‚Ź Q1, putting it into the conligurationshownirr Figure 10.5.Now the rnachineis in a state from Q; tr,ndwill work on the lower track. F\rrtlter details of the simulatiori are straightforward'

TheOff-LineTuringMqchine The genera,I tlefinition of an arrtomaton in Chir,pter I containcd an input file as wcll as temporary storage. In Definition 9.1 we disr:arded the input file fbr reasons of simplicity, claiming that this tnade no difference to the T\rring machinc concept. We now expand on this claim. If wc put the iuput fitc back into the picture, we get what is known as an off-line Thring machine. Itt such a machine, each move is governed by the intt:rilal state, what is currently re:ad from the inprrt file, and what is seen try the read-writc head. A schematic representation of a'n off-line machirrg is shown in Figrrre 10.6. A forrrral definition of an off-line Ttuing machirre is easily made, but we will ltrave this as an exercise. What we

Figure 10.6

Read-onlv input file


256

Chopter l0

Ornnn MoDFlr,s or..TuRnrcMACTTTNES

Figure 10.7

want to do briefly is to indicate why the cla-ssof off-line Thrirrg machines is equivalent to the cla^ssof standard machirrert. First, the behavior of any standard T\rring machine can bc simulated by some off-line model. All that needs to be rlone by the simulating machine is to copy the input from thc input file to the tape. Then it carr proceed in the same v/ay as the standarrl machine. The simulation of an off-line rnachine M by a starrclarcl machine fr requircs a lengthier description. A standard machine can sirnulate the computation of an off-line machirre try using the four-track arrangement slnwn in Figure 10.7. In that picture, the tape contents shown represent the specific configuration of Figure 10.6. Each of the frrrrr tracks of M plays a specific role in the simulatiorr. The first track has the input, the second marks the position at which tho input is read, tlNr third representsthe ta,pe of M, and the fourth shows the position of M,s rc:rrd-writehead. The simulation of ea,chmove of M requires a number of moves of _ffi. Starting from some starrderrdposition, say tha le.ft end, and with the relevant informatiorr rnarked by special end markers, M scarches track 2 to lor:a,te the position at which the input file of M is read. The syrnhol found in the cgrresponding cell on track 1 is remernbered by puttirrg the control unit of M i'to a state chosen for this purposo. Next, track 4 is searched for the position of the read-writc head of M. With the remembered input and the symbol on track !. we rrow know that M is to do. This irforrnation is again remembered by M with arr tlppropriate internal state. Next, all four tracks of M's tgqe are modified to reflect the move of M. Firrally, the read-write head of M returns to the standard position for the simulation of the next move.

1. Give a formal definition of a Ttrring rrrachine with a semi-infinite tape, Then


10,1 MINoR VlRhrIoNs

oN THE TuRINc M.q'cHII-InTHurae

257

prove that the class of T\rring machines with semi-infinite tape is equivalent to the class of standard Turing machines. Give a formal definition of arr off-line Ttrring machine.

s. Give

convincing argunrerrts that any language arcepted by an off-line T\rring machine is d,lso accepted by some standard machine,

4 . Consider a Ttrring machine that, on any particular mover cart either change the tape symbol or move the read-write head, but not both. (a) Give a formal definition of such a machine' (b) Show that the cla^ssof such machines is equivalent to the clags of standartl T\rring machines. ffi i}.

Consider a model of a Ttrring machine in which earh move permits the readwrite hearl to travel more than one cell to the left or right, the distance and direction of travel being one of the arguments of d. Give a precise definition of such an automaton and sketch a simrrlation of it by a standa,rd T\rring machine.

6 . A nonerasing Tirring machine irr orre that cannot change a nonblank symbol to a blank. This can be achieved by the restriction that if 6 (qr,a) = (qj, E, L or R) , then a must be !.

Show that no generality is lost by making such a restriction.

ffi ,.

Consider a Ttrring machine that cannot write blarrks; that is, for all d (gn,o) = (qi,b,L or R), b must be in f - {tr}. Show how such a machine can simulate a standard Thring machine. Suppose we make the requirement that a Ttrring marhine can halt only in a final state, that is, we ask that d (q, a) be defined for all pairs (g' a) with a â‚Ź | and q t' F. Does this restrict the power of the Turing nrachine?

9 . Suppose we make the restriction that a Tlrring machine must always write a symbol different from the one it reads, that is, if d (Sr,a) : (qi ,b, tr or R) , then a and b must be different. Does this limitation automaton? ffi

reduce the power of the

1 0 . Consider a version of the standa,rd Ttrring machine in which transitions can depend not only on the cell directly under the read-write head, but also on the cells to the immediate right and left. Make a formal definition of such a machine, then sketch its simulation by a standard T\rring machine.

1 t . Consider a T\rring machine with a different decision process in which transitions a,re made if the current tape svmbol is not one of a specified set. For examule 6 (qt,{a, b}) : (fi, c,,R)


258

Chopter | 0

Orrrnn Mor)ELS or,' TuRrr.rc Mecruruns

will allow the indicated movc if the currerrt tape symbol is neither a nor [r. Forrrralize this conccpt and show that thisr modification is equivalent to a standard Turing rnachine. ffi

M

T u r i n gM o c h i n e sw i t h M o r e C o m p l e xS t o r o g e

The storagc device of a standard Thrirrg mrrchine is so simple that one might think it possible to gain power by using more complicated storage clevices. But this is not the ca$(]rL$we now illustrate with two examples,

MultitopeTuringMqchines A multitape Thring machine is a T\rring rrrar:hine with several tapes, each with its own irrrlependently controlled read-write head (Figure 10.8). The formal delinition of a multitape T\rring machine goes beyond Definition 9.1, since it reqrrires a modified transition function. Typically, we define an rz-taperrrachineby M : (Q,E,I, d, g0,n, F), where Q, E,l, gs,F a,reas in Definition 0.1, brrt where d:QxI-"-Qxlu

x IL,RI"

specifies what lnppens on all the tapes. For example, If n : r:urrent configuration $hown in Figure 10.8, then

2, with a

d ( q n ,* , e ) : ( q r , r , y , L , R ) is interpreted as follows. The transition rule carr be applied only if the rnachirrc is in state Qsand the first read-write head seesarr a and the second a,n e. l'he synrt-rol on the first tape will then be replaced with an r and its read-write head will move to the left. At the same time. the svmbol on

Figure 10.8

Tapc 2


10,2 Tunrt'tc; M.q.cHIunswIrH MoRl; Corr,tplnx Sronacn

259

Figure 10.9 o.

tr

b

v

t

f

Tape 2

Tapc I

the sccond tape is rewrittt+n as'g and the read-writc head moves right' The control unit then chtr,ngesits state to 91 and the machine goes irrto the new conliguration shown in Figure 10.9. To show the equivalence between mrrltitape and sta,ndartl Ttring machines, we arguc tltat any given multitelpe Thring machintl M cau be simrtlated hy a statrdard Ttrring trachine M and, converscly, that any standartl T\uing rnachine can he simulated by a multitap() orle. The second part of this clirirrr treedsno elabclrtrtion, since we can alwa.yselect to run a,multitape 'Ihe simula,tiorr of a machirrc with only one of its tapes doing useful work' mriltitrrpe machine by one with a single tape is a little more complicated, but corrceptually stra,iglrtftrrward. Cotrsider, for example, the two-tape macldlle in the conliguratiotr depicted in Figlre 10.10. The simulatirrg single-tape machirre will have f'ogr

Figure 10.10 7

a

lr

t

)

E

"f

h


260

Chopier l0

Otrnn Moonls oF TLrRrNcMncnrr-rps

Figure 10.11

trat:ks (Figr.rre10.11). The fir:st track rcpresents the corrtcrrtsof ta,pe 1 of M. The nonblank part of the second track hrrs a,ll zeros, exctlrt firr a,single 1 tnarkirrg the position of M's read-write iread. Tfa,cks3 and 4 play a similar role for tirptr 2;f M. Figure 10.11 makes it clcirr that, for the relevirnt txrnfigurations of M (that is, the ones that have the indic:atedform), there rs a,unique corrtli;ponding configrrrirtion of M. The represerrtir,tionof a multitrrllc rnachine by a sirrgkl-ta,pernachine is similar to that ust:d in the sitttulation of a,n off-line rnachine. The actual steps in the simulatiorr irre a,lsomuch the same, the only differencebeing that thcrtl ilre more tapes to rxlnsicler.'fhe outlinn given for the simulation of offline tnachinescat:riesover ttl this casewith rrrirrrlrmodifications and suggests a, proceclure try wirich the transition function f "f fr can be constmr:tecJ fiom the transitirtn ftnction 6 of M. While it is rrot difficult to nrake the corrstnrction precise, it takes a lot of writing. Certainly, thu computations of M given the appeara,rrr:e of being lengthy and elaborattl, llrt this has no beallng on the cottclusion. Wratever can trt: done of M can also be done on M, It is irnporta,nt to keep irr rnind the followirrg point. When we c;laim that

l,lH-Jil"'m,il'1il;*"'"IJ"ilili,il#T:,il'Hl pa,rticularly,what la,nguages rnachines, can be a,ccepted, txqm,p[* ilCI,l'

Considertlte ltrngua,ge a laborious {a"'b"}. In Exampleg.7, we clescribecl

;1Tli:iJ'il1'#'JH.lHf, :1fl:illi,l";,H'-:;':i1?l,.,lJJi'

thir,t a,n initial strirrg e,t'bttt is writtun on tape I at tlx,, beginning of tlrc cornlltta,tiou, We tlrcn read all t,lte nts, copying Lherrrrlrrto tape 2. Wherr we rcat:h the end of tht: a,'s,we match tlrt: b's on tape 1 against the copied a's ou Lapc 2. This way, wc c:ir.ndeternine wlxlther there are arr cqrral number of a's and b's without rcpeated back-anrl-forth movement of the read-write head.

t


10.2 Turr,Ilrc MecHtut:s wIrH MoRn Colvtplnx SroRlGu

261

Remember that the various models of Thring ma,t:hinesare considered equivalerrt only with respect to their ability to do things, not witlt respect to eilsc of progra,mmirrg or auy othtlr efficiency mcasure we rnight consider. Wc will return to this importtrrrt point in Chnpter 14.

TuringMochines Multidimensionol A multidirrtensiona,l T\rrittg machine is otre in which the tape ca,rrbe viewed as extcrrditrg infiniterly itr tnore tharr one dimensiotr, A ditlgrart1 of a, twtr dimensiotnl Turing rnachine is shown in Figure 10'12. T[e formal rJe:firritionof a, two-dimensional Ttrring machitre involvt:s a transition function tI of the fbrm

, ): Q x f - I x I ' x { . L R , ,U,D}, where [/ anrl D specify rnuvement of t]rc read-writs ]read up tr,nd down, rcspectively. To simqlate this machitte on a stanrlrrrd't\rring machine, we (iaII use the two-tra,ck moclel depictcd in Figure 10.13. First, wc associa,teorr ordering or acldrnsswith the r:clls of the twtrditneusiona.l tape. This t:atr Lredone irr a lumher of ways, fbr example, in the two-dimt:rrsioual ftr,shion indica,ttxl in Figurc 10.12. Thc two-tra,ck tape of the silrulating rnirchiue will use clne track to stortl cell conttlnts and the other one to kcep the a^rsociated a,dclress. In thc scheme of Figure 10.12, the configrrration in whictr cell (1,2) contains a, ancl cell (10, -3) contains b is shown iil F'igure 10'13' Note orrclcornplica,tiorr:the cell a,tlrlresscan invcllve arbitrnrily large intcgers, so the address track caunot trse a fixed-sizc field to sttlre addresstls. Instead, we must rrstt a variable field-size a,rrrlrrgetnettt,using some spt:c:ialsymbols to delimit the fields, tr,sshowu itt thc picture' Lelt us assumethat, at tlre sta,rt of the simulation of eat:hrrrove,the rcadwrite heaci of thrl two-clirnensional ma,chintl M and ttre read-writcl head of the simrrlatirrg ma,chinc -ffi ut* u]{tw. oll corre$poncling cells. To simrrlate a move. the simulating tlachine M first comprrtgs the addressof the tnll to

Figure 10.12

Two-dimensiond addressscheme


262

Chopter l0

Outtlrr MoDELSor, 'I'uRrr*rcMecrrrr{r:s

Figure 10.It

which M is to rnove. using the two-dirr*,'nuional addrcss scheme, trriu is a sitnple computation. once tlrc trddressis computed, M Iinds the cell with this a.ddressorr tra.ck 2 and then changes the cell corrtents to account for tll rnove of M. Aga,in, given M, there is a stra,ightforwirrd construction fbr 44.

The purpose of much of ortr discussiorr of Ttrring rna,chines is to lend cre'rhrirrg's dence to thesis by showing how seemingly rrrrlre complex situations cirn be sirnulated on a starrda,rd T\rrirrg machine. unfbrtunately, detailed sirnuler.tions are v()ry tedious irnd conceptually uninterr:sting. In the exercises trelow, describe the sirnrrlrrtions in just enough depth to show that the details can be worked out. 1. A multihea,<lTrrring machirre can be visualized as a T\rring machirre with a single ta,pe and a single control 'nit but with multiple, irrdependent rearJwrite heads, Give a formal definition of a rrlrltihead T\rring machine, and the[ show how suc:h a machine t:an bc simulated with a standard Ttrring machine. ffi 2. Give a formal dcfinitiorr of a multihearl-multitape Turing rnachine. 'Ihen sh6w how such a r'achine ca' be simulated by a standartl r[rring rnachine. 3' Give a forrnal definition of a T\rring rnachine with a si'gle tape but multiple control units, cach with a sing;le read-write heatl. show how srrch a machine can bc sirrrulated wit,h a multitape rrrar:hine.

;,tr:

* 4. A queue automaton is an a,utomatorr in which the temporary storage is -, a clueue. Assume that sur:h a machifle is an on-li[e machine, that is, it has rro input file, with the strirrg to be pror:essedplaced in thc queue prior to the start of the cornputation. Give a formal definition of such an autorrraton. then investigate its power in relation to Ttrring machines. ffi * 5' Sho* that for every il\rring tnat:hine there exists an equivalent standartl Ttrring uachirre with no rnore than six states, 'lot"R*.iu.,.

the number of rcquirecl states in Exercise b as far as you can (Hint: the smallest possible nurnber is ihree).

* 7' A c.rrnter is tr stack with an alpha,bet of exactly two symbols, a stack start symbol and a courrter symbol. Only the counter symbol t:an be put on the stack or removed frorrr it. A counter automaton is a deterrrrinistic autorrraton with orre or more counters as storage. Show that any T[rring machine carr be simulated using a counter autornaton with four counters.


I O.3

NONDETER,MINISTIC IIURIIIIC MNCHII,.IES

263

g. show that everv cornputatiorr that can be done by a standard T\rring machine can be done by a multitape machine with a stay-option and at most two states.

ffi 9. Write out a detailecl prograrn for the computation in Example 10.1.

N o n d e t e r m i n i s tTi cu r i n gM o c h i n e s While Thring's thcsis makes it plausible that the specific tape structure is immaterial to the power 9f the T\rring machine, the sa,mecannot be said of nondeterminism. Since nortdeterminism involves an element of choice and so has a nonmechanistic flavor, an appeal tg T[ring's thesis is inappropriate. We must look at the effect of noildeterminisrn itt more dertail if we wa,nt to argue that notrdeterminism adds nothing to the power of a Turing machine' Again we resort to simrrlation, showing that nondetermirristic behavior can be handled deterministit:ally.

A nondetermirtistic Thring rnachine is an lrrtomaton as given by Definition 9.1, except that d is now a.function d:Qxf-+2Qxrxtr'fi}. As a,lwayswhen nondetcrminism is involved, the range of d is a set of possible transitions, atry of wlfch can be chosen by the machine.

*fimrttplClq;il

has transitionsspecilied6y If a Thring ma.c"hirre ,l (qo,a) : { (qr, b, R) , (q2,c, L)} it is nondeterministic. The rloves qouaal bq$,& and qa1znal qzlcaa are both possible.

I


264

Chopter l0

O.ruun Moonls or.''IuRrr-rcMACtrrNEts

since it is not clear what role nontlcterminism plays i'computing function$, nondeterninistic a,utomata are usually viewerl as accepters. A nondetcrministic 'rhrirrg machine is sairl to accept u if there is any possibre sequence of moves such that t111,ut i qq..r2, wiih q7 â‚Ź F. A nondeterrninistic machin{} may have move$ available that Iead to a nonfinal $ta,te or to an infinite loop. But, as always with nondetcrminism, these aJternatives are irrelevant; all wc ir,re interested in is the existerrce of some sequence of moves lcading to acceptance. To show that a nondeternrirristic Ttrring machine is no more powerful than a deterministic one, we need to provide a dcterministic equivalent for tht: nondeterrninisrn. We have already alluded to one. Nondetermirrism can be vicwed as a deterministic backtrrrcking algorithm, and a deterrninistic rrrar:hine can simulate a nondeterministic one as long as it can handle the bookkecping involved irr the backtrackirrg. To see hovr this can be done simply, let us mrnsider an altcrnative view of nondeterminisrn, one which is ustlfirl in rnany arguments: a norrdrlterministic rnachine can be seen a$ one that has the ability to replicate itself whenever neces$ary. when more than orre rnove is possible, the machine prt)duces as marry replicas as needed and givcs eaclt replica the task of carrying out one of the alternatives. This view of rrondeterminisrn rnav seern partit:ularly nonmechanistic. unlirnited replication is certainly not within the llower of present-day computers. Nevertheless,the processcarrbe simulated by a deterministic Thring machirre. Consider a Thring rnachine with a two-dimensional tape (Figrrre 10.14). Each pair of horizontal tracks representsone machirrc;the top track conta,ining the rrrachine'stape, the hottom one for indicatirrg its internal state and the position of thc read-write hczrd. whenever a new machine is to be created, two new tracks are rrtarted with the appropriate information. Figure 10.15 representsthe initial r:onfiguration of the machine in Example 10.2 and its succ:e.$sor configurrr.tions. The simrilating machine searches aJl active tracksl they are bracketed with special rna,rkersand so can always be found. It then carries out the irrdicated moves, activating new machines as needed. Quite a few dctails have to tre resolved befcrrewe can claim to have a,reasonableoutline of the simulation, but we will leave this to the reader.

Figure 10.14

Tape contents Internal state and position ofhead Tipe contents Internal state and position ofhead


10.3 NoNDETERMINIsTIcTunInc Mncutt'lps

265

Figure 10,11'r Sirmrlation of a rrorrtletenninistic fitove.

Based on this simulation, our conclusion is that for every nondeterministic: Thring machine there exists an equivalent deterministic one. Because of its irnportance, we state this formally. 'I\rring machines and the class of nondetermirristic The class of determirristic 'I\rring machincs are equivalent. Proof; Use the: corrstruction suggested atrove to show that anY nondeterministic Ttrring rnachine can be simulated by a deterministic one. r

I.

Discuss in detail the simrrlation of a nondeterministic Ttrring machine by a deterministic one. Irrdicate explicitly how new machines are created, Itow artive machines are identified, and how machines that halt are remove<l frorn further consideration.

2 . Show how a two-dimerrsional nondeterministic T\rring machine can be sirrruIated bv a deterrnirristic machine. 3. Writc a progranr for a trondeterministic Ianguage

T\rrirg

machine that a,ccepts the

7={ww:ue{a,b}+}. Contrast this with a deterministic solution. W 4. Outline how one would write a program for a nondeterministic chine to arcept the langrrage

-

L:

(

"

'a.bl+I.

. \ r U ' u t - " u : t r . t â‚Ź{ , .

5, Write a simple pl'ogram for a nondeterministic the language 7 :

{***"a

Tltring ma-

)

Ttrring machine that accepts

i 'r1a1ue {a, b}+ , l"l } lgl}

How would you solve this problenr deterministically?


266

Chopter l0

OrHnn Monnls oF TrrHrNC.t MAcHTNES

6 . Desigtr a rrorrdeterministic Turing rnat:hirrethat accepts the language L:

{a" : rr is rrot a prime uumber} . ffi

* 7 . A two-stack autornaton is a nondeterministic puslxlown arrtomaton with two indepcndent stacks. T<r clefine such an autornaton, we modify Dcfinition 7.1 so that d : Q x (tU{A})

x l' x f *

f i r r i t es u h s e t so f Q x l - x f . .

A movc dcpends orr the tops of the two stacks arrtl results in new values being pushed on these two stacks. Slrow that the class of two-star:k automata is equivalent to the class of T.uring rrrachines. NN

A UniversolTuringMochine ffiffiffifif;ffi Cortsider the fbllowing argurncrrt a,grr,instThring's thesis: "A Ttrring machine as preuented in Definition 9.1 is a special purpose cxrmprrter. Once d is defitred, tht: ma,chine is restricted to ca,rrying out one partir:rrlrrr type of computation. Digital computers, on the other hand, arc guneral purpose machines that can be prograrnmed to do rliff'erent jobs at differerrt tirnes. Consequently, Thring machines cannot be considered equivalent to gerrcral purposc digital computers." This ohjection can be ovr:rrxlme by designirrg a rcprogrammahle Ttrring machirrc, called a,universal T\ring machine. A rrnivt'rsill T\rring rnachine M, is arr autclma,ton thal, given ils input the descriptiorr of any T\rring machine Jl[ ir,nd a, string ?u, can simrrLlte the computatiorr of M on ro. To construct such irn M,r, we first choose a standa,rd way of descritring T\rring machines. We rnay, without loss of gerrerality, assume that Q = | q t , Q z , . . . , Q ,r r I with q1 thc irritial state, q2 the sirrgleIinal state, and f : { a r , & 2 , . . . , & r n, } where a1 rcpresents the blank. Wc then select an encodirrg in which q1 is representedlry 1, qz is representedby 11, and so on. Similarly, a1 is encoded as 1, a2 as 11, etc. The symbol 0 will be rrsedas a separator btrtweenthe J.'s. With the initial rrrrrl final state and the bltr.nk defined by this corrvcntion, ir,nyTtring rnachirre r:an be described cornpkrtcly with d only, The transition function is encoded actxrrding to this scherne,vrith the arguments and result itt sornc prescribed sequencc. For exa,mple, d(gt, uz) : (qz,a3,"L) might appear a,s

. . . 1 0 1 1 0 1 1 0 11 01 1 0 . . .


10.4 A Lilrvllnsnl

TuRrNc M.c(:Hr'lr:

267

Figure 10.16 Control unit of M,

Internal

state of ,44

Tape contents of ll4

It folklws frotr this thirt arry Thrirrg rrrachirrehas a fitrite ettcoditrg as a string trrr {0, 1}+, rrrrclthat, given any encodingof M, we carl clecodeit uniquely. Some strings will not represent any Turing machine (e.g., the strong 00011), but we can easily spot these, so they a,reof no concern. A universal T\rring rnilr:hirrc M,, thrlrr has itn irrprtt alplnbet tlrrrt includes {0, 1 } and the stnx:trrrc of ir.rrlrltititllt: milt:}rirrt:,its sltowtt in l'igure l 0.16. For atry input M and u, ta,pe I will keep a,n encxrdeddofirritiorr of M. Tape 2 will contain the tnpe cxrntentsof M, a,nd tapc 3 thc irrtertral state oI Il[. M* looks first ir.t t]rc cxlrrtcrrts of ta,pcs 2 arrd 3 to deternine the configurtrtion of M. It thc:rrcxrrrsultstape 1 to seewhat M would do in this c:onliguratiorr. Firrally, tapes 2 and 3 will be rnodified to reflect the resrilt of the rnove. It is within rea$en to txlnstnrr:t an actuar,lurriversal Turittg tnaciritre (see, for exiltrplt:, Dt:nrrirrg,I)enrris, and Qualitz 1978), but the process is 'I\uing's rrninterersting. Wrr prcfer irrstead to appeal to hypothesis. The irnplerrxlrrtrltirlrr tlc:irrly catr be done usitrg solne programming la,ngua,ge;in far:t, thtl l)r()grilrrrsuggesteditr Exercise 1, Section 9.1 is a rea,lizir,tionof ir, utriversal Ttrring machiue in a, higher level la,ttguage.Thereforc, wtl tlxpect that it catr also be done by a stanclard T\rring machine. Wetarc tlrt:rrjustified in claiming the existerrceof a,T\uing trachine that, givcrr any program, call carry out the comprrttr,tiorrssptrilied by that progl'an ancl tha,t is therefbre ir proper urorlt:l firr ir gcneral purpose cotnputer. The observation lhat every T.\rring ma,chinecan het rcprtlserrted by a string of 0's and l's ha,s iruporta,nt irnplications. Brrt bcfbre we explore these implications, we need to review some resrrltufrorn sct theory. Sone sets are finite, lrut most of the interestirrgst:ts (and languages)are infinite, For infinite set$, we distinguisir trtrtwccrrsets that are countable ancl sets that are uncountable. A set is snirl to trc coutrtable if its eletnents with the positive integers. By ca,nbe put into a,one-ttl_oncrxlrrt)tporrrlerrce this we mean that tht: clernerrtsof the set can be written in ${lme ordtlr, sirlr,rr11 stl thirt cvtlry elertretrtof tire sel has some fitrite index. Fclr ,.Tz,,T:1,..., cxilrnplt:,thc set of all even iutegerscau be written in therordcr 0,2,4,....


268

ChopterI0

Ornpn Morrtlt,sol''l'unrlc MAcHTNES

Figure 10.17

Since any positive integer 2n, occurs irr position n { 1, the set is countable. This shorrldnot be too surprising, but there are lrlore corrrplic:irttxlcxilrrrllltis, sorrre of whit:h ma,y seem counterirrtuitive. 'I'ake the set of all quotitlrrts of tlte fortrt ytf q, wlrt:rc p ilrxl q a,repositive integers. How should we order this set to show that it is courrtabk:'l Wc t:ilnnot write 1 1 1 I

I't'3'4'... becaust:thtrrr fr worrld never a,ppear. This does not irnply tlnt thc stlt is uncountable; irr this r:asu,there is a, clever way of ordering the set to show that it is in fact courrtablr:. Lrlok ir,t the scheme depicted in Figure 10.17, and write dowtt the elerrrerrtirr thc rlrder encountereclfollowing the arrows. This eives us

1 1 2 1 2

t ' z 'T ' 3 ' i ' " Here the elerrtcrrt f; or:r:rrrsin the eiglrth place, ancl every elerrrent has sotrc place in the seqrrcrrc:t:. Tht: set is therefore countable. We see from this t:xirrnplc thir"t we can prove that a set is courrtabltl if we can produce a rnethotl try whir:h its elements can be written in sorrre $equence.We call such a rnethod arl enumeration procedure. Since an proce$$1we can use a {lrurrnr:rationprocedure is sorne kind of rrrer:}rirrrit:al Thring trracltirrtl rnork:l to define it formally.

Let 5 be a set of strirrgs orr rromrtnlpha,bet E, Then an enurneratiorrproc:+ 'Ihrirrg dure for 5 is a rrrirc:hirxrthilt r:rr,ncarry out the secluenceof steps

qoni q"rr # ELi Q"'rz# sz ..., witlr 1llr1 e I'* - {#} , ", â‚Ź S', in such a way that any s in ,9 is prodrrt:rxl irr a, finite number of steps. Tlxr stir,tr:q." is rr,sta,tesignifying rnernbershipirr S; tha,t is, whenever q" is entercrl, tlxr strirrg folbwing f rnust be in S.


10.4 A Ur{rvnnsal.TunlNc Macnrr'm

269

Not every set is countable. As we will seein the next chapter, there are sorne uncountable sets. But any set for which an enurneration procedure exists is corrntable bccarrsethe enumeration gives the required sequence. $trir:tly speakirrg, iln (:rnrrnerirtiorr llrttcedur{l crrrrrrot be called att algorithm, since it will not termina,te when S is infinitel. NevertlNrluss,it r:an bc c:onsirlcrcrl it rnerrrrirrgfrrlllrot:cus, bc{:ir,u$eit prochx:cs wcll-rlc:firrcxlarrd prtxlir:tabk: rtlsults.

Let X : {a, b, c}. We can show t}rat the S : X+ is countable if we can find n,nenumeration prclcedurethrr,t prodrrcesits elermrrntsin sornr:rlrrklr, sa,yin the order in which they would appea,rin a dictiorra,ry. However, the order used in dictionaries is not sr.ritablewithout modifica,tion. In tr rlictiona,ry, all worcls beginning with a a,relisted befbre the stling b. But when there ale an infinite number of a worcls, we will never reach b, thus viola,ting the condition of Definition 10.3 that any given string he listtNl a.f'tcrir. finitc number of steps. Instead, we (:an rrse il rxldificd rlrdrlr, in which wtl takc the lerrgth of thc strirrg as thc first criteriorr, followed by an alphabetic ordering of all equal-length strings. This is a,n enumera,tionprocedure tha.t protlut:us thc sequence d, It,r, ua, ab,ac, ba,hb,bc:,ca, cb, cc,aae,r,,,, As we will have severa,luses for sur:h an orclerrirrgwcl will crrll it the proper order. I Arr irrrportant consequence of the above discussion is that T\rrirrg mir.chines a,recorrntable. The set of all T[rring nta,chines,tr,lt]xtug]rirrfirritc, is countable. Proof: We ca,n errcodeeir.chT\rrirrg rna,r:hirrtlusirrg 0 and 1. Witlt this encoding, we thrlrr rxrrrstnr(:tthc following euunreralion procedure. 1. Genera,tethe next string irr {0, 1}* irr properl orcler. 2. Chcck the getrerateclstring to see if it defines a, Tlrrirrg rnirchirre.If so, write it on the ta,pe in the form rr:tpircd by Definition 10.3. If not, ignore the string. 3. R.r:trrrrrto Stcp 1. Since every Tlrring rnilt:hirrc hirs a linite description, any specific milr:hirx: will eventrrir,lly lltl gtlrrt:rateclby this process, I

-


270

Chopterl0

Mncslrur:s OurnR MoDEr,soF TtrRrN(+

The particular ordering of T\rring machines clepends on l,he encoding we use; if we use a diffcrclrrt crrcxlrlirrg,wc rmrtlt oxllrx:t a rlifferent orrklrirrg. This is of rro conserlrrcrr(i(lrhowevcr, ir,rrd slxrws that the ordering itrrelf is utrimportant, What rrrattcrs is thc cxisttlrrr:t:tlf sorntl rlrrlt:rirrg.

1 . Sketch an algorithm that exarnines a strirrg in {t), I} I to tleterrnine w}rether or not it represents an encoded 'I\rring rna,chine,

2 . Give a complete encoding, using the suggestcclmcthod, for thc Ttrring rnachine with

d ( g r ,a r ) : ( S r ,a r , I t ) , , i ( g r, a : ) : ( g r , 4 r , 1 , ) , d (93,a r ) : (S.?, az, .L) . 3. Sketch a Turing rnachinc program that enumerates the set {0,1}t o..Je.. {h

irr proper

4 . What is the irrrlex of ()tlr in Exercise 3? D.

Design a Ttrirtg rnat:hirre that emrmeraterr the following set in proper ordcr, 7:

{anb" : ri } 1}

6 . For Example 10.3, find a function / (tt) that gives for ea,chtl its index irr the propel ordering. a .

L

Show that the set of all triplcts, (i.,j,h) with ,i,.7,k positive irrtegers, is countable. Suppose that Sr and 5'r are coutrtable sets. Slrow that then Sr U Sz and ,9r x ,92 arc also countallle. ffi

9 . Show that the Cartesian uroduct of a finite trurnber of countable sets is rrlrntable.

gpi.$ffiffi1f Lineor BoundedAufomqto While it is rrot possiblc tri qrxtendthe power of the standard 'I\rring machine by complicating the tirpc: stnrt:trrrt:, it is possible to limit it by restricting the way itr which tlrt: ta,1lcr:an br: rrsed. We have already seenan example of this with pushclowtrautornatir. A llrshdown automaton can be regarded as a, nondeterrninislic I'uring rrrachirrewith a, tnpe thar,tis restricted to being used like a stack. We catr also restrict thc tir,pe rrsrr,gein other ways; fbr tlxample, we nright permit only a finite part of tirc tirlrrl to tlc ustxl as work $pa,ce. It can be shown that this leads us bir.r:kto finite arrtomata, (see


10.5 LINITAR Brturunr:tr Au'r'oNl.q,'nc 27L

Ext:rr:isc3 rr,ttho end of this section)r $o we need not pursue this. Rut there is ir wiry of lirnitirrg taptl usr: tha,t lt:ir,rlsto rl rnore interesting situation: we allow the tnachine to use only that part of the tape occupied by tht: input. Thus, more space is available for long input sl,rings Lhan for short ones, generating another class of machines, the linear bounded automata (or lba). A linear bounded automaton, Iike a standard T\rring machine, has an unbounded ta,pe, but how much of the tape can be used is a function of the input. In particular, we restrict the usable part of the tape to exactly the cells taken by the input. To enforce this, we can envision the input as hracketed by two specia,lsymbols, the left-end marker ([) and the rightend marker (]). For ilrr irrput rl, tlrtl irritia,l rxrrrfigurilti<lrrof thc T\rring rnachine is given by the instarrtaneousdescriptiort qn[r]. The errd rrrarkers cannot be rewritten, and the read-write head cannot rnove to the left of I or to the right of ], We sometimes say that the read-write head "bounces" off the end markers.

r,St*.$[iP'i;ninPii1,1lffi A linear lrounded a,utomaton is a nondeterministic Tirring machine M : (8,E,l-,d,Qo,n, F), a,sin Definition 10.2,srrhiectto the rerstric:tion that X rmrst r:orrtairrtwo spct:ial synrbols land ], suclt tlnt d (qr,l) catr contain only elerrrentsof the forrn (qi,l, 1t), and d(qi,]) can contain only elementsof the form (qi,],.L).

i

i

l

A string 'uris a,cceptedby a linea,rbounded autornaton if there is a possible s(xllr()rrcc{lf movr:s

qnlrl i l*rqr*rl fbr some et E F,;t)1,n2 â‚Ź f*. of irll suc;harr:t:pttxl strings.

Tlxl lirrrguirgcacccpted by the lba is the set

Note tha,t in this clefinition a, Iinea,rhorrndcd irutornirtorl is assumed to he nondeterministic. This is not iust il rrrilttt)r of corrvetriencebut essential to tlxl rlisc:ussiorr of lba's. While one can deline deterministic lba's, it is not known whrrtlxlr thc:y ir.rtl ecprivalent to the nondeterministic version. For sorne exDloration of this. see Exercise I at the end of this section.


