n oa B on 8 Theory of © Recursive = © Functions and p @ Effective ee oo Computability Hartley Rogers, Jr. 1Theory of Recursive Functions and Effective Computability Hartley Rogers, Jr. The MIT Press Cambridge, Massachusetts ‘London, EnglandThird printing, 1992 First MIT Press paperback edition 1987 © 1987 Massachusetts Institute of Technology Original edition published hy McGraw-Hill Book Company, 1967. All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (ineluding photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. This book was printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Rogers, H. (Hartley), 1926- * Theory of recursive functions and effective computability. Bibliography: p. Includes indexes, 1. Recursive functions. 2. Computable functions. 1 Title. QA9.615.R64 1987 5113 86-83764 ISBN 0-262-68052-1 (pbk.) TO MY PARENTS Ti may be we shall touch the Happy Isles, And see the great Achilles thom we knew. Tennyson: UlyssesPreface to Paperback Edition (1987) The theory of recursive functions (more commonly known as recursion theory) emerged, as a distinct branch of mathematics, from formal studies in the 1930's by Kleene, Church, and Turing. In the 1940's, Post showed, by example, that the intuitive simplicity and naturalness of the concept of general recursive function permitted discourse and proof at a level of informality comparable to that occurring in more traditional mathemat- ics. In the 1950's, recursion theory began a remarkable period of growth and of interaction with other areas of logic and other branches of math- ematics. The informal approach of Post was carried forward and elabo- rated by Dekker, Myhill, and others. (The present book follows and builds on this approach.) In the middle and late 1950's, Friedberg, Spector, Sacks, and others discovered new methods and combinatorial principles, such as the priority method and the infinite injury method, which pro- vided important tools for further exploration and development of the the- ory. The decade of the 1950’s was a time of excitement, enthusiasm, and promise for students of recursion theory. The month-long Summer Insti- tute for Symbolic Logic at Cornell in 1957 marked an epoch for the theory (as well as for other areas of logic) at a moment when a significant num- ber of independent and largely self-taught researchers came together for the first time to share approaches and ideas. The present book was written during the pericd 1957-1967 and began as a set of mimeographed notes published by the M.LT. Mathematics De- partment in 1957. The growth of recursion theory during this time was so rapid that, even at its publication in 1967, the book could not claim to be a comprehensive survey. Its most significant omission, perhaps, was the theory of recursion for finite types as initiated by Kleene. The development of recursion theory during the two decades since 1967 has been extraordinary both in depth (where we have witnessed the so- tution of numerous open problems and the uncovering of new and unex- pected combinatorial structure) and in breadth (where new areas of the- ory have emerged and interactions with other parts of logic, mathematics, science, and technology have continued to grow). It is impossible in theviii Preface to Paperback Edition (1987) few paragraphs of this preface to do more than indicate some main fea- tures of the development since 1967. The related theories of recursively enumerable sets, of degrees of un- solvability, and of Turing degrees in particular, are a central concern of this book. Since 1967, these topics have continued as a rich field of study—richer and more active, perhaps, than might have been predicted in 1967. Many individuals have made substantial contributions, and it is invidious to mention names and results in the limited scope of these com- ments. Nonetheless, by way of example, one may note the work of Lach- Jan on the monstrous injury method, the work of Soare on maximal sets and automorphisms, the work of Shore in settling a number of interesting global questions about degrees, and the comprehensive study of degrees by Lerman. ‘A second group of topics, in later chapters of the book, has to do with generalizations of recursion theory. Extensive development in this area bas also occurred since 1967, with work by Sacks and others on a-recur- sion theory and with exploration of higher-type recursion theory as de- fined and studied by Gandy, Normann, Moschovakis, Slaman, and others. A variety of fruitful applications of generalized recursion theory have been made to set theory, to model theory, and to proof theory. Moreover, the theory of hyperarithmetical computability has emerged as a branch of recursion theory in its own right. One may note the separation theorem. of Louveau in this area. A third group of topics, mentioned only tangentially in the book, has to do with subrecursive computability and subrecursive hizrarchies. Discov- ery by Cobham, in the middle 1960's, of the natural and invariant notion. of computability in polynomial time has led to the study of computational complexity, a field of active interest in theoretical computer science with applications to number theory, cryptography, and other areas. A further field of computer science, also related to recursion theory, has arisen from the exploration by Scott and others of sets of continuous functionals as bases for a semantics of programming languages. Computer scientists have told me that republication of this book will be welcomed by them and hy their students. A basic reason, perhaps, for such a welcome may lie in the naturalness and appropriateness of the mathematical concept of partial function for the purpose of modelling the action of a computer. A partial function with a given argument may or may not be defined, just as a computer with a given input may or may not produce an output in a specified number of steps. Apart from the correction of several typographical errors, the text of this 1987 republication remains unchanged from the 1967 printing. Fur- ther updating or modification (beyond this preface) would be a formidable task, and I have not ventured to attempt it. Since the original publication in 1967, researchers and students have informed me from time to time of Preface to Paperback Edition (1987) ix statements or questions in the text (chiefly in the Exercise material) which are incorrect or now out of date. I am grateful for the interest and help of those who have communicated with me. Emendations are indi- cated following the Introduction. I am especially grateful to The MIT Press, to its Director, Frank Ur- banowski, and to its Science Editor, Laurence Cohen, for their initiative and encouragement in bringing about this republication. From 1974 to 1981, I had a close association with The MIT Press as chairman of its Editorial Board. I am honored that the Press should now wish to publish a work of my own. Hartley Rogers, Jr.Chapter 2 gli ga u3 gl4 515 31.6 §L7 us g9 $1.10 Chapter 2 92.1 §22 §2.3 §24 §2.5 52.6 Chapter 3 93.1 $3.2 $3.3 Chapter 4 44 H2 Preface to paperback edition (1987) Introduction: Prerequisites and Notation Emendations RECURSIVE FUNCTIONS ‘The informal notion of algorithm An example: the primitive recursive functions Extensionality Diagonalization Formal characterization The Basic Result Chureh’s Thesis Gédet numbers, universality, ¢-m-n theorem The halting problem Recursiveness UNSOLVABLE PROBLEMS Further examples of recursive unsolvability Unsolvable problems in other areas of mathematies Existence of certain partial recursive functions Historical remarks Discussion Exercises PURPOSES; SUMMARY Goals of theory Emphasis of this book Summary RECURSIVE INVARIANCE Invariance under 6 group Recursive permutations Contents Rae Belt sear a 21 26 3z 32 35 36 38 39 46 43 30 51xii Contents $43 §44 5 M6 Chapter 5 §6.1 $52 $5.3 $54 $6.5 §5.6 + 957 $5.8 Chapter 6 §6.1 §6.2 Chapter 7 W4 72 47.3 7.4 57.5 $7.6 97.7 87.8 §7.9 Reeursive invariance Resemblance Universal partial funetions Exercises RECURSIVE AND RECURSIVELY ENUMERABLE SETS Definitions Basic theorem Recursive and recursively enumerable relations; coding of ketuples Projection theorems Uniformity Finite sets Single-valuedness theorem Exercises REDUCIBILITIES General introduction Exercises ONE-ONE REDUCIBILITY ; MANY-ONE RED: UCIBILITY; CREATIVE SETS One-one reducibility and many-one reducibility Complete sets Creative sets ‘One-one equivalence and recursive isomorphism One-one completeness and many-one completeness Cylinders Productiveness Logie Exercises TRUTH-TABLE REDUCIBILITIES; SIMPLE SETS Simple sete Immune sets Truth-table redueibility Truth-table reducibility and many-one redueibility Bounded truth-table reducibility Structure of degrees 82 53 53 55 7 77 79 80 82 85 87 89 94 9 105 105 107 109 N2 lit 118 Contents $8.7 Other recursively enumerable sets §88 Exercises Chapter 9 TURING REDUCIBILITY; HYPERSIMPLE SETS §0.1 An example §9.2 Relative recursiveness §9.3 Relativized theory §9.4 Turing reducibility §9.5 Hypersimple sets; Dekker's theorem §9.6 Turing reducibility and truth-table redueibility; Post’s problem §0.7 Enumeration reducibility §9.8 Recursive operators §9.9 Exercises Chapter 10 POST’S PROBLEM; INCOMPLETE SETS §10.1 Constructive approaches $10.2 Friedberg’s solution §10.3 Further results and problems §10.4 Inseparable sete of any recursively enumerable degree §10.5 Theories of any recursively enumerable degree $10.6 Exercises Chapter 11 THE RECURSION THEOREM §11.1 Introduction §11.2 The recursion theorem §11.3 Completeness of creative sets; completely productive sets §11.4 Other applications and constructions §11.5 Other forms of the recursion theorem §11.6 Discussion §11.7 Ordinal notations §11.8 Constructive ordinals $11.9 Exercises Chepter 12 RECURSIVELY ENUMERABLE SETS AS A LATTICE $12.1 Lattices of sets §122 Decomposition $12.3 Cobosive sets §12.4 Maximal sets §12.5 Subsets of maximal sets §12.6 Almost-finiteness properties $12.7 Exercises 10 121 7 127 128 134 137 138 141 145 148 154 161 161 163 167 170 li 174 SEUSRER Exiv Contents Chapter 18 gat $13.2 $13.2 §13.4 $13.5 313.6 513.7 313.8 §13.9 Chapter 14 gla 42 $14.3 glad 4s $14.6 47 g148 g49 Chapter 15 $15.1 §15.2 $15.3 5154 Chapter 16 s16.1 $16.2 §16.3 gi64 $16.5, $16.6 $16.7 96.8 $16.9 DEGREES OF UNSOLVABILITY The jump operation Special sets and degrees Complete degrees; category and measura Ordering of degrees Minimal degrees Partial