272

Exanrplle,l,ilL4li

Chopterl0

Ornen Morrus on Tunrlc MAcsltlss

l'he Ianguage [,:{u,nltncn:z>1} is accepted by sorne linear bounded autouraton. This follows from the discussiorrin Example 9.8. The computation ouLlined there does not require spaceoutside the original input, so it can be carried out by a linear bounded automaton, I

6xil,mF,tl,ii$ti$,Find

a linear bounded autornaton that accepts the language L:{ont;rz}0}.

One way to solve the problem is to divide the number of tl's successively by 2,3,4,..., until we can either accept or reject the string. Ifthe input is in -L, eventually there will be a single a lefl; if not, at some point a nonzero remainder will a,rise.We sketch the solution to point out one tacit implication of Dcfinition 10.4. Tht: tapc of a linq+ilrboundtxl automatorr rnay be multitrack, and the extra tracks can be used as scratch space. For this problem, we can use a two-track tape. The first track contains the nurnber of a's lefb drrring the processof division, and the second track conta,insthe current divisor (Figure 10.18). The actual sohrtion is fairly sirrple. IJsirrg the divisor on tlte secortcltrack, we dividc the rrurntrer of c,t$ orr the first track, say l-ry retrovirrg all syrnbols exccJrt those at multiples of tlx: clivisor. After this, wc irrcrement the divisor by one, and continue rmtil we either find a nonzero remainder or are left with a single a. I Thc last twcl cxilrnplt:s suggest thirt lincir,r bounded automata, are mortl powerful than pushdown automata, since neither of the languages are context-fiee. To prove srrch a conjecture, we still have to show that any contextfiee language ctr,n be accepted by a linear bounded automaton. We will do this later in a somewhat roundabout way; a more direct approach is srrggestedin Exercises 5 and 6 a,t the end of this section. It is not so ea"Ey to rnake a conjecture orr tlrtl relatiorr betweerr Thrirtg rrtachirres arrd lirx:ar botrrrrlcrl arrtorrata. Problems like Exarnple 10.5 a,re invariably solvable by a littear borrnded aut{rilratorr, sirrce art atrount of scrirttlh spircc l)rol)ortiorral

I ' i g r r r cl { ) . 1 8


10.5 LINFTARFJor,iruuutAurc-ruere

273

to the length of the input is availatrle. Irr ftu:t, it is quite difficult to come up with a concrete and explicitly defined larrguage that r:rr,nnothe accepted by rrny linear hounded a,utomaton, In Chapter 11 we will show that thc r:la,ss of lirrcar brlrnded arrtomata is less powerful than the class of unrestricted Thritrg tttachirres, but il dc:monutra,tion of this requires a lot more work.

L Give details for the solution of Exat4rle 10.5. 2 . Find a solution for Example 10.5 that does rrot require a rrecontl track as scratch space. ffi

3 . Consider an off-line Turing machine in which the input can be read only once, rnovirrg left to rigtrt, arrd not rewritten. On its work tape, it can use at most n extra cells fbr work space, whete rt, is fixed for all inputrr. Show that such a machine is equivalent to a finite autornatot. 4.

Fittd lirtear bountlerl automata for the following languages. ( * ) . L: { a ' " : n : n f ,nr}l} (b) fI : {o," : n is a prirne mrrnher} (") .L : {ai'o : n, is not a prime number} (d) r:

{ww,w e {a,b}+}

( e )I =

{wn:w â‚Ź{a,b}+,n>1}

W

( f ) f : { r u r u r t y ' t : , u r e{ a , b } + } D.

Find an lba fbr the cornplernent of the larrguage irr Exarrrple 10.5, assurrring thatE:{a,b}.

6 . Show that for every context-free language there exists an accepting pda, such that the number ofsymbols in the stack never exceedsthe length ofthe input string try urore than .rtt*. W Use thc observation in the abovc exercise to show that anv context-free langrrage not containing A is accepted by some linear bounded automaton.

L

To define a detenninistic linear bounded autornaton, we carl use Definition 10.4, but require that the Thring machinc be deterministic. Examine your solutions to Exercise 4. Are the sohrtions all deterministic linear bounded a,utornata? If not, try to fin<l rrolrrtions that are.



A H i e r q r c h yo f F o r m q lL q n g u o g e s ond Automqtq

e ttow returtr our atterrtion tt) (lrr rnairr intelrc$t,the study clf fbrrrral Ianguages. Our irrrrrrediirtcgoal will bc to cxa,minethe la,nguages 'I\rring associat,edwith tnircltirrcsilrrtl srlrncof thcir rostrit:tions. Re'Ihritrg cause tnachiues can perforrn arry kirrtl of ir.lgrlrithrnir: r:ornputation, we expect to finrJ tira,t the family of languages a"ssociatedwith thern is rpitt: llroarl. It int:hrrlcs not only regula,r a,ncJcontext-free langua,ges,but also the various exarnples we have crrcourrtereclthat lic outsidc tht:sc fir,milies.The nontrivia,l questiorr is whether there are arzy languages that arc rxlt a,crx.'ptedhy some Tirring ma,clrine. We will answer this questiorr lirst by slxlwing thrrt tlxrrc a,remore la,nguagesthan T\rring rnachines, so that thtlrc rmrst llt: sornt: Iarrgrrilgcsfor which there a,re no Ttring rnachitres, I'he proof is slrort itntl c:lt:garrt,Ilrt norrr:r)n$tnx:tive,and gives little insight into the probleut, lbr this rcas()rr,wt: will cstalllisir tlx: cxistenrx: of latrguagestrot recogrrizablcLry Tlrring rnar:hinesthrrnrgh mrirc cxplicit exa,tnplesthat actually allow us t,o iderttify oric suc:h lirrrguagc:. Arxltlrcr avenue of investiga,tionwill be to look aL the relatiorr betweerrThrirrg rnachirrtlsirrrrl t:crtilin tyllt:s of grirmmar$ and to esta,blisha,connection between thtlsc grarnmilrs rlnd relgrrlrrra,nd r:ontext-fiee gra,mmars. This leacls to a

,-t I ti


276

Chopter I I

A HlrnnnclHv or FoRtr,tll Leucuacns axn Autoraana

hierir,rc:hyof grilmrnars arrd tlrrough it to a method for classifying ltrnguerge f'rrrrrilics.Sorrreset-theoretic cliagrarnsillustrate the relationships tx:twt:crr various language families clearly. Strictly spea,king,many of th{r irrgurnetttsitt this clmpter are valid only for la,ngua,ges that do not inchrrlc thc tlrnpty strirrg. This restrictiotr arises fiom the frr,ctthat T\rrirrg urac:hirrt:s,as we have defined theur, canuot accept the ernpty strirrg. To irvoitl havitrg to replrrase the definition or having to disr:lairncr, w{: rnake t}re tacit assutnptiotr l}rat the languages adcl errepr,ra,tncl cliscussecl in this r:haptor, rrnll:ssothcrwist: statetl, do ttot contain A. It is a trivia,l matter to rtrstir.tccvurythirrg so tlrrrt .\ is irrcluded, but we will leave this to tire rerrder.

Recursiveond Recursively irilmmffiffiilt E n u m e r o b l[eo n g u o g e s We start with some terminology for the languages associa,tedwith Thring rnachines. In cloing sor we rnust nrake the important distinction between Ianguagesfor which there exists an accepting Ttrring machine and klnguages for which there exists a,memhership algorithm. Ber:inrsea T\rrirrg rnat:hirrt: does not necussarilyhalt rlrr irrput that it rloes rrot accept, the first does trot irnply the second.

A language1, is said to be recursively enumerable if there exists a,Ttrring ma,chinetha,t accepts it.

This deflnition implies only that there exists a Ttring rnachine M, such lhat, fbr every ?0â‚Ź L, tyyut I y

:I:tqII:21

with gy il finill ster,tc. Thc dtrlirritiorr says rrothirrg about what lnppetrs for 'rr.rnot irr .L; it rnay btl that the rrrachine halts irr a rrotrfinal state or that it never halts and goes into an infinite loop. We can Lremore demanding irnd ask that Lhe rnachine tell us whether or not any given input is in its Ia,ngua,ge.


11,1 R,ECURsrvrrt AND R.ncunslvl:lv Exulannenlt

Larvcuacns

277

I |.2 ,rFirqflir$,Iriion A languirge tr on x is said to be recursive if there exists a 'rhrirrg machine M that ar:rx:pts L and tlut halts on every .ro in X+. In olher word$, a Ia,nguageis rrlcursive if and only if there exists a rncrnbership algorithur for it.

If a, language is lrx:ursive, then there cxists an easily cclrrstnrctedenuttterirtion procedure. Sr4lposethat M is rr T\uing machine thir,t rleternines mernbtirship iu a recursivtr la,nguage-L. Wc first construct anotlrrtr lhring rnac:hirre,say M, that gcrreratesall strings in X+ in proper orrler, let us sa/ '{t1,1.r)2,,,,,As thesc strings are generatrlrl,they becourethr: irrprrt to M, wltir:h is modifled so that it writes strings orr its tape only if thcy are in L. 'r'hat thcrr: is a,lsoa,nenurrrerationprocedure lbr cvery recursively enumera'ble languagc is not as easy to srlrl. We cannot use the erhoveargumerrt ir,sit stands, beciruseif sorne rui is rrot in 1,, the machirrc M, when startecl with ttj ou its tirltt:, may never halt rrrrrltirerefore never grlt to the strings irr r, that follow rr.'i irr the enurneratiorr. Trr make sure that this does not hrlppen, the cornllrtation is perforrnetl irr a, clifferent way. Wc Iirst get fr to glenera,teu1 arrrl lt:t M execute one rrr()veon it. Then we ltlt M generate ur2 irnd let M executt: one rnove on ur2, folklwod by the second ilr()vc on urr, After this, we gerrrlrate?lr3a,nd clo one stcp on ur3, the second step on ur2, the third step on [r1, rrrd sri on. The order of ltrlrftlrma,nceis depictcxl in Figure 11.1. Flom this, it is clca,r that M will never get into an infinitcr Ioop. Since any ur e tr is generatecl by -ffi ancl acceptecl by M in a, finite numlrer of steps, cvery string in .L is cvcntrra,llyproduced by M. It is easy to sec tha,t cvery language frrr which a,nenumeraticlrr pror:edure clxistsis recursively crurmerahle. We sirrrply cornpare the given irrput string agairrst successivestrirrgs gcriera,ted by the errurnera.tiorrprocedure. lf 'ur E .L, wc will eventually get a rna,tr:h,and the processr;nn be terminatecl. Defitdtions ll.J ancl 11.2 givc: rr$ very little insiglrt irrto the na,ture of either recursive or recur$ively enrurrerallkllanguages. 'I'hese delirritions a,ttilr:h narnes to languagc I'nmiliesa,ssociatedwith rtrring rnachines,but shcd

wl

Figure 11.1 First move

Secondmove Third move

u2

-j

uj


278

A Hrnnancny o!' FoRMALLANcuAGEsANDAuToMATA

Chopter I I

no light on thc nature of represerrrtativelanguages in these families. N6r do they tell rrs much about the rclationships between these languages or their conmlt:tion to the la,nguagefarnilies we have tlrrcountered before. We "Are there languages are thercfore immediately faced with question such as "Are there lanthat are recursively orrurnerable but not recursive?" and guages, describablt: sotttehow, that are not recursively emrmtlrable?tt While we will be able to supply some &nswers,we will not be able to produce very explicit examples to illustrate these questions, especia,llythc second one'

Enumeroble Longuoges ThqtAre Not Recursively We can establish the existence of latrguages that are rrot recursively enumerable in a variety of ways. Orte is very short and rrsesa very fundamentir,I and elegant result of matherrratics.

Let S tre an infi,nite countirble set. Then its powcrset 2* is not countable. Proof: I,et S : {"r, "r, s3,...}. Thtrn any element f of 2" can be represented by a sequenc:cof 0's attd L's, with a 1 in position i if and only if sa is in t. For example, the set {"r,.tR,,s6} is representedby 01100100...,while {rr,"r,.ts,...} is representedby 10101"" Clearly, any eletrerrtof 25 can be represcntcd by such a sequencc, arrd atry such sequentlc represents a unique clernent of 2s. Suppose that 2s were countable; then its elements could be writl,en iu some ordcr, say t!,t2,.,., and we could enter these into a table, as shown in Figure 11.2. In this ta.hk:,take the elements in the maitr diagonal, and compklrnerrt each entry, that is, replace 0 with 1, and vice versa. In the example in F-igure11.2, the elermentsare 1L00...,so we get 0011... as the resrrlt. The trew sequencerepreserrtssome element of 2D, say f1 for some i. But it camrot be f1 bet:ir.useit differs from fr through s1. For the same

Figure 11.2

0

'l

t-,

(t)

r

0

0

f 0 )

1

0

f 0 ]

\.,n

t

l

L

t

t

l

t l

\*../j

:


I 1, 1 R.ncuRsrvg AND R.ECUR,srvgLyEr,rurr,ren.c,eLH LANCIUAGES

275

rca$on it cannot tre f2, f3, or arry other fi. This contradiction crcates a logical impasse that can be rernoved orrly by throwirrg out the a,ssurnption lhat 2s is countable. I

This kind of arflument, becausc it involves a rrranipulation of tlxr criagonal elements of a table, is called diagonalization. The techniquc is attrilruted to the mathcrnatician G. F. Cantor, who used it to clernonsrrirre that thc uet of real nurnbers is not countable. In the next fcrw chapters, we will see a similar argurrrent in several contexts. Theorern 1l.l is diagonalization in its purest form. As an irnrnediate consequrlnt:eof this result, we can show that, in some serrue,there are fcwer T\rring nrachineu than there are languages, so that there must be sorne languages that ir,renot recursivcly enumerable.

For any nonempty x, there exist larrguagesthat are rrot recursively errurnerable. Proof: A language is a,srrbsetof E*, arrclevery such sutrset is a language. Thercftrre the set of all Innguagesis exactly 2E-. Since X* is infinite, Theorem 11.1 tells us that the srrt of all languagrn on E is not countable. But the set of all T\rring rnachincs t:tr,nbe enumerated, so the set of all recursively enurnerable la,nguagesis courrttrble, By Exercisc 16 at the end of this section, this irnplies that there rrlrst be some languageri on x thar arc rrot recursivelyenurnera,ble. I This proof, eilthough short arrd simple, is irr rnany ways unsatisfyirrg. It is completely rronconstructive ancl, while it tells us of the existence of sorne larrgua,gesthat are not rur:ursively enurnerable, it gives us no ferrling at all for whirt these la,nguagesrnight look like. In the next set of results, we investigate the conclusion rnore explicitly.

A longuogeThotls Not Recursively Enumeroble Since tlvery language that carr he described irr ir direct algorithmic fa-shion can be acct:pted by a I\rring rnirc:hineand hence is recursively enurneratrle, the description of a larrguagc that is not recrrr$ively enumerable must be indirect. Nevrlrtheless,it is possitrkl. The argumerrt involves a variation orr the diagonalization theme.

Tltcre exists a recursively enumerable lirrrguage whose cornplenrerrt is not recursivsly enurnerable.


280

Chopter | 1 A Htunnrrc:IIYoF FoRMAI Leucu'q.cus.twn AuroMere

Proof: Let E : {o,}, and consider the set of all Ttrring machirreswith this input alpha,bet. By Theorem 10.3, this set is t;ourrtable,so wc can associatc arr orcler Mt,M2,... with its c:lements,For cach Tirring rlachile Mi, there Conversely, for is an associated recursively enumerable lirrrguage L(Mt). qirch recursively etrumerable larrguage on X, there is some T\rring machirrt: that accelrts it. We now cotrsider a ncw language tr dcfined as follows. For each i ) 1, tlre string a'; is in.L if arrd oilly if ad E L(Mt). It is clcirr that the la.rrguage ,L is well defiiled. since the Btatement u' â‚Ź L (Mr)' and ltettce a' E tr' firust either btl true or fhlsc. Next, we corlsidet the complcment of -L,

T:

{ o o' a i 4 L ( M , ) } ,

(11.1)

which is rr,Isowell definctl but, a,swe will show, is ttot recursively enumerable. We will show this bv srtrtradiction, strrrting from the assumption that .L is recqrsively enumerable. If this is tro, then there must be some Ttuirtg machine, say Mp, sucltrthat

T: t lttt*1

(1r.2)

Consider the strirrg aft, Is it in .L or in Z? Suppose that ak eZ- By (11.2) this implies that a , ke L ( M i l . But (11.1) now irnplies that

o *f T . Conversely, if we assume tha,t ak is in .L, therr oo # L and (11.2) implies that

ak 4 t' ltvt*1. But thett from (11.1) we get that ak eT. The contradiction is inescapablc, and we must t:onclude that olr assumption that Z is rccursively enumerable is false. To complete the proof of the theorem a$ $tated' we must still show that .L is recrrrsively enumerablc. We can use fOr this the knowrt enumeration procedure for T\rring machittes. Given a', we first find i by counting the number of a's. We therr use the enumeration procedure for I\ring machirres to firrcl Mr. Finally, we give its description along with a" to a universal T\rrirrg rnachine M,, that simulates tlte action of M orr a'. If a' is in tr, the computatign carried out by M,, will eventually halt. The combirred effect of this is a Ttrring tnachine that accepts every o,' e tr' Therefore, by Definition 11.1, L is recursively emrmerable' I

-


11.1 RFrcuRSrvE ANDR,ECTTRSIVELy _Er-ruunR.anln LRrucu.qcns

ZBI

The proof of this theorern cxplicitly exhibits, throrrgh (11.1), a welldeflned Ianguage that is not recursively enumerable. Thiu is not to say that there is arr easy, intriltive interpretation of Z; it would be difficult to exhibit more than a ferw trivial members of this language. Nevertheless, Z is llroperly defined.

A Longuoge Thotls Recursively Enumeroble But Not Recursive Next, we show there are some languagersthat are recursivrlly enumerable but not recursive. Aga.in, we need do so in a rather roundabout way. We begin by establishirrg tr.srrbsidia,ry result. If a lattgutlge tr and its complernt:nt Z are both recursivrlly enumerable, then both languagesare recursive, If.L is rcr;lrr$ive,then Z is also rccursive, and consequently lloth are recursively errrrnrerable. Proof: If tr and -Lare both recursively enunrerable,then there exist TLrring machirres M and M that scrve a,s enumeratiorr procedures for -L and Z, respectively. The first will produce rily,rlt2,... in L, thc uecond fr1,fr2, ... in tr. Suppose now we are given any ?r,â‚Ź E+. We first letff generate ?x1and compare it with ru. If they are not thc same, we Iet M gerrerate fi1 and compare again. If we need to continue, wc: next let M generate u2, then M gerrcrate fi2, and so on. Arry ru â‚Ź X+ will be gerreratgd either by M or ,4.{, so eventually we will get n match. If the rrmtching string is procluced by M, zrrbelongs to -L, otherwisc it is in Z. The process is a membership algorithrn for both -L and .L, so they are both recursive. For tho (ionverse, assume that tr is recursive. Then there exists a membership algorithm for it, But this brxxrmesa member$hip algorithm fbr Z by simply corrrplt:rnenting its conclusion. Therefbre Z is recursive. Sirrce any recursive language is recursivcly enumerable, the proof is completed. I Flom this, we cxrnclude directly that the family of recursively enurrrertr,blelanguages and tlx: f'rr.milyof recursive larrguages a,renot identical. The language tr irr Tlrrx>rem11.3 is in the first but not in the second family. There exists a recursively enumerable languaga that is not recursive; that is, the family of re:cursive languages is a prolrer subset of the family of recrrrsivelyenurnerablelangrrages. Proof; Coruider the la.ngrrage1, of Theorerrr 11.3. This language is recursively enumerable, but its complement is not. Thereftrre, by Theorem 11.4, it is not recursive, giving us thr: kroked-for example. I

-:


282

Chopter I I

Ln'lqcuecnsnNrr A(r'ro\'Iann A Hrun.qnc;Ilvorr FoRIrnAt.

Wc cotrclude from this that there tr.roirrdeed well-defincd latrguages fbr which one cantrot crrrrstruct a memhcrship algorithrn.

1, Prove that the set of all real mrrtrbers is uot countable. 2. Prove that the set of all languages that are trot recursively enurnerablc is not countable. ffi 3. Let -L bc a finite larrguage. Show that then .Lr is recursively enumerable. Suggest arr etrumeration procedure fbr -L+, 4, Let -L bc a context-free language. Slxrw that l+ arrd suggcst a,n entrrttetation procedure ftrr it.

is recursively enumerable

5. Show that, if a language is not recursively enumerable, its complement carrrrot be rer:ursive. 6. Show that the family of recrrrsively enumerable languages is closed under union. ffi 7. Is the family of recursivcly enumerable languages closed urrder intcrsection? 8. Show that the fatrily of recursive lartguages is closed under union and intersection. L

S[ow that the farrrilies of recursively emrrnerable and recursive larrguages are closed under reversal.

10, Isr the family of recursive larrguages closed under concatcuation? 11. Prove that the cornplcment of a t:otrtext-frce language rnust be recursive.

ffi

12. Let -Lr be recursive and -L2 recursively cnurnerable. Show that ,Lz - -Lr necessarilv rccursivelv ernrrtetable, 13. Suppose l,hat L is such that there cxists a'firring rnachine that emrmerates the elements of tr in l)roper ordcr, Show that this rneans that .L is recursive. L4, If L is ret:ursive, is it necessarily true that 1,+ is also recursive? W L5, Choosc a particula,r' encoding for T\rring rrrachines, and with it, find one clement of the larrguagc Z ir't Theorerrr I1.3. 16. Let ,9r be a countable set, 52 a sct that is not countable, and Sr C 52. Show that S'r must then r:orrtairran infinite number of elements that are not in,Sr' 17. In Exercise 16, show that in fact Sr - Sr cannot be countable, 18. Why does the a.r'gumcnt in Theorerrr 11.1 fail when S is firrite'l 19. Show that the set of all irrational mrrnbers is not countable'

ffi


11.2 UNRESTRTCTED GRaMunRs

283

U n r e s t r i c t eG d rsmmdrs Trl investigate thc connection betweerr recursively enumerable languages and grarnrnars, we return to the genarnl definition of a grarnrnilr in Chapter 1. In Definition 1.1 the production rules were allowed to take any form, but various rcstrictions were later made to get specilic grirmmar types. If we take the gcrreral form and itttposc no restrictiorn, wc gert rrnrestricted grammals.

iiilffifr,ffi$ffi,fi{,ffi,{tilli l'iirj A grammar G : are of the form

(V,T,S,P)

is called unrestricted

if all the produt:tions

U+U,

where z is in (V U ?)+ ancl u is in (y U 7)..

In an unrestri<:tedgrammar, essentially no conditions are irrrposedon the productions. Any rrurrrber of varia,bles and tertrinerls ca,n be on the left or riglrt, a,nd these can occur in ilny order. There is orrly one restriction: .\ is not allowtrd a,sthe left side of a prodrrction, As wc will see, unrestricted grrrmmars are nruch rnore powerfirl than restricted forrns like the regular and contr)xt-fiee grammars we have strrdied so far. In fact, unrestricted grarrrrnars txlrrespond to the largest faurily of languages so we ca,n hope to recognizc try mechanical rneansl thir,t is, rrnrestricted gralrrrnars gcnerate exactly the farnily clf recrrrsively enurnerablc languages. We show this in two parts; the flrst is quite straightforward, but the second involvcs n lengthy constructiorr.

Anv language getreratcrl by tln unrestricted grirrrrrnirr is recursively enurnerir.ble. Proof: The grarnnrar irr cffer:t defines a procedurr: frlr enumerating all strings in the language systt:matically. For exarnple, w(.' (:ernlist all ur in .L such tha,t S+u, tha,t is, ur is derived irr orre step. Since the set of the productions of the grammar is finite, thcrc will he a finite number of srrr:hstrings, Next, we


?.84

Chopter I I

A HtnneRcHy or.' !'on,uel

Lancueons

AND AUToMATA

Iist all w itt L tha,t can he dcrived itr two steps

,9+r+tr, and so on. We can simulate these derivations on a, Tltring rnachine ancl, therefbre, have atr enumeration procedure for the langrrngtl. Hence it is recursivelycrrutnerable. r

This pnrt of the correspondenccbetweett recursively emrmerable langueges and urrrestricted gra,mmars is rrot surprising. The grammar generates strings by a well-defined algorithmic process, so tlte derivations can he dclnc on a T\rring ma,chine. To show the converse, we describe how any Tlrring machine can be mimickccl by an uurestricted grarrrmar. and want to We are given a Ttrring mat:hineM : (Q,E,f,d,go,n,F) prodnce a grammar G such that .L (G): L (M)' The idea bchind the constrrrctiorr is relatively simple, but its itnplemeutation bet:tttresttotatioually cumbersotne, Since the computatiorr of the T[rring machine carr be described by the sequerrceof instantancous clescriptions

qou I rqfA,

(11.3)

we will try to arrangerit so tltat the corresponding grarnmar has the property that

Qxu tf'I:Qy!,

(11.4)

'Ihis is not hard to do; wltat is more difficult to if and only if (11.3) holds. see is how to rnake the connection hetw(:t:rr(11.4) and wha't we rea,llywirrrt, namelv, S4u' ftrr all tu satisfying (11.3). To irc:trievethis, we construct a grarrrrrlarwhich' irr broad outline, has thr: following properties: 1 . ,5 carr derive qou for all ur e X+.

2 . (11.a) is possibleif and only if (11.3) holds.

3 . Wherr a string rqly with qy â‚Ź lr is gerrerated,the grammar,rtratrsforttts this string into the origirral'u.


11.2 UrunnsrRrcrED Gnalruans

285

Tlrt: rxlrnlllete $equen(ieof derivations is then

$ 4 qorrl

xq.yy1w.

(i1.5)

Tlrc thirrl step in the a,bovederiva,tionis the troublesome one. How can the grrrmmar remember u if it is rnodified dr.rringthe secondstep? We solve this Iw enurding strings so that the coded version originally has two copies of u.'. Thc first is saved, while the second is used in the steps in (11.4), When a fitral corrliguriltiorr is tlrrtcred, the grammar erases everything except the saved'ur, 'Io produce two copies of tr and to harrrllrl thc statt: syurllol of M (which eventually has to be removed by the grammar), we introduce variables Vo6 ir,ndVoi6f'or all a € XU {n}, h € I-, and all rl such that qa e Q. The variable ydb ()rr(j()(l()tr tlrt'two syrnbols rr,tr,nclb, while V.i6 encodesa and b as well as thc statc: r;;. Tlx: first step irr (11.5) can berac:hieved(in the encoded fbrm) by

5 - VrrSlSYnll?l

(11.6)

7' -

(11.7)

TV,rolvooo.

for a,lla e E. '.fheseprocluctions allow the glarrunar Lo gerreratean encoded version of any string q0urwitlr a,n arbitra,ry number of leading and trailing blarrks. lbr the second sttp, for cirdr trarrsitirlrr

,l (qu,,J : kt.i, ,1,, R) of M, we prrt into the gra,mmtr,rproductions W.Vprt --, WdVrtq, for irll n, p € EU {l},

( rr . s )

q e f. For cirr;}r d ( q 0 , " ): ( q j , d , L )

of M, we inch-rdein G VpqVo,t"-

VpiqVoa,

( 1 1 .)e

for all a, p e }JU {tr}, q e I'. If in thc st:t:ttrrrlstc:p, M t:tttr:rs a firral stattl, the gratrrrnar tnust thert get rid of everyLhingexcept ur, which is saved in lhe first indices of the V's. Therefore, for every ei E F, we include productions WiA + et

( 11 . 1 0 )


286

Chopter I 1 A HInn,.tncIIY oF Fon,nrlal Lnr'{cuncos awn Auttlltane

f'rrrnll a â‚Ź X U {tr}, b e f. T}ris createst}re first terrninal in the string, which tltcrr causesa rewritirrgirr the rest,by cWt Wtc for all a, câ‚Ź EU{tr},

(11.ri) (11 . 1 2 )

CQ,

a(,

lre f. We need one more specrialprodrr:tion E -

(11.13)

.\.

This last production ta,kescare of the ca,sewhen M moves outside thnt pa,rt of the ta,pe rlr:t:rrpicd tly thc input 'ur. To rnirkc things work irr this casc, wc rmrst first rrsc (11.6) rrnrl (11.7) to generate E...lqprrrn...!, representing all the tape region used. The extraneous blanks are removed at the erd by (11.13). The following example illustrates this complicated construction. Carefully check each step in the exa,mpleto see what the variorrs prodrrctions do antl whv thclv itrt: rrecrlcd.

lliHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH mt M : (Q,x, f , d,qo,n, F) uea Turingmachine with I : {qo,qr} , 1.:{a,b,tr}, )J : {rr,b} , t' : {qt} ' itncl

d (so,o) - (q11,a, R) ,

d (qo,tr) : (gr,!, r) . Tlris rttachirreacccptsL(uu*). Clonsirler rrow tlte tlttrnuutittiorr esaal aesal aaesl F aqla[

(11,14)

which acct:1ltstlrt: strirrg na. Ttl tklrivc this strirrg with (J, wtl first us(rnrlc$ of the form (11,6) and (11.7) to get the appropriate starting string, S =+ 5'Vll

=+ TVaa +'lV""Vaa

4 VoooVooVll.

I'he Iast sentential fornr is the starting point for the part of the derivatiorr that mimics the cornputation of the'I\rring machine. It contains the original


I1.2

UNRnsTRrcTED GRarrrlrans

287

input aaE in the sequellce tlf first indices and the initial instantarreous description esaal in the remaining irrdiccs. Next, we apply - WoWoo, VoooVoo and VoooV;,z * Voolltr0tr, whiclt arc spcr:ifir:insta,ncesof (11.8), and Vo*l/tr0tr -

WruVaa

corning frtlur (11.9). Then the next steps in the derivatiorr are V,,0,,.V,.uVaa * VouVoooVVa)VooVooVlon

+ VooV,srV11.

The sequenceof first indices remains the same, always rernembering the itrititrl input. The sequenceof the other indices is 0aan, aOa[, aa0!, a1aE, which is equivalent to the sequenceof irrstarrtanrxlrsdesr:riptionsin (11.14). Finallv, (11.10)to (11.13) are used itr tlte last steps Vu"VtuVaa 4 VooaVag + V""al

+ aal + aa.

Thc cxrnstnrctiondescribedin (11.6) to (11.13) is the basis of the proof of thc following result.

T

For every recursively enurnerable languagc.L, there exists an unrestricted grammar G, such that ,L : L(G). Proof; The construction described suarantees that if

rla, then e ( r )+ e ( y ) , whercc (z) denotesthe encodingof a string accordingto the givenconvention. By an induction on ttre rmmberof steps,we cirrrthen showthat e (So?,')5 " (g) ,


288

Chopter I I

A Hrnnnncnv or Fon,ltnl, L.q.NcuacnsArttr Atlr'ovane

if and only if q1uri Y' Wc also rnust show tha,t we can gcrrt:rate every possible sta'rting t:ttnfiguratiorr and that tu is properly rerxlrrstnrctt'd if and only if M enters ir lirrirl configuration. The cleta,ils,whir:h arc rrot too difficult, a,releft irs iilr cxercise. I

These two theorems esta,hlishwhat we set out to clo. Thev show that the family of langua,gesa,ssor:iirtt:rlwittr utrresl,ricted gra'mma,rsis idt:rrtitlirl with the family of rec:rrrsivclyetrumerablelauguages.

1. What language does the utrrestrictcd grammar Jj + 5'rB '5r - a$rb bB -- bbbB aStb '

aa

B - A

clerive? ffi

2 . What difficulties woultl arise if we allowed thc empty string as the left sicle of a production in an unrestrictcd grammar?

3 . Consider a variation oII gl'alnmars in which the starting point for arry derivation can be a finite set of strings, rather tharr a single variable, Formalize this concept, then investigate how such gratnma,rs telate to the unrestricted grarrrfirars we havc usccl here. ffi

4 . In Example 11.1, prove that the constructerJ grarrlrrrar cannot generate anv sentence with a b in it. 5 . Give the details of the proof of Theorern 11.7. 6 . Construct a Turirrg rrrachineftrr tr (01 (01)-), thcn find an unrestrictetl grarnrtrar for it using the coustruction in llheorem 11.7. Give a derivation fbr 0101 using the resulting gramnrar. t .

Slxrw that fbr everv unrestrictcd grammar there exists an etpivalent stricted grammarr a,ll of whose protluctioris have the fortn 'u+at with z, u â‚Ź (Y U ?)+ and lzl < lul, or .4 +.\,

withAâ‚ŹY. m

unre-


I1.3 Corur'r:x'r-Snrqsrrrvn Clnauuans .q,rup Larucuacns

289

8. Show that the conclusion of Hxercise 7 still holds if wc add the further conditiorrs zl < 2 and l"l < 2. 9. Somc authors give a definitiot of urrtestrictetl grarnrnars that is not qrrite the same a,sour Dcfinition 1..|.,3.In this alternate definitiorr, the productions of atr utrrestrictetl Hrarnrnar are rerruired to be of the form T+U,

where

,dâ‚Ź(vu").v(vu7)., a,nd

aâ‚Ź(vu7)", The diffctcncc is that hcrc thc lcft side rnust have at least orre variable. Show that this alternate definition is llasir:ally the sarne as the one we use, in the sensc that for cvcrv grammar of one t.ype, there is an equivalent grammar of the other typc.

C o n t e x t - S e n s i t i vGer o mm or s ond Longuoges Bctween the restricted, context-fiee gra,mmarsancl the general, unrestricted grarrlrnar$r ir grt:at vir,rir:ty of "srlrncwtrilt restricted" grammitr$ t:tr,nbe defined. Not all c:irscsyit:ltl irrttlrcstirrg rrnultu; among the ones thtlt rlo, the cotttcxt-st:rrsitivtl grilmma,r$ ha,ve relcc'ived considerahle a,ttention. These grarrrrrrirrs gtlrrcrilte lrr,nguer,ge,'s rr,ssrlr:ia,ted with a, restricted clir-rs of T\rring macltirres,lirrc:itrtrourrdcxlirutornata, whic:h we irrtrorhrt:edirr Ser<:tion 10.5.

N$ffi$fl tl.'$$fr,ffi,fi N,\tthitffi ,l' A grammar G: (V,'I',5,P) tions are of the forrn

is said to Lrecontext-sensitive

if all produc-

where tr, a E (V U T)+ ancl

l z l5 l s l ,

(il.15)


290

Chopter I I A Hrnn-,wclHyor FoRrr,telLauc;uacns eNn Aurol"tn'r'A

This definition shows clearly one asper:t of this type of grammar; it is noncontracting, in the sense that thc lcngth of successivesentential forms can never decrea^sc.It is less obvious why such grammar$ should be called rxlrrtcxt-serrsitive, but it cau tre shown (see, for cxarrrple, Salomaa 1973) that all such grarlmars can be rewrittt:rr irr a rrormal form in which all productions are of the form xAY + YPY. This is equivalent to saying that tlre prodrrr:tion

A-tu cirn be applied only in the situation where -4 occur$ irr a context of ihe string z on the Ieft ancl the string 3yon tlrc riglrt. While we use the terminology arising from this particular interpretation, thc forrn itself is of little interest to us here, and we will rely entirely on Dtllirritiorr 11.4.

Context-Sensilive Longuogesqnd LineorBoundedAutomotq witlt As thc terrrrirrologysuggests,context-sensitivegrammar$ are a^ssociated a latrgua,gefamily with the sarnc rrarrle.

if thcre exists a context-sensitive A language ,I, is said to be contcrxt-st:rrsitivc g r i r m r n ? r(rJ , s u c ht h a t - L : L ( G ) o r L : , (G)U {A}.

In this definition, wo' reirrtroduce the empty strittg. Definiiion 11,4 implies that r + A is not allowerl, so that a context-sensitive gra,mmar can never gernr:rittc a lirnguagtl containitrg the empty string, Yet, every contr)xt-frrxl larrguage without .\ can be generated by a special ca-seof a context-sensitivc grarnrrlar, say by otre in Chomsky or Greibach normal ftrrrn, both of which satisfy the conditions of Definition 11.4. By inclrrding thc ernpty strirtg itr the definition of a context-sensitive langrtage (but not in the grammar), we can claim that thc frrmily of corrtext-free languages is a subset of the family of context-sensitivc languages.


11.3 Colrrlxr-SnNsrrrvr Gnntttrrtns.q,rqu LeNcuecns

. . . Hm6,tttfbleI t,2

291

The language L : {artfirlsn: n > 1} is tr, context-sensitive language. We slrow this by exhibiting a context-sensitive grirrrrilrirr fbr thel la,ngua,ge.One sut:h gramma,r is

S - tthclaAbc:, Ab - bA, At: + fi6"", bB --+Rb, aB --+aalaaA. We can see how this works by looking at a derivation of o:lblcil. S + aAbc+ abAc+ abBbct, + aRbbcc + aaAbbcc+ aubAbt:c + aabbAcc+ aabbBbtcc + aab Bbbccc+ aaB bbbccc. + aaabbbccc. Ihe solutiorr clIectively rrses the varia,bles A and B as rnessengers. Arr ,4 is created orr thc luft, trirvels to the right to the first c, where it creates another b arrd r:. It therr sends the messenger B back to the Ieft in order to create the corresporrdinga. Tlrc pror:essis very similar to the way one might program a Thring machine to accept tlrc lir,rrgutr.ge tr. I Sirrrxl tlxl lir.ngrragein the a,bove example is not context-free, we see that the farrrily of rxrrrtcxt-free ltrnguages is a proper subset of the farnily of context-sensitive larrguagcs. Exarlple 11.2 also shows that it is not an easy rnatter to find a context-stlrrsitivc grirmmerr even for relatively simple examples. Often the solution is rnost cirsily obtained by starting with a Ttrring machine progra,In, then finding atr equivalerrt grarnrnar frrr it. A ftlw oxa.rnpleswill show that, whenever the language is context-serrsitive, the corresllorxlirrg T\rring machine ha,spredictable space requirementsl in particular, it carr trclvirlwrxl as a,lirrear hounded automaton.

Frlr every crrntext-sensitive language .L not including .\, there exists sorne lirx:irr brlrnded a,utomaton M such that tr : L(M\. Proof; If -L is contexL-sensilive,t,henthere exists a corrtcxt-sc:nsitivt:grammar fbr r-{A}. We show that derivations in t}ris grarrurrarcan be sinrulatecl trv a, Iinear bounded autorna,ton. The linear bounded automaton will have


292

Chopter I I

A HreRRRcHyoF FoRMAI Lnmcuacns eno AurouAlA

two tracks, one containing the input string ur, the othrtr r:clrrtainitrgthe sentential forms derived r.rsing G. A key point of this argumetrt is that no possible sentential fbrm can have lcngth greater than lul. Another point to notice is that a,lincar bounded automatotr is, by clefinitiotr, nondeterministic. This is rrct:cs$aryitt the argument, since we can claim that the cclrrect productiorr carr always be guessed and that no unprodrtctive alterrnativcs havc to be pursued. Therefore, the computation described in Theorctn 11.6 can be carried out without using space except tha,t origirrrrlly occupied by ur; that is. it can be done bv a linear bounded automaton. I If a language tr is acc:eptcrl by sorne litrear bounded automaton M, then thcrc r:xists a context-sensitive grammar that generates tr. Proof: The coustruction here is similar to tha,t in Thcorern 11.7. All productions generated in Theorcm ll.7 are rroncontractitrg except (11.13),

tr-*4. But this production can be omitted. It is necessaryorrly when the Thring machine moves outside the bounds of the original irrput, which is not the case here. The grammar obtir,ineclhy thc corrstructiorr without this umrecessary produr:tion is nttncorttractittg, cotnpleting the argument. r

RelotionBetweenRecursive ond Context-sensitive [onguoges Thrxrrr:m 11.9 tclls us that cvt:ry cotttext*sensitivelatrguageis accepted by somtr T\rrirrg rnirr:hitre:arrtl is tlterefore recursively enunerable. Theorem 11.10 follows easily frorn tlfs.

Every r:rlrrtext*setrsitivelattguage -L is recursive. Proof: Consider the context-sensitive la,nguage.L with an tr,$sociatedcontextsensitive gramrnar G, and look a,t a,derivation of ro 5+11 =+r?'+"'+rn+ut' We can a,ssumewithout any loss of generality that nll serrtcrrtial forrns in a single derivation are diff'erent; thirt is, *.i t' x.i for all i t' j. The crux of orrr argrrment is that the numbcr of steps in arry tlerivation is a bounded firnction of kpl. We know that

l*il5 l"i+rl,


11.3 CoNTExr-Snmsrrrvn GnavveRs ANDLANcuAcr,ls

293

because G is noncontracting. The cxrly tlfttg we need to add is that there exist some rn,, dcperrdirrg otily otr G atrd 'ur, such that

l z . i< l lri1.,l, for all j, with trl : nL(lrl) u bor.urded functionof lV u 7l nnd lz.'1.This follows becausethe finiteness of lV U ?l implies that thcrc are only a linite number of strings of a,given length. Thcrcfore:,the length of a derivation of w E L i s a t m o s t l r u l r n( l r r r l ) . This observation giveu rrs imrncdiately a mernbership algorithrn for .L. We check all derivations of lcrrgth up to l'ulrn (l'rrrl). Since the set of productions of G is finitc, there are only a finite number of these. If any of them give ur, then u â‚Ź L, otherwise it is not. I

There exists a recursive langurr.grlthat is not context-sensitive. Proof: Conuider tlrtr set of all context-sensitive grammars en 7: {a,b}. We t:tlrr r$() a convention in which each grammar has tr,vtr,ritr.blcset of the f'orm V : { V o , V t , V z , . . . }. Evcry context-sensitive grammar is completely spet:ifitrl by its productionsl we can think of them as written au a single strirrg l U 1+

AtlfZ

+

UZl ...

' ,I r n . +

Jlrrr.

To this string wtl now apply the hotnomorphisnt ft'(a) : 616,

h (b): o1zo, h (-)

: 61:t6,

lz(;):s1+6' h'(V):01i+50' Thus, any context-sernsitive grarntrtar call be represented rrniqucly by a string f'rom I ((011-0).). Furthermore, the representatiorr is itrvertible itr the sense that, givcrr arry such string, there is a,t most ont: corrtext-sensitive grarrrrrrarcorresponding to it. Let us introduce a proper orcleringon {0, 1}+, so we ca'nwrite strings in tlre order uJltlrzt etc. A given string ui rrlay ntlt define a context-sensitive gramrnar; if it does, call tho grirrnrrritrGi. Next, we defiue a' la'nguage-L tly L:

(J4 alrd *, # L (G,)}. {u.i: trr4definesa context-sensitivegrammar


254

Chopter I I A HtnnencHv or FoRvel, Larucuncus ANDAuroMNr'A,

1, is well defined and is in fact recursive. Tcrseethis, we construct a membership a,lgorithm. Given ?o?,we check it to scc if it defines a context-sensitive grammar G,;. If not, then ruaf L. If the strirrg rlclrlsdcfine a,grammar, then L((;t) is rer:ursive,and we can use the membership algorithm of Theorem 11.10to find orrt if rui € I'(Gt). If it is not, then u,.bekrrrgsto.L. Elut .L is rrtlt rxlrrto.xt-sensitive. If it were, there would exist sornrr rl; suclr that L : L (Gi) We ca,nthen a,skif uri is in .L (Gj). lf *" ,t.s,r*L tlrat trri ( L(Gj), tlrcrr by clefinitionuj is not in.L, But L: L (Gi), so wtr Irave a corrtrrrrlir:tion. Conversely,if we assurnethat,ui # L(Gi), then by definition wi € L and we havr: irnother contradiction. We must therefore conclude that .L is ttot mrntoxt-sr:nsitive. t

The result in Theorern 11.11 indica,testhat linear bounded automata are indeed less powerful than TLrrirrg urachines, since they accept only a proper subset of the recursive languagt:s. It f'ollows from the same resrrlt that linear bounded automata are rrrore powr)rfirl than pushdown automata. Context-free languages, beirrg gr:nerated by context-free grarnmars, art: ir,srrhset of the context-sensitive languages. As various examples show, they arc ir proper suhset. Because of the esserrtial equivalence of linear bounded arrtorrrattr tr.ndcontext-sensitive languages orr one harrd, and pushdown autornilta ilnd context-free la,nguageson the othcr, wu see that anv Ianguage accepterl by ir,llrshdown automaton is also accepted by stlrne linea,r bouncled automaton, but that there a,re la,nguagesaccepted by sorntl Iinear bounded automata for which there are no pushdown automata.

* 1 . I.'incl context-sensitive granlmars for the following {anguages. (u)

{an+rbnr"

1:ri}1}.

( b ) Z , : { t t " h " a . z " :n , > 1 } (c) I:

r n ] l,rn, > 1] {e,rtfir'txrt4rrt

(d)l,:

{ww,w €{a,b}+}

ffi

* 2 . Find context-serrsitive fJrammars for the following languages. (a) I:

{ t n : n * ( ' u r ): 7 r , ( u , ) : n . ( t u ) }

(b) t : {w t no (r) : rro (u,')< n,, (ur)} 3, Show that thc family of corrtext-serrsitive languages is closcd under union. 4. Show that the farnily of context-sensitivc languages is closed under reversal.

ffi


11.4

'I'HE

CHor"tsxv HIURARC:HY

295

5 . Frrr m, in Theorem 11.10, give explicit bounds for m, as a function of lrrrl and

lv u71.

Without explicitly constructing it, show that therc cxists a contcxt-frcc grammar for the language 7 : {uul i 1t.t11t, E {a, b}+} ffi

TheChomskyHierorchy {ffiffiffiffiffi We have now encountered a nurnber of language families, arrlong them the recursively enr.rmera,hle languages (Lnu), the corrtext-sensitive la,nguages (Zcs), tho rxlrrtcxt-fieerlir,ngrrages (.L6,'p),and the regrrlar lilngrrargers (Ln.nc). Orre way of exhibitirrg the relatiornhip between these farnilies is by the Chomsky hierarchy. Noarn Chorrrsky, a founder of fonnal language theory, provided an initial classification into four language types, type 0 [o type 3. This original tenninology lns persisted and one firrds frequent refererrcesto it, but the numeric types are actually different rlarnes for the language families we have stuclied. Type 0 languages are those generated by unrestricted gramurars, that is, the recursively enumerable languages, Type I consists of the context-sensitive languages, type ? consists of the context-free la,nguagesand type 3 consists of the regular la,nguages.As we have seen, each larrguage farnily of type i is a proper sulrset of the family of type i - 1. A diagrarrr (Figure 11.3) exhibits the relalionship clearly. Figure 11.3 shows the original Chomsky hierarchy. We have also met several other language families that can be fitted into this picture. Including the families of deterministic context-free languages(L nc p) , and recursive la,ngua,ges (I'nac), we a,rrivea,t the extended hierarchy shown in Figurc 11.4. fa,miliesr:rr,nhe defined tr,ndtheir pltx:c:irr Figrtrc 11.4 Other la,ngr.ra,ge sturlirxl, altirough thcir relatiorrships do trot always ltave Lite neatly trestecl structure of Figures 11.3 and 11,4, In sorne instances, the relationships are not cornpletely unclerstood.

Figurc I I.3

--------7;'----


296

Chopfer I I

A HrpRaRcuyol FoRraer,Lnmcuacns enn AuroMArA

Figure 11.4

ll\]i$'$,.ilWil$\$

language we havepreviouslyintrod'cedthe context-free L:

{ w : n " ( w ) : n 6( u ) }

and shown that it is deterministic, but not linear. On the other hand, the language 7,: {a"bn} U {a'b?'} is linear, but not deterministic. This indicates that the relationship between regular, linear, determirristic corrtext-free, and rrondeterministic context-free languagesis as shown in Figure 11.5. f

l'igrrre 11.5


11.4 Tsr: Clsorr,rsxv Hlpn,an.crrv

257

There is still a,n unresolved issue. We introduced thu trrrx:cl1ltof a deterministic linear boundecl automaton in Exercise 8, $ec:tiorr10.5. We can now ask the question we a-sked in r:onnection with other autorrrata: What role does nondeterminism plav here'l Unftlrtrrrratclly,there is tro easy answer. At this time, it is not known whether thc tarnily of lattguages is ir, propcr subset of accepted by deterministic linea,r bounded automrr,ter, the context-sensitivela,ngtrages. To slrmmarizc, w() have explort:d thcl rclationships betweeu several langurlgclfarnilies arrd their associatedautomata. In doing so, we established a hierarchy of languages and classified autornata, by their power ir^slanguage accepters. Ttrring machines a,re more powerfirl thtln lirrcar botutded autouata. These in turn are more powerful tha,n pushdown inrtomata. .At thc bottorn of the hierarchy are finite accepters,with which we begarr our study.

I , Collect examples given irr this book that dernotrstratc that all the sub.set relatiorrs rlepit:tetl irr Figure 11.4 are indeed propcr ones. 2

Firxl two exarrrples(excluding the onc in Example 11,3) of languages that are Iirrear but rrot deterministic context-frcc.

3 . Firrtl two exalrrples (excluding thc one in Exa,mple I1.3) of larrguagesthat are deterministic context-free but not linear.



L i mi t s o f A l g o r i t hmi c C o mp u l q t i o n

rrvirrg trr,lked a,bout what Thring nrachincs r:iur rlo, w() now look at what they cannot do, Although Thring's tlxnis letadsus to believe that thcre are f'ew limitations to the power of a T\rring ma,chine, we hitvc c:lnimed on severa,loccasions that there coukl rrot uxist a,nv a,lgorithmsfor the solution of ccrtain problems. Now we rnake more explicit what we mean by this claim. Sorneof tlxr results ca,meabout quite sirnply; if a langurrge is nonrecursive, then by definitiorr thurc is no membership algoritlrrn ftrr it. If this were a,ll there was to this issuc, it would not be very interestingl norrrec:ursive:lirrrgrur,geshave little pra,ctical value. But the problem goesdeeper. For r:xa"rnplc,we hirve sta,ted(but nol yet proved) that there exists no algoritlmr to detc:rrrirut whether a context-free grammar is rrnambiguous, This question is clearly of prrr.crtical significancein the study of progra,mrning languages, We first define the concept of decidability and computability to pin rlown wha,t we mean when we say that sorrxlthing cannot be done by a, Thrirrg urar:hine. We then look at several classical prcltrk:msof this tvpe, alrrong thern tlx: well-known halting problern for Thring rnachirres. FYornthis follow a nuuber of rrllatrxl prohlems fbr T\rring machines and recursively

299


300

Chopter l2

Lrurr$ oH Ar,coRt'ruMIC CoMpurATIoN

cnumerable languages. After this, wc look at sotne questions relnting to context-free langua,ges. Hcrc we find quite a few importa,nt ltrobletns for which, unfortunir.tr:ly, thc:rc are Ito algorithms.

S o m eP r o b l e m sT h o t C o n n o fB e S o l v e db y T u r i n gM o c h i n e s The argurnent that the power of met:hilnir:irl txrrnputations is limited is not surprising. Intuitively wc krrlw that many vague and speculative qutrstiorrs require special insiglrt and reasotritrg well heyond the captr,city of arty cotnplrter that wc carr now cottstruct or even pla,r.rsiblyforerstx:.WInt is more irrterestirrg to computer scientists is that thcrc arc questiotrs that can be clearly and simply stated, with irrt apparetrt possibility of an a'lgorithmir: solution, br.rt which are known to lltl rrttsolvableby atry computer.

ond Decidobility Computobility In Definitirlrr 9.4, wrr stated that a function / on a, certrr,indornitirr is saicl if there exisls a Tlrring ma,chinetha,t computcs t]re value to be c:omputrr.bltl tlf / frlr all arguments in its doma,in. A firnctirin is rrrx:rlrnputableif no such 'I\rring machine exists. Them miry be a Turitrg rnachine that can compute ,f rlrr pnrt of its dorrtairr, but we call the function comprrtable only if there is a,Tlrrirrg rnac]rine t]ral computes the frrnction on ther wtxlltl of its dotttaitr. We see fron this that, when we cla,ssify a, ftrrrt:tiorr rrs cortrputable or not cornputable, we rnust be clear on wira,t its rlomairr is. Our conceru her:e will be the sorncwhat sirnplilicd setting where the result of a comlrutatiotr is a, sitrlrkl "yr)$" r)r o'rr)." lrr this case, we talk about a problem being decidable or undecidable. By a proble.mwe will understir.nda set of relatecl staterrents, each of which must he either trtrtt or false. l'or example, we consider the sta,tement "Fer il t:ontt:xt-frtxl graIIIInaI' G, the langua,geI, (G) is rr,mbigrrorrs."Fur sttrtreG this is true, for others 'I'he probletn is to it is fa,lse,brrt cleir,rlywc rnust have otte or tlte other. decide whelthrlrthrl stirtcrrtetrtis true for any G we are given. Again, there is iln rrnrk:rlyirrgtlorrrain, the set of all context-free gramma,rs. We say thtlt a prublc:rnis decidable if there exists a,Tirring macirine that gives tlxr txrrrcct irnrJw()rfilr t:very statetnetrt in the domain of the problem. Whetr we state decidability or: undecida,bility results, wc rntrst always krrow whal the domain is, becausethis ma.ya,flectthe:c:onrlhrsiotr.The problem rnay be decidable on some rlomairr llrt rrclt ort atrother. Specifically, a single instance of a problem is ir.lwirysrkx:idablc, sirrce the allswer is either "ttue" true or fa,lse. In the first cirtrcra TLrrirrg rnachitre lhat always answers gives ttxl rxlrr(x:t irr$wer, while in the second case one tha,t a,lways answers "firlse" is appropriate. This may seenrlike a fa,cetiollrtrn$w()rrtnrt it ernphasizes atr ittrporlatrt point, fhe fact tha,t we do not know what the correct'


12.1 Solrn Pn,otl,EMS THAT Carvruor Bu Sol,ven ey TuruNc; M.c,cHtvt;s

301

alnwcr is makes tro differerrcc,what rna,ttersis that tlxlrc exists some'Ihring machinc that cloesgive the corr(x:t re$ponse,

TheTuringMochineHoltingProblem Wt: begin with sotne problurns tha,t have sornr: historica,l signific:ance arrrl thir,t rr,tthe same tirne givc: rrs a, starting point frlr clevelopinglater results. Tlrc best-known of thr:sr:is the Ttrring rnachirrc:halting problem. Sirrrply stated, the prol,rlernis: given the description of a. Thring machine M irrrd rr,ninput 'ur,does ,4,f,whcn startecl in Lhe iritia,l configuration {6ro, perform rr.c:omputation that t:vcntrrerlly halts? Using irrr tr.bhreviated way of talkirrg aborrt the problem, we rrskwhelther M applied to .ur,or simply (M,,u), halts or docs not iralt, 'fhe dorrrairrof this problem is to bc tir,kenas the set of all l\ring machirrcs ir,nd a,ll ru; tha[ is, we arc looking for a single T\rring marchinethat, given thu rlt:scription of an arbitrirry M and tl, will predict wltctlNlr or not the cornputatiori of M applied to to rvill hrrlt. We cannot firrrl the iruiwer by simulating tlrc a,c:tionof M on tu, say by perfbrming it ol a urrivt:rsill Thring machine, br:(:ir,rrse there is no limit orr the length of l,he cornltrrta,tion.If M en[ers an irrfinite loop, then no matter how long we wait, w() (iirrrrrever be sure Lhat M is irr ftrct in a loop. It rnrry simply be a caseof a,vcry long cornputation. What we need is an algoritlrrn tha,t can detennine thc r;rlrrect answer for any M arrrl ur by perfbrming sorne arrirlysison the machine's clesclilltion a,nclthe input, But as wc now show, no such algorithm exists. For subsequent discussirirr,it is convenient to hilvr: ir, precise idea of wlra,t we mean by tlu: ha.lting problem; for this rcilrJorr,we ma,ke a specific tlcfinition of whal, we statc:rlsrlrnewhat looselv above.

r'Rfi$,fiinnrnmtmW,Ni'..8Nlli|li Let up1 be ir,string that describesa,T\rring machine fuI : (Q, )_1,l,d-,r1o,n, F), atrcl let,'u bc a strirrg in M's alphabet. We will imsllme tha,t'ur,q a,nclru are encoded as a string of 0's irrrrl l'$, a,ssuggestedirr Secltirln10.4. A solution of the halting problt:rn is a T\rring machine H, wlfch firr a,nv ?rA,rand ?r, llcrforms the computatiorr i *rqo*r, Q(Ju.:11'w if ,tr' applied Lo ,urhalts, anrl qo'tlt A,tIt) i,rr rrl u'.yr, if M applirxl to zr does not halt. Here q, arrrl g", are both fitral states of ff .


302

Chopter l2

CoururNt'lol Lrlurs or- Ar.c-+orrtrtIr,tIc

Thurc cloes not exist any Ttrrirrg rnachiue F-l that bchirves as required by Definition 12.1. The haltirrg problem is therefbrc urrdecidable. Proof; We imsurtte the contrary, namttlv, that there exists a,n algorithtn, and consequently sorne I\rring ma,chine ff , thrrt solves the halting protrlcrn. The input to ff will be tlte string u,14ru. The: rcxpiretnent is then thtr,t, givctr any wMwl the T\rrirrg machine,tl w^ill htr.lt with either a yes or no arswer. We achieve this trv askittg that 11 ha,lt in ttnt: of two correspottding firrrrl states, s&y, {ri or qr,. The siluation ca,nbe visualized }ry a block ditr,gra,rn like Figure 12.1. TIte intetrt of this dia,grarrtis to indicate that, if M is startsl irr state {s with input rit14tr, it will eventually ha,lt in state q, or 4r. As reqrrired by Definitiotr 12.1, we want ff to operate according to thc following rules: Qowtwuisngyn2, if M applied to ur halts, and qDwMu.tf n lltQ,,,!!2, if M appliecl to tl does mrt halt. Next, wermodifu fI to produce a,T\rring rnitt:hirreiI' with the strrtctrrrtl showrr irr Figure 12.2. With thc a,ddeclstates itr Figure 12.2 wc warrt to colrvey that the trtt,nsitiorrs tretween state qo atrd tlrtl rrtlw states q* and q6 are to be rnade, regardlcssof the tape syntbol, in srrt:hrl way that lhe tape remains unchanged. Thc wtw this is done is stra,iglrtforward. Comparing iI and II' we see tha,t, irr situations where fI retit(lhtl${rr alrd iralts, the rnodified machine fll will t:rrter att infinite loop. Formirllv, the action of 11' is descrihed by qilt)Nt'tu irr,*, if M applied io 'u,halts, and qowM u) i rt,,U't(1,,.'!!2, if M applied to ru dot:s rrot lrrrJt.

Figure 12.1


12.1 SoMEPRoBLEMS Tg.q,rCaruruorBp Solvno ny TuRrrrrcMacnrNns

803

Figure 12.2

FYom f/'we construct arrother Ttrring machine f/, This new rnachine takes as input ur14 and copies it, ending in its^irritial state qs. After that, it behaves exactly like I/'. Then the ar:tion of fI is such that qow*t l ii Q11wx4u)14 Ff, *, if JI,/ applied to w71halts, and qowm ifi qouMlilM f p Afl,,Az, if M applild to wp1 does not halt. Now fI is a Ttrring machine, so it has a description in {0, 1}*, say, ri. Tlfs string, in additiorr to being the description of ,[y',also can be used as input string. We can therefore legitimately ask what would happen if ,F is applied to fi. Flom the ir,bove,identifyirrg M with Il, we get * gsfi l-6 oo, if fr applied to fi halts, and

qofri n AtenUz, if fr appliedto fi doesnot halt. This is clearlynorsense.The contradiction tells us that our assumption of the existence of 11, and hence the assumption of the decidability of the halting problem, must be false. I

One rnay ohject to Definition 12.1.,since we required that, to solve the halting problem, "I1 had to start and end in very specific configurations. It is, however, not hard to see tlnt these somewhat arbitrarily chosen conditions play only a minor role in the argument, and that essentially the same reasoning could be used with a"rryother starting and ending configurations. We have tied the problem to a specific definition for the sake of the discussion, but this does not affect the conclusion. It is important to keep irr mind what Theorem 12.l says. It does not preclude solving the haltirrg problem for specific cases;often we can tell by an analysis of M and ur whether or rrot the T[rring machine will halt. What


304

Chopter | 2 LIvIrs oF AlcloRl'rHr,tIcCotueurrrlom

the therlrem says is thrrt this caunot always be done; thtlre is no algorithrtr thrr,t catr make a correct decision for all toy and ur. The argumerrts for proving Theorem 12.1 vrcre given becarrsc they are classical anrl of historical interest. The corx:lusion of the theorem is actua,lly inrplied irr previous results as the following argument shows' If the haltirrg problem werq tlecidable, then (lvety recursively crrurnerable language would be recursivc. Consequentlv, the halting protrlcrn is undecidable. Proof: To seethis, Iet .L be a recursivelv errurlerable languageon E, and let M be a Ttrring mac:hine that accepts .L. Let 11 be the T\rring machine that 'We cotrstruct from this the following proceclure; solves the halting problem. ','no," thett by definition rrr is not in .L. 1. Apply H to w14tt.If ,F1says 2. If H says "ycs," then apply M to u. But M rnust halt, so it will eventually tttll us whether tu is irr .L or not' This constitrrtcs a trtembership algorithm, making -L recursive. But we: already know that there are recrrrsively enumerable lattguages that Arc rtot recursive. The contradiction implies that I1 cannot exist, that is, thrrt the halting problem is undecidable. t

The sitrplicity with whit:h the halting problerm can be obtaincd from Theorem 11.5 is a consequenceof the fact that the halting problettt and the memtrtlrship problem ftrr recursively enumerable languages artl rrearly identical. The only different:c is that i1 the haltirrg problem we do trot distinguish between haltirrg in a final and rrotrfinal state, whereas in the membership problcrrr we do. The proofs of Theorems 11.5 (via Theorem 11.3) and J.2.1a,roclosely related, both being a version of diagonalization.

Problemlo Another Reducing One Undecidqble The above argument, t:ttrrrrecting the halting problem to thc ntembership probletn, illustrates the very important teN:lrrriqueof reductiorr' We say that a, problem A is reduced to a problem B if the decidability of A follows from the decidabilitv of B. The1, if we ktrow that A is urxlecidable, we carr conclude that B is also undecidable. Let us do a few exa,mplesto illustrattr this idea. The state-entry problern is as follows. Giverr atry Ttrring machine M : (8, X, f , d, 40,[, F) and arry q â‚Ź Q, w â‚Ź E+, <lecidewhether or not the state q is ever entered wherr M is applied to tr. This problem is urrdecidable.


12.l

SoME PRoBLEtMsTHnr Calrlror

Bn Solvnt

ey Tunrlrc M.q,crrrlus

305

Trr reduce the halting problem to the state-entry problem, suppose that we have an algorithm A thrrt solvesthe state-entry problem. We could then uscrit to solve the halting prohlem. For exanlple, given a,ny M and ,u, we first modify M to get M irr urrcha way that M halts irr state q if and only if M halts. We can do this simply by looking at tlxl transition function d of M. If M halts, it cloes$o beca,usesome d(qi,rr) is rmrlefined. To get fr, we chrr,ngeevery such undefined d to , ) : ( q , a ,R ) , . l ( q na where q is a Iinal state. We apply the state-entry algorithn eto ('rt,1,u,). If A rr,nswers yes, that is, the starteg is entered, then (M,.rrr) ha.lts. If A says no, tlrcrr (M,*) does not halt. Tltus, the assumption that tlrt'sta,te-entry problem is decidable gives us an algorithm fbr the halting protrlcur. Because the halting problem is undecidable, the state-entry problcrn must also be undecidable.

T

Bxfiilnpls t fi.il

The blank-tape halting problem is anotlrcr problem to which the halting problem catr be reducerl. Given a Ttrring urachirrc M, determine whether or not M halts if started with a blank tape. This is undecidable. To show how this rtxhrr:tion is accomplished, assumc tha,t we a,regiven sornt,' M and some tu, We first rxtnstruct from M a new rnirr:hine Mthat starts with a blank tape, writcs ur on it, then positions itsclf in ir, configuratiorr qoil). After that, M. acts like M. Clearly M- will halt on a blank tape if arrd only if M halts on ru. Supposenow that the blank-tapo halting problem were decidablt:. Given any (M, ru), we first construct Mr,,, thon apply the blank-tape haltirrg problern algorithm to it. The conclusiorr tr:lls us whether M applied to tl will halt. Since this can be done for any M and ur, an algorithrn for the blarrktape haltirrg problem can be converted into an algorithm for the haltirrg problem. Sirrcc the latter is known to be urrdercidable,the same must be true for the blarrk-tape halting problem. I The construr:tion in the argurnents of thesc two exelmples illustrates an approach cornurorrirr cstnblishing undecidability results. A block diagram often helps us visualize the process. The constnrc:tion in Example 12.2 is summarized in Figure 12.11.In that diagrarn, we first us(] an rrlgorithm that transforms (M,w) irtto M,,,; srrch an algorithm clearly exists. Next, we use the algorithm for solving the blank-tape halting problem, which wrr a*.JrJume exists. Putting the two together yields an algorithm for the haltirrg problem. But this is impossible, and we can rruncludethat A cannot exist.


306

Chopter l?

I,rrurn'sou AlcoRlrHMI(t CoMPUTATIoN

Figure 12,5 Algorithrn ftrr halting problerri.

A decision problem is cffectively a' fitnctiorr with a ra'nge {0, 1}, that i$, ir true or false ilnlJwcr. We cau loclk ir,lso at rnore genertr,lfurtctions to to do so, we f'olltlw the established trcthocl and stx: if they are comprrter,blc; rerhrr:cthe halting problcrrr (or any other prtlblern knowu to be rrntlecidable) to the problern of computitrg t,he fr,rnctitlrrin question. Becatrscof Tbritrg's thesis, we expect that functions t:rrcoutttereditr prar:titlal circurnstanceswill he crtrnputable, so for exermplcs of uucotnprrta,llltl funclions we must Iook a little firrther. Most exa,rnllltls of uncomputablc ftrtrctions are a,sstlciated with tr,ttcrnpts to preclict thtr bchavior of T[rring rrrirr:irines.

Let f : {0, 1, n}. Considor t}rc function / (n,) whose value is the mtrxirrrutn rrumber of moves thilt catr be made by any z-state Ttrring milchirre that halts when started with a blattk tapc. This function, as it turns out, is not cxlrnputable. Before we set out to dt:rnonstrate this, Iet trs rrtake sure that .f (n) is dclirred for all rz, Noticg first that there a,reorrlY a finite number of T\rring machines with n statcs. This is because Q arxl 1' are finite' so t)-iras a finite domain a,nd ritrrge. I'his in turn implics that there are only a linite rrurnber of different dts and Lherefore a firrite rrurnber of difl'ercrnt rl-state T'uring machinrrs. Of a,Il of the n-state machintls, there are some that always halt, for exarrrple rnachines that ha,vrlorrly final sta,tesand therefore tnake no mov()lt. Some of the z-state tnachineswill rxrt halt when started with a blank tape, but they do not enter the definition of /. Every ma,chinethat does halt will execrrttl a certaiu number of rntlvesl of these, we ttr,ktrthe Iargest to give

I (n)-

'l\rring Tirkc irny ma,chine M and positive numbcr ttl. It is easy ttr modify M to produce M in sttch a way that the la,ttcr will always ha,lt with one of two answers: M a,pplietl to a blank ta'pe htrlts irr rrd more than rzr nrovcs, or M applied to ir, hlarft tape makes mtlrtl tltatt ?Iumoves. All we hirvc to do for this is to hirvr: M count its moves arrd termiuate when this r:outtt exceeds rn. Assrurlc rlow ihat / (tz) is <xlrnputable by some T\rrirtg


12.1 sorvrsPRonlnvs T'rrn*rcaNNor Bp solvnn nv Tun,lxc.rMAcHlNns

sOz

Figure 12.4 Algorithrn Iirr blank-tape halting problern. Mdoes not halt in rn steps

rrtirr:hineF. We carr then put fr ancl F together as shovrrrin Figure 12.4. l-irst we compute /(lQl), where Q is thc ster,teset of M. Thiu tells us the mir,ximum number of moves that ,\4 r:arr ma,ke if it is to halt. The value wc get is then usecl as rrr ttl r:onstruct fr as outlinecl, a,ncla clescription of -ffi is givt:n to a,universal T\rring ma,chinefor extx:rrtiorr,This tells us whether M applied to a,trla,nktape halts or rloesnot halt in lcss tha,n / (lQl) steps. If we find thal ,4.fapplied to a blank tapc makes rnore tharr /(lQl) rnoves,then trtx:tr,useof the definitiorr of /, the irnplicirtion is that M nevc:r ha,lts. Thus we have a solutiorr to the blank tape hrrlting problem. r'he irnpossibility of tlrc conclusion forces rrs to accept that ./ iu not cornputable. I

I'

,

Dcscribe in detail how H in Theorern I2.1 ca,n bc modified to prodrrce Ii' Suppose we r:hangeDefinition 12.1 to require that qo,rrrnr,u I q.o,,t, o, qow*tw i g,tu, depertding on whether M applied to rl ha,ltsor not. R.eexarnirrethe proof of Thcoretr 12.1 to show that this diflererx:e in the clcfinit,iot does rrot affect the proof in any significarrt way.