degrees ‘The Medvedev lattice Further results Exercises THE ARITHMETICAL HIERARCHY (PART 1) The hierereby of sets Normal forms The Tarski-Kuratowski algori ; gorithm Arithmetieal representation The strong hierarchy theorem Degrees Applications to logic Computing degrocs of unsolvability Exercises THE ARITHMETICAL HIERARCH 'Y (PART 2) The hierarchy of sets of sets The hierarchy of sets of functions Functionals Exercises THE ANALYTICAL HIERARCRY The analytical hierarchy lytical representation; icati i fosbotical epresentatfon; applications to loge Tij-sete and Absets Geacralized computability Sperdegrees and the hyperjump; Z}sets and Absets Basis results and implicit definability yyperarithmetical hi The bye erarchy Bibliography Index of Notations Subject Index 254 234 262 265 273 276 279 282 289 295 301 305 307 312 314 316 318 323 331 335 385 346 353, 367 373 873 392 597 402 409 418 A434 459 473 tation Introduction: Prerequisites and This book is intended for use as a senior undergraduate or first-year graduate text. It assumes a knowledge of basic set-theoretical terminology and techniques such as might be obtained in an undergraduate course in modern algebra. In most parts of the book, no knowledge of logic is assumed, but the reader will find some knowledge of logic helpful. We shall use a few notations from elementary logic, ss described below. The literature of recursive function theory unfortunately lacks a single, commonly used terminology and notation. We present our choice of basic terminology and notation in this introductory section. The reader is urged to give this section a careful, preliminary reading and thes to return to it from time to time as may later prove necessary. Initial effort by the reader here will, hopefully, be rewarded with clarity and facility in the main text. For the most part, we deal with nonnegative integers, sets of nonnegative integers, and mappings from nonnegative integers to nonnegative integers. Unless specifically indicated otherwise, we use the words number and integer to mean nonnegative (rational) integer. N is the set of all integers. A,B, C, .. . (Latin capitals, early in the alphabet) denote subsets of N. Gis the empty set. 2, y, z,... (Latin lowerease, Iate in the alphabet) denote members of N, i.e., integers. ‘We use the following set-theoretical notations. A = B means that A and B are identical as sets, i.e., have the same members. x € A means that z isa member of A. {| } is the notation to indicate set formation. {2| -- + 2+ + -] is the set of all z such that the expression «++ z+ +> is true when “2” is interpreted as the integer z. The universe from which 2 set is formed is indicated by the style of symbol appearing before the vertical bar. Thus {z{ - - -} must be a set of integers. AW Bis the union of A and B, that is, {z|z € A or 2 € B or both}. AQ Bis the intersection of A and B, that is, {zjx€ A andz€ B}. Ais the complement of A, that is, {z[nol zs A}. A CB means that A is a subset of B; that is to say, for all z, ifz GA thenz@ B. A> B means that BCA. A is a proper subset of Bif A C Band not A= B. (Thus AC Band BC A imply A = B) ‘We occasionally denote a finite set by an expression in braces listing its members, in any order. For example {2,5,3} is the set of the first three primes. We sometimes suggest certain infinite sets by # “listing’’ in braces. For example, {0,2,4, . . . ,2n, . . .} is the set of even integers.xvi Introduction: prerequisites and notation Given z and y, is the ordered pair consisting of x and y in that order. Similarly <21,72,... ,%3> is the ordered k-tuple consisting of ay... 5 min that order. A x B is the cartesian product of A and B, that is, (|z € A andy © B}. Similarly Aix Arx ++ XK Aum f[t1 € Arand «> + a E Ait. The cartesian product of A with itself k times is denoted as A*. P, Q, R, .. . Qatin capitals, late middle alphabet) will stand for relations on N, ie, subsets of N* for some kK >0. If RCN*, BR is called a keary relation. Let F be any k-ary relation. We say that R is single-valued if, for every <2. . - ,t,a>, there exists at most one zsuch that <2, . . . ,241,2> © R. If B is single-valued, we say that its domain is { there ts az such that © R}, and we call this collection domain R. Clearly a single-valued k-ary relation may be viewed as a mapping from its domain into W. For this reason, instead of saying that R is a single-valued k-ary relation, we shall sometimes say, synonymously, that F is a partial, function of k -1 variables. (Here the word “partial” suggests that the domain of R may not be allof NW!) We shall'use y,¥, . . . (Greek lowercase, late alphabet) to denote partial functions, and we shall often use ordinary functional notation with these symbols; thus 9(2,y) = 2 will mean that € ¢. The reader should keep in mind that, fundamentally, a partial function is to be construed as a relation. (We thus identify a partial function with its “graph.”) We shall most often be concerned with the case k = 2, ie. Partial functions of one variable, and partial function will mean partial function of one variable unless otherwise indicated. With partial funetions, the functional notation can result in ambiguity. For example, the ascertion that o(z) not = y might mean that not © , or it might mean that there is a z such that € ¢ and znot = y. Our intended meaning in such situations will always be clear. i ¢ is a partial function, we say that @ is defined (or convergent) at 2 if z © domain o; otherwise g is undefined (or divergent) at x. (Similarly for partial functions of more than one variable.) In case a partial function of k variables has all of N* as its domain, we call it a function, or, oceasionally for emphasis, a total function. We use F,9:h, . . . (Latin lowercase, early middle alphabet) to denote functions. As before, f(z) = y will mean Ef. @, B, ©, . . . (script capitals, early alphabet) denote either sets of sub- sets of N or sets of relations on ¥. ‘The range of a partial function 9 of k variables is {2| there exist a1, . . «4% such that <2... ,tu2> Ey}, and we denote it range p. Members of Tange » are called values of ¢. If y(t, . . . ,2:) = y, ¥ is called the value of ¢ corresponding te argument . A partial function is onto if its range = N. A partial function » is one-one if, for every y, there is ab most one k-tuple such that g(x... .t) =¥- Ca Will Introduction: prerequisites and notation xvii be the characteristic function of the set A; hence ca(x) = 1 if x € A, and ca(z) = 0 if x not E A. Occasionally, a set A is represented by & (non- unique) funetion f such that A = {zlf(z) = 0}. Such an f is called a ‘esenting function for A. Piet veh be an expression such that given any integer in place of “2,” the expression defines at most one corresponding value. (For exampie, the expression ‘‘z? + 2” defines a value for every integer; while the expression “jeast proper prime divisor of 2” defines a value for integers which are not prime and are different from 1.) Then \z[—2—] denotes the partial func- tion: {I[—2—] defines the value y when “2” is interpreted as the inte- ger x}. This is Church’s lambda notation for defining partial functions. For example, given ¢; and gz, then Az[y1(2) + ¢2(z)] is the partial function ¥ such that domain = domain yx domain ys and Wz) = gs(z) + pa(2) for all xin domain y. We also use the lambda notation for partial functions of & variables, writing Amis «> + a: in place of Ax. , If ¥ and ¢ are partial functions, ¥ indicates their composition, i.e., the partial function {<2,y>|there ts a z such that <2,2> € g and E #}. (Note the reversed order of y and y so that y(x) can be expressed as ¥(e(2)).) Other common notations include ¥! = {| € ¥]; WA) = {ylfor some x, x GA and Wiz) = y}; YMA) = [2] for some y, y EA and ¥(z) = y}. . | For binary (that is, 2-ary) relations, the terms transitive, reflexive, equivalence, linear ordering, portial ordering will be given their usuel mean- ings (which we do not define here). Partial orderings will sometimes be strict (<) and sometimes nonstrict (<). | ‘We use certain notations and conventions from elementary logic: “and” will sometimes be abbreviated as “&”; “or” will be used in the inclusive (and/or) sense and will sometimes be abbreviated as “WV”; Meg? wil abbre- viate “only if”; “ey” will abbreviate “if and only if”; “>” will abbreviate “not” and will be placed before the statement which it negates. We some- times combine “4” with “G” or “=” as “@" or “+.” Brackets will be used to indicate grouping of statements, except that (contrary to logical usage but following common mathematical usage) “(i) = @) arcs (n)” (where (1), (2), - - . , (n) are statements) will abbreviate “[[() = 2) & (2) = @] & + - - [(@ — 1) = (W)IL,” and “() @ @) @ «- - = in)” will abbreviate “[[(1) = @)] & (2) @)]& - + [m@— 1) eM)” “VW and “J” are called universal and existential quantifier symbols, respectively. “(Wz)” is read: “for allz.” “(Az)” isread: “there exists an x auch that : nan Groups of symbols such as “(Wz)” and “(dz)” are called, respectively, universal and existential lifters. The above logical symbols en as convenient abbreviations of ordinary mathematical language. For example, the meaning of A c B can be expressed (V2)|z € A => 2 € Bl, and the definition of domain » can be expressed {z|(4y)[e(z) = yl}.xviii Introduction: prerequisites and notation Xo is the cardinality of N. 2” is the set of all subsets of N; we sometimes call this collection 1. 2*° denotes the cardinality of 2%, ie., the cardinality of the continuum. sz! ---2-- -] is the least integer z such that the expression - + - z+ + is true when “2” is interpreted as the integer z, if this least integer exists. Various other general and special notations will be introduced as the book progresses. Both logic and recursive function theory lack a universally accepted sys- tem of notation. Our choice of logical abbreviations is not uncommon. A choice of notation for recursive function theory presents some difficulties, especially in a treatment that covers a variety of areas. Part of the eurrent literature uses Greek lowercase letters for sets. We do not do so because of the almost universal use of these symbols for ordinals. The reader should note that by zumber-theoretic predicate, Kleene and others mean relation on N. The reader should also note that some writers use f, g,... to denote partial functions in general, rather than only total functions, and that in a significant part of the literature, Greek lowercase letters carly in the alphabet are used for total functions. Tn the assertion of a theorem or Jemma, unless otherwise specified, we follow the usual mathematical convention that all unquantified variable symbols (representing integers, sets, or relations) are to be taken as operated on by unexpressed universal quantifiers standing at the beginning of the entire assertion. Tn the course of a proof, certain variable symbols may appear as uni- versal variables, e.g., “Let z be any integer such that --- .” Other vari- able symbols may be introduced as temporary names, e.g., “Choose 29 to be some fixed integer greater thanz ....”