3 . Show that the following problerrr is undecida,blc, Given any T\rring ma,chine M, a E f, and.rrr â‚Ź X+, detcrrnine whether or not thc symbol a is ever writtcn when M is appliecl to 2.,. W 4 . Irr the genera,lhalting problern, we ask for an algorithrn that gives thc corlect answer lbr arry M and u, We can relax this gencrality, for exarnple, by looking for an algorithm that works for all ,41 }lrt only a single ,il. We say that such a problern is decidable iffor every u there exists a (possibly different) algorithnr that detcrmines whether or not (M, tu) halts. Show that even in this restrir:ted setting the problem is unrler:idable.


308

Chopter l2

LItr,tIrson Al.GoH,rrttr,tIcCoururun'tol,l

show that there is no algorithrn to decide whether or not an arbitrary Tirring machine halts on all input. 6 . Corrsider the questiorr: "Does a Turing trarhine iu the course of a computation revisit the starting cell (i,e, the t:ell under the read-write head at the beginning of the cornputation)?" Is this a decidable questiorr'i D.

7 . Show that there is no algorithrrr for dcciding if arry two Ttrring machines M1 arrtl Mr accept the sarne language. ffi L

How is the conclusion of Exercise 7 affected if Ml is a finitc automaton?

9 . Is the halting problcm solvable for deterministic pushdown automata; that is, givetr a pda as in l)efinition 7.2, can we always predict whether or not the automatorr will halt on input Trr?

1 0 . Let M be any T\rring tnachinc and r arxl gr two possible instantaneous descriptions of it. Show thatr thc problern of determining whether or rrot

niu u is undecidable.

ffi

1 1 . In Example 12.3, give the values of / (1) and / (2)' 1 2 . Slxrw that the problem of tletermining whether a T\rring machine halts orr arry input is undecidable.

1 3 . Lct B be the set of all Ttrring machines that halt when started with a blank tape. Show that this set is recutsively cnumera,ble' brrt not recursive. 14.

ffi

Considcr the set of all tl-statc Trrring machirtes with tape alphabet | : {0, 1,D}. Give an expression for rn(rr,), t}re nutnber of distinct Tbring machines with this f,

1 5 . Lct I' : {0, I, !} and let b (n) be the maxirrrurn nurnber of tape cells exatnined by any rz-state Ttrring rnachine that halts when started with a blank tape. Show that b (rr,) is not t:onrputable, 1ti. Deterrnirre whethcr or not the following statement is true: Arry problem whose dornain is finite is decidahle. ffi

ffi

U n d e c i d o b l eP r o b l e m sf o r R e c u r s i v e l y E n u m e r o b l[eo n g u o g e s

We have cletermined thrr.t there is no membership algorithm for recrrrsivtrly tlrrurnerable languages. Thc lack of an algorithm to decide on some propertv is not an exceptionirl state of affairs for recrrrtrivcly enumerable lan;lrages, but rather is the gcrreral rule, As we now show, there is little we caII say about these la.rrguages. Recursively erlrrnerable languages are so general that, in e$$errce,any question we arrk about them is undecidablet. Itrvariably,


12. 2 UNDECTDABLTI PRonlnus

FoR R.ECUR$IvRr,yEr*rurrrnRasr,ELaNcuncns

309

whetr we ask a question about recursively enurrerirble languages, we find that there is sorneway of reducing the halting problern to this question. We give here sonre examples to show how this is done and frorrr thcso examllles derive a,nindication of tlre seneral situatiorr.

Let G bc an unrestricted grailrmar. whethcr or not

Then the problem of determinirrg

L(G\: O is undecidable. Proof: We will reduce the menrbership problem for recursivcly enumerable languages to this problem. Suppose we are given a Thring rrra(:hineM and some strirrg zr. We can modify M as folkrws. M first savesits inprrt on some special part of its tape, Then, whcrrerverit enters a final sta,te,it checksits saved input and accepts it if and only if it is ur. We carr do this by changing d in a simple way, creating for r:irr:hur a machine M,,, srrch tha,t L(M",):

L(M) n {zr}.

Usirrg Theorern 11.7, we tlrcn construct a corrcsponding grammar G-. Clearly, thc rxlnstruction leading frorn M and tl to G. carr alwtr,ysbe done. Eqr.rallyclear is that.L(G.,) is nonernpty if and only If w e L(M). Assurne now tha,t there exists an algorithm A for deciding wlrt:tlrcr or not tr(G) : E.If we let 7 denote an algorithm by which we gerrcrateG,u, then we carr prrt ? and A together as shown in Figure 12.5. Figure 12.5 is a T\rring rnac:hine whiclr for any ,44 arrd ,rr.rtells us whether or rrot rt is in L(M). If sur:h ir T\rring machine existed, we would have a mernbership algorithm for any recursivcly cmrmera,blelanguage, irr dirs:t contradiction to a prcviously esta,blishedresult. Wu conclude therefore that the sta,ted problem "L(G):0" i s n o t d e c i d a , b l e .I

Figure 12.5 Mernbership algorithm.

L(G-) not empty ueL(M)

eL(M)


310

Chopter I2

Coututnrlot't LIuIrs oF Al.colt,I'r'ulr,tIc

Lct M be any Ttring machirre. Then the qrrestion of whether or not L (M) is finite is undecidable. From M we construct arrProof: Consider the Blting problem (M,*). First, the halting states M that does the following. T\rritrg machine other input is accepted hy Mif one is reached, all M a,rc clnttged so that any of This can be done by having any halting configurtrtiorr go to a final state. Sccorrrl, the original mat:hirre is modified so that the new machine ff first generates tr on its tirpe, thetr performs the sarrrecomputations as M, using the newly crcer,tecluratrd some otherwise urrused space. In other words, the moves rrrade by M after it has written u on its tape are the salne as would have been tnade by M had it started in the origirral configuration gnrr. If M halts in any configrrrrrtion, thgp fr will tralt in a final state. Therefore, if (M, ur) halts, M will rea,chir. final state for all input. If (M,u) tloes not halt, then frwiil not halt either a,nclso will accept nothing' In other words, M either accepts the infinite language X* or the finite Ianguage 0. If we trow assulne the existence of an algorithm A that tells us whether or not , (fr) is finitc, we can construct a solutiorr to the halting problcm as shown in Figure 12.6, Therefore no algoritlrrn for deciding whether or not tr (M) is finite can exist. I IE Notice that irr the proof of Theorem 12.4, the specific nature of the question a,sked,narnely "Is L (M) finite'i", is imrrraterial. We can change the nature of the problem without signifitnntly affecting the argrrrnent.

Show that for an arbitrrrry Thritrg machiue M with X : {4, b}, the problem .L(M) contains two different strings of the samc length" is undecidable. To show this, ye use exactly the sarnc approach as in Theorerrr 12.4, except that whert M reaches a halting configuration, it will be modilied to accept the two strings a and b. For this, tht: irritial input is saved and at the

Figure 12.6


12.2 UwpBcTDABLE PRonlu,rs non R.ncuRsrvr:uyENulrnRanl,pLer,rcuAces

Bl1

errtl of the cornputatiorr compa,redwith a arxl h, a,cceptingonly these two strirrgs. Thus, if (M,w) hirlts, M will accept two strings of equal lcngth, otherwise fr wiil accept rtothing. The rest of thc ir.rgument then procxrercls as irr Theorem 12.4.

I In exactly the sarrrc ilIirnner, we can substitute other questiorrs slch as "Does -L (M) contairr tr.rrystring of length five?" or "Is -L (,4,f) regular?" without affecting the argurnent essentially. Thcse questions, as well ar-s similar tluestions, are all undecirlrlble. A general resrilt fbrma,lizing this is known as Rice's theorem. This tlreorem states that any nontrivial property of a recursivcly enumerable latrguir,geis undecidable. The adjective ((nontrivial" refers to a property possessedlly sorrre but not all recursively enumeratlle Ianguages.A precise statetnent arrd a,proof of Rice's theorem can be fourrd in Hopcroft arrd llllman (1979).

1 . Show in detail how the rrrachine .[f in Th"o."rrr I2.4 is constructed.

,

Show that the two problems mentionetl at the end of the preceding section, namely (a) .L (M) contains any string of length five, (b) .L (M) is regular, are undecidable. LeL Mt and Mz be arbit,ra,r'y T\rrirrg machines. " L (Mt) C L (Mz)" is rrndecidable. ffi

Show that the problem

4 . Let G be arry rrnrestricted grarnmar. Does there exist an algorithm for determining whether or rrot .L (G)E is recursively enumerable? 5 . Let G be any unrestricted grarnmar. Does there exist an algorithrn fbr determining whether or rrot .L (G) : f (G)R?

6 . Let Gr be any urrrestricted gramrnar, and Gr any regular grammar.

Slrow

that the problcrn .L(Gr)n L(Gz): is urxlecidable.

a

ffi

Show that the question in Exercisc 6 is undecidable for any fixed G2, as long a^rtr (G2) is not empty.

L

lbr an unrestrictecl grammar G, show that the question "Is tr (G) : L (G)*?" is undecidablc, Argue (a) frorn Rice's Theorem and (b) from first principles.

ffi


312

Chopter l2

LItr,tIrs oF ALcoRI'I'UMIC CoMPUTATIoN

Problem fiWilffiThe PostCorrespondence The undecidability of the halting problem has many (toruequencesof prirctica,l interelst, particularly in tlrc area of context-fitNl languages. But irr rrratry instanrxrs it is cumbersomrr ttl work with the hrr,ltirrg problem dircctly, and it is corrvcrrient to establish $ome intermediate rcsults that bridge the gap hetwecrr the halting probkrrn and other prohltrrns. These intermcdiate results fbllow from the rrndr:r:idability of the lrtlltirtg problem, brrt are more closcly related to the protrlerrrs we wa,nt to study atrd therefbre make the argurrrerrtseasier. One srrchintermediate result is the Post correspondence problem. The Post corrcspottdence prohlem t:itrt be stated as follows. Given two sequence$of z stritrgs on som{} alp}rabet E, say A:

u)Lt,LDzr,,,,Il)n

Ld B :

UtrI)2r...rI)r1

we say thnt there exists a, Post correspondencesohrtion (PC-solution) for pair (A, B) if there is a nonelmpty sequenceof integers i,i,...,h, such tltat U ) i ' t . t . t ". i' u ' k

:

UiUi'''

1.)1r.

The Post correspondence probk:rn is to devise an algoritlun that will tell us, for any (A,B), whether or not there exists a,PC-solutiotr'

Eifrrlplq t{;S

{0,1} and take A and B as

Let x:

r u r : 1 1 , ? r 2: 1 0 0 r u t r: 1 1 1 ur:111,ue:001,?s:1I For this case, there exists a PC-solution as Figure 12.7 sltows. If we take

?01 Ul

u2

ul

Figure 12.7 1

I ,1

0

= 00,rlz : 001,urs: 1000 :0,u2:11,us:011

u.

0 v2

I "J


I2.ll THr';Posr CoRRnspoNDENCD Pn,oeLEM

313

there cannol, be arry PC-sohrtion simply becausc a"ny string composed of tlltlments of ,4 will be lorrger thnn the corresponding strirrg frrlrrr B. I In sptx:ific insta,nceswe rray be abk: to show by explicit constructiorr that a llair (4, B) permits a PC-solutiorr, ()r we may be able to argue, as we did abt)vc, that no such solution can cxist. Brrt in general, there is tro algoritlun ftrr dor:iding this qr.restionunder irll r:ircumstances. The Post corresponderrceprrlllklm is therefore undecidable. To show this is a sourcwhat lengthy process. For tlrt: sa,keof clarity, we break it into two parts. Itr tlrc first pa,rt, we introduce the modified Post correspondence problem. Wc ua.ythat the pair (A, B) has a modilied Post correspondencesolutiott (MPC-solution) if there exists a sequrrnr:eof intrlgursi,.i, ...,fr, such that ' W 1 ' W i ' U i" " t t t k = U 1 1 ) i 1 ) i ' ' ' U 1 x ,

In the modified ltost corrt:sllorrdence problern, the first elernt:rrts of the se-

ilirr,f,:fi ",f.xTIffi ilTl-il'll;",tJ*Ti.:K'J#liJJT

an MPC-sohrtion, then there is also a ltC-sohrtion, but the converseis not true, The rrrotliliod Post correspondence problern is to devise an a,lgorithm for deciding if arr artritrrlrv pair (4, R) admits an MPC-solution. This problem is a,lsor.rndecidable. Wcrwill demonstra,te the undecidability of thc modified Post correspotrdence problerrn by reducing a known undecidablc llrotrlem, the membership problem for rrlcursively enumerable languages, to it. Ttr this end, we iutroduce the followirrg r:onstruction. Supposewe are giverr arr unlestrictcrl grammar G: (V,'l',5,P) arrrl a targr:t string ur. With these, we create thc pa,ir (4, B) as shown in Figure 12.8. In Figurc 12.8, the string .F,9 =+ is to be taken as'u1 irrrrl tIrc string F as u1, The order of the rest rlf the strings is immaterial. We want to claim eventually that rp e L (G) if and only if the sets A artd B txinstructed in this way have an MPC-sohrtion. Since this is perhaps trot irnrntrditr,telyobvious, Iet us illustrate it with rr simpkt example.

Exom$lf I f,,d

Let G : ({A, B, Cl ,Ia,b, c,} ,5, P) with productiorrs

S - aABblBbb, Bb-C, AC --+ uat:,


314

Chopter l2

CoMPUTATIoN

B

d

Figure 12.8

LIIrllrs ol AlcoRtrHMIc

F.C --\

Fis a symbol not in FU 7'

foreveryaET

a

v

tl

E

JroE

ti

for everyVreV

E is a symbol not in I/ U ?

for every *, *1, in P

x,

---}

D

and take ut : aaac. The sequencesA a.rrdB obtained from the suggested is in.L(G) and con$tructiolr are given in Figure 12.9. The string ru : a,a,(Lc lus a derivation S+aABb+aAC+a,e,e,c. How this derivation is pa.ralleled by an MPC-solution with the corrstructed sets can be seen in Figure 12.10,wltere the first two steps in the derivation are shown. The integers above and below the derivation string show the indices for z.' and u, respectively, used to create the string.

Figurc 12.9

T

u,

I

FS+

F

a

tl

tj

b

2 3 + 5 6 7 8 9 10 11 72 13 l4

vi

t

t

A B

A B C

E d'{Bb Bhh U

-j

c ,s

ailt

s s s

Bb AC

addcb


12.3 THE Posr Con,n,pspoNt)ENcE Pnoer.nlr

wl

Figure 12.10 f

il!

%to

s

a

B

d

,7

*10

s ?10

d

n

vz

il.

1J), ^

B

b

u1o

F

s =

a

vI

"10

,2

il14

u2 tus u12

=

4

rt

ur4

*z

-s

C

il'rz

-1

Figure 12.11

b

?10

ur F

S15

i

b

D

l

v12

+ ilt^

@r2 ur4

C

a

v-

'13

u2 a

u.a

il

t

E

vg

Exarnine Figurc 12.10 r:a,refully to see wlnt is htr,ppening. We want to construct an MPC-solution, so we must sta.rt with tu1, that is, f.S +. This string contains 5, so to rnatch it we have to use u1s or u11. Irr this instance, we u$e tr10;this brings irr'rr.'16, lending us to the secondstring in the partial derivatiorr. Looking at several rnore $teprJlwe see that the string u.,rur;ur1.., is always longer than the corresponding string a1uiaj...t and that the first is exactly one step ahead in the derivation. The only exception is the last step, where uo rnlmt he applied to let the u-string catch up. The complete MPC-solutiorr is shown in Figure 12.11. Tho c:onstruction, together witlr the example, indicate tho lines along which the rrext rcsult is established. I

Let G : (V,T,S,P) be any unrestricted gra,mmar, with u any string irr T+ . Let (A, B) be the correspondencepair c:t)rr$tructeclfiom G and trr be the process exhibited in Figure 12.8. Therr thc pir.ir ('{, B) permits an MPC-solution if arrd only if w e L (G). Proof: The proof involves a forrrrir,linductive argument based orr the outlined reirsonine. We will omit the details. I

Witlt this result, we ca,rtreduce the rrretrbership problem for recursively enurnerabkl lrr,ngua,gesto the rnodified Post correspondence problem and thereby dernonstrrrtethe undecidability of the lattcr.


s16

Figure 12.12 Membershil.r algorithm.

Chopter l2

LItr.nrs on Al(loRIrttNatc CotrrlpurerloN

ueL(G)

Construct,4 and B as in Figure 12.8.

vte L(G)

The modified Post correspondenceprohlem is undecidable. Proof: Given arry urrrestricted gra,mmar G: (y,T,S,P) and ur â‚Ź T+, wt: construct the sets A and B as suggestctlabove. By Theorem 12.5, thc lrair (A,B) Iras an MPC-solution if and only if w e L(G). Suppclsc'lnow we assume that the modifiecl Post correspondencleprobItlrr is decidable. We can then cttnstruct atr algorithm for the merrrbership problern of G as sketched in Figure 12.12. An algorithm fbr constrttctitrg A front B from G and tu clearly exists, but a membership algorithttr for G and u.' does not. We must therefore c:clrrcludethat there ca,nnot be arry algorithm for deciding thc rrrodified Post correspondorrt:ttprobletn' r

With this prelimina,ry work, wc are rlow ready to prove the Ptlst corresporrdenceproblem in its originirl ftrrm.

The Post corresponclenr:cprotrlern is undecidable. Proof; We argttc that if the Post corresponden(icl)roblem were decidable, the modified Post correspotrdence problcm woulcl be decidable. A : 'u)71Lu2r.,., tu* and B : 1,t,\)i1...1'ttn. Srrpposewe are given sequence$ on lrorrrealphabet X. We then introdu(le rrew symbols g and $ and thtl rrtlw sequeilces C:Ao,Aft-.-tUn*tt D :

z o t Z t t , , , tZ n l t ,

defincd as follows, For i : 1,2,...,n, li :

wilQu)i2Q'''

zi :

. . Lti,;, l,ai1Qui2Q.

LDi,n,i$,

where ru4i irrrtl u4i denote the jth letter of u); arrd ?i, respectively, a,nrl rni : lu.il, r'1: lual. In words, gri is creir,ttxlfrorn trri by appending I to cac:h


12.3 THE Posr ConnnspoNDEN(iEPRosr,urra

Figure 12.13 MPC algorithrn.

317

MPC-solution No MPC-solution

charat:ter, while aa is obtainecl by prefixing each chirrnr:tr,'rof u; with f . To cornplete the deflnition of d and D, we tako A 0: g Y r ) 3/n,+I: $r f() *

3lr

2",+1* #li. Consider rrow the pnir ((J, D), and suppose it has a PC-solution. Bcr:arr$el of the placernent of q arrd $, such a solution must have g/0orl Lhe left rurd Un1l olr tlrtl riglrt and stt must look like ' 'Qwtt(, , 'Qw;1' . Q1.1111Qw12'

, q \ : g z t 1 1 t l u 1.2. L . u . i . r | .. . ( . u t ; t. . . Q E .

Ignoring tlte charactcrs r/ arrd li, we see tha,t this inrplies ' W 1 U l . j ' ' " l i l 1 r ':

U1'11.i''' Ulrj

so that thc pair (,4, B) pcrmits a,n N{PC-solution. We can trrrrr tht: ilrgrrrnerrt rlrclund to show that if there is an MI'Csolution for (4, B) therr tlierc is a PC-solrrtion for the pair (C, D). Assurne now that tlrrl Post r:orrespondenceproblem is deciclable. We r:an then construct the rnachineshown irr Figunl 12.13. This mar:hineclea,rJv tltx:idesthe modified Post correspotrcletrce problcrn. Brrt tlrc rnorlifiexlPost cortesprlrrrlcrrtxrprohlern is undecidable, consequently, we camrot havel arr algorithtrt for decidirig thc Post rxlrrosp{nclenceproblerrr. t

1

,

L e t A : { 0 0 1 , 0 0 11 , I 1 , 1 0 1} a n d I J : { 0 1 , 1 1 1 ,1 1 1 , 0 1 0 } . Does the pair (4, B) irave a PC-sohrtion? Does it havc an MPC-solution'l ffi Provitle the rletails of the nroof of 'l'heorcm l.2,5,


3L8

Chopier l2 Lrvrrs or, AlcoRrrHurc Conrur.qrtox

3. Show that for lll : 1, the Post correspondeuce problem is decidable, that is, there is an algorithm that can decide whcthcr or not (4, B) ha-sa PC-solution ftrr arry given (,4,-B) on a singlc-lctter alphabet. ffi 4, Suppose we restrict the domain of the Post correspondenceproblem to include only alphabcts with cxactly two symbols. Is the resulting correspondcncc problem decidable? 5, Show that the ftrlkrwing rnodilications of the Post correspondence protrlerrr a,re undecidable. (a) There is an MPC-solution if there is a scqucnce of integers srrch that tuatrri"'utNwT = uiuj "'utrur. ffi (b) There is an MPC-solution if there is a $equenceof irrtegers such that ultrr2zriurj "' u)k :'Dt'Dz'tti'uj'' "D hr. 6. The corrcspondence pair (4, B) is sairJ to have arr euen PC-solut'ion if and only if there exists a nonempty seqlren(:eof everr integers 'i, j,...k such that wnwj , , ,wk : Itn'u j. . . 'r.'5.Show that the problem of deciding whether or not an arbitrary pair (,4, B) has an even PC-solut,ion is undecidable.

ffiilffiffiUndecidobleProblemsfor Context-Free [onguoges The Post correspondence problem is tr convenittrrt tool for studying undecidable qrrestions f'rrr txrntcxt-f'ree larrguages. We illustrate this with a few seler:tedre$rrlts. There exists no algorithm ftrr decidirtg wltetlter any given context-free grammar is amhiguous. (wt,1J)2,...,t1r)tr,ndB : Proof: Consider two sequencesof strings A: (ut,u2,...'ur) over some alphabet E. Choose a, new set of distirtct syttrbols & I r & 2 r . . .a, r , S U C h that {or,or,...,&r}flX:O, and qlnsider thtt twcl larrguages Lt1 :

{wiwi

- . - wtuth&ka4. - . aiai.l

and Lp :

luitt.i

. - . 1)luke,ka,I. . . &jilil .

Now look at the context-free gramma,r G : ( { S , S r a ,5 n } , } J U { a 1 , a 2 , . . . a n }, P , S )


12.4 UuoncrDABLEPRonlnlrs FoR,CoN'rEXr-FRueL.q,ncuncns

3L9

wltere the set of productions P is the union of the two subsets: the first set P4 corrsists of S-Sa, S.q.-

i : I,2,...,n,

wi]eatlwiai,

the second set P6 has tltc lrro<1rr<:tions S __+Sr, Sn -+ uiSnailuiat,

'i: I,2,...,n,.

Now take G a : ( { S , S a } , I \ ) { a 1 , a 2 ,. . . ,& , , 1, P e , S ) and

G n : ( { S , S n } , X U { a 1, a 2 ,. . . , a n } , P n , S ) . Then clearly Ln:L(Gn), Ln:

L (Gs),

and L(G): Ler Ln. It is easy to see that Ga arrd G6 tly themselves rlre rrnamhiguorrs. If a given string in .L (G) ends with aa, then its derivation with grammar Ga rrrust have started with 5 + u)isai. Similarly, we can tell at any later stage which rule has to be applied. Thus, if G is ambigrrourr it mrrst be because there is a ur for which there are two derivations J :* Jg )

^

wi5nai

*

4

wiwi " 'wtrak " 'eiai

:

111

and 5 =+ ^9gg + ui]Bai 4

uiui ' ' 'ukek'''aiai

:

v1.

Consequently, if G is anbiguous, then the Post correspondence problenr with the pair (4, B) has a solution. Conversely,if G is unambiguous, then the Post correspondenceproblem cannot have a solution. If there existed an algorithm for solving the ambiguity problem, we could adapt it to solve the Post correspondence problem as shown in Figure 12.14. But since there is no algorithm for the Post correspondenceproblem, we conclude that the ambiguity problem is undecidahle. I


320

Chopter l2

Coupuranron Lluns or Ar,c+oR,rrrrrurc

Figure 12.14 PC algorithm.

There exists no algorithm for decidirrg whcther or not

r ( G r )n L ( G 2 ) : n for arbitrary context-frcc grarrllllars G1 and G2. Proof: Take as G1 the gralnlnar Ga and as G2 the Hrammar Gr a*sdefined in the proof of Theorem 12.8. Srrppostrthat tr(Ga) and L(G") have a common element, that is S a 3 t r t i t u.i' , ' u t r a k ' , ' a j a . i and S 6 1 u i u i . . . u k e k" ' e j a , i . Then the pair (A, B) has a PC-solution. Converscly,if the pair does not have a PC-sohrtion, then L (Ge) and I (G6) cantrot have a common element. We conclude that ,L (Ga)n L(Gu) is nonempty if and only if (,4, B) ha-sa PCsolution. This reduction Drovesthe theorem. r

There is a variety of other known results along these lines. Somtr of them can be reduced to the Post correspondence problem, whilc otlters are more easily solved by establishing different intermediatc rc'sults first (see for example Exercises 6 and 7 at the end of this section). We will not give the arguments here, but poirrt to sorrreadditional results in the exercises. That there are many undecidable problems connected with context-fiexr languages seems surprising at first and shows that there are limitations ttt computations in an area in which we might be tempted to try an algoritlrrrric approach. For example, it would be helpful if we corrld tell if a programrnirrg language defined in BNF is ambiguous, or if two different specificatiots of a language are in fact equivalent. But the results that have beerr established tell us that this is not possible, and it would tra a waste of titne to Iook for an algorithm for either of these tasks. Keep in mirrd that this does not rule out the possibility that there may he ways of getting the answer for specific


12.4 UNDECTDABLE PnonlnrulspoR,Cor'r.rr.;XT-[,'REE LeNcuecns

321

casesor perhaps even most interesting ones. What the urrdccidirbility results tell us is that thert: is no txrmpletely general algorithm and that no matter how many dilTerent cases a rncthod r:an handle, there are invariably some situations for which it will breah tlowrr.

I.

* ?

Prove the claim made in Theorem 12.8 that Ga and G.e by therrrselves are unambiguous. Show that thc problem of detertrining

whether or not

1,(Gr)c t(Gr) is trnrler:idahlefor context-free grammars Gr, Gr. t 3. Show that, frrr arbitrary context-free grammars Gr and G2, the problem "L (Gr) n L (Gz) is context-free" is undecidable. * 4. Show that if thc languagc L(Ge)nL(Ge) in Theorerrr12.8 is regular, then it must be empty. Use this to show that the problem "I (CJ) is regula.rJ is urxler:idable for context-free G. * 5, Let -Lr be a regular larrguage arrd G a context-free grammar. Show that the problem "lt C L (G)" is undecida]-rle. * 6. Let M be any Tlrring rrrachine. We can as$lrme without loss of generality that everv computation involvcs an even number of rnoves. For any such comnutation gourF fll F 12 F ." [- ff,, we can then construct the strins qot F *fl F rc, F :r.f | ..'l

xn..

I-his is called a valid computation, Show that for every M we carr cortstruct three context-free gramrnars Gr , Gz, GB, such that (a) the set of all valid rnmprrtations is /, (Gr ) n L (Gr), (h) the rret of all invalid comprrtations (that is, the complement of the set of valitl t:ornprttatiorts) is I (Gl).

* 7 . Use the results of the above exerciseto show that "-L (G) : X*" is undecidable over the domain of all context-free grarnrnars G.

* 8 . Let Gr be a context-free grarnmar and Gz a regular gramnrar. Is the problem I ( G r )n L ( G z ) : o decidable?


322

Chopter l2

* L

Lrurrs or Ar.c;on.rrrnrrc CoNrpurerroN

Let Gr anrl G2 be grammars with Gl regular. Is the problern L(Gr) = L(Gz) decidable wherr (a) Ge is unrestricted, (b) when G: is context-free, (c) when Gz is regular?


O t h e rM o d e l s of Computqtion

Ithough T\rrirrg rnar:hinersa,re the most general models of cornputation we carr corrstnrt:t, thev a,re not tlre only ones. At variotrs times, other rnodels hirvt: llccrr proposed, some of which at flrst glance seernedto tre raclically different f'rorn Thring machines. Eventually, however, all thr: rnorlcls were fbund to be equivalent, Much of the pioneering work in this area was tlorrr:in the period hetween 1!130and 1940 and a nurnber of rnathernaticiirrrs,A. M. Ttrring among them, contribuled to it. The results that were fourrtl slrrxl light not only on the concept of a mecha,nicalcornputation, but orr rnatlxlrnaticu a-r a,whole. T\rring's work was published in 1936. No t:rlrnrn(:r(:ilrlrxlrnprrterswere availahle a,t tha,t tirne. In fact,, lhe whole idca had bccn rxrnsideredonly in a very pt:ripherir,l way. Although T\rring's ideas eventually becarnc vcry importattt irr cornputr:r sr:iunr:e,his origina,l goal was not to provicle a fotrrrda,tion for the study of digital r:ornputcrs. To understand what Thring was trying to do, we must brielly look at tlxr statc of ma,thematicsa,t that time. With the discovery of diffr:re:rrtiill irrrtl integra,l r:ir,lcrrlusbv Newton and Leibniz in the seveuteenth and eiglrteenth c:etrturic:s,irrtelrcst in ma,thematics inr:rea,sed and the discipline entered an era of explosivr:grclwth. A mrmber of

323


324

Chopter l3

Motnls on Clolrlru.rart(ltt O'r'truR.

different areas were studied, a,nrl sigrrificatrt aclvauceswere made in alrnost all of them. By the end of the nirxlteenth century, the body of mathcrnatical knowledge hacl become quite la,rgrt. Matitettnticiatrs also had becromeurrfficienlly sophisticated to recognizc'that sorne logical difficulties had a,rison that required a more ca,refulapproach. This krd to a concern with rigor in reasoning and a consequent examiniltiorr of the foundations of rna,thema,tical knowledge in the proce$$. Ttr see why this was llecessary, considtlr what is involved in tr, tvpit:a,l proof itr just about every book and ptr.ptrr dealittg with mathematical subjects. A sequenceof plausiblc cla,irnsis rrrade,inlerspersedwith phraseslike "it ca,nbe seen casilytt irrrd ttit follows from this." Such phrases a,re (pnv()rrtiorritl, irrxl what one Ineans by them is thtr,t, if challengedto do so, one rxrrrldgivt: rrtclredetailed reasoniug. Of course, this is very dangerous, since it is possiblc to overlook things, use faulty hidden assumptions, or make wrong infi:rcnt:tls. Whetrever we see arguments like this, we cannot help but wonclelrif tlrtl llroof we are given is indeed correct. Often there is no way of telling, irrrd lortg ancl involved proofs have heen published and fbr,rnclerr*one()uiJorrly after a considerable amount of time. Becarrseof pra,t:ticalliruitatiorrs, Itowever, this type of reasoning is a,txxrptcd by most rnirtherrraticians. Ihe arguments throw light on the subim:t attd at least increaseour confidencethir,t tlrc rt:sult is true. But to those demanding cornplete reliability, they are unacceptrr,ble. One alternative to such "skippv" rnathtlrlrirtitls is to formalize as fa,r as possible. We sta,rt with er,sct of assutned givetrs, called axioms, and prec.isely definecl nrlels for logical inference and cleduction. The rules tlre used in a $eqrr()n(:oof stclls, circlt of whiclt takes us from one proven fa,ct to a,nother. Thc: nrlr:s rnust bel such Lhat the correctnessof their a,pplictr,tiorr carr be checkeditr a routitre atrd completely mechanicalway. A propositiorr is irr a Iirrittl sequeltce considered proven true if we can derive it fiom the ar,xiorrrs of logical steps. If thr: propositiorr corrflicts with another proposition that t:iln b(: llroved to be true, then iL is considered false. at t]te !'indirrg such lortnal systetns was a ma,jor goa,l of mer,t]rcrrratic:s crrd of the nineteenth century, Two concerns immediately arose. Tlrrl first was that the system should be consistent. By this wc rrx:irrrtltat tirere should trot be atry proposition tha,t ca,n be proved to bc tnxl by orre sequellce of steps, then shown to be ftr.lseby irrxlthtlr ctpirlly valid argumeut, Consistency is indispcnsablc irr rnathernatics, aud atrything cleriveclfrom an inconsistent $y$torrrwoultl tltl cotrtrary to all we agree on. A secoud concern wils whctlrtlr a systern is complete, by which we mea,Irthat a,nypropositiorr cxpressible in the syst,eurcan be provecl to be true or fh,lse.Frlr sorne tirntr it was hoped that consistent and complete systomtr ftrr rrll of tnalhernatics could be devised thereby opening the door to rigorous but cotrtpletely mechanical theorem proving. Rut this hope wrs rlir,shtxl by the work of K. Godel. In his famous Incornpleteness Theorem, Gcidel showed that any interesting consistent systeiurmrrst tltl irrrxlrrrlrlcteltlnt is, it rnust cou-


1i1.1 R,ncunsrvn!-uncrroFls

3?5

tain soure unprovable propositions. Gtjdel'u rnvolutiotrary conclusion was published in 1931. Gcidel's work left rrnanswurudthc tluestiorr of whether the unprova,hle sta,tementurxnrld sornt:ltowbe distinguished from the provable oncs, so that thcrc was still sotrtehope that most of ma,thematit:stxltrld be tnade precise with mechanically verifiable proof's. It was this problettr Lhat T\rring and other mathematicians of the time, prr.rticrrlarlyA. Church, S. C. Kleene, arrtl E. Itost, addressed. In order to studv the tprcstiott, a variety of forrnir,l rrxrtlels of computation were esta,blished. Prorrrirrc:rrtirIIIoIrg them were tht: rccursive futrctious of Church and Kleene a,ntl Post systems, but there arc rnarry other such systems tha,t havc br:crr sttrdied. In tiris cha'pter we briefly review some of the idea.sthat arosc out of tltese studies. There is a wealth of materia,l here thnt wc carrrtot cover. We will give onlv il vtlry bricf presrlrrtatiorr,referritrg the reader to other ref'ercnctlsfbr detail. A qr.rite accessibleaccount of recursiverfirrrt:tiorrsarrtl Post systernsca,nbe firrrnd irr Denning, Dennis, a,nd Qualitz (1978), while a good disclnsiorr of vitrious other rewriting systems is given in Salornair(1973) and Salomaa (1985). The utodels of computation we study here, as well as others tha,t ha,vt: bccrr proposed, have diverse origins. Brrt it was evetttually found tha't thev wcrel all equivalent in their power to c:itrry ottt cornputations. ilhe spirit of this observation is genera,llyr:a,lkxlChurchts thesis, This thesis statcs that all possible models of ulrnputir.tion, if they are sufficiently hroa,tl,rrrust be eqrrivaient. It also irnplies lhal there is a,ninherent limitation itr lhis atrd thir.t there are futrctiotrs that cannot bc cxprcssed itt atry way tha,t gives iln explicit nrethod for thtlil cotrrputation. I'he clainr is tif t:tltrrsevery closely related to Thring's tk:sis, and the combined notion is sotrretines called the thesis. It provides a genertll prirrciple for algorithmic: Church-Tlrring cornputation and, whilc not provable, gives $trong e:virlt:trcethat uo more powerful models cart lle fcrutrd.

RecursiveFunctions 1gffiffifi$ffiffi Tlrc rnrrcept of a function is firndeurcrrtalto rnuch of nta,thematit:s.As surrrmilrizcd in Section 1.1, a,flnction is a rulc tlnt assigusto a,trelcrlcrrt of otre st:t, called the domain of tlx: firrrc:tiotr,a unicluevalue in a,ntltlrt:rset, called thc range of the function. This is very broad and generrnlirrrd itnmediately raisesthe questiorrof how we can explicitly reprtrst)rrttltis association. There are many wa,y$irr whidr furrctions can be definelcl. Sorrte of thetn we use frecluently, while otht:rs are less colnmolt. We are a,ll fnmiliar wilh functional rrotatiorr irr wlticlt we write expro$sions like

I U i : r z 2+ l ,


326

Chopter l3 Orrrnn, MoDEr,soF CoMpuTATroN

Tlfs dcfinrls the function / by means of a recipe for its computation: given any value for the a.rgrrment n, multiply that value by itself, and therr add one. Sirrce the function is defined irr this explicit wa,y,we can compute its valucs in a strictly mechanical fashion. To cornplcte the definition of /, we also rmrst specify its domain. If, for exarrrplc, we take the domain to be the set of all integers, then the range of / will be some subset of the set of positive integers. Since ma,ny very complicated furrctions (:an be specified this way, we rnay well ask to what extent the notatiorr is universal. If a function is dcfined (that is, we know the relation bctween the elements of its domain atrd its rartgc), can it be expressed in such a functiorral fclrm? To answer the question, we must first clarify what the permissible forms are: for this we irrtrodrrce some basic functions, together with nrles for building from thenr solne rnort) cxlmplicated ones.

PrimitiveRecursive Funclions To keep the discussiorr sirnple, we will consider only functiorrs of one or two variables, whose dornairr is t ither I, the set of all non-negative integers, or I x I, and whose range is irr 1. In this setting, we start with the ba-sic: firnctions: 1 . Tlre zero function

z (r) :0,

for all r â‚Ź 1.

2 . The successor function.s(r),

whose value is the integer ne.xt in sequence to r, that is, in the rmual notation, s (z) : z * 1.

3 . Thct projector

functions P n ( n 1 ,r z ) : r k ,

h : I,2.

There are two wa,yrrof building more complicated functions from these: 1. Composition,

hy which we construct

I @ , a ): l t ( n ( * ,a ) , s z( r , a ) ) frorn defined functions gL)gz,h. 2. Primitive through

recursion,

by which a function can be defined recursively

:';,i,i; (n,u)), ,,.',-Tll ,v),r from defined functiorrs gt, gz, and h.


13.1 Rncunsrvr Fr,rlcrloNs

327

We illustratc how this works by showirrg how the l-rasicoperatiotu of integer arithrnetic can be constructed in this fashion.

Additiorr of integers r and E can be implernented wiih the function add(r,y), defined by ttdd(n,0) : ,r,, adrl,(r,3/ + 1) : add (r,t/) + 1. To adcl 2 and 3, we apply these rules successively:

a d d ( 3 , 2 :) a d d(.3 , I ) + I

: ( a d d ( 3 , 0+) 1 )+ I :(3+1)+1 :4*1:5. I

Using the add function defined in Example 13.1, we cirrr rr()w dt:fintl rrrrrltiplication by mult(r,0):0, r n u l ' L( n , Y + 1 ) : a d d ( n , m u l t ( r , Y ) ) . Forma,lly,the uerxrndstr:p is arr applicatiott of pritnitive recursion, iu which h is itlrrntilicd with Lheatld, function, and 92(r,g) is the projector ftnction Pt\r,tl). I

Substraction is not rpite so obvious. First, we rnust deflne it, ta,king into rr,ccorrntthnt rrcgativtl nutrrbers are not perrnitted iu our system. A kind of subtra(:tiorr is dcfined from usual subtraction by : t - ! ! : . t :- g l f x : 7 A , :t-"u:0ifc<E' The operator - is sornetirrxlu cir,lltxl thc: monus; it defines subtraction so that its range is 1.


328

Chopter | 3 Ornon Monnls or, Conrpureuor-r

Now we defirre the predecessor function P r e d ( 0 ): 0 ' Ttred(Y*I):'c, and from it, the subtracting function subtr (r,0) : r, subtr (r,t/ + 1) : pred(subtr (*, y)) . To prove that 5*3 : 2, we reduce the proposition by applying the definitions a rrumber of tinres:

subtr(5,3): : -

pred,(subtr (5,2)) pr ed(ytred (subtr (5, 1)) ) pr ed(pred (pred(subtr (5,0)) ))

pred(ered(ered(5))) - pred(pred$)) : pre.d(B) _,

T In much the same way, we can define integer division, but we will leave the denronstration of it as an exerci$e. If we accept this as given, we see that the basic arithmetic operation$ are all constructible by the elementary processesdescribed. With the algebraic operations precisely defined, other rnore complicated ones can now be constructed, and very complex computations built from the simple ones. We call firnctions that can be constructed in such a manner primitive recursive.

i'r,. fil i,il$' lrr ,irfr,.t,.l A firnction is called primitive recursive if and only if it can be constructed from the ba*sicfirnction$ s? ,e,pk, by successive composition and primitive recursion.

Note that lf h, gz, and lr are total functions, then / defined by composition and primitive recursion is also a total function. It follows from this that everv primitive recursive function is a total function on I or I x 1.


13.1 RncunsrvnFUNCTToNS 3?9

The expressivepower of primitive recursive functions is considerable, and most common functions are primitive recursive. However, not all functiorrs are in this cla*ss,its the following argument shows.

Let f' denote the set of all frrnctions from I to I. function in .F that is not primitive recursive.

Then there is some

Proof: Every primitive recursive function can be described by a finite string that indicates how it is defined. Such strings can be encoded and arranged in standald order. Therefore, the set of all prirnitive recursive functions is countable. Suppose now that the set of all functions is also countable. We can then write all functions in some order, say, f t, f2,.... We next construct a function g defined as

s ( i ) : f i ( i )+ \

i : 1,2,....

Clearly, g is well defined and is therefore in F, but equally clearly, g differs from every fi in the diagonal position. This contradiction proves that F cannot be countable. Combining these two observations proves that there must be some function in f' that is not primitive recursive. I

Actr.rally, this goes even firrther; not only are there firnt:tions that are not primitive rccursive, there are in fact computable functions that are not primitive recursive.

Let C be the set of all total c:omputable furrctiorrs frorn 1 to 1. Then there is some function in C that is not primitive recursive. Proof: By the argument of the previous theorem, the set of all primitive recursive functions is countable. Let us denote the functions in this set as rr,rz,...and define a firnction .gby

sQ):rift)*-By construction, the function g differs fiom every ri arrd is tlterefore not primitive recursive. But clearly .g is r:onrputable, proving tlte theorem. r

The nonconstructivc proof tlnt there are computable functions that are not primitivc recursive is a fairly simple exercise in diagonalization. The actual construction of an example of such a function is a much more complicated matter. We will give here one example that looks quite simple; however, the demonstration that it is not primitive recursive iu quite lengthy.


330

Chopterl3 OrHnn Mounls or.'CoNtrurnrtotl

Ackermqnn'sFunction Ackermann's function is a function from 1 x 1 to 1. defined bv

A(o's):3/+1, ,4 (4a,0) : A(:r - 1, 1) , A ( u , t 1 * 1 ) : A ( r - 1 ,A ( t : , y ) ) . It is not hard to seethat ,4 is a total, computable function. In fact, it is quite elementary to write a recursive computer program for its computation. But in spite of its apparent simplicity, Ackermann's function is not primitive recursive, Of course, we cannot argrre directly from the definition of A. Even though this definition is not in the form reqrrired for a primitive recursive function, it is possible that an appropriate alternative definition could exist. The situation here is similar to the one we encountered when we tried to prove that a language wa$ not regular or not context-fiee. We need to appeal to some general property of the class of all primitive recursive functions and show that Ackermann's function violates this property. For primitive recuntive functions, one such property is the growth rate. There is a limit to how fast a primitive recursive function can grow as i --+ oo, and Ackermann's function violates this limit. That Ackerma,nn's function grows very rapidly is easily demonstrated; see,fbr example, ExercisesI to 11 at the end of this section. How this is related to the limit of growth for primitive recursive functions is made precise in the following theorem. Its proof, which is tedious an<l technical, will be omitted.

Let / be any primitive recursive function. Then there exists some integer n such that /(i) < A(n,i,), fbralli:n.n+1..... Proof: For the details of the argument, see Denning, Dennis, and Qualitz ( 1 9 7 8 ,p . 5 3 4 ) . r E If we accept this result, it follows easily that Ackermann's function is not primitive recursive.

Ackermannts function is not nrimitive recursive. Proof:

Consider the function

s ( i , ): A ( i , i ) .


13.1 R.ncun.srvp Fur-rcrronrs 331

If A were prirnitive recursive,then so would g. But then, accordingto Theorern13.3,there existsan n suchthat

s(i)< A(n,,i,), ftrr all rl. If we now pick i : rr', wt] get the contradiction S(n): A(n,n) < A(n,n,), proving that A cannot be primitive recursive, I

Recursive Functions 11,To extettd thc idea of recrrrsive functions to covcr Ar:kermann's function and other computable furrctions, we must add somethirrg to thc rulcs by which suc:hfincltions can be constructed. Orrclway is to introduce the p or minimalization operator, defined by pA (g @,?)) :

snrallcst 37such that g (r, y) : O.

In this definitiorr, w(l iL$$rrme that g is a total function.

Ler llHr*sNE[U*il*,$* g(x,'a):n*u-3, which is a total functiorr. If u { J, fhsn U:3-r is the result of the minima,lization, but if r ) 3, then therc is rro y â‚Ź I srrt:h that z I y - 3: 0. Thercfrrrc, for r: { 3 P'a(g @,u)) : 3 - r, : unclcfined,for r ) 3. We see from this that everrthough g(r,ll) is a total function, pA(S@,A)) mav only be partial. I As the abovc t:xamplr: rthows, the minimalization operation opens the possibility of definirrg partial firnctions recursively. But it turns out ihat it


33?

Chopter l3

Orul:R MoDELSoF CoMpurATroN

also extends the llrlwt:r to defitre total functions so a,$to int:hrtlt: all computa,ble fi.rn<:tiorrs. Agaitt, we tnerely quote the major resrrlt witlt references to the litcrra,turtlwirere the deta,ilsmay be fortnrl.

lili.ffiRq,,frtiffift, Nhft'ilrii\ A firnction is sa.idto br: p-recrrrsiveif it can be constructed fiom thc birsis functions by a,seqrxlnr:t:of irpplicatiots of the p-operator and the opera,titlrrs of compositiotr and primitivc rct:ursiort.

if a,rrdorrly if it is cotttputalrle. A lunction is trr,-recursiver Proof: -Fbr a proof, see Denuing, Derrnis, ir,rrtlQuiilitz (1978, Chapter 13). I

I'he p-recursive finctions therrrftlrc givc rrs itrxrtlrtlr rnodel for algoriihmic computation,

t.

Usc thc definitions in lixamples 2+3:6.

,

Define the lunction

and lll.2 to rrrove that 3 + 4 -

Ill.I

g r e a t e(r* , v ) : 1 i f r > U :0ifr<U Show that this function is prirnitive recursive.ffi

3 . Ct-rnsiderl,he luttctiou (t:,'g) : I rtyt,a,l,s

:0

if r : u ,

1 I x t ''c.

Show that this function is prirnitive ret:trrsive. 4.

Let / be defined by 'f(n,tt):n :0 Show that this futctiot

ifr*u, i fr :

is ;rrirnitivc rccursivc.

U.

7 and


13,1 RncunsrvnFrrrucrlons

333

rk 5. Integer division can be defined by two functions iCzuancl lern: diu (x,'g): n. wlrere n is the largest integer such that x j rrg, anrtl rem,(t,'a):

n - nA.

Show that the functions d,i,ua:rd re.nr ate prirnitive recursivc, 6. Show that

f ( r t ): 2 " is primitive recursive. 7. Show that the furtction g (n,u) : x is primitivc recursive. W 8. Write a cornputer l)rogram for computing At:kermann's function. Ilse it to evaluate A (?,5) and A (3,:3). L

Prove the following fbr the Ackermann lunction.

( a )A ( 1 , s =) s + 2 f f i (b) A (2,a) :2u-r3 ffi ( c ) A ( 3 ,u ) : 2 ' + 3

* 3

10. Use Exercise I to compute A (4, I) and A (4,2). 11. Clive a gerrcral expressiorrfor A (4, g). L2, Sltow the sequencc of rccursive calls in the cornputation t-rfA (5,2). 13, Show that Ackernrann's functiorr is a total function in 1 x 1. 14. Tfy to usc the plograrrr t:orrstrrrctedfor Exercise B to evahrate ,4 (5,5), Cat you explain what vou observe'l 15. lbr cach g below, r:orrrputeltakl@,37)), and tletertrine its domain, (a) .s (r, :a) : :ra

(lr) .q(", '',):2"' -r u - 3 ffi (c) s (r,u) : intescrpartof (* - 1)I fu +t7 (d) g(n,g) :u

mod(s*1)

16, The definition of pre.tl in Example 13.11,althorrgh intuitivcly clcar, does rrot strictly adhere to the defirrition of a prirrritive reclrrsive function, Show how thc definitiorl t:an be rewritten so that it has the corret:t form.


334

Chopterl3

MM

OrunR Monnls or Covpurnrton

PostSystems

A Post system lotlks very much like tln utuestricted grammar' colsistiug of an alphabet rr,rxlsorne production nrltls by which successivtl strings can be derived. But there are significant difl'crettces in the way irr which the productions arc applied.

s.sf,ffi ,,, fiilitftilll\l,l$,1,, A Post systeur II is defined by

fI : (c, v,A,P) , where C is a finite set of constants, consisting of two disjoint sets (,'1,', called the nonterrninal constants, and C7', thtl st:t of terminal constants, V is a finite set of va,ritr,blt:s, A is a linite set fiom (J*, cir,lledthe axioms, P is ir"fittite set of uroductions.

The prodrrctions irr a Post system mrrst satisfy certaitr restrictions. Thcy must bc of the form f r 1 V 1 f r 2 . - . V r T , * -, y t W t U z - . - W m , ! l m - l t ,

(13.1)

where ff,;, yt â‚Ź C*, and Va, Wi EV, sutrject to the requirerncrrt tltat atry variallle can appear at most on(:c orl the left, so that Vfviforilj, and th.r,t eac:hvirriable on the right mrrst irppear ou the left, that is

U*rU"

i=t

t=l

Srrpposewe have a string of tcrrnirrals of the forrn rru)trzl:tz' ' 'rDnfrn+tt wlrere the substrings rt, Tz.'. rrrirtchthe correspondingstrittgs itt (13'1) arrd'ur1â‚Ź C*. We can then rrrirkethe identification u)1 : V1,'trz : V2,..., and substitute these va,lrresfor tlte W's ou the right of (13.1)' Since every


13.2 Posr Sysrpus

335

W is some yd that occurs on the left, it is assigned a unique value, and we get the new string lltwiUzwj . . .Antt. We write this as I1W1I2W2,,,

frnlt

+

!yUa!2Wi,

..

Arnlt.

As for a grammar, we carl now talk about the larrguage derived by a Post system.

The language genera,tedby the Post system II : (C, V, A, P) is I, (II) :

{w

e Ci.: ,u.rs 4 u for some zuoe ,a} .

irrt C7: {a,b}, Cu =4, V : {Vr}, A : {I}, and production Vt -

aVrb.

This allows the derivation A =+ ab + aabb. In the first step, we apply (13.1) with the identification 11 : A, Vr : tr, rt : A, At : &, Wt : Vt, and 3r2: b. In the second step, we re-identify Vt : ab, leaving everything else the same. If you continue with this, you will quickly convince yourself that the language generated by this particula,r Post system is {attbt' : rz } 0}. I

irn C 7 : { 7 ,* , : } , Cw:A,

V : lVr,Vz,VsI, A:{1*1:11},


336

Chopter l3

OrImn Moont s ot'' CoupurartoN

and productions Vt*Vz:

V 3-

V r 7I V z :

Vrl'

Vt*Vz:

V 3-

VtIVzI:

Vrl'

The system allows the derivation 1*1:11+11*1:111 + 1 1+ 1 1 : 1 1 1 1 . Irrterpreting the stritrgs of .['s as ullary representer,tiorrsof integers, t]re derivation can be written as 1+1:2=+2+1:3+2*2:4. The language gcrrerated by this Post system is the set of all identities of integer additiors, such as 2 * 2 :4, derived fiom the axiom 1 * 1 : 2.

_l

Example 13.6 illustrates in tr sittrple manner the original intent of Post systems a^r a rttecltanism for rigorously proving mathematical statt:rnents from a set of axioms. It also shows the inherent awkwardness of such a completely rigorous approat:h and why it is rarely used. But Post systems, even though they are t:urnbersome for provirrg complicated thtloretns' are general models for computation, as tlte next theorem shows. A language is recursivcly enumerable if and only if there exists sotne Post system that generir,tcsit. Proof: The argurnents here are relatively simple and we sketch them briefly' First, since a derivation by n Post system is completely mechanical, it can be carried out on a Thring machine. Tht:refore, any langua.gc getrerated by a Post system is recursively enumerable' Frrr the converse, remember that auy recursively enumerable language is generatod by some unrestrit;ted grarnmar G, havirtg productions all of the form t+lt

with r, y e (V U T)-. Given any unrestricted grs,mmar G, wc create a Post system II : (Vn, C, A, P11),where Vn : {yt' Vz} ,Cw : V,Cr : T, A : {S}, and with productions V1rV2 + VrUVz, for every production fr -i A of the grammtrr. It is then an etr$Ymatter to show that a ur carr be generated by thc Post system II if arrd only if it is in the Ianguage generated by G. t


13.3 RnwnlrrNc SYSTTTMs

1 . For X:

337

{u,,h,c}, find a Post systern that generates the ftrllowing languages

(a) L(a.b 1a['*c)

,

(b) r:

{urtu} ffi\

(c) l,:

lanb"t!'j

Find a Post s.ysterrrthat generatcs t:

{***,

u,e {a,b}-}.

3, Ftrr X -- {o}, what language rloes the Post systern with axiom {a} ancl the following production generate?

Vr * I4Vr W 4. What language tloes the Post system irt Exercise 3 gencrate if the axiqm set is {a, ab} ? Find a Post system for proving the identities of irrteger multiplication, startirrg from the axiom 1 + 1 : 1. ffif,$

6 . Give the details of the proof of Theorem 13.6. t .

What language does the Post systcm with VnaVV and axiorrr set {ab} generate'l

8. A restricted Post system is onc on which every production r * g satisfies, in addition to the usual reqrrirements, the further restriction that the number of variable occurrcnces on the right and left is the safire, i.e., n : nr in (13,1). Show that for every language.L gerrerated by some Post systern, there exists a restricted Post system to generates Z.

R e w r i t i n gS y s t e m s The virrious grammars we havcl studied have a nurnber of things in common with Post systems: They anl all based on sotne alphabct fiom which one string can be obtained frorn irnother. Even a 'Ihring rnirt;hine can he viewed this wrr,y, since its instantarreorrs description is a string tha,t completely defirrcsits configuration. The program is then just a set of rules fbr producing one suc;hstring from a previou$ one. These observations carr ller formalized irr tlNr concept of a rewriting system. Generally, a rewriting system consists of arr tr,lphabet E and a set of rules or productions by which


338

Chopter l3

O'rseR Monnls or,' Cotrpurnuon

a,strilg in E* can produce another. What distinglishcs one rewriting svstern from a,nqtht:r is the nature of X atrd restrictions for the applictr,titlrr of the productiorrs. The iderr.is cluite broad a,nd allows atty numtrt:r of specific ca$osirr adtlition to tfie orrcs we have alretr.tly etrcountered. Htrrc we briefly intrt)thrce sonre less well-ktrown ones that are interesting arrtl also provide general tnod(1985)' els fbr rxrrnputatiou. For dotir,ils,see Salomaa (1973) and Sa,lomeur

Mqtrix Grommqrs Matrix grammars differ from the grarnrrtars we lrave previously studied (which are often r:alled phrase-structure grammars) irr how the prodrrctions can be tr,pplied. For matrix graIIIIIlars, the set of productious consists of subsets P't, P2,..., Pn, ea,chof which is an orderul sequellce 1.Zt+'AIt12+!2r,-.-

Whenever thc first production of some set .P,,is applied, we must next ir,pply the secondorte to the string jrrst created, then the third otre, and so on. We cannot tr,14rlythe first production of 4 unless tr.ll otlter productions in tltis set can also be applied,

Consider thrl rnatrix grammar P1 : ,9 --+ ,91,92, P2: 51 + 0,5I' Sz + bSzt:, P 3 : 5 ' 1- t r , , 9 2 - t r . A derivation with this grammar is S + S1S2 + a,9rbSzc+ aaSlbbS2cc+ &abbcc. Note that whertever the first rule ttf 1'2 is used to create arl ar the second one also ha^rto be used, producing a corresponding b and c. This rnakes it ea,sy to see that the set of termirur.l strings generated try this matrix gramrnitr is 7:{a"b"c":n}0}.

T Matrix grillrtlnars contain plrrase-structure gramttrars as a special ctt*se in which each 4 coltains exa,ctlv one production. AIso, since tnatrix grrlrnm&rs repre$crrt algorithmic prot:t:sses,they are governccl by Church's thesis. We cxlnclude from this thir,t rnatrix grammar$ trrrtl plrrase-structgrtl graurmars trave the same power ir,srnodels of computa,tion. But, as ExamJrle 13.7 showtl, sornetimes the use of a,rrtatrix grammar givcs a much simpler solutiotr thart we cau a,chievewith arr uurestricted phrase-structure grammar.


I3.3 Rnwnluuc Sysrnlrs

339

MorkovAlgorithms A Markov algorithrn is a rewriting systt+mwhose productions n+ U are consideredordered. In a tlcrivation, the first applicable prodrrction must tre used, Furthermore, the lcftmost occurrerrt:eof the substring r must be replaced by g. Some of the productions rnay be singled out ir,sterminal productions; tlNry will be showrr a,s r + .,!. A derivation utarts with sorne string ,ur â‚Ź x tr,ndcontinues urrtil either a tcrminal produr:tion is used or rrntil there arr: rro a,pplicableproductions. For languageat:ceptance,a set ? C X of terrnirra,lsis identified. Starting with a terminal string, productiorrs a,re applied rrrrtil the empty string is Drodrrced.

I

Lct M be a Markov a,lgorithrnwith alphabet E arrd terminals ?'. Tlrcn the set L(M):

{",

e f. ;t,5.1}

I ,it_;! t

is tlre languageacceptedby M.

-R;;' -'Fr"

t*qmFld lS.$

Considerthe Markova,lgorithnwith X *T : {4,6} and productions ab-4, ba - ,\. Every step in the clerivation a,nnihilatesa srrbstring ob or ba, so L (IV):

{ r r â‚Ź { o , b } . : r t , ,(,t u ) : " o ( r ) } .

I


340

Wnd

Chopterl3 OrHnn Motlpl,s op Courur.q"ttotl

a' Mtr,rkov algorithm f.r I. = {trLfirl : rz > 0] . An arrswer is ab- S, tz5b -+ S, 5 --i .,\. If irr this last examplc we take the first two productiorrs and reverse thtr left nnrl right sides, we gc:t a context-f'rt)(t gralnntar that gcrterates the la,rrgua,gcl.L. 11 a certain uense,Markov a"lgorithms are simply phrase-structure grailrrrrars working brr,t:kward. This ctr,rrrtotbe takett totl literally, since it is nqt rlc:irrwhat to do with the Iast produt:tiotr. But the observation does prtr viclq ir.starting point for a proof of thc ftrllowing theorem that cha'ra'cterizcs tlx: power of Markov algorithms.

I

A ltr.rrguageis recrtrsivt:ly errumera,bleif arrd only if thertl cxists a Ma,rkqv tr,lgorithrn for it. Proof: SeeSalomaa(1985,p. 35). r

L-Systems The origins of L-systems a,rc quile clifferent f'rorrr what we miglrt expect' Their clervcloper,A. Lindentrir,yer, used them to rttodel the growth pattern of certairr organisms. L-systcrrrs are essentially Ttarallel rewritirrg systems. Bv t|is we mean thnt irt each step of ir.tlt:rivatioll' everY syrtrbol has to bc rcwritten, For this to rnake sense,thtt llroductions of tr,rrL-system must bc: of the form 0,+'u)

(13.2)

where a â‚Ź X a,rrdz â‚Ź X*. Wht:rr a stritrg is rewritterr, otre such productiol nust be a,pplitxl to every symbol of the string bcfore the new strirrg is geuertrtetl.


13.11Ruwnruuc Sysrr:MS

341

Letx: {a}anrr i\iHl$[Nt$[$W,toi,, define an L*system,

startirrg

fiom the string tr,, we carr rna,ke the derivation

a + aa + aaaa + a,a,a,a,0,o,a,a,

The set of strings so derived is clcarlv 7,:{u.2":n>0}. Note agairrhow such special rewriting systemsare able to deal with prohlems that are quittl difficult for phrase structure gremrnars.

I It is known that L-svstems with productiorn of the forrn (13.2) are rrot sufficierttly general to provide for all algorithmic r:omputations. An cxtension of the idea provides the necessary generalization. In an extenrlrxl L-systern, productions irrel of the forrn \ixlatA) +

u1

wlrer(ra e E and 'J:1y 1u E E* , with the interpretaticln that a carr bc replaced by z only if it occurs as part of thc string rag. It is known that such extended L-systems are gerrcrell models of computatiorr. For cletails, sec Salomaa(i985).

IxEncrsES I-

Find a matrix grarnura,r for 7 : {uzn: ut â‚Ź {a,b}-}.

2. Wlut

m

language is generated by thc matrix grammar P1 :,9+S1.S2, Pe: Sr -

a S L b , 5 2+ h S z a ,

Pr:5'r-tr,S'r*tr. 3. Supposc that in Examplc 13.7 we changc the la^,;tgroup of procluctions to P 3 : , 9 1-

4 , , 5 2+ , 9 ,

What language is generated l-rythis matrix grarnmar? 4. Why does the Markov algorithm in Examplc 1B.g not accept abab'l


342

Chopter l3

OrHon Monpls

oF CoMPUrerIor'l

5. Find a Markov algorithrn that derives the language L

{a"hncn : n } 1}'

ffi * 6. Find a Markov algorithm that acccpts L:

{a"b"'un* i n}

1 , r z l} 1 } .

7. Find an L-system that generates tr(oo-). 8. What is the set of strings generated by the L-system with productions o, +

aa1

o, +

o,o,0,1

when started with the strirrg a'l ffi


A n I n t r o d u c t i o tno co mp u t q t i o n oI co mp l e x i t y

rt strrdying algoritlrrns a,nd computirtions, we have so f'ar pa,id Iittle atterrtiorr to what actually can be expcctcd when we apply these ideas to rual compulers. Wu have been alrnost exclusively corrr:ernecl with questions of the existt:rrc:eor nonexisterrr:e clf aleorithrrrs frrr certain problems. This is an appropriate starting point firr a theory but cleirrlv of limited lrrat:tica,l significant:c. For actual c:ornprrtations, we rrt)tlrl not only to know that a,problern can br: solved in principlc, but we also rnust be rrble to construct algorithms that c:rr,nbe carried out with reasonabkr efficiency. Problerns Lhat can be solved effectively a,recalled tractable, a descriptive term that will Lregivcn a nrore precisc mea,ningin this chapter. In the pra,clical world of software dcvelopment,, efficicncy ha,s many facets. Sometimes,we are cuncerneclwith thc efficient use of thc r:omputer's resources,srr(:tras processor tirne and mernory spa,ce, At other tirnes, we may be rnore concemed with how quickly softwa,re can be crca,tecl, how effectively it ca,nbe maintairx,rd,or how relirrble it is. At still tlther times, we may ernphilsize the efficiency with which a us(lr'$ problems can trtl solved. AII this is mrrr*r too complicatcrl to be captured by any abstract thcorv. All we can do is to ftrcus on sorne of the more tangitlle issues and create the

343


344

Chopterl4

Arv lurnonuc:rtot'tto Couru'lA'l'IoNALCoMPLEXITY

appropriate irbstract fiamc:work for thesc:. Most of thc tesults tha,t have been tlevelopecl arlclressthe sllirt:c and time cfficiency of ir, computatitllr, Ieading to the important topig of complexity theory. In the stucly of t:ornplexity, the primary colcern is the efficiency of a computir.tioll as met$rlred by its time arrd space requirerttettts. Wt: refer to this irs the tirne-complexity and the space-complexity of algorithms. computational complexity theory is ilrr extetnive topic, most of which is well outside the scotretlf this text. Thcre are some rt:sults, however, that are simply statecl and errsily appreciatQtl, and tha,t throw further liglrt on the nature of languagesarrd compUtations. In this set:tiotr,we givera brief overview of some compklxity results. Fclr the most pirrt, proofs a,rtrtlifficult ald we will clispense with thern hy rcference to appropriate sources. Our intent heru is to present the flavor of the subject mirtter and show how it relates to what we knqw about languages and autorrrata. For this reason we will allow ourselvesir great deal of latitude, hoth irt the selectiorrof topics rlrrd in the formality of the discussion' we will lirnit our dis<Irssion here to issues of time-cornplexity. There are similar results fbr space-cotnplexity, trut time-corrrplexity is a,little rnore accessible.

E f f i c i e n coy f C o m p u t o t i o n I,et rrs start with a ctlncrete examplc. Givetr a list of orre thousand integers' wc want to sort tltem iu som() way, say, in irscendiug ordtlr. Sorting is ir sitnple problcrn but also onc that is verY furrdamental in cortrputer science. ,,How long will it take to do this task'i" we see If we now ask the questi0il immtxliately that mrx:h more inforrration is needtxl before we (laII answel it. clearly, the numtrerrof itenrs in the list plays arr itnportant role in how much time will be taktlrr, but there irre other f'actors. There is thc question of what computer we use and how we write the program. Also, there are a number of sorting rrrethodsso that selection of thc algorithm ilr irnportant. Tfuere are probably a few more things you carl t]rink of that lleed to be looked a,t btlfore you can evctt tttake a rorrgh guess of the tirne requirements. If wc have arry hope of producing sorne general picture of sorting, rnost of these issueshavc to be ignored, arrd we must corrcentrateon those that are most fundtr,metrtal. For our rliscussion of computational complexity, we will rnake the folIowing simplifving assumptiorts. 1. The nodel fbr our study will be a'ftrring mat:hitre. The exact type of Ttrring machirre to be used will be discussedbelow. 2. The size of tht: problem will be denoted by n. For our sorting problem, rl is obviously the number of iterns in the list. Although thc: size of a


14.1 Epr.,rclruNcrr oF CoMprr.r'A.Ftoru 345

problem is rrot rr,lwaysso easily c:harrr,cterizecl, we can geuerally relate it in some way tcl rr positive integer. 3. Irr analyzing atr algorithm, we are less intcrcsted in its perforrrrarrr:rr on a sper:ific case than irr its general behavior. We are particularly conctlrrrttdwith how Lhealgorithm behaveswltt:rrthe problem size irrr:rerlres. Becaust+of this, tire prirnirry qrrestion is with how fh,stthe resourrxlrequirenents grow as rz becornr:sltr,rge. Our inrtrretliategoal will therr llc tcl characterizt:the time requirement of a, problern il^sa fhnction of its size, rrsing a Thring rnalchineas the cornputer model. First, we give some mea,ningt,othe concept of time for a Thring machine. 'I\rrirrg Wtr think of a ma,chine as making one rrove per tirrrc rrnit, so the tirne taken by a courllrta,tion is the nurntlt'r of moves nade. As stated, we wtrnt to study how the cornputational rrxlrirements grow with the size of tlrc problem. Norrnally, in the set of all protrlems of a given size, there is sorrrc vtr.riation. Here w(r irre irrterested orrly irr the worst case that hils the hight:st resource requirclrncrrts. By saying thrrt a, computation has ir timecorrrplcxitv 7(rz), we rrrearrthat the computation firr any problern of sizc n can be tlclrnpleted in no rnore thirrr 7 (rz) moves orr rrome Ttrring rnachine. After scttling on a, specific typc of T\rring rnachirre as il computa,tional model, we coukl a,na,lyze algorithrrn try writing explicit progrrrm$and counting the nurnber of stcps involvecl in solvirrg the protrlem. tsut, lbr a va,riety of reasons, this is rrclt ovrlrlv useful. First, the number of operatirlrrs perforrntxl may vary with the srnelll details of the progra,m and so iltay depcrrd strclrrgly on the prograrrirrr{)r. Second, frorrr a pra,r:tica,lstandpoint, wc irre itrtercstcd in how the algoritlrrn perfbrms in the real world, which may difftlr considelably fionr how it does orr a T\rring machine. TIrc.best we can hope fbr is that thc T\rring machine analysis is representative of tlxr major aspects of the real-lif'eperformance, for cxa,mple,the asyrnptotic growth rate of the time complexity. Orrr first attempt at rrnderstanding the resour(:e requircrncnts of an algoritlun is thcrefore invariably an ord,e,r-of-magniLude analysis irr which we use the O, O, nnd () notation introdrrr:ed in Chapter 1. In spite of the apparent inforrnality of this approach, wc often get very useful itforrniltion.