{ In general, but not always, we shall use subscripted symbols in the latter case. (Subseripted symbols may also be used in the former case when sufficiently many distinet variable symbols are required.) Occasionally, for emphasis, we shall also use letters in the middle of the Latin lowerease alphabet, i, j, k, m, 7, ... , with or without subscripts, as temporary names of integers. Occasionally, for emphasis, we shall use letters late in the Latin eapital alphabet, X,'Y, Z, as universal variables for sets. In this respect our usage will not always be consistent, but it will be elear from the context whether we adbere to these conventions or deviate from them. Chapters are divided into scetions. Thus §7.4 is the fourth section of Chapter 7. Theorems receive roman numerals, beginning anew in each chapter. Theorem 7-VI is the sixth theorem in Chapter 7. Exercises receive arabic numerals. Exercise 7-14 is the fourteenth exercise at the end of Chapter 7. When reference is made to a theorem or exercise occurring f In the terminology of elementary logic, a unizersal variable is 2 variable symbol intro~ duced by unixersal specification, and a temporary name is variable symbol introduced by existential specification (see Suppes [1957]). Introduction: prerequisites and notation xix in the same chapter, the chapter number may be omitted. The conclusion of a proof is indieated by the symbol BL Reference to entries’ in the bibliography is made by giving the author's name and a bracketed date. The item will be found under that name and beside that date in the bibliography.Emendations This section lists some substantive corrections to the text. It also men- tions, for some questions posed in the text, answers which have been com- municated to the author but which have not, to the author's knowledge, appeared in the literature. It does not consider questions, like those on pages 228 and 229 and in Chapter 13, which have been treated (in some cases, extensively) in the literature since 1967. In the references that follow, “page 228, line 12,” for example, refers to the twelfth line on page 228, not counting the line of the running head at the top of the page, while “page 301, line 4b” refers to the fourth line from the bottom on page 301 including the lines of the footnote on that page. Page 41. Page 43. Page 55, line 12. Page 122. Page 129, lines 13b-5b. Page 155. Exercise 2-10 (c) is incorrect. Condition 2 is not a sufficient condition as claimed. (R. Byerly, G. Riccardi.) In exercise 2-14 (f), the characterization is too narrow. The class a, must also satisfy the closure rule: if g and ’ are in a), and Misa periodic set of natural numbers, and if f = g on M, and f = A elsewhere, then fis in a,.(D. Siefkes.) A partial answer to this question has been ob- tained by Y. Vladik. Exercise 8~7 (c) may be incorrect. A simpler oracle-machine for this purpose can be defined: Fo1BQe, BGG WB1ge, q21Rge, BR Qs, G21 Baa. Exercise 9-5 is incorrect. (D. Hickerson, D. MacQueen, D. Morris.) See also lines 5-7 on page 132. xxi Emendations Page 156. Page 228, line 12. Page 291, lines 12b-8b. Page 301, line 4b. Page 325. Page 362, line 20. Page 394, line 17b. Page 450, Page 451. Exercise 9-38(i) is incorrect. One can find a recursive f and a hyperimmune A, where dAP>0, such that (vzId(x},) = Ol CW. Jackson.) For “an element” read “a non-unit element”. The proof of Theorem XXV requires modifi- cation as follows. Replace step (g) of the con- struction with the following: “(g) For each x which was thrown out of some 2'-list during (e) and (/) but which did not re- ceive a plus, add x to the z-list of each z such that: G) m(x)>z, Gi) x is used negatively in b“(z), Gi) the computation $,(z) has not become obsolete, (iv) if z, = wo’ [x is in the z'-list after comple- tion of step (f)] and if any such z, exists, then 2 as a tape with a string of x, +1 consecutive 1’s followed by a B followed by tat] consecutive V's followed by a B - - - followed by 2, + 1 consecutive Vs; and by again starting the machine in state go on the leftmost cell containing 1. ‘The reader can verify that, for a function of & variables, the particular machine described in the preceding paragraph yields Xz. - - - a(2m +a: °° + t+utk—V . . Bach stage in a Turing-machine calculation can be described by giving B iti ii) the internal state of the Turing machine, @ the condition of the tape, (ii) the intemal c and (ii) the location of the cell being examined. We say thet such infor- mation determines an instantaneous description. This information can be expressed in the form tegen, where --- >> - is a string of adjacent 1’s and B’s from the tape which includes all that portion of the tape that is not blank, where g; is the current state, and where “g,” is inserted immediately to the left of the symbol for currently being examined. | rama i the above machine for Az[2z] is applied to the input tape for 2, then the following instantaneous descriptions give the first few steps in the computation. gold Bll gli eB ql 1g.B16 Recursive functions At first glance, the class of algorithmic partial functions given by Turing machines might appear rather limited. Nevertheless, the definition does provide us with P-symbolism and L-P specifications. Each set of quad- ruples defining a machine may be viewed as a set of instructions P. As basic symbols for our P-symbolism we need only ¢, 1, B, B, 1, and digits for numeral subscripts. We may take the computing agent L to be human. ‘The L-P spcoifications are the simple rules according to which 2 suecession of machine-tape configurations is determined from the initial tape and from P. The relation between the Turing-machine definition and the discussion of §1.1 will be considered further in §§1.6 and 1.8, The Kleene Characterization We next look at Kicene’s formal characterization. Consider recursive relations of the general kind employed in §1.2 to define the Ackermann exponential. A set of instructions P consists of a set of such “recursion equations.” A computation is a finite sequence of equations, beginning with P, where each equation after P is obtained from preceding equations either by the substitution of a numeral expression for a variable symbol through- out an equation or by the use of one equation to substitute “equals for equals” at any occurrence in a second equation or by the evaluation of an instance of the successor function Az[z + 1}. In P, we allow auxiliary function symbols, in addition to the main fune- Hon symbol in whose evaluation we are interested. Thus the set of equations 40) = 0, g(t) = f(z) +1, F@+)) = 9) +1, with g 9s auxiliary function symbol and f as main symbol, determines the function X2[2z], 28 can be verified, Three related difficulties arise in connection with this notion of com- putation: (1) the course of a computation is not uniquely determined by the input and by P; (2) it is possible that two different outputs may be obtained from the same input (by different computations}; (3) it is possible that no output may be obtainable from a given input.t, We avoid difficulties (1) and (2) in the following way. We say that an equation is dedueible from 9 given P if it is obtainable in the course of some computation from P. tis possible to describe » uniform procedure ¥ Kleene, in his original characterization, simply accepts difficulty (3) end goes on formally to define algorithmic functions as those functions which are obiainable from sets of instructions which heppen not to be subject to diffeulties (2) and (3). The fret that there is no evident way of identifying which those sets of instructions are is the Price paid for avoiding diagonalization trouble. £1, in the Turing characterization, the consistency restriction were dropped, then the Turing characterization (thus modified) would be subject to a difficulty similar to (2), §1.5 Formal characterization Wr secording to which, given any P, we can effectively generate a list of all equations deducible from P. (The procedure makes, in effect, an exhaustive list of all possible computations and is similar to the procedure discussed in §14 for listing all possible primitive recursive derivations.) Formal details of such a procédure would be complex, and we do not give them. Assume that we have made a fixed and permanent selection of such a procedure, Then given any P and given any input, we define the principal output for that input as the first output associated with it in the standard list of deducible equations generated from P. Each input can yield at most one principal output; thus the relation between inputs and principal outputs defines a partial function. We can hence associate a partial func- tion with every set of equations P; and we have a formal characterization for a class of algorithmic partial functions. The P-symbolism in this case can be described in greater detail as follows. Take as basic symbols f, 9, z, +, =, parentheses, comma, digits for ordinary numerals, and digits for numeral subseripts (for use with g and x). Z, Zo, £1, . . . are called variables. ; _ f,% 94 91, - - - are called function symbols, and f is the principal func- tion symbol. Terms are defined inductively as follows: A variable is a term; An ordinary numeral is a term; If ris a term, r followed by “++ 1” ig also a term; : / Let « consist of a function symbol followed by parentheses containing a string of terms which are separated by commas; then ¢ is also a term.t The expression formed by placing “=” between two terms is called an equation. , Any finite set of such equations constitutes a set of instructions P. The L-P specifiestions describe how the uniform list of all derivable equations (from any given P) is to be generated. We omit details. | The reader will note, incidentally, that every primitive recursive