Givr:rra,set of rz numbers ir:r1r'21...,r, and a key mrrnher r, determine if tlrt: set txlntrr,irrsr, Unlcss the set is organized in some way, the sirrrpklst nlgorithm is just a l'inear seu,rch, irr which we corrrtr)ar(r r successivelyagainst :r:11.T21,.,, until we either firrcl a, ma,tch or we get to tlrc la,st element of the set. Since we may find a rrrir,tr:hon tlte first cornprrrisorror on the last, we carrnot predict lxrw much work is involved, but we know thtlt, in the worst case, wcl hirvel


346

Chopter l4

Ar'r Inrnonucrrlow ro Coururh-l'loNAl CoNrRLuxtrv

to rriake rz comptr,riscltrs.We ca,n tlx:n say that the titne-complexitV of this Iirrear search is O(tl), or even bt:tter, O(n). In ilraking this analysis, we rnacleno specifit: ir.ssurnptiousa,ltclrt what machine tlfs is run on t)r how the rrlgorithm is implernented. I

LL supposc you are givetr a set offl, rrutttbers l.^r1frz1.'.irD alld are asked trt tleterminc whether this set conta,insarry duplicates. (a) Suggest an algorithrn and find an orrler-of*rnagnitude exJrression for its tinre-comJrlexity. (b) Examine if the irnplernentation of the algorithtn on a T\rrirrg rnachinc affectrr your conclusions. 2, Repeat Excrcise 2, this tirne dctcrmining if the set containrt arry triplicates. Is the algorithm as efficient as possible? 3. Review how the r:hoice of algorithrn afl'ects the efficient:y of sorting.

ffiffi

T u r i n gM o c h i n e so n d C o m p l e x i t y

In Chtrpter 10 we a,rgrtedthat the various types of T\rring machirres were equivalerrt in their powt:r to solve problems. This allowed us to take whattuvcr type was most t:orrverfettt for an argurnent and even use programs in higher-level r:orrtputer langua,gesto avoid some of the tedium involvtxl in using thc standard Thring machine model. But whetr we tnake cornplexity an issue, the equivalence btttween the various tVpes of T[rring mAchifles no longer holds.

Irr Example 9.4 wc t:onslructed a, singltl-tape guage

'ftrring machine for the lan-

7={anb":rr,}1}. A look at that algorithm will show that for ut : unb" it takes roughly 2n stepu to rnatch each a with the correspondirrg b. Therefbre tlte whole comprtation takes O (n2) titne. But, as we later indicated in Exa,mple 10.1, with a two-tape machine w() carr use a different algorithm. We first copy all the a's to the second


14.2 TuRtNc MncHrruns AND CoMpLEXrry

347

Figrtre 14.1 Tape 1

Tepe I

Tape 2

Tape 2 (a) Initial tapes.

(b) Tapes aftcr copying of a',s.

tirpe, then rnatch tlrcm tr,ga,instthe b's orr the first, The situation before and afher the copying is slrown in Figure 14.1. Both the copying arrrl the rrtatr:hing can be donc irr 0 (n) time atrd wc see that a two-tape malt:hine has tirne-corr4rlcxity O (n). I

f;tumpl*

l4,S

Irr Sections 5.2 arrcl6.3 we cliscusseclthe rncmbership problem for t:onrextfree larrgrrages.If we take tlx: Iength of the inprrt string tt as the problr:m sizen, then the exhaustive st:arr:hmethod has compklxitv O (n M) , where M depends otr thcl gramma,l:. T'he nore elflir:ientCYK algorithrn him urmplexity O ("'). Both thcse algorithrns are cletcrministic. A nondeterrnirristica,lgorithrnfor this problem proceedsby sirnply guessing which sequenc()of productions is applir:d in the derivalion of rr. If we work with a gralrlllrirr th.r.t has no unit or .\-productions, the Ierrgth of the dtlrivtr,tionis essentially l,rlrl,so we have an O (n) algorithm. I

Fr(qffiFl$ l4;4

We uow introcltrc:cthe satisfiability problern, which plays an irnporta,nt role in complexity thexrrv. A logic or boolcla,n,,.rnsta,rrtor variablc is one tha,t can take orr exar:tly twtt vtrlues, true or firlse, which we will denotc: tly I and 0, respecl,ively. Botlletr,noperators are thcrr used to combine booklan rxrnstantsa1d variafles irrttl bclolea,nexpressiorts. The simplest boolearr oprltltclrs are or, denoted by V a,rrddefined by 0V1:1V0=lVl:l 0V0:0,


348

Chopter l4

Ar'{ IrurRonucrIoN l'o CoMPITTATIONALCotr,tplnxlrv

rr,ndtlrtl artd operation(n) defineclby 0 A 0 : 0 n : L: 1 A 0 : 0 , 1z\1:1, dettotedby a ba,r,and defineclby Also neededis ne.gat'ion, 0: 1, t:0. iu corqiunction normal form. In exprt:ssitlrts We considerrurw boolea,rt frotrt variablâ‚Źsfi1,1r2,...,rrr, starting with this fbrm, we createexpressions (::ftAtrA...At7.. Tlre terms ti,ti,...,fk nrrgation, that is,

(14.r)

trr() t:reated by or-ing together variables tr,ilrl their

Ii :.s,V.sm,V...V.$pr

( r4 . 2 )

wlrere each El.srr,...,.9pstands for somtl variable or the tregtr,titlrrof a variable. The satisfiability problem is therr simply sta'ted: givcn atr expression e in coniurrctive norma,l fbrm, is there an assignrnent of values to tire va,riahlels that will tnake ttrtl value of e true. Rtr rr specific case, lclok irt rt,Tz,...,t' e 1: ( 7 1 V r 2 ) n ( r 1 v r 3 ) . Therassigntnentrt : 0, fi2 : 1, t3 : I na,kes e1 trtrc:so that tllis expressiclrr is satisfiable. On tht: otlter haucl, c2: (r1Vr2)AzrArz is not satisfirr,blt:btlcause every a,ssignrnc:rrtlbr l,he variables flr irlrd z2 will make e2 fa,lsc. A dr:tr:rrninistic algorithm for the satisfiability problcxn is easy to discover. Wu take:all possibleva,lltcsftll the variablesrr , ff2, ..., t'r, alrd evaluate the exprrlssiotr. Siuce there a,rtl2t' such choices, this t:xhiruslive approa't:h iras r:xporrentialtinre compk:xity, Agaitr, the nondeterurirristic:approaclt simplifics rrtatlers, If e is satisfi'flils e. is essentially irrr itble, we guessthe va,ltrcof citch r i and then evir,hrir.tc: O (n) algorithrn. As irr Exarnple 14.3, we hirvt: ir deterttriuistic exhaustive sea,rcha,lgrlritlrrn whose complexity is exponetrtial and a lirrear uotrdeterrninistic orrt:. However, urrlike tht: llrtlvious exa,tttllle,we dtl rr<ltktrow of any norr()xl)()rtentialdeterministit: algclrit,hm, I


14.2 'I'uRrNclMecHrrueserur Col,tprnxrlv

349

These exarnples suggc:stthrlt complexity cluestions are affectcrl try the type of Ttrring tnar:hirrt: w() ll$e a,nd tha,t the issue of deterrnirrisrrr v()r$llti noncleterminism is a partir:ularly r:nrr:ial one. Exa,rnple 14.1 suggeststhat a.lgorithrns for a rnultitapc rnirr:hine may be reasonably close to what wt: uright use when we prograrrr irr ir rxlrnputer language. For this reason, we will rrse a, nrultitape T\-rring rnachinrl iLsorrr model for stuclying cornplexity issutls.

For the tlxur(:ise$in this set, assume that the Thriru rrrirr:hirrcsinvolved are all deterrrfnistir:. I . Find a, linear-timc algorithrn lor rnernbers]rip in {tlur: ru E {a,b}.} using a two-tape Turing machine. Wha,t is thc Lrest,you could expect on a one-tape macirinc? a)

Show that any cornputation that can be performed on a singlc-tapc, off-line 'llrring machine in titre 0 (T (rr)) also r:an be perfornred on a standard Turing rnar:hinein timc O (? (n)),

s. Show

that any txrrnglrtation that ca,n be pcrformcd on a standard Tfring tnachitre in tirne O Q (n,)) also can be performcd on a Turing machine with one semi-infinitc tapc in tiue O (T (n)).

4.

lihow tltat any t:ornJlrtation that can be performcd on a t,wo-tape rrrachirre in timc O(Z(n)) catr be perforrnerl orr a standarcJ'l\rrirrg machinc in time

o ("' (n)). 5 . Rewrite the boolcan cxprcsslon (rr Arr) Vnil in conjunctivc nornral lbrrn. 6. Dctertuine whether or rrot the expression (rr vr.; vrr) n (r1v:r2 vn 3) n (nr vnr vn3) is satisliable. 7, Irr Example l4.2 we claimcd that thc first algorithtn harl tirne trrmplexity 0 (rr,') antl the $et:onrl O (n,). ()a,n wc. bc morc precise and rdairrr that T (n):(.) (tl2) ftrr the first t:ase,and T'("): O (n) for thc scconcl'lIIow this strcngthen the argutrretrt in Exarnple I4.2?


350

Chopter l4

Aru IN'r'Rot)uc'I'Ior'tro Co\,IIUTATIONAL Corvet,r:xIrv

L o n g u o g eF o m i l i e so n d C o m p l e x i t yC l o s s e s In the Chomsky hicrirrr:hy for language classifica,tion, we assctc:iatelanguage fa,milieswith clir.ssosof inrtornata, wltere each cla,ssof arttomir.til is deflned by the natrrre of its tt:trrporary storage, Another way of t:lir,ssifyirrglatrguages is to usc a Thrittg tnachiue of a 1ra,rticrrlartypc but corrsidertiute complexity rr distinguishing factor, To do so, wc first clefitre the time complexity of tr language.

ir lolllll."f'i,,ir ritsof'1.1ini, Wu sa,ythat a Thring machine accepts a langrrrr,ge.L irr tirrre T (n) if every ?u in.L with lrl < r, is acceptedinO(T (z)) moves. If M is trondeterministit:, tlris implies tha,t for ()v{}ry 'u â‚Ź L, t}rere is at lea,st one $eqlrenctlof rrrovesof lcrrgth O (Z (ltul)) that leads to accepta,nrx.

A lrrrrguage.Lis said to be a tnember of the cla,ssDTIME (f (")) if there exists a deterrrrinistic mrrltitape Tirrirrg rnachirre tltat accepts .L in time T (n\. A lir.nguirgc.L is said to be a member of the cla,ssNTIME (f (n)) if there t:xists a rrorrdeterrninistic rmrltitape T[rring mar:hine that atrepts -L in tirne T (n).

Some relatiorrs between theserrxlrnplc:xityclassessuch as DT|ME (r('))

c NTIME

(7-(")),

and T 1 ( n , ): O ( T . . 2 ( n ) ) irnplies

DTIME (ri (")) c DTIME (Tz(n)), n,reobviolrs, llrt frorn here the situation gets obscure quickly. Whtr,t wt) r:arr say is th.r.t a,stlrtl order of T (z) increases, we take in progressively morr: languirgc:s.


14,3 LalrcuecnFauu,rpsenn CourlExrrrr Classns

851

For everyintegerA > 1, DTIME ('"*) . DTIMn (t**t). Proof: This follows from a result in Hopcroft and Ullman (1979, p. 299). r

The conclusion we can clraw frorn thiu is that there are sorne larrguages thrlt ca,n be accepted in tirne z? for which there is no linear merrrberslrip nlgorithm, that there are languagcs in DTIMH (n3) that are not irr DTIME (n2), and so on, This gives us iln inlinite number of nested com* plexity classes. We gct cvcn more if we allow exponerrtial timt+ complexitv. In fhct, there is no limit to this; no matter how rapidly the cornplcxity firnction 7(n) grows, there is alwrrys sornething outside DTIME (f (")).

There is no tota,l Ttrring computable function / (n) srrt:hthat every recursive larrguagcris in DT I M E (f (n)) . Proof: Consirlt:r the a,lphabet X : {0, 1}, wit}r all strirrgs in X+ arranged in proper otder'u1,trzt....AIso, assumethat we have a proper ordering for the T\rring rnachinesit M1, Mz,.... Assume now that the function / (n) irr the statement of the theorem exists. Wt: ca,n then define the Ianguage 7 - {tut: M, does not accept ,rli in f (lur;l) steps} .

(14.3)

We cla,im that .L is recursive. To see this, r:onsider any ur f tr and compute first / (lurl). By assuming that / is a tota,l T\rring computable function, this is posuible. We next find the position i of u,'in the sequence101,r1t2,.... This is also possibk: bet:arr$ethe sequenceis in proper order. Whcn wr: have i, we find M. and let, it operate orr to fbr /(lr,l) steps. This will tell us whether or not zu is in -L, so is recursive. But we carr n()w show tha,t.L is not irt D'I'IME (/(")). Supposeit were. Since.L is recrrrsivt:,thert: is some M6, such that tr: L(Mn). IslL.ln in,L? If we clairn that'ur1. is in -L, then M6 accepts trrp in /(ltrr6l) steps. This is becauseL â‚Ź DTIME (/ (")) and every w e L is accepted by M6 irr time / (ltrrl). But this corrtradir;ts(14.3). Conversely,we get a contradiction if we assume that ** # L. The irurbility to resolve this issue is a typical diagonalization result and leads us to conclrrde that the origina,l assurnption, namely the existence of a computable / (z), must be false. I Theorern 14.1 rr,nd14.2 allow us to make various clairrn, for example, that there is a language in DTI M E (na) ttrat is not in DTi M E (nB). etthough this may be of thcorctical interest, it is not clear that such a result

#


352

Chopter l4

Arq INrnotrtrc'tIouro ClonrurlTtoNAI-Coltplnxlrv

ha,s any pr:r,ctical sigtrificance. At this point, we have no t:hrtl wltat the fu DT'IME (na) might bc. We can get a little r:Irirracteristicsof a,la,ngrra,gt: rrroreinsight into thc matter if we rela'tethe c:omplcxity classificationto tht: languages in the Chornsky hierarchy. We will krok irt sotne simple exa,mplc:s tlrrrt give sorne of the mrlrc cltrviousresults.

e * o m p tie; , $ ,

Ilvery regular language can be rerxlgrriztxl by a det,ermittistic finitt: autornatorr irr titrte proportiona,l to tlrc lcrtgth of the input. Thert:ftrre Lanc; C DTIME

(rt).

Brt, DTIME (n) includes mtrc:hrttore lhatr L6a6. We havc rrlready estabIislxxl in Example 13.7 that tht: c:ttrrtext-freelauguage {a"b' : rr.> 0} can be rclcogtrized in time O (n,). Tlrc argtrtnetrt given there r:tlrr lltl rrsed for eveu rnore complicated la,ngrrilges. I

f *innpidlilUildlt The

non-contcxt-f'reelanguage7': {wu: rr.râ‚Ź {n,b}.} is fi NTIME(n,). This is straiglrtftrrward, as we calt recognizt:strirrgs in this langua,geby thtr algorithm 1. Ct4ry tlte itrput from the inprrt filc to tape 1. Nondetertninistictr.llvgut:ss thc rnirldle of this string. 2. Ctlpv the secoud llart to ta,pe 2. 3. Compare the symbols orr tape I and ta'pe 2 one bv tlrrtl. Clearly all of tlre steps r:an bc dorrein O (lul) time, so L e NTIME (n). Actua,lly, we cirrr show tltal L â‚Ź DTIME (n,) if wtr c:arrdevise au algorithm for flnding tlx: rnitltlle of a string in O (n) tirntr. TIds can be done: we look at ea,chsyrntlol orr tape 1, keeping a count on talrc 2, but couuting only every seconrl syrrrbol. We leave the details tls iln t:xercise. I

It follows fiorn Exarnple 14.2 that [,cp { DTIME

(n:])

Lar E NTIME

(n).

and


14.4 THn Coupr,nxlly Cl.,l,ssesP nrul NP

353

Consider now the family of context-srlrrsitive languages. Exhaustive searc;lr parsing is possible here also since at every step only a; limited number of productions are applicirtrkr. Therefore, every string of lcngth n can be parsed in time n,M, where M clcpcnclson the granurrar, Note, hrlwever, tha,t we ca"rrnotclaim from this that Lcs c DTIME

(nM)

becausewe cannot prrt arr upper bound on M. I Frorr these exa,mples we note a trrlrrrl: as 7 (n,) increases, more and nrore of thc fermiliesL,R1G, Lcn, Lcs arcl txlvered. But tlte connectiorr between the Chomsky hiera,rchy and the corrrplcxity r:lilssesis tenuous and not verv clear.

1. Cotrplete the argumcnt in Example 14.5. 2 . S h o wt h a t . L :

{wwdw:w

â‚Ź {u,,4}+} is in DTIME(n).

3 . Show that 1, : {www : tr â‚Ź {a, b}+ } is in DTI M h) (n), 4.

Show that thcrc are langutr,gesthat are not in NT'IMfl

(2").

TheComplexityClossesP ond NP ffiffiffi*ffiffi Sincc the a,ttempt to produce meaningful hicrarr:hicsvirr,time-complexities witlt differcrrt growth rates a,ppearsto lre unproductivrl, lct us ignore some f'actors that are less irrrportirnt, ftrr example by removing sorne urrirrtercsting distinctions, such as that betweert DTIME (nft) and DTIME (rzh+l). We (rfirrargue that the difference between, say, DTIME (n) and DTIME (n2) is trot fundamcntal, since some of it depends on the specific modtrl of TtringE machine we have:(rr.g.,how ma,ny tapes), and it is not a priori clear which model is most appropriate frrr ir real computer. This leads us to consider the famous corrrplclxity c:lir^rs

p-UDTrME(nt). rl) I

This r:la,ssincludes all languages that are accepted by some deterministic T\-rring machine in polynornial tirne, without arry rcgard to the degree of the polytrotnial. As we htr,vea,lready seen, .L466; and .Lcrp are in P.


354

Chopter l4

Couplnxtrv Ar'r lrutRoouctIoN To CoMPITTATIoNAL

Since the distinction bctween deterministir: irrrd noldeterministit: txtrnplexity classesa,ppear$to bc fundamenta,l, wtl irlso int,roduce

NP : ! I'rzrua ("t) . Ohviously PgNP, but whtlt is rrot known is if this c:orrtainment is proper. Whilc it is generally believrxl that there are sorne ltr.rrguirgesin NP that are not irr P, no one has yct fourrd an exanrple of this. I'he interest in these complexity classtltt,particularly in the clilss P, come$fiorn arr at,tempt to distingrrish between realistic and urrrcalistic cotnprrtatiorrs. Certain comprrtir.tiorrs,alLhoughtheoretictr,llvpossible,have such high resourcerequiremrlrrtsthat in practice they mrrst tlc:rejected as unrealistic on existing computtrrs, irs well as on supercomptrtt:rsyet to be designed. Such problems are sorrxrtirncscalled intractable to irrtlicate that, while in prirrciple cornputable, thrrrc is rro realistic hope of tr, prirctical algorithm. To utrdersLandthis bettelr, c:ornputerscientists have ir.ttt:rnptedto put the itlea of intracta,bility orr ir formal basis, One atternpt to defile the term intractable is ma.dcirr what is generally called thc Cook-Karp thesis. In the Cook-Ka.rp thesis, a problem that is in P is called [racta,ble, and onc thnt is not is said to be intra,ctnbkt. Is the Cook-Knrp thc:sis a good way of separir.tirrgproblems we c&n work with rea,Iistic:a,1lyfrotn l,hose we ca,nnot'l The answer is not clearcut. Obviouslv, rrny rxrrnputation that is not in P hirs tirtre complexity that grows faster tha,n ilrry llolytrornial, and its require:trcrtts will increase very quickly with the probletn size. Even for a, firnctiort like 20'1", this will be excessive fbr large ?r, say rz ) 1000, and wc rrriglrt feel justified in ca,lling a.prohlcrtt with this complexity intra.c:table.But what a,bout problcrns tlrat are il D'I'IME (n,too)f While the Cook-Karp thesis t:a,llssuclt a probIem tra,ctable, rlrrc srrrely cannot do much with it even for small n'. The jrrstificntion for tlte Cook-Karp thesis sccrnsto lie in the empirical obscrvatiorr tlrat tnost practical problems irr P are h DTIME (n), DTIME (rf), or DTIME (rz3), while thosc outside this class tetrd to havc exponential crrtrplexities. Among practica,l prcibletns,a clear distinction exists between llroblems in P and those not irt P. The study of the rela,tiorr between the complexity cltt^sscsP and NP has generated partir:ular irrterest alnoltg computer scientists. At the root of this is the question whcthtlr or trot P:NP. This is one of the funda,mental urrsolved problems iu the theory of cornputation. To explore it, cxrmputer scientists have introducerrl a variety of


14.4 THU Col,lplnxrry

Cllessns P aNn NP

355

related concepts and questions. Orre of them is the idea of an NP-<:omplete problern. Loosely speaking, an NP-complete problem is one that is as hirrcl as any NP problem and in sorrre sen$e is equivalent to all of thern. Wha.t this rrreans ha,sto be explained.

llM A ltr"nguage,Llis said to be polynomial-time reducible to sornelirrrgrra,ge .L2 if there exists a determirilstic T\rring ma,chineby which an1' u1 irr thr: alphabet of "L1can be transfonned irr polynomial time to a tu2 in the alphabet of .L2 in such a way that w1 â‚Ź L1 if nnd only if w2 â‚Ź. L2.

Florr this we see that if .Lr is polyrromial-time reducible to .L2, arrd if L2 e P, therr .L1 e P. Similarly, if Lz e NP, thcrr trr E NP.

Definition

14.4

A Iangrrage /, is said to be NP*corrrplct<: if I e NP and if every .L/ e NP is polynomia,l-time reducible to .L.

It follows easily frorn these definitions that if sorne .L1 is NP-txrmplete

ruljn",T,'1l-H:til,"*j#l"ll'-';*.-'r'J::'::,T tinte algorithm f'or any NP-cornplete larrguage, therr cvcry la,ngrragein NP is also in P, tha,t is P:NP. This puts NP-cornpltlterrttssin tr,central role for the study of this tpcstion.

a' be vir:w.rl a,sa language problem, We e'cocle specific instances as a string that is actrrpted if and only if the expressionis satisfinble. This problem is NP-complett:. The staltementtha,t the satisfiability problem is NP-complete is known as Cookts theorern, rr,discussion of which t:irrrbe fcrund in Hopcroft and Ullman (1979). I In addition to the satisfiability problern, a largc rurrrrberof other NPcomplete probltrms have heen found. For all of thenr we carr lirrd txponuntirrl


Chopter l4

356

Cotr,lpr,nxIrv ArrrlrutRopucuoN ro CoN4r,urATIoNAL

algorithms, but for none of them has anyone discovered a llolynomial-time algorithm. These failrrres ltlad us to believe that prohably PINP, t-rut until someone produces an actual language in NP that is not on P or, alternatively, until someone proves that no such language exists, the question remains open.

1. Prove the statement that if a larrguage .Lr is NP-complete time reducible to tr2, then -L.: is also NP-complete. **

and polynomial-

2, Consult books on complexity thcorv, and compile a list of problems that are NP-cornplete. 3. Is it possible that the question P : NP is undecidable?


S o l u t i o n sq n d H i n t s

for Selected Exercises

ChopterI Section 1.1 5. To prove that two sets are equal, wc mrrst show that an eletnetrt is in the first set if arrd only if it is in the secorrd. Suppose n .€ 51 U Sz, Then r f 51U,92, which means that r catrnot be in 51 or in 52, that is r € ,91tl 52. Conversely,if z € ,9r t'l ,92, then r is uot in 51 and r is not in 52, that is r e 5r U 52. 6. This can be proven by an induction on the rrumher of sets. Let' Z : 51 U 52...U Sr. Then,9r U,5'z-..U 5'. U 51",+r: Z l) 5,,.+t. By the stattdard DeMorgan's law

M:E)8,,+t. With the inductivc assurrrption, thc rclatiorr is true for up to rz sets, that is,

Z :5, n Fr rt ...n-.9*. Therefore ZnE,,+l : Sr f-'l5z fi "'15' I F,+r, completingthe ilductive step. 357


358

Alrswnns

8 . Suppose 5r : 52. Then 51 l'-l$2 : ^91tl 5"2: ^9rfl^9r = O and tlte entire expression is the empty set. Srrppose now that St f Sz and that there is an element r in Sr but not in,5z. Thcn r e Fz so that SL1E1 + fr. Thc c:ornpleteexpression carr then not be equal to the empty set. L2, If r is in ,91and u is itr ,92,then u is not in (,51U Sz) - Sz. Becauseof this, a rrecessaryand sufficient condition is that the two sets be disjoint. 1 5 . (c) Since

21, nl nn-L n n n n n n is the product of factors lesstlurr or equal orre. Therefore, n! : O (n"). 27. An arguilrent by contradiction works. Suppose that 2 - 14 were ratiorral. Ttren n

2-rt:

WL

gives

\/t = 2 mn1,- n contradicting the fact that V2 is not rational. By induction. Suppose that every integer less than z can he written as a product of primes. If z is a prirne, there is nothirrg to prove, if not, it carr be writterr as the oroduct R,:

?'tyTL2

where hoth factors are less than n. By the inductive assumption, they both can be writterr a*sthe prochrct of primes, and so can rt,. Section 1.2 R 2. Many string identities can be proven by induction. Supposethat (zu) : ufuft for all u â‚Ź E* and all u of length rz. Take now a string of length rr { 1,,say ur : ua. Then

(.,, )*:r#;;;,,il",,T,J,,,_ ;::_"

By indrrction then, the result holds for all strings.


Solu:uorvs nNn llrnrs pon Snlncrnr Exlln,crsus

S59

4. Sitrce abaabu.o.uba.u. t:ilrr bu decomposed into strings ab,e,e,baa,ab,aa, eaclr of which is iri -L, thc string is in tr*. Similarly, baaaaabaais in L*. However, there is no possiblclrlt:r:ompositionfrrr haaaaabaaaab,so this string is not in I*. 10. (d) Wc first go'nerrrtethree c,'s,then add an arbitrary nurnber of a's ancl bts arrywhcrtl. S --+AaAaAaA

A - aAlbAlA The first productiorr grlrrcratcsthree a's. The secorrdcan generate any number of ats and b's in atry posititln. This shows tha,t the grammar can geuerate atry string trr â‚Ź {4, b}* rrs klrrg ir.sn,, (ru,)} 3. 11. S+aA+abS+o.baA+abubS fiom which we see that L(G) : {(ob)" : n,> 0} . 13. (a) Generate one b, then an equal number of ats and bts, finally a.srrrarry more b's as needei. S --+ AbA A -+ aAblA B -, bAlA 13. (d) The arrsweris ca.sic:rto stxrif yorr notice that Lt : {1""-l:1b"' : ttt } 0} , I'his leads to the easy solutiorr S-

aaaA

A -

aAbl,\

14. (b) The problem is simplified if you break it into two cases,lrulmod 3 : 1 rr.ndI'rirlmod3 :2. Tlrc first iu cxrveredby St + aaaStl(t t thc sec:ondby Sz -

aaaSzlao,

TIte two carr be t:orrtbined irrto a singlc gramrnar by 5 -

51152.


360

ANSWERS

ro. (a) We can use the trick and results of Exantple 1.13. Let .L1 be the language in Example 1.13 and modify that grammar so that the start syrnbol is 51. Consider then a $tring u â‚Ź L. If this string start with an a, then it ha,sthe fbrm ur : o:tltt where trrl e .Lr. This situa,tion can be taken care of by S -+ rr,$'r.If it starts with a b, it can be derived hv S --r ^9r5. Section 1.3 1. integer --+ sign magnitrrdtr sign + + l-lA magnitude - digit I digit magnitrrde

digit --+0l1l2l3l4l5l6l7l8le Tlds can be considered an ideal version of C, ir^sit puts no lirnit on the Iength of an integer. Most real crompilcr$, though, place a Iimit on the number of digits. 7. The automatort lus to remember the input for one time period so that it r:arr bc reproduced for output later. Remembering t:an htt done by labeling the state with the appropriatc irrforrnatiotr. The label of the state is then produ<:erda*soutput later.

/)uo /-\-''

,*, -(

^)

7q l \

+ \Vi\ r b/a

)---\./ Y , t )

.

\-)uh

10. We remember inprrt by labcling the states mnemonically. When a set of three hits is dene, w{: produce output and return to the beginning to profft$$ tlx: rrext three bits. The following solution is partial, but the completion shtiulrl trc obvious.


SolurroNs ANDHrNTspoR Spl,pcteD ExERcrsEs

11.

In this case,the transducer must rememberthe two precedinginput symbols and make transitions so that the neededinformation is kept track of.


362

ANSwERS

Chopter2 Section 2.1 2. (c) Break it into three caseseach with an accepting state: no o'il, t)tltl a, two a's, three a's. A fourth a will then send the dfa into a non-accepting trap state. A solution:

n' \n' a' \n' n"o t \t I, \t

\t +ti

]t+il

\/o o.

,

\/"

Alf+il {

A ll-+ilA

\y'"

ti+t

\y'd

l

\*J

(a) The first six symbolsare checked.If they are not correct, the string is rejected. If the prefix is correct,we keeptrack of the last two symbols read, putting the dfa in an acceptingstate if the suffix is bb.

7. (a) UsestatesIabeledwith ltulmod 3. The solutionthen is quite simple.


Soluttor'rs aNn Hn-rrs FoR SELECTEIExnRcrsns

363

(d) For this we usenine state, with the first part of eachlabel no (ru)mod B, the secondpart rz6(tr)mod3. The transitionsand the final statesa,re then simple to figure out.

9. (a) Count consecutive zeros, to get the main part of the dfa.

(}--O-r-O:{} Then put in additional transitions to keep track of corsecutive zeros and to trap unacceptable strings.


364

ANSWFTRfi

(d) Here we need to remembcr all contbinations of thrcc bits' This requires I states pluu sorrrcstart-up. The solution is a little long but rrot hard. A uartial sketr:h of the solutiou is below.

-{;n \

1 3 . Ttrl tlirsiest way to solve this probll:rn is to cotrstruct a dfa, fbr L {u"' : n,:4}, then complerlt:rrtthe solution.

2 1 . (a) By rxrrrtradictiorr. Suppose G1a has no cycies in arry patlt from the initia,l sta,tc:to any final state. Then every wa,lk ha*sa, Iirrite trumber of stcps, arrrl so every accepted string ha,sto berof firrite length. But this irnplies that t,helanguage is finite. (h) AIso by cortttaclictiou, Assutne that (J,14ha,s$otrc c:yclein a path 'Wt: carr then use the cycle f'rornthel irritial staLeto sotne accepting utir,tt:. with atr acceptedstring. But to generatean arbitrarily long wa,lkLr,berlrxl a finite language ca,nnotc:ontairra.rtritrarily lotrg stritrgs,


Sot,tj'uor'rsano Hlr'r'r'snoR Snlnc;rnr FixnRcrsns

365

24. Tltcre il,re many diffcrr:nt solutions. Here is one of tlxlrn.

Section 2.2 4 . d * ( q o , a ) : { s o ,e t , e z l t , d(*Q r , t r ) : { q u , q r } . 7. A four-sta'tesolutiorris trivial, but it takesa little expcrirnernting to get a tlrrr:rr-sta,te one, Herr:is one answer: r.L

/

A a / - \ *V-\'_/-\-i

r n

\

No. T'he strirrg a,hclia,s three diffi:rent symlrols arrcl tlxrre is no way this r:iln he accepted with f'ewerthan tlrrcc utates, 1 5 . This is the kind of problem in which ylrr jrrst ]rave to try diffcrcnt ways.

Prohalbly rnost of youl tries will not work. Here is one that doos.

17. Introduce ir sirrglesta,rting sLatep6. Then add a transitiorr d ( P 6 ' . \ )= g u . Next, rerrrtlvcster,rtingstate status frrim Qe, It is straightfrlrward to see thnt the new nfa is rxlriva,lent to the origina,l one.


366

ANSwERS

20. Introduce a non-accepting trap state and make all undefined transitions to this new state. Solution:

Section 2.3 2. Just follow the procedurenfa-to-dfa. This givesthe dfa

7. Introduce a new final state p1 and for every g â‚Ź F add the transitions

d (q,A) : {prr . Then make ?7 the only final state. It is a sirnple matter then to argue that if d* (qs,w) e F originally, then d* (qo,r) : {pf } afber the modification, so both the original and the modifies nfa'$ are equivalent. Since this construction requires tr-trarrsitions, it cannot be made for dfa's. Generally it is impossible to have only one final state itr a dfa, as can be seen by constructing dfa's that accept {I'o}.


Sol,uttolrs er'ro Hrr-rTSron Snr,ecrpo ExeRcrsns

367

8. Getting an answer requires some thought. One stllution is

_ffi a,b

h

lL. Suppose that .L : {trrr,wz,...w^}. Then the nfa acceptsw,,

- - - - - - - - = + { tfr

\r'

tl

accepts -L, so the language is regular. 14. This is not easy to see. The trick is to use a dfa for tr and modify it so that it remembers if it has read an even or an odd number of symbols. This can be done by doubling the rrumber of states and adding O or E to the labels. For example, if part of the dfa is


368

ANSwERS

its equivalentbecomes

Now replace all transitions from an E state to an O state with .\transitions.

With a few examplesyou shouldbe able to convinceyourselfthat if the the new automatonwill accepttra2.\aa'.., originaldfa acceptsa4o,2&3o'4, and thereforeeaen(L). 15. Supposewe have a dfa that acceptstr' We then (a) identify all statesQ ttrat can be reachedfrom gs, reading any twosymbol prefix u, that is Q:{SeQ:d*(go,u)*qi. (b) introduce a new initial state p0 and add d (Ps,,\): Q. It should not be hard to seethat the new nfa acceptschopz(L). Although the constructiorris plausible, a complete answer requires a proof of the last statement.