deriva- tion is expressible in the P-symbolism in an obvious way. The derivation for \z{2z] in §1.2 would become a set of equations beginning q() = 2, g(a) =a +1, Ga(@ryt2,22) = La, Tt is easy to show that the function determined by these equations in the new formalism is identical with the function determined by the original primitive recursive derivation according to the procedures of §1.2. } An expression such as“z + 2” can, in effect, be used as a term but must be written as“z-b 1-1".18 Recursive functions §1.6 THE BASIC RESULT In §1.1 we raised the question of whether or not a characterization can be found which supplies a satisfactory formal counterpart to the informal notions of algorithm and algorithmic function. In §14 we indicated difficulties that any such characterization must face; in particular, we mentioned (§1.4) the possibility that there might be no single, maximal, formally characterizable class of algorithmic functions. In §1.5 we gave the formal characterizations of Turiag and of Kleene. During the 1930's and since then, separate formal characterizations have been proposed by Chureh [1936], Post [1936], Markov (1951), and others. These charac- terizations have varied widely in form; each, however, can be represented as a certain choice of P-symbolism and a certain choice of L-P specifications. How satisfactory are these various characterizations? What is their relation to each other? How successfully do they avoid the difficulties of §1.4? Extensive work has been done on these questions, We sum- marize this work in the following Basic Resuli, which is fundamental for the remainder of the book. Basic Result Basic Result, Part 1 By means of detailed combinatorial studice (cee, for example, Turing [1937] and Kleene [19364],) the proposed characteriza. tons of Turing and of Kleene, as well as those of Church, Post, Markov, and certain others, were all shown to be equivalent; that is to say, exactly the same class of partial functions (and hence of total functions) is obtained in each case. Definition The functions falling within this class are called recursize Junctions. The partial functions of this class might, naturally, be termed “recursive partial functions.” It has become standard usage, however, to call them partial recursive functions, These equivalence demonstrations can be generalized to show that over certain very broad families of enlargements of these formal characterizations the class of partial functions obtained remains unchanged. (For example, if we allow more than one tape, or other symbols than 1 and B, in the definition of Turing machine, the partial functions obtainable are still partial recursive functions; see Turing [1986]. For a development based on. programming, techniques, see Shepherdson and Sturgis [1963].) In fact, if certain general (and reasonable) formal criteria are laid down for whas may constitute a P-syrabolism and L-P specifications, tt ia Possible to show that the class of partial functions obtained ts always a subclass of the “maximal” class of all partial recursive functions. Basic Result, Part HM A wide variety of particular partial functions, cach agreed to be intuitively algorithmic, have been studied. Each has been §1.6 The Basic Result i demonstrated to be a partial recursive function; that ts to say, @ set of snsiruc- fions for it has been found within one of the standard formal characterizations, A variely of useful principles and techniques have been developed for making these demonstrations. (Parts I and II thus provide strong empirical evidence that the formal characterizations are sufficiently inclusive.) Basic Result, Part HI The proofs for the results in Part I have the following common structure. In every instance, the fact that one formally characterized class of partial functions is contained in another ts demonsirated by supplying and justifying a uniform procedure according to which, given any set of instructions P from the first characterization, we can find « set of instructions P from the second characterization for the same partial function. Although i does not operate on integers, this uniform procedure tiself happens, in each instance, to be algorithmic (in the unrestricied informal sense of that word, with no restriction to numerical inputs and outputs). Comment. In Parts I and II, emphasis was extensional, i.e. on the class of algorithmic partial functions defined rather than on. the class of “algorithms” defined. Parts I and II show that there is a sense in whieh each standard formal characterization appears to include all possible algorithmic partial functions. Part III, taken together with Part i, now shows that there is a sense in which each standard characterization appears to include all possible algorithms (for partial funetions). For, given a formal characterization of the kind mentioned at the end of Part I, there is a uniform effective way to “translate” any set of instructions (.c., algorithm) of that characterization into a set of instructions of one of the standard formal characterizations. (We discuss this further in §1.7; see also Exer- cise 2-11.) ; Some of the detailed work upon which the Basic Result rests will be found in Davis [1958] and in Kleene [1952]. (These references include several principles and techniques of the kind mentioned in Part IL of the Basic Result.) In work on the Basic Result, the Turing definition has been especially useful as a standard to which other characterizations can be reduced. . From a mathematical point of view, the formal characterizations of §1.5 are noninvariant, i.e., they are dependent on arbitrary choices. This is true of all known characterizations. The Basic Result shows that the characterizations nevertheless provide a natural and significant class of partial functions. It has been remarked that this class is one of the few absolute mathematical concepts to originate in work on foundations of mathematics. f eat js ‘cular t Where abode means “existing apart from and largely independent of partioul symbolic formulations.” Quantificational provability (for classical two-valued logic) is another such concept.Recursive functions §1.7 CHURCH’S THESIS The claim that each of the standard formal characterizations Provides satisfactory counterparts to the informal notions of algorithm and algo- rithmic function cannot be proved. It must be accepted or rejected on grounds that are, in large part, empirical. (That the claim for one charac- terization is equivalent to the claim for another follows from Parts I and IIT of the Basic Result.) The Basie Result provides impressive evidence that tke class of partial functions defined is « natural one (Part I) and that it is sufliciently inclusive (Parts I and II). The Turing characterization pro- vides convincing evidence that every partial function in the elass is com- putable by a procedure that is, intuitively, “mechanical.” (In §1.10 we shall discuss further the possibility that the formal class is too inclusive; see question *10 in §1.1.) On the basis of this evidence, many mathema- ticians have accepted the claim that the standard characterizations give a satisfactory formalization, or “rational reconstruction,” of the (necessarily vague) informal notions. This claim is often referred to as Church's Thesis. Church’s Thesis may be viewed as a proposal as well as a claim, 3 proposal that we agree henceforth to supply certain previously intuitive terms (e.g, “funetion computable by algorithm”) with certain precise meanings. In recent theoretical work, the phrase ‘Church's Thesis” has come to play a somewhat broader role than that indicated above. In Parts IT and III of the Basic Result, we noted that a number of powerful techniques have been developed for showing that partial functions with informal algorithms are in fact partial recursive and for going from an informal set of instructions to a formal set of instructions. These techniques have been developed to a point where (a) a mathematician can recognize whether or not an ulleged informal algorithm provides a partial recursive function, much as, in other parts of mathematics, he can recognize whether or not an alleged informal proof is valid, and where (®) a logician can go from an informal definition for an algorithm to a formal definition, much as, in other parts of mathematics, he can go from an informal to a formal proof, Recursive-funetion theory, of course, deals with a precise subject matter: the class of partial funetions defined in §1.5. Researchers in the area, however, have been using informal methods with incressing confidence. ‘We shall rely heavily on such methods in this book. They permit us to avoid cumbersome detail and to isolate crucial mathematical ideas from @ background of routine manipulation. We shall see that much profound mathematical substance can be diseussed, proved, and communicated in this way. We continue io claim, however, that our resulis have concrete mathe- matical status as results about the class of partial functions formally charac- terized in §1.5. Of course, any investigator who uses informal methods and makes such a claim must be prepored to supply formal details if challenged, §1.7 Church’s Thesis 22 Proofs which rely on informal methods have, in their favor, all the evidence accumulated in favor of Church's Thesis. Such proofs will be ealied proofs by Church's Thesis. ‘We meet our first examples of such informal methods in the remaining sections of this chapter. Almost all the proofs in this book will use Church’s Thesis to some extent. The analogy to informal methods of proof im other parts of mathematics is instructive. In both cases, the use of informal methods is a matter not of extremes but of degree. The degree of formaliza- tion of a proof usually depends upon the complexity and abstraction (what might be called the “‘danger”) of the argument. The degree of formal detail we employ in this book will similarly vary with circumstances. ‘The beginning reader, who does not possess first-hand knowledge of the evidence for Church’s Thesis, may be troubled by our arguments. To whatever extent he experiences doubt, we urge him to use the books of Davis and Kleene, in which he will find the tools needed to formalize our arguments fully. . §18 GODEL NUMBERS, UNIVERSALITY, ¢n-n THEOREM We have adopted the Turing-machine characterization as basic. We saw in §1.5 that a set of instructions is a set of quadruples satisfying the consistency restriction. It is possible to list all sets of instructions by a procedure similar to that indicated in §1.4 for listing all primitive recursive derivations. This procedure is itself algorithmic (in our first, unrestricted, informal sense of that word). It can be viewed as a procedure which asso- ciates with each integer x the set of instructions falling at the (x + 1st place in the list of all sets of instructions. We assume now that we have selected one such listing procedure. We keep it fixed for the remainder of the book. We do not give formal details. Definition P, is the set of instructions associated with the integer z in the fixed listing of all sets of instructions. = is called the index or Gédel number of P.. v-® is the partial function of # variables determined by P:. 2 is also called an index or Gadel mumber for ¢.. (We shall drop the superscript (k) when its value is clear from eontext or when k = 1. We shail be most often concerned with functions of one variable.f) Clearly the listing procedure gives us doth (a) an algorithm for going from any to the corresponding P., and @) an algorithm for going from any consistent set of quadruples P to a corresponding integer 2 such that Pis P.. {The notation {x} (for our ¢-) also appears in the literature. We use the gv, and P, notations to emphasize further the distinction between extension and namo, i.e. between partial function and set of instructions.22 — Recursive functions Theorem IV Our fixed choice of Gédel numbering will be used throughout the book. It appears to be a rather noninvariant feature of our theory. We shall see, however (Chapter 4 and Exercise 2-10), that our results possess an invariant significance independent of this choice. In this respect, our use of a pat- ticular Gédel numbering is much like the use of a particular coordinate system to establish coordinate free results in geometry. Several facts, already implicit in the Basic Result, can now be stated. Theorem I There are exactly Ny (a countable infinity of) partial recursive Functions, and there are exactly No recursive functions. Proof. All constant functions are recursive, by Church’s Thesis. Hence there are at least No recursive functions. The Godel numbering shows that there are at most No partial recursive functions. Theorem H There exist functions which are not recursive. Proof. By Cantor’s theorem, there are 2» (continuously many and therefore nondenumerably many) fonctions. The theorem follows. The next fact appears to depend, for its proof, upon our particular characterization and Gédel numbering. Theorem TIT Each pariial recursive function has No distinct indices. Proof. We need only show that there are at least Nz indices. Let a Partial recursive function g., be given. Let m be an integer greater than the subscript integer of any internal-state symbol occurring in P., Then if P,, is modified by adding to it the quadruples Gok 1g, Gartilgmer, - + + Qnsellgnse, the partial function determined remains unchanged, since none of the states gm, ... , dniz can be entered. As & varies, this gives Xo distinct sets of instructions for ¢-,.5) ‘Theorem I] is not an accident of our chosen formal characterization and Gédel numbering. Exercise 2-10 will give it invariant significance. Observe that the following is, informally, an algorithm for a partial function y: given any input , find P, (by generating the standard effective list of all sets of instructions until P, is reached), then apply P, to input y to caloulate y.(y); if and when ¢,(y) gets an output, take this as output value for ¢(z,y). Therefore, vay) = { ely), if e<(y) convergent; divergent, if pa(y) divergent. By appeal to Church’s Thesis, we conclude that y is partial recursive and that ¥ = ¢:,® for some z;. (Formal details of P., can be found in Davis {1958]}.) Stating this as a theorem, we have Theorem IV. _. Theorem IV There exisis az such that for all x and y, v.(z,y) = o2(y) of vs(y) ts defined, and y.(z,y) is undefined if p(y) is undefined. The ¢, of Theorem IV is called a universal partial function for the Partial recursive functions of one variable. P, is @ single Turing machine §L0 Gidel numbers; s-m-n theorem 23 which can be used to duplicate any partial recursive function of one variable. ‘Theorem IV, sometimes called the enumeration theorem, was a main result of the early (and more formal) work on recursive function theory. Clearly, the proof of Theorem IV can be generalized to yield, for each k > 1, a function of k ++ 1 variables which can serve as a universal partial function in the obvious sense for the partial functions of k variables. Theorem IV is the case k = 1. Theorem IV has a nontrivial practical significance. It shows that, for computing partial functions of one variable, there is s critical degree of “mechanical complexity” (that of P.) beyond which all further complexity can be absorbed into increased size of program and increased use of memory storage. (In Exercise 2-5, we shall see that the restriction to partial func~ tions of a particular fixed number of variables is not essential.) P, is called a universal machine. If we compare our formal characterization with questions *6 to *10 given in §1.1, Theorem TV shows that the computing agent L need not be human, and it substantiates our claim in §1.1 that L can be severely limited in its “abilities” ‘Thus our formal Turing-machine concept is compatible with our answers to questions *6 to *9. It is also compatible with our response to *10 in that it makes no restriction of the kind that an affirma- tive answer to *10 would require. We shall look further at “10 in Theorem XI and in Exercise 2-8. Theorem V, now to be given, also appears to depend on the particular characterization and Godel numbering used. Like Theorem III, however, it has invariant significance. In conjunction with Church’s Thesis, it will be basic in later work. Theorem V For every m,n > 1, there exists a recursive function s." of m +1 variables such that for all z, yx, ~~, Ym Der + ele Ga. amt oe nll = OMe. sne* Proof. Take the case m = n= 1. (Proof is analogous for the other eases.) Consider the family of all partial functions of one variable which are expressible as Xz{y.((y,2)} for various z and y. Using our standard formal characterization for functions of two variables, we can view this as a new formal characterization for a class of partial recursive functions of one variable. By Part III of the Basic Result, there exists a uniform effective procedure for going from sets of instructions in this new charac- terization to sets of instructions in the old. Hence, by Church’s Thesis, there must be a recursive function f of two variables such that ales (y2)] = vrea- This f is our desired 3.1.00) The informal argument by appeal to Church’s Thesis and Part III24 Recursive funetions Theorem VI of the Basic Result can be replaced by a formal proof. (Indeed, the fune- tions 8," can be shown to be primitive recursive.) We refer the reader to Davis [1958] and Kleene [1952]. Theorem V is known as the smn theorem and is due to Kleene. Theorem V (together with Church's Thesis) is a tool of great range and power. We give one illustrative application here and shall see a further application in §1.9. On many occasions in later chapters the s-m-n theorem, like Church’s Thesis, will be tacitly used. Theorem WI There is a recursive function g of two variables such that for all zy, Peon = Pty (Ley = Aalvaley(e))D- Proof. By Chureh’s Thesis it is immediate that, for any given z, y, 1 = gag, is a partial recursive function. It remains to get the recursive function g; i.., we must. show that a Gédel number for 7 can be found in a uniform effeetive way from 2 and y as z and yvary. The reader who has reflected on the Basic Result will regard this as quite likely. Theorem ¥ gives a tool for proving it. Define ane) = velo) = enltealae)), where gs, is the universal funcifon of Theorem IV. By Church's Thesis, ¢ is partial recursive and has an index wp, Applying Theorem V, we have ty = el ve(2Y,2)] = gasverzms and dzy{s:"(ws,2,y)] is our desired g. $1.9 THE HALTING PROBLEM Is there an effective procedure such that, given any x and y, we can determine whether or not .(y) is defined, i.c., whether or uot P. applied to input y yields an output? By Chureh's Thesis this question can be put in the following equivalent and precise form. Is there a recursive function g such that 9(z,y) = 1 if ¢x(y) is convergent and g(z,y) = 0 if gay) is divergent? We answer the question in the following theorem. Theorem VIE There is no recursive function 9 such that for all x, wy all if ¢-(y) convergent; sea) = la if ely) divergent, Informal proof. Assume there is such a recursive g. It can be used to define a new partial function ¥ as follows: afi if g(z,2) = 0; va) | Scergont if g(z,a) = 1. Since g(x,c) must = 0 or 1, this gives an algorithm, and by Church's Thesis, §L9 The helting problem 25 ¥ will be a partial recursive function. (When g(z,2) = 1, the instructions for ¥ can make use of some infinite cyclic condition—a pair of quadruples qullgn, GBBq, will suffice—to guarantee that ¥(z) be undefined.) Let yo be a Gédel number for ¥. Then, by the definition of ¥, ¢,,(ya) convergent © g(yo,yo) = 0; but, by our initial assumption about g, g(yo,yo) = 0 vu(yo) divergent. This is a contradiction, and hence there can be no such recursive g.{&] The proof just given uses Church's Thesis in a proof by contradiction. The reader may well ask what formal counterpart can exist to the use of Church's Thesis in such “hypothetical” context. It is instructive to trace the path that a more formal proof takes. More formal proof. Define y by 1, if ea lea) = 0; ¥G2) = | divergent, if g(z,2) + 0 or divergent, By Church's Thesis, ¢ is partial recursive.