SoLurroNs eNn HINrs rron Sunc.r.l;D Exnncrsns

369

Section 2.4 2' (c)

_@_"*64__r_@___@*G) Tlris is rninimal for the firllowing reasorr. % f F and qa â‚Ź F, so (3 and ga are distinguishable. Ncxt, d* (Sr,o) # -F tr.ndd* (qo,a) â‚Ź h-, sr) e2 and Qa are distinguishable. Sirnilrr,rly,d* (qr, uu,) f F and d* (q,uu) e F, ur) Q1and {3 are distinguishable. Continuing this way, we see that all statcs ir,redistinguishable and therefore the dfa is minirrml. First, rerrrovethe inaccessiblestates {2 and 4a. T}ren use the procerhrre marh to firxl the indistinguishablepairs (go,gr) arrd (g3,gr), This thcn gives the rrrinimal dfa.

6 . By contracliction. Assume that -il[- is not minirnal. Then we can corrstruct a srnallcr dfa M that acceptsT. tnffi, cornplerntrntthe final state set to give a dfa fbr .L. But this dfa is smaller than M, contradictiru the a,ssumptionthat M is minirnal. 1 0 . By contradiction. Assume that q6 ir,ndq. are indistirrguisha,hle.Since go a.nd q6 are indistirrguishable and indistingrrishability is arr equivalence relatiorr (Exercise 7), qo ancl g,, must be irrdistirrgrrishable.

Chopfer3 Section 3.1 2. Yes,treca,use ((o+ t)(0+ 1).). denotesarrystring of 0's ancl1's. $o does(0 + 1)-. 5. (a) Separate irrtrlruNes???:0, 1,2,3. (Ienerate4 or rrrorea'$,fbllowed by the requisite nunrberof b's. Solutiott: uaa(1,(1,* (A + b + bb+ bblt). (r:) The complernerrt of the languageirr 5(a) is harder to find. A utring is not in L If it is of tlxl fbrm o,t'b"",with either rz { 4 or rrz> 3, brrt


370

ANSWERS

this cloesrrot txrrrrpletelydescribe Z. We rmrst also ta,kelirr the strings in which a b is followed by an a. Solution: (.\ + a + u(r + aaa)b* + u*bbbbb** (4,* b)* ba (a+ t). . 9 . S p l i t i n t o t h r e e r : i r s e sn: t ,: l , z ) 3 , n Z 2 , m , 7 2 , a n d n , : 1 , r t t ') 3 ' Ea,c:hcase has a,strrrightforwarrl solution, 1 2 . Enumerate irll caseswith I'rrl: 2 to get aa (a * b)* aa I ab (a -f b). o,b* ba,(a 1 b)* ba + hb(a * b). bb. 1 4 . (c) You just have to get in each symbol at lcast once. The tcrrn

(a + b * c)*a (a * b * c)*b(a + h+ c)*c(a + b + rNwill do this, but is not enorrgh since the a, will precede tht: b, etc, For the completc solution you rilust genera,teall permutations of the three symbols, givirrg six terms that cau be adrltxl. The answcr, although quite lorrg, is conceptua,llyrrot hard' 1 5 . (c) Create two 0's, interspersedwith l's, then rcpeat' But don't fbrget the casewhen there are no 0's at all. Sohrtion: (1*01.01.)* + 1*. 1 6 . (a) Create all strirrgs of length tlree and repeer,t.A short sohrtiorr is

( ( a+ b + c ) ( a + b + r ) ( a * b + c ) ) . . 18.

(c) The staternent ( r ' 1* 1 2 ) . = ( r ' 1 * r ' 2 * ) * is true, Bv thc given rules (r1 +rt)* denotes the langua,ge(I(r1)u of eletr ("r))., that is the set tlf all strings of arbitrary cont:er,tcrrations ( " r ) ) ( I ( " r )).)., ( ( r , d e n o t e s ( r 1 ) ( r r ) . B r . r t ( r 1 ; r 2 * ) * anrlL m e n t so f . L which is the same set.

2L, The expresSiorrfor an infinittl latrguage mgst involve at least rlrre starred subexpressiorr,otherwise it r:arronly denote Iirrite striugs. If there is one starrecl sullexpressiontha,t tlenotes a non-errtpty string, then this string ca,nbe rcpeated as often as desired and thcrefore denote arbitrarily long strings. 2 3 . A closecl gorrtour will be gcntlrated by an expressiou r if tr.nd only if n1 (r) : n", (r') and nu (r) : na (r).


Solurlons .c,ND Hrr-rrsFoR SELECTEo ExeRcrsns

BZl

25. Notice several things. The bit string must be at least 6 bits long. If it is longer than 6 bits, its value is at least 64, so arrything will clo. If it is exactly 6 bits, then either the second bit fiom the left (16) or the thircl bit from the left (8) must be 1. If you sec this, then the solution

( 1 1 1+ 1 1 0+ 1 0 1()o + 1 )( 0+ 1 )( o+ 1 )+ 1 ( 0 + 1 )( 0 + 1 )( 0 + 1 ) ( 1+ 0 ) ( 1+ 0 ) ( 1+ 0 ) ( t + 0 ) . readily suggests itself. Section 3.2 S. This can be solved from first principles, without going through the regular expression-to-nfa construction. The latter will of course work, but gives a more cornplicated answcr. Solution:

4. (a) Start with

Then use the nfa-to-dfa algorithm in a routirre rrranner.


372

ANSWERS

7. One case is

n

l l t \ ir'-----\ }{ ( rr )--"-----*( I ) \+/ 11

Since there is no path from qj to qa,the edgesin the general case created by such a path are omitted. The result, gotten by looking at aJI possible: paths, is

Ou.o

\ / t /

qlL **19 The other case can be analyzed in a similar manner. E. Removing the middle vertex gives

bb+ah

The language accepted then is ,L (r) where r : a* (a + b) ab (bb * ab* o.a*(a I b) ab) . 10. (b) First, we have to modify the nfa so that it satisfies the conditions imposed by the construction in Theorern 3.2, one of which is qs f F. This is easily done.


SoLUTroNsnr'ro HrNrs r,.onSnr,ncrnl ExnRcrsns

BZB

Therr rernove statrl 3.

aa+b

Next, removclstate 4. (ab) + (aa + h)(ba)t bb

0

-ai}-r-o \---l

\/

The regular exprelssionthen is r' : (ab * (aa + tr) (Oo)-OA- - . ) 17' (a) This is a hard pr.blem until yorr see the trick. start with a clfa witlr states eotett,.,t irrrd introduce a "parallel" autorrraton with states Q0,4t,.... Then arrange matters scl that the spurirlrs symbol nondettlrministically transfers f'rom any state of the original automaton to the corresponding state irr the parallel part. For exarnple, if pa,rt of the original dfa looks like

tltett the dfa with its para,llelwill be arr rrfh whose corresponding part is

It is not hard to nrakc the argurnerrt that the original dfa rrccepts .L if arrd only if the corrstnrcted nfa accepts i,nsert(L).


374

ANSWER,S

Section 3.3 a. Right linear grarrrrrrar: S-aaA A -

aAlB

ts -

bbbG

C -

bClA

Left Jinear gramma,r: S -+ Abbb A'-+ AblB B --+aaU C --i a(.J1.\ t .

we (:arr show by irxlrrction that if tu is a sententirrl form derivtxl with G, thcrr t tR can be <lcrivedin the sirrrrenumber of steps by G. Beciilrse ur is cretr,tetl with left lincar derivations, it Inust have the form ur : Awr, with A â‚Ź V andrur e T*. By the inductive assumptiorr wR : u,FA can be derived via G. If wc now apply A --+ Bu, then w 1 Btnttl, But 8 contairrs the rule A ---,uRB, so we (larl tnake the dcrivation ll,R -- wfuRB : (Buw)R completing the inductive steP.

l o . Split this into two cases: (i) z and rn are both even and (ii) rr and rn are both odrl. The solutiorr then falls out eir"sily,with 5 --+ aaSlA A + bbAll tnkirrg care of casc (i). rZ. (a) First constnrct a dfa fbr -L. This is straiglrtforward and gives transitious such a^s d (qu,o) : q1,d (qu,b) : ,1, d (qr, o) : (1o,5(qt, b) : qr d (q2,b) * Qo d (qr, o) : t1:1, d ( g s ,a ) : e 2 , 6( q r ,b ) : g r


SoLU'r'roNs nr'rn HrNrs poR Sur,eflret ExnRcrsns

378

with qs the initial and final state. Therr thc' construction of Theorcm 3.4 gives the answer qs _, aq1 lbg2l,\ qr --+ bqnlaqo qz -

aqtlhrlo

Qz-

tulzlbQt

16. Obviously, 51 is regrrlar as is 52. We can show that their union is also regular by constructing tlrc f'ollowing dfa.

\ Ff^f"'L(qtl The condition that V1 ancl V2 should be disjoirrt is cssential so that the two nfa's are distinct.

Chopter 4 Section 4.1 2. (rr) The construction is straightforward, but tedious. L ((a + b) a.) is given by d ( q o ,a ) : q r ,

d ( q o , l r ): g r ,

l i ( g r ,a ) : q r ,

A dfa for

6 (h,b) : qt,

with q1 a trap sta,teand final state 91. A dfa for ,L(baa.) is given by d (po,n) : pt,6 (po,b) : pl,6 (p1ra) : pz, ,f (pr, b) : Ft,6 (pz,a) : pzl6 (pz,b) : p, with Iirral stir.te p2. Flom this we find d ( ( s o , p o ,) a ) : ( h , p r ) , . 1 ( ( q , r , p o, )b ) : ( q r , p t ) , d ( ( q t , p t ) , a ) : ( q t , p z ), d ( ( s r , p r ) , a ) : ( q r , p z ), etc. When we complete this construction, we see that the only firral utate is (qr,pr) and that L ((u + b) a.) n L (baa*) : baa*.


376

Ar-rswHlrs

7. Notice that nor (I'1, Lz) : L1l-) L2. Tlrtl result then follows frorrr closure under irrtt:rsection and complcrntltrtir,tiorr. 12. The an$wcr is yes. It can be olrtairreclby sta'rting fiorn the set identity

Lz = ((Lr u.Lz)n Zi) u (r' 1rz) . Tlre key observer,tiorris that sitrce tr1 is Iirrite, L1 tt l'2 is finite atrd therefore reElrlirr for all tr2. The rest ttrett follows easily fiorn the known closures unrlcr uttiott antl compl(lrrxlrrtatiotr' 14. By closurc urtcler reversa,l, .Lft is regular. The resr.rlt tht:n follows fronr closure untler cottcateuation. X*, which is regula,r.Tlrt: 16. Use trr : X*. Therr, fbr aIV tr2, LylL2: given statement would then imply tha,t :r,ny.L2 is regular' 18. We cim use the following construction. Find a,llsta,tcsP suclt that there is a path fiom tlrt: ittitial vertex to somt: elertrent of P' and fitlrn tltat elenent to a, final state. Then make every tllrlrrrent of P a fina,l sta,tt:. a n d G z : ( V z , T , 5 2 , P 2 ) - W i t h o r r t l o s so f 2 6 . S u p p o s e( J r : 1 y t , 7 , 5 r , P r ) genertr,lity,we calt a"ssumethat yl arrd V2 are disjoint. Corntlirre the two gra,mmars aild (a) Make 5 the new sttr.rt sytrtbol and add prodrrt:tiorrs.9 - 51 lS2' (b) ln P1, replace cvury production of the flrrrn A+ frt with A € y1 atrdr€?*,by A+:rSz. (c) In Pr, replar:t:every production of thc forrn A+ rt with A e V1, ir,ndz € 7*, by A - r:Sr. Section 4.2 1. $ince by -Exarnple4.'l L1 - L2 is regular, there exists a' rrtetnbership algorithrn for it. 2, If Lt c .L2,thetr L1l L2: Jlz. Sint:cL1l L2 is regular and wtl have att algoritlrrn fbr set equality, werillstt have att algorithm for sct irrclusion, 5. Frotn the clfa ftrr.L, cotrstruct the tifa for -Lft, usitrg the constnrt:titlrt suggestedin Thcorern 4.2. Then use thc rxprality algorithm in Thcxrretrt 4.7. 12. Htrrc you need a little trick. If .L contains Ilo even length strings, theu L n L ((u.u1 ab * ba I bb).) = a. The Ieft side is regula,r,$o we can use Theorem 4.6.


Soluuons .qxo Hrw,r.sr..onSelncrrnn Exnncrsns

377

Section 4.3 2. For the dfa ftrr -L to l)r(x:e$rJthe middle string u reqrrirt:s ir. wa,lk in the transition graph of lt:rrgth lul. If this is longer than thc rurrnber of sta,tesin the clfa, Lherermrst bc ir,cryclelabeled E in this walk. But r:leir,rlythis cycle can be repcatrxl a,soften as desired wilhout charrgirrg the acceptability of a strirrg. 4, (a) Givcxrrn, pir:k rt.t: ettzb't'a2"',The strirrg,ryrmrst then he aa and the putnpcxl strings will be 'ut; :

a,m't

(i'* I)k r,rnt O2rn,

If we ta,kei ) 2 then nL+('i - i)ft > rn, then u; is not in /:. (e) It does not seerneasy to apply tht: prrmping lerrrma,directly, so we prcltxlerdindirectly. Suppose that 1, wert: rrlgrrhlr. Then by the closure of regular languages under c:ornlllrlrnonttr,tion,Z would also be regular. Brrt Z : {to ; n,o(w): no (u)} whit:h, rrsis ea"rilyshown, is not regular. By contradiction, .L is not regular. 5. (a) Ibke p to bt: tlxr srna,llestprime number greater or tlqrrirl to rn rr,nd clroose'Lt): trP. Now ig is a,string of a's of length fr;,so lhat wi :

aP+\i-t)h.

I f w e t a k e i - 1 : p , t h t , r rpr + ( i * Irtplr is nclt irr ttn lirngrrir,ge.

1 ) f u : p ( h + 1 ) i s c o m p o s i t ea r x l

E . Tlre prolrosition is false. As a courttc:rtlxilmple,ta,keL1 : {a"b* : n, 1 m,} tr,rrrlL2 * {att6nt ; n, > rn}, bot}r of whic}r arrl rulrr-rcgrrltr,r.But Lrt-lLz : L(a,*h*), which is regular, 9 . (a,) The Iauguage is regular. This is nxrst qilsily seen by splitting the

p r o b l e m i u t o c a s e ss u c h a s l : 0 , h : 0 , r t cottstnrr:t rcgrrlilr expressions.

> 5, firr which one can ea,sily

(lr) Tlris lir.ngua,ge is not regula,r. If we chooseu : a(Laauab-a,-, our opporx.'rrthrr,ssevera,lchoices. If y consists of orrly o's, we use i : 0 to violatr: thc txrndition n ) 5. If the opponent choosesgr as consistirrg of b's, we carr thclrrviolir,tothe condition ft < L 1 1 . tr is regular. We see this frorn L : h

rf .Lf rrnd tlx,r known closures fbr

regular Ia,nguages. 1 S . (a) The larrguagc is rcgular, since anv string tha,t has two consecutive symbols the sarne is in the languagc. A regrrltr,r expreusion for "L is (a + h) (a * b)- (aa + bb)(a + b) (a + b)..


378

ANSwER,S

(b) The language is uot regrtla,r.Take '7,t: (ab)'"'aa(ba)"'. Tlte adversary llow has severir,lr:hoit:cs,such as y : (.ab)kot ,'tJ: fub)k'a, In the first case uro = (nb)--k aa(ba)"'' . 'tl6 is trot in 1,. Since the ouly possihlt: idtlntification is ururft = hta,o,ltl , so it r:irtrnot Lrein 1, With the second choirur,thc lettgth of u0 is odd, either, 1 5 . Take .Li : tlb' ti : 0, 1, ,.,. For ea,chri, -t; is firrite atrd therefbre regtrlirr, brrt the utriotr of all the langtrilgcs is the notr-regula,r languir.gc tr : { a ' b " : r z> 0 } . L 7 , No, it is not. As cotrttterexantple,take the lir,nguirges Li:

l z} ,2: 0,1,2,... { u i u u f ; : l t r i l= i } u { r . ; r f : l u d <

We claim that the rrrriorrof all the .Li is the set {ttrtrrft}. To justify tlris, take any strirrg z : lttu)Rt with ltul : n,. If rr. ) ti, then a E { u i u , u f ; : l o i l = i } a , r r r l t } r c r e f o r e i n . LI af .r z < z , t h t r nz e { t t i u f ; l u l l < z } , z: {0, 1,2,...} ilntl so also itr -L,;.Consequently,e is irr the uniou of a,ll t,ht: Li. Corrversely,take any string.r of lt:rrgtlt rtl that is in a'll of tht:.La. If we take i greater tha,n m,, .s (:irrrrlotbe in {tizuon: ltr,l = i} trec:auseit is not Iong enough. It must tht:rcfrrrctbe in {lr1u,f ; luil < i}, so t}trt it has the form tutlft. As the final step we must slrow that for each i,, Liis regultr.r.This follows from the fa,ct that for circh i tltere are onlv a' finite mrrntttlr of substrings ui.

Chopter5 Section 5.1 a. It is qrritc otrvious that atry string genera,texlby this grarnmar has the silmc rrurlrber of ots as b's. To show thtr,t thtl llrelix conclition no (u) > n6 (u) holds, we carry out an induction on thtt lcrrgtlt of ihe derivation. Suppose that for every sentential fbrm dt:rivcd frotrt 5 itr rt, steps this corrdition holds. To get a sententirr.lforrn irt rl * 1 steps, we ca'n a,pplv S - .\ or 5 - 55. Since neitlrtlr of these chatrgesthe nurnber tlf a's arrd b's or the location of those irlrt:ir.rlythere, the prefix condition txlrrtirlrt:s to hold. Altcrrratively, we apply S - a5b. This ndds arr extra a attd an extra b, tlrt sirrce the addecl a is to the left tif tht: adtled b, the prefix r:orrdition will still be satisfied. Thrru, if tlrtl prefix cotrdition holds after z steps, it will still hold a,fJern { 1 stt'ps. Obviously, t}re prefix conditiorr holds after one step. so we havc a trirsisarrd the induction succeeds.


SoI,urroNseun Hrrurs poR SslucrHD FlxgRcrsEs

379

7. (a) First, solve the ca$e n, = m * 3. Then add rnore b's. This can be done by S -

ttaaA

A + aAblB B - - +B b l I But this is incomple,te since it creates at Ieast tlrree a's. Trr tir,ke care of the casesz : 0, 1,2, wt: add

5 - AlaAlaaA (d) This ha^sarr unexpectedlysimplesolution S + aSbbla^9bbblI. Theseproductionsnondeterministicallyproduceeither bbor bbbfor cach generatecla. (a) For the first c&serz : ?n arrd ft is arbitrary. This c;rn be achievedby St: AC A + aAbl\ C --+Ccl,\ In the secondcase,n, is arbitrary and m 5 h, Here we use Sz --+BD B -+ rr,BlA D - bDclE E - Ecl),. Finally, we start productionswith S -t Srl,Sz. (e) Split the probltlrninto two ()ii*,Je$: n,:h*m and rn: h+rz. The first caseis solvedby 5 - a,9cl51l.\ Sr - aSrbltr. L2. (a) If S derives.L, then ,5r -

55 derives-L2.

1 5 . It is normally not lrossihleto use a grammar for .L directly to get a

grammar for Z, so we rreedanother, hopefully recursivedescriptionfor


380

ANSWER,S

Z, ttris is a little harrl to stle here. One obvious subset of 1, contains tire strings of odd length, trrrt this is not all. Suppose we hrr,veirrr cv(llt length string tha,t is rrot of the form uuft. Working fiom thc r:crrtcr to the left and to the right sirrrultaneously, compare corresporrrling symbols. While some part irroutrd lhe center ca,nbe of tlte forttr wwR , at some point wc get aII a otl tire left and a b 'I'he string must in ther rxrrrcspondingpla,ceon the right, or vice versa. tlrcrefore be of the fortn utturRbu rsr ubwwRau with lul * lul. Orrct: we see tlris, we can then corrstruct grailtlna,rsfor these typtls of strittgs. C)nesohrtion is

S - ASAIB A -+ a,lh B -

b](rla]b

C -

aCa |bCblA.

The first two productiotrs geuerate the u and u, thc third the two disagrecirrg symbols, atrd the last the innttrmost palitrdrome. 19. The only possible derivations stirrt witlt S + aaB + aaAa + aabBba + aabAabu.. But this scrrterrtial form has the suffix abc,so it cer.nnotpossibly lead to the senterntlcaubbabba,

22. E --+E * E lE.EIE.l(E)llls. Section 5.2 2. A solution irr S --+aA, A --+a.ABlb,ts Note that the more obvious grammar 5 --+ a^9rB

5r --+a5rBltr B-b is rrot arr s-grammar. 6. There are two lef'tmost derivatiotrs for w : aab. 5+aaBlaab S+AB+AaBlaaB+a,ab.

b.


Exnn.crsns Souluoxs ar'rl HrNrs FoR,SELECTET

L

381

Ftorn thtl dfa for a rcgrrlilr limguagc w(] cirrl gct a rcgulirr gritrnrnat by the rrtt:thorl rlf Th(xlrcrn 3.4. Thcl grirrnrnirr is itrr s-grittrtrnar except for {l - .\. But this rulc rlotls rrlt crclatc irrqy arntrigttity. Sirrce the dfa never has a choice, there is never arry choice in the procluction that can be applied.

r4. Arntriguity of the grir,mmilr is obviorrs fiom thu rklrivatiorrs S+aSb+ah S +,5,9 + abS+ ab. An orltivir,lont rrrril,rnlligrttlrrtt gri.rrnrnirr i$

5 - Al,\ A - aAblablAA. It is not ea,syto see tha,t this grammar is unambiglr()lrir. To rnir.kt:it plansible, consider the two tvpical situations, ut = aahb, whictr r:ir.rrorrly be derived by starting with ,4 -n aAb, and ti., * uhab, whitfi carr otrly be derived stir,rting with A --+AA. Morc cxrrnplicated strings are built frorn these two situations, so they can be parsed only in one way. 20. Solution:

S - aAlaAA A - bAblhh.

Chopter6 Section 6.1 3 . Use lhe rule in Theorem 6.1 to substitr,rtefbr B irr tirc first grantrnar. Then B becomesuselessand the associatedprodtrt:tiorrsc;ilrrbe retrtoved. By Theorems 6.1 and 6.2 thc two grilrnmirrs are equivalent. The only nullable variable is A, so removing A-prorlx:tiorrs gives S - - +u A l u l a B B

'?

,li

A - aaAlau B --+hClhbC

'i*hl h

,,fi

'

r |

i",d

f i A

\,s

b

"*\> '8

yl

a?

C-8. C -

B is the only unit-prodrrction rr,ndllrnovitrg it results in

fi --+uAlalaBB A + ae,Ala,a B --+b}lbbl C --+bClbb1.


382

ArqswnRs

Finally, B and d ere useltltss,so we get S - uAl(r A -- aaAlaa. The language generated hv this grarrrrrrar is I ((aa)- a) . 14. An exa,mpleis S+aA

A- BB B - aBbl\. Whenwe removetr-productions we get S-

aAla

A - BBIB B - aBblab. 1 6 . This is obvious sincxr the rernoval of rrseless Drocluctiorrs rrever adds anvthing to the grammar. 2 1 , The grammar ^9 --+ aA; A --+ a does not have any useless productions, any unit pr()ductioruJrer any A-prodrrctions. Brrt it is not minimal gince S - aa is an equivalent grammar.

Section 6.2 5. First we must eliminateA-productions.This gives S - - +A B l B l a B A-aab B - bbAlbb. Tltis has introduced a unit-production, which is not acceptablcin the constructionof Theorem6.6. Removalof this unit*productionis easy. S - AB lbbAlaBlbb A - aa.b B + bbAlbb. We can now apply the construction and get S --+AB lVhV,AlV"BIV,V, A --+V,rV6V1, B + V7,V6AIV6V1,


Souluoms eNl Hrr'rrs poR,Srlncrno Exnn,crsns

383

a,nd

S - ABlV.AlWBlVbVb A -

VaVa

g - V,.AIV,V, V' - V,V, Vt - VnV, W * a Vn-b' 8. Considerthe generalfonn for a production in a linear grarnmar A -

a t a z . . . aB nb1b2...bru.

Irrtrodrrce a new variable Vr with the productions Vr + az...anBbft2...b-, a,nd A -

atVr.

(lontirruc this process, itttroducitrg V2 alrd Vz --+QB:..(trn Bb1b2...br,, irrrrl so rlrr, until no terminals remain on the left. Then use a, simila,r I)ro(lessto rernove terminals on the right. 9. This normal fbrm r:arrbt: reached easily frorrr CNF. Productions of the form ,4 --+BC a,reperrrnittr:d since a : .\ is possible. Fbr A - a, create new varitr.bleluI/r, Vz attd productions A - aVtVz, tr/l - tr, Vz - A. 12. Solrrtions:5' --+aV6laSluV"S, W -

a, Va -

b.

15. Only A --+\ABC is not in the required form, so we introduce A - bAV and V + BC. The latter is not in correct form, but after substituting filr B, wtl hirvc S-aSA A-bAV V-bC C -

aBC.

Section 6.3 2. Sirrce aab is a prefix of the string in Example 6.11, wc catr use the yij colrtputed there. Since S â‚Ź Vrir, the string arlb is irt the latrguage generated by the grammar and can thercft)rc be parsed.


384

ANSWF]Rs

Frrr ptr,r$ing,we determine the prodrrt:tions that wttrc trscrl irr jrrstifirrg5eV13: S e V r r b e c a , u sSe -

AB,with 4 € V1l and R eVzt

A €, V11 b e c a u s eA -

a

B €V23 becauseB -

AB, with A € Vzz,B e Vn

A € V 2 2 becauseA-

a

B e V r r bet:a,trse B + b.

This showsa,llthe productionsneedecl to justify membership;thesecan thcrr bc usetlirr thc parsirrg S+AB+aB+aAB+aaB+aab.

ChopterZ Section 7.1 2. The key to the argument is the switch frorn q6 to q1, which is done nondeterministically and need not happen in the middle of the string. However, if a,switch is made at some other point or if the input is not of the ftrrm ur?rrft,iln a,cr:eptingconfigura,tion ca,nnot he reached. Suppose tlre ccrrrtcrrttlf thcl stirck ir,tthc timc of the switt:h is n112...np.z.Trr arx:ept a string we rnust get to tlte cotrfiguratiott (r71,.\,.e).By exarrrirrirrgthtr transition functiotr, we see tlut we carr get to this tlorrfiguratiort orrly if at tlris point the utrread lrirrt of tltc irrptrt is r1:r2...rp., that is, if thc original input is of the forrn 'ururEand the switch was rnade cxactly irr the rniddle of the input string. a. (a) The solution is obtained by letting each a put two markers on the stack. while eaclt b corsurrresorre. Solution: d ( g o ,I , z ) : { ( q 1 ,a ) } d ( q o ,a , z ) : { ( q 1 ,1 1 a ) } , , 1 ) : { ( q 1 ,1 1 1 ) } d (qoo d ( q r ,b , t ) : { ( S t , I ) } d ( , 1 rt,r , u ) : { ( , 1 i ," ) } . (f) Here w€)u$e nondeterminism to.genera,teone, two, or three tokens by d ( q o ,a , a ) : { ( q r , 1 " ) , ( q r , 1 1 a ), ( q 1 ,1 1 l a ) } arrtl , , a ) : { ( q r , 1 1 ), ( g r o ,1 1 1 ), ( q 1 ,1 1 1 1 ) }. d (qoo


Solu'r'rons Ar'rDHlNrs rron Snl,l:c.r'r;rr llxnRcrsns

385

The rest of l,]resolrrtiorris then essentially the $a,mea"s4(a). L This is n pda that makes rro us(] of the stack, so that is, irr trfl'ec:t,a, finite acx:t:pter. The state transitiorrn r:a,nthen be taken directly from the pda, to givc ,l (go,a) : qr r) (qo,b) : gt d ( q r ,a ) : c r d (gr, 6) : ,70 11. Tlace through thc: pror:ess,taking one path rrt ir time. The transition fiom qs to 92 can bc rnade with a single a. The altcrnative path requires olle a,, followed by orrrl or more b's, terrninated by irrr n. These are the only choices. The pda thercfbrc a,cceptsthe Ianguage L:{"}DL(tbh*a). 14. Here we are not allowrrd enough states to track thrl switr:h from n,'s to b's and back. T(r ovtrrcome this, we put a syrnbol irr the stack that rernemberswherc irr the sequerrcewe are. For exarnplrr,tr,solution is d ( s o ,a , u ) : { ( S o ,1 ) } , 6 ( e o , a ,t ) : { ( S n ,t ) } , d (,10,lr,1) : {(So,2)} , d ( r 1 1a,, 2 ) - { ( q o , 2 ) }, d ( Q oA, , 2 ) : { ( q l , Z ) } . We have only two states, thc initia.l state qs and the acceptirrgstate gr. What wor.rld normally be trackrxl hy different states is now tracked by thtr symhol in the stack. 16. Here we use internal states to rcrnember symbols to be put on the stac:k. For cxilmple, 6 ( t u . i , a , h:) { ( q i , c d e ) } is repla,t:edby 6 ( q o , o , b :) { ( q i . , d e ) } d (qj", tr, d) : {(qi, cd)} . Since d carr hirvrl only a finite number of eletrerrts and each can only add a finite arnount of infbrmation to the stack, this rxrrrstnrction ca,n be ca,rriedout for any pda.


386

ANSWERS

Section 7.2 'I'heorem T.l tlr you can trotice that 3. Ytnr carr follow the constnrt:tiorr of tlre language is {a"+?bzntl :'n, > 0}. With t}re latter observatiorr wc get a solution d ( q i 1n, , z ) : { ( S r ' a ) } d ( q r ,o , a ) : { ( q z ,z ) } 6 ( q r , o , a ) = { ( q 2 ,1 1 . z ) } 6 ( q r , o , 1 ): { ( q 2 , 1 1 1 ) } 6 ( t 1 2 , bt ), : { ( q r , I ) } d ( q r , l , ,1 ) : { ( q r ' A ) } ,l (,1*,,tr,a) : { (q.r, a) } where qs is the initirr,l state atrd q1 is the finer,lstirtc:. 4. First txlrrvert the gramtnar into Gritlbach trortnal form, giving 5' aS,5^5';,l - aB; B - b. Then ftrlkrw the cotrstruction of Thslrr:rrt 7,1' d ( q t ' ,A , a ) : { ( q 1 , 5 ' r ) } d (qt , o, ,5) : { (q1, SSS) , (qr , B)} d(qr,b,f) : t(,1r.,I)] d (q', .\,") : {(qr. e)} , 'Iheoreur 7. Florn 7.2, givcn arry trpda, we call constrtrt:t arr equivalent cotrtext-free grammar. Frorrr that gramlnar we can tlrtlrr construct an equivalent three-sta,trlrrlltla, usitrg'Iheorem 7.l. Be<:auseof the trausitivity of equiva.klrrcltl,the original and the fintr,l rrpdrr's are also equivaIent. L We first obtain a grirnrrrrarirr Greibach normal filrm for tr, for exantple S + o,liBlb,B - b. Next, we apply thc corrstructiou in Theortrrn 7.1 to gct art npda with three statcs, (lo)(h)qI. I'he state tJ1ttirrl bc clittritratecl if we use a special stttck syrrtbol z1 to rnark it. A r:ornplctesolution is .)-(qn,tr, s) : { (qo,5'sr) } , i ( , 1 no, , S ) : { ( q o ,S B ) } d (r;n,b, S) : { (qo,I) } 6 ( q o , h , B ): { ( q n , I ) } d ( q o ,A , a r) : { ( s i , I ) } . 11. Thcrc rnust be at least one a to get started. After tha,t, d(g6,a,A) : {(Sn,A)} simply reads o,'s witltout changing the stir,t:l<.Fitrall.y, when


Soluuow$ er'rnHlFrrsnon SnlncrnD ExliR{trsus

387

the first b is encountered, Lhe prltr goe.'sinto state q1, from whir;h it ca,n only ma,kea ,\-transitiorr to tlxr final state. Therefore, a strirrg will be ac:t:cptedif and only if it consistsof one or more {r,'s,followed try a single

Section 7.3 4, At first glauce, this tna.vseerrrto lle a noncleterminislic larrguir,gc,since the prcfix q, ca,llsfor two different types of suffixes. Nevertheless,tht: Ianguageis tlcltr:rrnirristic,as we can constrrrc:ta rlpdtr,.This dpda, goes into a final statc wlxrrr the first input sytrbol is arr a. If more symbols follow, it gclrrsorrt of this state and thcrr ir:rnpts at'bt. Complete solutiorr: d ( q o ,a , a ) : { ( q 3 ,1 z ) } d ( q Ba, , i ) : { ( q 1 ,1 1 ) } b ( Q r , a , 1 ): { ( q 1 , 1 1 ) } d(q',b,1):{(s1,r)} , , a ) : { ( q z ,a ) } d ( , 1 rA whereF:{q2,qJ}. 9. The solution is straiglrtforwrrrrl. Prrt a's and b's on the stack. Thc: c signals the switch from savirrg to rnr.rtching,so everything can be dorrc dctcrrnirristically, 11. There are two stiltcs, the initial, non-accepting statc 96 ir.nd the final sta,te q1. The pda will bc in sta,te qr unless a a is on top of thc sl,ack. Wlrcrr this happens, the pda will switc;h stirtcs to q0. The rest = is essetrtially tlrtl sitrnc nu Exa,mple 7.3. 'I'hus we have d (t111,a,,2) : : ( s 1 , O t ) } d a , , 0 ) ( q r , , t r , e t c . w i t } r d " ) a ) , {(qr, t(s,,00)}, {(so, }. When you write this all out, you will see thai the pda is deterrnirristic. 15. This is ohvious since every regular languirgc r;rlrr be tr,r:ceptedby a dfa artrl such a dfa is a dpcla with an unused stir.r:k. 16. The basic idea here is to corrrbirrcrr,dpda, with a dfa along the lines of thc rxrnstrrrctionin Theorem 4.l, with thc sta,r:khnndled as it is for.L1. Il should rmt tre too ha,rclto see that the result is a durlir. Section 7.4 2. Consider the strings aabb and aabbbbua^In the first case,tire derivation rrrrrststelrt with 5 + aSB, while in the sccxrrxl5 + SS is the necessary Iirst stt:p. Brrt if we see only the first four syrnbols, wo r:ilnnot decicle wlric:lr t:ir"st:a,pplies. The grammar is thereftrrtt rrot in LL(4). Since


388

ANSWTTR$

similar exa,mplcsc;arrbe made for arhitrarily long strings, the grirrntnar is not /,I (k) fbr any A. aab, ct aba, then the 4. Ltrok rrt the lirst three symbols. If they are (r,tl,(t,r are abb' then lf the first three svrnbols c:an only be in.L(a.bn). strirrg atty parsa,blestring rmrst be in .L (abbb.). Ftlr t:ach case' we can find a,rr -L.L grammar anrl the two can be combintxl irr an obvious fa,shion. A sohrtiorr is .5 - 51152 ,91-'+ nS1lba Sz + ahbB

B - bRlA. Looking a,t the first tlrree symbols tells rrs if S + necessrrry.The gramtnar is thereftrrt: LL(3).