- Applying Theorem vy, a Gédel number for X2{y(z,z)] can be found uniformly effectively from 2; ie., there js a recursive function h such that for all z, gaa = Asiv(z,z))- (del[s:'(o1,2)} is such an A, where w, is a Gédel number for ¥.) Now assume g = ¢,, for some z». By the definition of h, gicg(z) con- vergent + (2,2) = 0. Substituting A(zs) for z, we have vauo(h(ze)) convergent + ¢,,(h(20),h(20)) = 0. But then ¢,, cannot be g; for, with input , ¢., either is undefined or gives erroneous information. | We see that no partial recursive function can satisfy the Tequirement for g, and our proof is constructive in the sense that for any z, h(z) provides a specific example where g, fails. ‘The following corollary to Theorem VII is immediate. Corollary VIE There is no recursive function f such that 1, galt) convergent; IQ) = ty &f exla) divergent. Proof. By proof of the theorem.) a Our original question {in the first paragraph of this section) has been called the halting problem, where the word “halting” means “having an out- put.” The fact stated as Theorem VII is known as the recursive unsolva- bility of the halting problem. It has precise content in terms of our formal characterization.t The Basic Result of §1.5 gives this fact fundamental { “There is no Turing machine which can solve the halting problem (ior Turing machines). Indeed, given any candidate ‘solver’ ‘Turing machine, we can exhibit an instance of the halting problem upon which it fails; in fact, there is a single Toring machine which will compute, from the deseription of sny given candidate, an appropriate counterinstanee.”26 «= Recursive functions Theorem VITI significance. The halting problem was one of the first “natural” combina torial problems to be shown recursively unsolvable. Demonstration of the existence of easily described recursively unsolvable problems is one of the more striking achievements of twentieth-century mathematics. Prior to such demonstration (in the 1930's) many mathematicians were unwilling to eoncede thet there could exist easily described combinatorial problems (such as the halting problem) which had no algorithmic solution. We give a more general discussion of such unsolvable problems in Chapter 2. The results of the eatly formal work on recursive function theory (prior to 1940) can be summarized, by and large, under (1) the Basic Result ($1.6) ; (2) existence of universal funetions (§1.8); and (3) unsolvability of the halting problem (§1.9). Turing’s basic paper [1936] is chiefly concerned with these matters. We conclude this section with another unsolvability result, It was first obtained by Kleene [1936], and it is a natural consequence of our dis- cussion of disgonalization in §1.4, Theorem VIE There is no effective procedure for deciding, ginen any 2, whether or not ¢, is a total function. That is to say, there is no recursive func. tion f such that a=!h Wes 4s totai; IO 19 iy ec is nol total Proof. ‘We give an informal proof. The iransition to 9 more formal Proof is similar to that in Theorem VII (see Exercise 2-7). Assume such a recursive f exists. Then define g by 90) = puffy) = 1, gf + 1) = ayly > gfe) and f(y) = J). Since f(y) = 1 for infinitely many y (Theorem 1), g is a total function. By Church's Thesis, g is recursive. Now define i by h = dalvca(z) + 1 By our assumption on f, h is total. By Church’s Thesis, h is reeursive, Let h =. By definition of g, g-Xeo) is uniquely defined; call it, yy, Then Myc) = gsco(¥o) + 1 by our definition of h; but A(ys) = Pate(Yo) by our definition of yo. Since h is total, this is a contradiction. $1.10 RECURSIVENESS In §1.5, we formslly characterized » class of partial functions, known as the partial recursive functions. Henceforth, when we say that a partial funetion is “effective,” “computable,” “effectively computsble,” “recur $1.10 Recursiveness 27 sively computable,” “mechanically computable,” or “algorithmic,” we shall mean that it falls within this class. The property of being a member of this class is called recursiveness. (Some mathematicians refer to recursive- ness as “general recursiveness”; others reserve the phrase “general recur- siveness” for total functions and refer to the recursiveness of partial func- tions as “partial reoursiveness.” In cither case, “general” serves to empha+ size that functions from the broad class of §1.5 are in question rather than functions froma narrower subelass (suchas the primitive recursive functions).) In the present section we discuss further aspects of recursiveness. In particular, we consider (1) extension of the concept to nonnumerical inputs and outputs (codings); (2) certain structural features of the partial recursive functions and the relation of these features to question *10 in §1.1 (ihe mu operator); and (3) the nature and possible usefulness of our future theory Cfinal comments). Goals and applications of the theory will be discussed in greater detail in Chapter 3. Codings The partial recursive functions are mappings from integers to integers, and their algorithms carry us from notations for integers to notations for integers. The original, unrestricted, informal notion of algorithm concerns procedures with more general kinds of symbolic input and output; ¢.g., the algorithm for differentiating polynomials. We have already used such unrestricted algorithms as intermediate stages in algorithms defining numeri- cal partial functions (and have then used Church’s Thesis to conclude that the numerical partial functions in question were partial recursive functions); eg., in our definition of a universal function, we used an algorithm (with nonnumerical outputs) that carried us from an integer 2 to the sct of instruc- tions P,. We now ask: is there any way to include such broader non- numerical algorithms within our formal theory? Two approaches can be made to this problem. The first approach is as follows. Given a class of nonnumerical inputs or outputs, choose some fixed one-one mapping from this class into the integers. Henceforth, for theoretical purposes, zdenlify each symbolic entity in the nonnumerical class with its corresponding integer “label.” Such a standard mapping is called a coding, and the integers used as labels are called code numbers. The coding is chosen so that (a) it is itself given by an informal algorithm in the unrestricted sense; and (b) it is reversible; ie., there exists an informal algorithm (in the unrestricted sense) for recog- nizing code numbers and carrying out the reverse “decoding” mappings from code numbers to nonnumerical entities. Furthermore, it is stipulated that a coding shall be used only when (c) an informal algorithm exists for recognizing the expressions that constitute the uncoded, nonnumerical class. By thus idendifying expreasions (of a nonnumerical class) with integers, we can bring the discussion of algorithms (for such a class) within our28 Recursive functions Theorem VHT formal theory of algorithms for integers} We have alzeadyseen an example of this in §1.8. Our fixed Gadel numbering for the partial reeursivefunctions is a coding from sets of instructions onto the integers. (Codings are fre- quently referred to as Gédel numberings, and code numbers are then called Gédel numbers.t) ‘The use of codings raises an immediate question of invariance. Once a coding is chosen, will the formal concept partial recursive function om cade numbers correspond to the informal notion algorithmic mapping on the wncoded expressions? As the latter action is informal, the answer must be, in part, empirical. Church’s Thesis provides an affirmative answer. Let C be the uneoded class (see the diagram below). Let y be the coding map from C into N, and let us assume, further, that takes C onlo N (modifies- tion of the argument, for the case where y does not map ono N, is straight- forward.) Let yt be the decoding map from N onto C. and y-! are Gnformally) algorithmic, by our definition of coding. Let y be any partial recursive function. » is (both formally and informally) algorithmic. Hence the mapping 5 = y~'yy is an (informally) algorithmic mopping from C into C. Conversely, let 8 be any (informally) algorithmic mapping from C into C; then ¢ = 7577" is an (informally) algorithmic partial fane, tion, and by Chureh’s Thesis, » is a partial recursive function. ‘Thus every formal ¢ has a corresponding informal 3, and every informal 6 has a corresponding formal ¢. The second approsch to-a formal treatment of nonnumerical algorithms is as follows. The formal characterization of §1.5 is broadened to include directly, as inputs and outputs, expressions from wider “nonnumerical” Glasses. The Turing-machine characterization is especially convenient for {Tho philosophically minded reader may sek why we take codings to be mappings into indegers (mathematical objects) rather than into numercls for integers (symbolic objects), since algorithms for partial recursive functions must themselves operate on some form of numeral. Tt is true that use of numerals would be closer to our motiva- tion; however, the distinetion is unimportant for our purposes. The use of integers as labels has greater theoretical convenience and is the general practice 1 The first use of» coding was made by Gadel, who chose » fixed coding from the formulas of number theory into the integers; he was thus able to study the formulas and proof logic of number theory within number theory itself. An attempt to find selfvreferential paradoxea via this coding yields the first incompleteness theorem of Gédel, ‘Tarski made a similar early construction of this kind. §1.10 Recursiveness 29 i e.f It requires only that the expressions of the wider classes Sie as finite strings in a fixed finite alphabet. of basic symbols (other than B and 1). The basic operations of Turing machines are extended to include printing and erasure of symbols from this new alphabst. ‘A nonpumerical mapping is then defined to be recursive (or partial recursive) if a Turing machine exists for carrying it out. After the formal charac- terization is so broadened, Parts I, II, and IIT of the Basie Result (§1.6) can themselves be modified and broadened to apply to this broader concept recursiveness. “ The second approach is evidently more direct. We use the first ap- proach, in order to limit our subject matter and to emphasize it as a formal discipline about mappings on integers. For results obtained in this book, it makes no difference which approach is used. The Mu Operator ; The operator » was defined in the Introduction. Theorem IX Let f be a recursive function of k+ 1 variables; then dey + mluylfGr, ... mY) = L] ts @ partial recursive function of k variables, . Proof. Immediate by Chureh’s Thesis; for let y= dm aaluylf@y ~~ - tay) = Ul then to compute ¥(zi, . . . ,%4) we need only compute, in succession, FG + 0), Sey... sted), fy. «+ ty), + If and when we find a y such that f(z... tay) = 1, we take it (the first such y) as value. ‘The subcomputations for f always terminate, since fis a recursive function.&] ; Theorem EX is known as the mu theorem. In Exercise 2-13 we shall ace that Theorem IX does not hold in general when f is replaced by a partial recursive function of k + 1 variables. . There is a sense in which each partial recursive function can be obtained from (total) recursive functions by a single application of gz. Theorem X gives this. Theorem X_ There exist fixed recursive functions p and t of one and three variables, respectively, such that for all 2, o = dalpQeylé.,9) = 1))- inati i have been studied by f Other extended formal characterizations for this purpose 8 Fost [1943) and Smullyan [1961] (see also Asser [1960}, Curry [1963], and Turing [1936)).oO Recursive functions Corollary X Proof. Define the function s as follows: seayu) = {1s if P, with input z yields output g, in fewer than w steps; ons 0, otherwise, By Chureh’s Thesis, s is recursive. Define p and g by = dalexponent of 3 in prime decomposition of x + 1), g = dalexponent of 2 in prime decomposition of z + 1}. Define i by t = dery[s(z,2,p(y),q(y))]. By Church’s Thesis, p, g, and hence {, are recursive. ‘The theorem is now immediate from the definitions of s, p, ¢, and £.&] Theorem X is known as the Kleene normal-form theorem, Once Theorem X is established, Theorem IV (the enumeration theorem) follows as a direct consequence. It is possible to show that both p and ¢ are primitive recursive. Theorem X can be stated and proved for partial functions of & variables, k > 1, by introducing appropriate recursive functions s, and % of k +3 and k + 2 variables, respectively. Corollary X There exist recursive functions p and t, such that for all 2, pe = day + mlpluyli@as, . . . 24) = ID} Proof. ‘As just indicated. A proof of Theorem X can, of course, be based on the Kleene formal characterization. After 5 somewhat different sequence of definitions, a function 7 can be obtained snalogous to our i, The relation T = {ltey,2) = 1 is called the Kleene T-predicate, and the assertion that © T is commonly abbreviated T(z,y,2). The Kleene T-predicate occurs fre quently in the literature. The question of a stronger form for Theorem X, with p eliminated, naturally arises. Is there a recursive function i* such that for all z, ee = dafuple"(z,z,y) = 1]? In Theorem 2-III, we shall show that no such é* ean exist, In fact, we shall find a partial recursive y of one variable such that for no recursive f is it true that ¥ = Azfaylfz,y) = iJ). The mu theorem emphasizes that our formal characterization permits a computation to take (in effect) the form of an unbounded search for an integer satisfying some given effective condition. The mu theorem is correlated with our failure to give an affirmative answer to question *10 in §L.L It is a priori possible, of course, that some meaningful positive $2.10 Recursiveness 31 answer to question *10 might be deducible from the formal cheracterization. Unfortunately, this does not turn out to be true. Theorem XI and its proof show that any reasonable affirmative answer would permit a diago- nalization leading to contradiction. The proof of Theorem XI is given as Exercise 2-8. Theorem XI There is no recursive function f of two vartables such that for all x and z: P, applied to z yields an output = P, applied to x yields an output in fewer than t{z,x) steps.t The nonexistence result in Theorem XI is similar to the unsolvability results of Theorems VII and VIII. Like Theorems VII and VIII, it is a consequence of the breadth of the formal notion of recursiveness. Because of such results, some mathematicians have argued that the formal notion of recursiveness is too broad to bo a counterpart to their private informal notions of algorithmic computability. Be this as it may, recursiveness does express one notion of algorithmic computability and is sufficiently natural to merit further investigation in its own right. Final Comments ‘The concept of reeursiveness has virtues (breadth and clarity) and defects (unsolvabilities) that are characteristic of the theory now to be developed. Diagonal methods like those in Theorems VII and VII will play an important role in the theory. It is not inaccurate 10 say that our theory is, in large part, a “theory of diagonalization.” Our theory is of limited practical usefulness at present. It is concerned with questions of existence or nonexistence of computational methods rather than with questions of efficiency and good design. Questions of the latter kind appear, not in our theory, but in more complex theories based on narrower concepts than recursiveness. Our theory can be viewed as a limiting, asymptotic version of these narrower, more difficult theories. ‘As such, it has some practical value. The inclusive concept of Turing machine, the concept of universal machine ($1.8), and other combinatorial resulis and methods of our theory have been found useful and suggestive in work on computer programming. Perhaps the most direct practical applications have come from nonexistence (Le., unsolvability) results, since these results carry over, a fortiori, to narrower theories. Ht must be stated, however, that at the present time our theory derives its principal significance from its relevance to pure mathematics. It pro- vides structures which possess considerable intrinsic beauty and naturalness. It gives new, and often deep, insights into other areas. ‘These insights have been especially helpful in mathematical logic, and they have been increas- ingly useful in more classica] areas as well. { If wo use the function 4 from the proof of Theorem X, Theorem XI ean be restated: there is no recursive f such that for all x, y, ¢, (Gw)[s(z,2,y,t0) » 1] + of2,zu flee) = 1]2 Unsolvable Problems $2.1 Further Examples of Recursive Unsolvability: 32 §2.2 Unsolvable Problems in Other Areas of Mathematics 35 $2.3 Existence of Cortain Partial Recursive Functions 36 §24 Bistorieal Remarks 38 §2.5 Discussion 39 42.6 Exereisea 40 §2.1 FURTHER EXAMPLES OF RECURSIVE UNSOLYABILITY Theorem 1-VII, Corollary 1-VII, and Theorem 1-VII gave examples of recursive unsolvability. Each concerned a “problem” which could be stated as the problem of effectively recognizing the ‘wagmbers of a certain set ‘or relation. For example, in Corollary 1-VII, this set was {z!g2(z) conver- geri}. Each of the three results showed that the set or relation in question failed to have a recursive characteristic function, We summarized each result by saying that the problem was recursively unsolvable, We now give several further examples of recursive unsolvability. We show, in each case, that a certain set (or relation) fails to have a recursive characteristic function. The “problems” are interesting and natural ones from within recursive function theory itself. All have to do with the general question: how much can be effectively determined about the behavior of a partial recursive function from a set of instructions for it; ie., how much can be found out about the behavior of ¢z from the index 2? There are two different: methods for demonstrating recursive unsolva- bility. The first, direct method is to make an argument, usuaily diagonal in flavor, showing that solvability would lead to contradiction. The second, indirect method is to take another problem already known to be unsolvable and then show that solvability of the problem under study would imply solvability of the problem known to be unsolvable, The latter method is called the method of reduction; in it, one shows that the known problem can be reduced to the problem under study (and hence would be solvable if the problem under study were solvable). ‘The reduction method is often more convenient than the direct method. ‘Theorems 1-VIJ and 1-VIIT 32 §21 Further exemples 38 were obtained by the direct method. The examples now to be given use the indirect method. In each case, the problem of Corollary 1-VII will be used as the problem already known to be unsolvable. We consider the following problems: ; (a) The problem of deciding, for any x, whether or not g, is a constant function. a (0) The problem of deciding, for any z and y, whether y is in the range of os. (c} The problem of deciding, for any z, y and z, whether o.{y) = z @) The problem of deciding, for any z and y, whether ¢. = () The problem of deciding, for any x, whether ¢, has infinite range. (S) For each fixed yo, the problem of deciding, for any given x, whether is in the range of ¢.. . " (g) For each fixed x», the problem of deciding, for any given y, whether y is in the range of ¢z,. Theorem 1 Problems (a), (0), (c), (d), and (e) are recursively unsolvable. For every choice of yo, problem (f) is unsolvable. Problem (g) may or may not be recursively unsolvable, depending on the choice of xo. ; Proof. (a) Let g be the characteristic function of {zly. 4s @ constant function}. We wish to show that g is not recursive. — Define ¢, 2 partial function of two variables, by the instructions: given input , find P, and apply P, to input 2; if and when this converges, give output 0. By Church’s Thesis, ¥ is partial recursive. Clearly ¥ satisfies the condition a, if o(z) convergent; WOM = | divergent, if g.(z) divergent. By Theorem 1-V (the sm-n theorem), there is a recursive function Ay such that dgl¥(z,y)] = gar. (Ar ts Aalsi4(z0,2)], where zo is an index for ¥) ‘Thus 9, if (2) convergent; en) = | divergent, if ga(z) divergent. Therefore, eco is # coustant funetion if and only if y(2) is convergent, Now if g is recursive, then gha(= Az|g(hx(z))]) is a recursive function, and 1, if gals) convergent; mul) = | it gc(a) divergent. But gh: would be a recursive characteristic function for {zleste) convergent}, contrary to Corollary 1-VII, Hence g cannot be recursive.4 Unsolwable problems Theorem I (0) Let f be the characteristic funetion of {]e. has y in its range}. Let Ay be the recursive function defined in the proof for (a). If fis reeursive, then Az[f(hi(z),0)] would be a recursive characteristic function for {x|y.(z) convergent}, contrary to Corollary 1-VIE. (¢) Let f be « characteristic function for {|¢.(y) = 2}. If 7 is recursive, then Az{f(hi(2),0,0)] would be a recursive characteristic function for {zle.(x) convergent}, contrary to Corollary 1-VII, where again Ay is as in the proof for (a). @) Let f be the characteristic function for {le. = g,}. Take yo to be some fixed index for Az{0], then Az|f(hi(z),¥0)} would be a recursive characteristic function for {x|g.(z) convergent}, contrary to Corollary 1-VII, where h, is as in the proof for (a). (e) Let f be the characteristic function for {zly. has infinite range}. By methods similar to those in the proof for (a), a reeursive function hy ean be defined such that a) = { w if ¢.