51 or S =* 5'z is

7. For a detcrmirristic CFL there exists a, dpda, When this dpdir. is converted into ir,grarnrrar, the gra,nlmtrr is urrirtnbiguous.

s. (a) S - a S cl 5 1A l '5'1--+ bS1cl.\. This is almost a,n s-grammar. As lotrg as the currentlv st:arrrredsytnbol is rl, we must apply S - aSc, if it is b, wc rnust use 5 - 51 , if it is r:, wc carr otrly use 5 - .\. The gra,mrrraris .L.L(1).

ChopterI Section 8.1 3. Take ut: u,"'b"'brrlartlattlbttr.The ir,dvcrsarylrow has several r:htlit:tls tltat have to be consiclered.If, ftrr exaurple, u: ak and't7: tl{, wit}r u atrd E located in the prcfix a*, tltett 1!)O :

Q,''*h-L6nt6m

OrnOut6rtt,

which is not in -L. There a,reil rurrrlberof olher possible choitxls,but in all casesthe string ca,nhe purnped out of the langua,ge. The only ?. (a) Use the pumping lcrntna. Givetr rrz, pick u., :4,t26m. choice of u irrrd g that treeds any serious tlxarnirratiott is tr : aft and


Solurrons

AND HrNTS FoR SELECTEnExnRcrsps

A : b I , w i t h k a n d I n o n - z e r o .S u p p o s et h a t l : 1 . so that tr2 ha^sm2 + h a's and rn * 1 b's. But (rn+I)2:mz

389

T h e n c h o o s ei : 2 ,

+Zn-t,*I

>-mz+k' Since ru2 is not in the language, the language cannot be context-free. Similar arguments hold a fortiori for I > 1. (f) Given m, choo$erp *qtnfitn+l"m*Z, the language.

which is easily pumped out of

E . (b) The la"nguage is not context-free. Use the pumping lemma with I.u: e,^bn'an'b* and examine variou$ choices of u and .,r7. L 0 . Perhaps surprisingly, this language is corrtext-free. Construct an npda that counts to some value h (by putting h tokens on the stack) and remembers the h-th symbol, It then examines the fr-th symbol in ru2. If this does not match the remembered syrnbol, the string is accepted. If w C L there must be sorne h for which this happens. The npda chooses the fu nondeterministically. L2. Use the pumping lemma ftrr linear languages. With a given rn, choose u) : a*bz*a*, Now u and y are entirely made of a's, so ,ruis easily pumped out of the language. 15. The language is not linear. With the pumping lemma, use u , : ( . . ,( o ) . . . )+ ( . . .( o ) . . . ) where (...( and )...) stand frrr rn left or riglrt parentheses,respectively. If lzl > 1, we can easily purnp so that for some prefix z, n1 (u) < ny (u) which results in an improper expression. Sirnila.r arguments holcl for other deurmpositions. 20. Use'u): qrtQ,wherep and q are prirnessuch that p> m arrd g > rn. If

luul: k, then

lni+rl: pS+ik. If we cho11se i : pq, then wi+L:

which is not in the larrgrrage.

aps(l+li)1


390

Ar-rswurrs

Section 8.2 1. The complemerrt is context-fiee. The complement irrvolves two caseu: n ,, (w) # nu (*) and ?2,](*) # n. (ru). These in turn can be broken into n " ( w ) > n o ( w ) , n " ( w ) 7 n " ( w ) , n " ( w ) < n o ( u ) , a n d n o ( ' u r )< n " ( u r ) . Each of these is context*free as can be shown by construction of a CFG. The full Ianguage is then the union of these four case$arrd by closure under union is c(lrrtext-free. d 5. Given a context-free gra,mmar G, construct a context-fre* gru*** by replacitrg every prochrt:tiorr A - r by A -- xR ' We can then show by an induction orr the number of steps in a derivation that if tu is a sentential fbrm for G then uft is a sentential form tbr d. 9. Given two linear grammars G 1 : (V1,7, ,91, P1) and G2 : (Vz,T, Sz, Pz) with VrftVz : @, form thc combined grammar G : (Vr\)V2,7,^9, Pl U Pz US -i Sr l,Sz). Then d i, Hnnu. and .L (t) : r, (Gr) u L (Gz). Ttr show that linear langrrages are not closed under concatenation, take tlre litrear language 1': {a"b": n > 1}. T}re language .Lz is not linear as can be shown by an application of the pumping lemma. t S . L e t ( J r : i y t , ? , S r , P r ) b e a l i n e a r g r s , m m a rf o r . L 1 a n d l e t G 2 : (Vz,T,Sz,Pz) be a left-linear grammar for L2. Construct a grammar 82 from Gz by replacing every production of the form V + rtn e T* with V --+ Sr?:. Combine gra,mmar$Gr and d2, choosing Sz as a start symhol. It is then easily shown that in this grammar SzlSrwlutu if arrd only if u E L1 atrd tu â‚Ź tr2. {u'nbtrt"rttl are both urlam{arrbtlc.Ill}and .L2 biguous. But their intersection is not evt)rr cotrtext-free.

1 5 . TIre languages -Lr :

2T, I e , (G) if and only if 5 is nullable.

ChopterI Section 9.L 2. A three-state solution thtr.t scans the entire inprrt is

d ( g oa, ) : ( u , u , R ) d (qr,a) : d (sr,h) : (q1,a,R) d (sr,tr) : (gz,!, l?) with F:

{Sr}.


Sot,urrolrs arul Hrmrs l.on Snr,ncrnDExERCISEs

Bg1

It is also possible lo get a two-state solutiorr by just examining the first sytrbol and ignoring the rcst of the input, ftlr example, d (qo,a) * (gz,a, lt) . 7. (a) d (go,a) : (8r, a, ft) , l ( S t , l t ) : ( q z , b ,R ) 6 (qz,a) : (qz,a, R.) 6 ((tz,b): (gs,b, fi) with .F : tqi]. (b) d ( q o ,a ) : d ( s o ,b ) : ( q 1 ,D , f l ) d (qo,tr) : (qz,n, ft) d ( q r ,a ) : d ( q r ,b ) - ( q s ,n , . R ) with .F: {qr}. 10. The solution is conr:eptua,llysimple, but tedious to write out in detail. The general scherntrlooks something like this: (i) Place a marker symbol c at cach end of the strirrg. (ii) Replace the two-symbol combination ca on the left by ac arrcl the twcrsymbol combinatiorr ar: on the right by ca. Repeat until the two c's meet in the middle of the string. (iii) R.emove one of the cr's arrd move the rest of the string to fill the gap' Obviously this is a lorrg joh, but it is typical of the cumbersone wayir in which Thring rrrachirrersoften do simple things. 12. We cannot just search in one direction since we dorrtt know when to stop. We rrrust proceed in a back-and-fbrth fashion, placirrg markers at the riglrt and left boundaries of the searched region arrd rrrovingJthe markers outward. 19. If the final state set .F corrtains more than one elernent, introduce a new firral state g.i and the trarrsitions d (q, a) : (q.r,a, R) forallqâ‚ŹFnndaâ‚Ź1,


392

ANSWHRfi

Section 9.2 s. (a) We catr think of the machine as constitrrted of two main parts, an add,-onerrtachinethat jrrst adds one to the input, and a multiplier that multiplies two numbers. Stltetnatically they are cornbined in a simple fashion.

o.

8.

(c) First, split the input into two eqrral parts. This can be done as suggestedin Exercise10, Section9.1' Then comparethe two parts' symbol by correspondingsymbol until a mismatch is fouttd' A solution: d (go,a) : (er., a, R) , : .l (qo,c) (Qo,c, lt) for all c e X * to} , d (so,l) : (qi,n,.R) . The state qs is any state in which the searchrightinstruction rrray be applied.

Section 9.3 2. We have ignored the fact that a Ti.rring machirre, as defined so far' is deterministic, while a pda can be non-deterministic. Therefbre, we cannot yet claim that Thring machines are more powerful than a pushdown autolnata,

Chopter| 0 Section 10.1 a. (a) The rrrachine has a transition function d:Qxf-8xfx{.L,rt,5} with the restriction that for all transitions d (qz,a) : (qi ,b, L or .B), the condition a: b must hold. (b) To simulate 6(si,u) : (Qj,b,.L) with a'+ h of the standard machine, we introduce new tratrsitions d (qa,a) : (rljz, b, S) and 6 (qiz,b) : (si,b, L) for all c â‚Ź f , and so on.


Sol,u'r'ror.ls exn Hrr-{rsFoR SELECTEIExnncrsns

393

6 . We introduce a pserrdo-trlrr,nk F. Whenever theroriginal machine wants to write n, the new rnachine writes B. Then, fbr each r) (qi,ll : (qi,b, L) wt':ilckl .l (qi, fr) : (qi,b,tr), rrrrclso on, Of course, the origintr,l trarrsitirrnd(S,,!) = (Qi,b,,L) must be retained to handle blarrkstha,t are origirra,llyon the tape. L This dotls rrot limit the power of the rrrachine. For each symbol {t â‚Ź l, we introdut:tr a pseudo-symbol, say A. Whenever we need to preservtr this a, wt: first write A, then return to the cell in question to replace .rl by a. 1 1 . Wc repla,ce

6 (s,,{o,b}) : (qr.c,rt) by 6(qe,d): (qi,c,R) fbralldef-{a,b}. Section 10.2 L. For theforrnad l e f i n i t i o nu s e f 7 ' : I ' x f x... xl andd: Q xlr 8 x fz x {L,nJ"', where rn is the rrurrrbcr of reacl-write heads. One issue to consider is wha,t happens when two reatl-write heads are on the same cell. Thc frrrmrr,ldefinition must provide ftrr the resolution of possible conflicts. To simulate the original rnachirrc (OM) by a standard'furing machirxr (SM), we let 5M have rrr f 1 trirr:ks. On one track we will keep the tape contcnts of the OM, while the other rrutracks are used to show the position of OM's tape heads. tape content ofOM position oftape head # 1 position oftape head # 2

,9M will simulate each move of OM by sca.nningand updating its active area, This exercise shows tlnt a qrrr,'uema,chine is equivalent to a standard Ttrring machine and tlrrt therefbre a queue is a rnore powerful sttlrage


394

ANSw}]RS

'Io simulate a standard TM by a queue machine' device thirn a stack. we can, fbr example, keep the right side of the OM in the front of the queue?the left side in the back. read-write head ,'

c

b

c

d

d

c

f

f

g

tapeof OM

x

b

Simulation by queue

A right move is easy arJwc just remove the front symbol in the queue arrd place something in the back. A left movtt, ltowever, goes against the grain, so the querre t:tlrttents have to be circrilated several times to get everything in the right place. It helps to use additional markers Y itnd Z to denote boundarics. For example, to simulate d (qr,c) : (qi, z, L) we carry out the following stcps. (i) Remove c fiorl the front and add .sY to the back. (ii) Circulate contents to get bzYdef gXa. (iii) Add Z to the back, then circ:rrlate,discarding Y and Z as they crlnte to the front, 8. We need just two tapes, one that mirrors the tape of the OM, the second thir.t stores the state of the OM.

e

b

c

d

c

configuration ofOM

configuration ofSM

SM needs only two states: an accepting and a non-accepting state'


Sor,uuolrs aNn HrNr.sroR SnlncrnD ExERcrsEs

BgE

Section 10.3 S.

(i) fJtart at the left of the input. Rememlrer the symbol hy putting the rnar:hinein the appropriatc state. Then replace it with X. (ii) Move the read-write head to the right, stopping (rrorrdeterministically) at the center of the irrput. (iii) Compare the syrrrbol there with the rerrrernberedone. If they . match, write Y irr thr: cell. If they don't rnatc:h,reject input. (iv) With tlrc rxlnter of the input rrrirrkedwith Y, we carr now proceed deterministic:allg a,lternatively rnoving left and right, corrrparing symbols. For ir,completely deterrnirristic solution, we lirst find the center of the input (e.g. by putting markers at cir(:hend, and rnoving theur inwards until they nrcct).

6. Nondeterministically choose a, value for n. Dt:termine if the length of the input is a rnultiplc of rz. If it is, accept. If an â‚Ź L, then there is $ome n for which tiris works. Section 10.4 3. An algoritlrrn, in outline, is as follows. (i) Start with a copy of the preceding string. (ii) l'ind the rightmost 0. Clrrrrrgcit to a L. Then change all the J.'s to tlrc right of this to 0's. (iii) If there are no 0's, chirrrgcall I's to 0's ancl add a 1 on the left. (iv) Repeat frorn step (i). L L e t 5 1 : { s 1, , s ? , . , , }a n d , 9 2 : enumerated by

{ t r , f 2 , . . . } T h e n t } r e i r u r r i o n r : e r nb e

5 1 U 5 2 : t . s r ,f 1, , s 2f,2 , . . . ] . If sotnc si. : tit we list it otrly orrc:t:. The union of l,]re two s.ts is thereforr: r:clrnta,ble.For 51 x ,92,usc thc ordering in Figure 10.17. Section 10.5 2. First, divide the input by two arrd rnove result to one part of trr,pe.This frcc sperr:einitially occupied by tlrtr input. This space carr thcn he used to slore sur:cxrssive divisors,


396

ANSWITRS

+. (e) Use a three-track machine as shown below. On the third track, we keep the current trial value for lrul. On the second track' we place dividers every ltr.'lcells. We then compare the cell contents between the markers. inFut dividers trial ralue of lwl

g . Use Exercise 16, Section 6.2 to find a grammar in two-startdard form.

Then use the construction in Theorem 7.1. The pda we get from this consumes one input symbol on every move and never increasesthe stack contents by more than one symbol each time. T. Example:

Configurationof OM

-*l-"1 H

-fl

r-l lno l l fI

rr

l e I Steckl

Stack Configurations

[l

---l Stack2

Stachl c:ontains the symbol under the read-write head of the OM and everything on the left. Stack? contains all the information to the right of the read-write head. Left arrd riglrt moves of the OM are easily simrrlated. For example,6(ei,e,): (qi,b,,L) can be simulated by popping tlre a off Stackl and puttirrg a b on StachZ.

Chopter| | Section 11.1 2. We know that the union of two counttr,hlesets is countable and that the set of all recursively enumerable larrguages is countable. If the set of


Sol,utlolrs altn HIt'Irs ron Snlncrno

ExpRcIsns

397

all languages that are not recursively enumerable were also countable, then the set of all langrrages would be cxruntable. But this is not the ca.ge,as we krrow, 6 . Let ,L1 and t2 be two recursively enumerable languages and M1 and M2 be the respective T\rring machines that accept these two languages. When represented with an inprrt u.', we nondeterministically choose M or M2 to processzrr.The result is a T\rrirrg machine that acr:eptsLtQLz. 1 1 . A context-free languageis recursive,so by Theorem 11.4 its complement

is also recursive. Note, however, that the complement is not necessarily context-free. 1 4 . For any given 'u ( L+ , consider all splits 'tD: 7ltLlz...'tlrr,. For eaclt split, determine whether or not wi e L. Since for each'ur there are only a finite number of splits, we can decide whether or not w â‚Ź L+ . 1 8 . Thc a"rgurnent itttcrnpting to show by diagonaliza.tion that 2s is not countable for finite 5 fails becauselhe table in F'igure 11.2 is not square, having lzsl ro*s and l5l columns. lSl columns

When we diagona,lize,the result on the diagona,lcould be in one of the rows below.

Section 11.2 1. Look at a typical derivatioru Sl

aStbB + aaSlbbB1a"S1b"B t on*r6n-7R+ an+rb"+l.B+....

Flom this it is not hard to conjecture that the grammar derives L : { a " + r b " t k , i l } 1 , f r : - 1 , 1 , 3 ,. . . } .


398

ANSwHRS

Fornnlly, the grammar carr be describedby G: (V u r)+ and L(G):

{ r e T * : . s* s

(V,S,T,P), with 5 f

r for any s â‚Ź ,9}.

The rrnrestricted gra,mma,rs in Dtllirritiorr 11.3 irrt: txpivalent to this exterrsiorr becarrse to arry given unrestricted grarrurrar we can always add startirtg rrrlrr.rSp --+ .5, f{rr all .saâ‚Ź S. 7. To get tlfs forur for unrestricted grammars, insert dummy variables on the righi whenever l"l > lul. For example,

AB-C r:anhe repla,cedby A.B --+CI) D - .\.. Thtl equivaklrr(jc irrgunro\nt is straightfrlrwarrl.

Section 11.3 1. (c) Working with context-sensitive grammars is not always easy. fhe idea of a messenger,introduced in Example 11.2, is often usefirl. In this problern, the first stcp is to crtlatt: thc scrrttlrrtial fclrrn o.nBt:nD. The variables -B and D will act as rrrarktlrs irrrcl urcsscrrgcrsto assrrrc that the correct nurnber clf b's and dts are r:reaterl irr the riglrt places. The first part is achievtxl casily with the produr:tiorrs S + aAcDlaBcD A + a.AclaBc. Irr the rtext sttlll, thc: B travcls to thc right to nxret the D, bv Bc --+cB Bh --+hB. When that happens, we can create one d and a return messengerthat will put the b in the right place a,rrdstop.

BD - Erl, cE-Ec bE-Eb aE - ub.


Sol.urroNs alo HrNrs ron Snrlcrrr

Exnncrsls

399

Alternatively, we create a,d plus a,marker D, with a diff'erent messenger

thatcr.*t.sa b,but**" tT;:T;;*, tF

4.

--+ Fr

b.F -

I-b

aF -

ubB.

'l'he

easiest argument is from an lba. Suppose that a language is 'I'hen context-sensitive. there exists an Iba M that acceots it. Given ?u, we first rewrite it as tuft, then apply M to it. Because LR : ( LR . 'I'he rnachine {w : wR â‚Ź. L}, M accepts trrn if and only if w tha,t reversesa string and applies M is a,nlba. Therefore -Lft is contextsensitive.

6. Wc (:arr rlrglle frorn a.n lba. Clearlv, there is an lha tha,t can recognize arty strirrg of thc ftrrtr u.r'url. Just stirrt irt oppositt: enrl$ a,nrl cr)rrrpiirc syrnbols rrrrtil yorr grtt ir mat<:h. Sirrce there is tln lbtr,, the larrguage is cxrrrtext-sen$itiveemd a. rxrntext-sensitive grammar must exiut.

Chopter| 2 Section 12.1 3. Given M and tu, rnoclifv M to get fr, which ha,lts if ancl only if a sllcr:ial syrntxrl, say arl irrtrrlrluc:ctl syrntrol ff, is writterr. We carr do this by changing the halting configurations of M so lhat every one writes f, then stops. Ihus, M halts implies the M writes f, and M writes 'Ihus, if we have an algorithm that tells us f implies that M halts. wlrether or not a specified syrnbol o is ever written, we apply ft tu fr with a: f . This would solve the halting problem. ?. Given (M,w) rnoclify M to fr so that (M,w) halts if and only if fr accepts some simple language, say {a}. This ca,nbe done by M first checking the input irnd rurncmbcring whcthe:r thc input was a. Then M carries out its rrorrrral cornputatiorrs. When it halls, check if the input was a. Accept if so, reject otherwise. Therefore M accepts {c,} if and only if M halts. Now construct a simple Ttrring ma,chine, say M1 , that accepts rz,.If we had an algorithm that checks tbr the equality of L(M). If L(fr\: t w o l a n g u a g e sw, e c o u l d u s e i t t o s e ei f L ( f r \ : \ , / \ , / L(Mr) therr (M,.u.,)halts. If I (,fr) # L(M) then (M,ru) does not halt arrcl we havc a solutiorr to tli

trattirrg problem.

10. Given (M,w) we modify M so that it always halts irr the configuration eyw. If the giverr problern was decidable, we could apply the supposed algorithm to the rnodified rnachine, with configurations qgu, and qf?r,. This would give us a solution of the halting problem.


400

ANSwER,S

1 3 . Take a universal T\rring machirre and let it simulate computations on an empty ta,pe. Whernevtlr the sintulated computations halt, accept the T\rring rnac:lftrc:bcirrg sittrulated. The universal Ttrring machine is therefore atr accepter for all Ttrring machines that halt when applicd to a blank tape. The set is therefbre recursively enumerirbk). Suppose now the set were recursive. There would then exist rr,ntrlgrr rithrn A that lists all Ttrring urachines that halt on a blank tape input in some order of increasing lengths of the program. See if the.original Ttrring machine is amongst the T[rring rnachines generated hy A. Sincc the length of the original program is fixed, the compa,rison will stolr when this length is exceeded. Thus, we have-a sohrtion to tlx: blirrrk tape halting problem. 1 6 . If the specific instances of the proble[ fl,rap1,i?2,...,pn, wQ corr$trlr(]ta Turing machine that behaves a^rf'ollows: if problem : gr1thr)n rcturrr firlst: if problem - pz then return true : if problem : p, then return true Whatever thc truth vahres of the various instances are, there is always some T[rrirrg machirre that gives the right answer. Remember that it is rrot nccessary to know what the I\rring machine actually is, only to guarantee that it exists. Section 12.2 3. Suppose we had rr.n ir.lgoritltrrr to decide whether or not L(ML) C L(Mr).We could therr construct a machine M2 such that ,L (Mr): a and apply the algorithm. Then L (Mr) Q L (Mz) if and only it L (Mr) : fi. But this contradicts Theorerm 12.3, sint:e v/e can construct M1 frorn anv given grammar G. 6. If we take ,L (Gz) : E*, the problem becomesthe question of Theorem 12.3 and is therefore undecidable. E. Since there are $omc grirmrrrars for vrhich -L (G) : L (G). and some for which this is not ${r, the undeciclability follows from R.ice's theorem. To do this from first principles is a little harder. Take the halting problem (M,*) and modify it (along the lines of Theorem 12.4), ry that if (M,*) halts, M will accept ia]" and it (M,w) does not halt, M accepts fl. FI'om M get the grammar G by the construction leading to

Theorem 1 1 . 2I.r t ( f r )

= { o } * ,r h e n . L ( d ) = r ( d ) . :

{a}*.But


Solurrons errp Hrrurs noR SnlncreD ExERctsHs

401

fi L (fr): o, then , (t) : @ and , (")- : {A}. rherefore, if this problem were decidable, we could get a solution of the haltirrg probletn. Section 12.3 r. A PC-solution is u3u4w1 : uzutat. There is no MPC-solution becarrse one string would have a prefix 001, the other 01. 3. For a one-letter alphabet, there is a PC-solution if and only if there is some subset J of {1, 2,...,n} such that

flrrl:fluil. JEJ

JE,I

Since there are only a finite numher of subsets, they can all be checkctl and therefore the problem is decidable. 5. (a) The problem is undecidable. If it were decidable, we would have an algorithm for deciding the original MPC-prohlem. Given u)Lj'trz...) n'n, we form tuf, w{...,u,f afld u$e the assumed algorithrn. Since wrwi...uk: (w[...wfu,f)f , tte original MPC-problern has a solution if and only if the new MPC-problern has a solution.

Chopter| 3 Section 13.1 2. llsing the firnctiorr subtr irt Example 13.3,we get the solution greater(*,A) : subtr(1,subtr(I,subtr(r,y))) ' 7.

s f u , a ): m t t l t ( * , 9( * , g -r 1 ) ) , g(r,0):1. s' (a) A ( 1 , y ) : A ( 0 , A ( 1 ,-s 1 ) ) :A(1,y-1)+1 :A(1,y-2)+2 : . 4( 1 , 0 )+ E :v12-


4OZ

AuswnRs

(b) With the resultsof part (a) we can useinductionto provethe next identity.Assumethat for A : I,2,...,D- 1, we haveA (2,a) :2U + S. Then A ( Z , n ' ) : A ( 1 ,A ( z , n - I ) ) :A(1,2n+1) :2n* 3, fiom part (a), Since A(2,0): A(1,1) -3, wtt havqtir basisand the equation iu true fbr all y. 15.If 2"'IA-3:0, t h e uA : 3 2 ' " . T h e o n l y v a l u e so f r t h a t g i v ea positivegrare 0 and 1, so the doma,inof p is {0, 1}, giving a minimum value of g : 1. Tlrt:rcfrrrc pa(2"*s-B):1. Section 13.2 1. (b) UseCr : {a,b,c}, Cx7: {ru} and A : {c}. The non-terminalr is usedas a boundary betweenthe left and right side of the target string and the two ru's are built sirnultaneouslyby V1rV2--+V1ar.V2{rlVrhrVzblV crVpl At the end, the r: is removedby VnVz + VVy 3. At every step, the only possibleidentification of Vr is with the entire derivcd string. This results in a doubling of the string and n:

{a'",2 > 1}. t )

5. A solutionis V 1, r V 2 : V r - V 1 I + V 2 : V t V z V1+ V2 : Vt -+ Vt * VzL: VsVr. tFor exanrple 1 * 1 : 1 + 1 1 +1 : 1 1+ 1 1* 1 1:. 1 1 1 1 , and so on.


Sol,urtot-rsann Htrvrs noR Snl,ecrnD ExERcrsEs

403

Section 13.3 1. P1 ^9* ^9r^92 P2 5r --+a5r, Sz + aSz P3 ^9r-

b,9r,Sz - bSz

Pa ,91 --+ .\,,92 --+ .\. 5. The solution here is reminiscent of the use of messengerswith contextsensltlve grammars, ab-+n nb+br rc -

,\.

8. Although this is not so easy to see, this is one way to solve Exercise 7. Take any string, say a255. This can be derived from a127 by applying & + &&& once and (t + ee, 1?6 times. Then a127 can be derived from a63 in a simila.r way, and so on. Thus every string in L (aa*) can be derived.



f o r F u r t h e rR e o d i n g

A. V. Aho and J. D. Ullman. 1972' The Theory of Pars'ing, TTanslation, and Compi.ling. Vol. 1, Englewood Cliffs, N.J.: Prentice Hall. P. J. Denning, J.B. Dennis, and J. E. Qualitz. 7978. Machi,nes,Languages, and Computafion. Englewood Cliffs, N.J.: Prentice Hall' M. A. Harrison. 1978. Introduction to Formal Language Theory. Reading, Mass.: Addison-Wesley. J. E. Hopcrofh and J. D. Ullman. 1979. Introduction to Automa,ta Theory, Languages and Computafion. Reading' Mass.: Addison-Wesley. R. Hunter. 1981. ?he Des'ign and Constructi,on of Comp'ilers' Chichester, New York: John Wiley. R. Johnsonbaugh. 1996. D'iscrete Mathematics' Macmillan.

Fourth Ed. New York:

Z. Kovahi. 1978. Swi,tching and Fi,ni,te Automata Theory. Second Edition. New York: McGraw-Hil1. A. Salomaa. 1973. Formal Languages. New York: Academic Press. A. Salomaa. 1985. "Computations and Automata," it Encycloped'ia of Mathernat'i,cso,nd lts Appli,cations. Cambridge: Cambridge University Press.

405



A accepter,26 Ackerman'sfunction,330 algorithm,2,246 alphabet,15 ambiguity, 136 of a grammar,141 inherent,144 automaton,2, 25 deterministic,26 nondeterministic,26 axiom$,324,334 B Backus-Naurfcrrnt,146 baseof a cycle,I blank, 2?3 blank-tape halting problem, 305 C Cartesianproduct of sets,5 child-parentrelation in a tree, 8 Chomskyhierarchy,295 Chomskynormal forrn, 149,165 Church'sthesis,325 Church-Ttrringthesis,325 closure,99 positive,18 star, 18 closureproperties of context-freelanguages,2L3 of regular languages,100 complement of a set, 3

of a language,18 completesysterns,324 complexity,343 of a grammar,163 space,344 tirne, 344 cornplexity classP, 353 complexityclassNP, 354 composition,326 computability, ?99 computablefunctiotr, 233 computation,228 models,323 valid, 321 corrcaterration of languages,18 of strings,15 configurationof an autclmatorr,25 conjunctivenormal form, 348 consistent$y$tems,324 context-freegrammars,126 context-freelangrrages,125 deterministic,195 context-sensitivegrammars,289 context-sensitivelanguages,290 control unit of atr automa,ton,25 Cook-Karpthesis,354 Cook'stheorem,355 cycle in a graph, 8 simple,I CYK algorithm,172 D dead configuration,52 decidability, 299 DeMorgantslaws, 4 407


408

Iwnex

dependency graph, 154 derivation, 21 leftmost, 129 rightmost, 129 derivation tree, 130 partial, 131 yield, 131 deterministic finite accepter, 36 dfa, 36 diagonalization,279 disjoint sets, 4 distinguishable sets in a dfa, 63

dpda,195 E empty set, 4 end markers for an lba, 271 enumeration procedure, 268 equivalence, 7 of automata classes,250 of dfa's and nfh's, 55 of grammars, 24 F family of larrguages, 42 final state,36 finite automata, 35 formal languages, 2 functions, 5 computable, 233 domain, 5 partial, 5 range, 5 total, S G grammar, 19 context-free, 126 context-sensitive, 289 leftJinear, 89 linear, 91 regular, 89 right-linear, 89 simple, 140 unrestricted, 283 graplr, 7

labeled, 7 Greibach normal form, 149, 168 H halting problems for Turing machines, 301 halt state of a T\rring machine, 224 hierarchy of language families, 275 homorrrorphir: image of a, language, 103 homomorphistn, 103

I

incompletenesstheorem, 324 indistinguishable states in a dfa, 66 inherent ambiguity, 144 initial state, 36 input file, 25 instantarreous description of a pushdciwn arrtomaton, 179 of a Thring machine, 226 internal states of an automaton, 25, 36 intractable problenrs, 354

L lambda-productions, 156 language, 15, 17 accepted by a dfa, 38 accepted by a dpda, 196 accepted by an lba, 271 accepted by an nfa, 51 accepted by a Ttrring machine, 229 ociated with regular expressions,73 generated by a grammar, 2L generated by a Post system, 335 language families, 42 lba, 270 left-linear grammar, 89 leftmost derivation, 129 linear bounded automata, 270 linear grammar, 91 Ll-grammars, 201 L-systems, 340


It'tlpx

M Markov algorithrtt, 339 mir,trix gramrnar, 338 membership algorithm, 111 for context-free languages, 172 for context-ser$itive languages,

293 minimal dfa, 67 minimalization opertrtor, 331 monus, 327 move of an a,utorntrtotr, 25 MPC-solution, 313 mu-rtrcursivefutrctions, 331 N nfa,48 non-contracting grarrtmars, 290 nondeterministic finite accepter, 47 nondeterminism, 52 nonterminal consta,nt,334 normal form of a grarnmar, 149, 165 NP-complete problems, 355 rrpda, 177 null set, 4

o order proper, 269 relation in a tree, I P parsing, 136 exharrstive search, l3(i top*down, 136 path in a graph, I labeled, 8 sintple, 8 pattern matching, 85 PC-solution, 313 pda, 175 phrase-structure gramrrrar, 338 pigeonhole principle, 114 polynomial-time reduction, 355 Post corresporrdenceproblem, 312 modifierd,313

409

Post system, 334 powerset, 4 primitive recursiorr, 326 prirnitive recursive functions, 328 primitive regular expressions,T2 productions of a grarrrrrtar,19 program of a Ttrring macltirre, 224 projector function, 326 proof techniques, I contradiction, 11 induction, 9 proper order, 269 propcr subset, 4 pumping lemma for context-free langirages, 206 for linear larrguages,21O for regular languages,115 pushdown automata, 175 deterministic, 195 nondeterminiutic, 176 !1 e: R read-write head of a T\rring machine, 222 recur$ive futtction, 325 recursive la,nguage,277 recursively erturnerable languages, 276 reduction rrf states in a dfa, 62 reductiorr of undecidable problems, 304 polynomial-tinre, 355 regular expressions, 7^ regular grammar, 89 regular language, 43 relation, 5 reverse of a language, L8 of a string, L5 rewriting systems, 337 Rice's theorem, 311 right-linear gra.mmars, 89 rightmost derivation, 129 right qrrotient of languages, 104 root of a tree. 8


410

1 t r q h 5 t E tl

,J'

J

,l

satisliability protrlcrn, i347 sernanticsof prograrnrning langrrages,I48 sentence,17 sentential forrn, 21 set, 3 countable, 267 uncounta,ble,267 set operations, 3 s-gra,trmilr, 140 simulaLion, 251 upace-complexity, 344 sttr,r:k,175 alphabet, 177 start symbol, 177 standard representation for regular languages, 112 sta,te-entryproblem, 304 storage of a,n automa,ton, 25 strirrg, lS ernpty, 15 Iength, l5 olletrir,tions,l5 prefix, I6 urrffix, 16 srrbset,4 proper, 4 substring, 16 successor func;tiorr, 326 symmetric difference of two sets, 109 syntan of a prograurrning langua,ge,

r47

V, R FJ

T ilfi|

I(r) 5 Io 'l T*

.t J

Ilt nr:x

r* 3.a I-E..

tape alphabet,223 tape of a Ttrring rnachirre, 222 terminal constant, 334 terminal symbol, 19 time-complexity, 344 tracks rlrr ir ta,pe, 253 tractatrkl prohlems, 343, 354 transducxlr, 26 transitiorr firnction, 25, 36

exterrdcd, 37 transition graph, 36 genera,lizecl, 8l trap statc, 39 trulu, 8 T\rring-computable function, 233 T\rqing machine, 221 multidimensional, 261 with multiple tracks, 253 mrrltitape, 258 trondeterrnirristic,26i3 off-line, 255 with scmi-infinite tape, 253 starrdirrd, 226 with sta.y-option,251 rrniversa,l,266 T\rring's thesis, 244 U urtdcxlitlir.bleprohlem, 300 for cotrtcxt-frtxl lir,ngrrages,3 II for recursively enurrrerable lir,ngrrages,308 trnit llroductions, 158 urrivtlrutll set, 4 rrrtivt)r$rrlTtrring machine, 266 urrrtlstrir:tedgrammar, 283 uscltlJsproductions, J.53 V vilriahle of a grarrtrrrar, 19 nullable, 156 start, 19 useless,153

w walk in a graph, 8 Y yield rlf a derivationtree, 13L

z zero function, 326


Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.