(x) convergent; Pints) divergent, if g.(x) divergent. If j is recursive, then Az{{(tn(2))] would be a recursive characteristic fune- tion for {x|y.(z) convergent}, contrary to Corollary 1-VII. (A) Let yo be given, and let f be the characteristic function for {zlys has yo in its range}. Thon if f is recursive, d2{f(te(z))] would be a recursive characteristic funetion for {2l¢s(z) convergent}, contrary to Corollary I-VIE, where he is the recursive function defined in the proof for (e). (g) Choose x so that ¢., is the identity function, that is, g.,(y) = for ally. Then {yly., has y én its range} has the recursive function dz[1] as its characteristic function. In Exercise 2-19 we shall see that values of 2 also exist for which {yly., has y in is range} fails to have a recursive characteristic function In each of (2) to (f), we have demonstrated that the problem in question is unsolvable by showing that the halting problem of Corollary 1-VI can be reduced to it, i.e., by showing that if we could effectively solve the problem in question, we could use this to get an effective method for solving the halting problem of Corollary 1-VII. The unsolvable problems in parts (a) to (f) of Theorem I (as well as ‘Theorem 1-VII, Corollary 1-VII, and Theorem 1-VIID are special cases of the following result, due to Rice [1953]. Let © be a collection of partial recursive functions of one variable. Then {x{p. € @} has a recursive char- Gcteristic function if and only if either © ie empty or © consists of all partial recursive functions of one variable. The proof of Rice’s theorem is given us Exercise 2-39. Generalizations will be given and proved in a later chapter. Theorem I is given as above in order to provide specific intro~ duetory examples of reducibility and reducibility proofs. $2.2 Problems in other areas 35 §2.2 UNSOLVABLE PROBLEMS IN OTHER AREAS OF MATHEMATICS A number of classical mathematical questions concern: the existence of algorithms for solving certain “problems.” By means of codings, as dis- cussed in §1.10, these questions can be restated as questions about the existence of recursive functions. In this latter, more precise form, a num- ber of these questions have been answered negatively. Some of the first results of this kind occurred in logie, through work of Gadel, Church, and Turing. The results of Gédel and Church concerned the existence of algorithms (“decision procedures”) for identifying the provable theorems in certain formal logical systems. We discuss these results briefly in §2.4 and more extensively in a later chapter. Other unsolvability results have been obtained in number theory and algebra. Consider polynomials in any number of variables, with rational- integer coefficients. Consider the problem of deciding, for any such poly- jnomial, whether or not that polynomial has a set of simultaneous real ots. ( is a set of simultaneous roots for the polynomial intr, . . - 2), if p(ra, . . . 4) =O.) This problem is solvable. Famil- far methods of analysis (including, for example, Sturm’s theorem) provide ‘an algorithm for making this decision (see Tarski [1948]). Is there an algorithm for deciding, given any such polynomial, whether or not that polynomial has a set of simultaneous roots in the rational integers (“dio- phantine roots’)? The “problem” concerned in this latter question is often referred to as Hilbert’s teath problem. At the present time it is not known whether this problem is recursively solvable. Davis, Putnam, and J. Robin- son [1961] have shown that a closely related diophantine problem is recur- sively unsolvable, namely, the problem of deciding, for any equation between exponential polynomials, whether or not that equation has a solution in the nonnegative integers. An exponential polynomial is a polynomial-like form, with nonnegative coefficients, in which variables can occur as exponents. The theory of groups provides an example of an unsolvability result in algebra. A presentation of a group is a Snite list of generators and relations determining the group. (We do not define these terms here.) The problem of deciding, for any presentation and any string of generators, whether or not that string can be transformed to the identity by the relations of that Presentation is called the word problem for groups. Novikev [1955] and Boone [1957] have shown this problem to be recursively unsolvable. In fact, they have each exhibited a single fixed presentation for which no algorithm for thus recognizing the identity exists. Their results, giving the recursive unsolvability of the word problem for groups, climaxed a sequence of results, by themselves and others, in which the word problems for various weaker kinds of algebraic structure were shown to be recursively%6 — Unsoluable problems Theorem I unsolvable.+ We remark in passing that the problem of deciding, for any presentation of a group, whether or not that presentation has a recursively solvable word problem, is itself known to be recursively unsolvable (see Exercise 2-29 for an analogous result). Unsolvability results have also been obtained in topology. The problem of deciding, when any two triangulations for four-dimensional manifolds are explicitly given, whether or not the manifolds are homcomorphic is recursively unsolvable (see Markov [1958]}. (The problem for two-dimen- sional manifolds is solvable by well-known methods.) In the case of problems from outside recursive function theory, the method used to show recursive unsolvability is almost, always that of reduc- tion, The problem in question is related to an unsolvable problem of recursive function theory—usually some form of the halting problem. Thus, for their unsolvability result on exponential polynomials, Davis, Putnam, and J. Robinson [1961] exhibit an effective procedure whereby, given any z, an exponential polynomial equation can be obtained—call it Ez—such that H, has a solution in nonnegative integers if and only if y.(z) is convergent, Description of such a reduetion procedure, and demonstration that it has the desired reduction property, can be accomplished only by detailed treatment of the recursive function formalism and by study of deeper facts about the particuler mathematical objects (e.g., exponential poly- nomial equations) occurring in the given problem. (One must show, in effect, that the objects of the problem are sufficiently flexible to “express” all instances of the halting problem.) A number of unsolyability results from logic, number theory, and algebra are given in Davis {1958] together with detailed reduction proofs. Such results constitute an interesting and vital application of recursive function theory; however, we do not give them further detailed treatment in this book. For examples of recursive unsolvability, we confine ourselves to “problems” from recursive function theory itself—problems where Church's ‘Thesis ean be conveniently used to demonstrate unsolvability. §2.3 EXISTENCE OF CERTAIN PARTIAL RECURSIVE FUNCTIONS In §§2.1 and 2.2, we were concerned with the existence or nonexistence of certain recursive characteristic functions. We now turn to several questions that have to do with the existence or nonexistence of partial reoursive functions. 7 In weaker structures, which do not necessarily possess an identity, the word problem can be formulated as the problem of deciding, for any two strings, whether or not one string can be transformed to the other by relations of the presentation. §2.3 Existence of partial recursive functions 37 First, can every partial recursive function be extended to a total recur- sive function? Theorem II gives a negative answer to this question. Theorem Il There exists @ partial recursive function such that for no yeeursive function f is it true that f is an extenston of y, i.e., true that (W2)[$(2) convergent = y(z) = f(z)]. Proof. A diagonal argument yields the result. Define y by the instruc- tions: to compute ¥(z), find P,, apply P, to x; and if and when an output occurs, take (2) + 1 as the value for ¥(z). By Church’s Thesis, y is partial recursive (¥ is dzl¢,,(z,2) +1], if ¢., is the universal function of Theorem 1-IV). Thus () = {[HO+1, if ex(2) convergent; ¥@) = | divergent, if gaCz) divergent. Let f be any recursive function, and let y be an index of f; that is f = y,. Since f is total, f(y) = yxy) is defined. Hence p(y) is defined, and vy) = f@) +1. Hence f cannot be an extension of ¥.2] Comment. Let P., be the instructions for y in the above proof. It is informative to examine what occurs when P,, is applied to a, Le. when an attempt is made to compute ¥(z,). P,, applied to z instructs the éomputing agent first to find instructions with index z, and apply them to x; But this latter situation is again the application of P., to 2. AD infinitely regressive sequence of similar situations is thus generated, each requiring the application of P., to x1; and the computation cannot termi- mate. ¥(z;) is for this reason undefined. Theorem 1-X, the normal-form theorem, suggests several further ques- ‘tions about partial recursive functions. Can each partial recursive function be obtained from some recursive function by a single application of »? In fact, can the function p be omitted from the statement of the normal-form theorem? Theorem III gives a negative answer to both these questions. Theorem HI There exists @ partial recursive function $ such that for na recursive f #9 tt true that y — rz{uyff(a,y) = 10. Proof. Define the partial function ¥ as follows. To compute ¥(z): find P., apply P, to z; and if and when an output occurs, set $(z) = 2. ¥ is evidently partial recursive. Thus a, if y.(x) convergent; ¥@) = | divergent, if ¢.(2) divergent. Assume there is a recursive f such that for all z, #(z) = wyLf(z,) = 1]. Then, yz(2) convergent = f(z,2) = 1; and ve(x) divergent = (Vy)if@y) + = fG@,x) + 1.38 Unsolvable problems Define g by Theorem IT af) ites =1; ae) { if fea) +1. g is recursive, since f(z,z) is defined for all x. But then g is a characteristic function for {z|y.(z) convergent}, contrary to Corollary 1-VIL3] Further results on partial recursive functions are given in Exercises 2-13 and 2-30 to 2-38. §24 HISTORICAL REMARKS The concept of recursive function has played a major role in recent developments in logic and foundations of mathematics. Some of the chief accomplishments of modern work in logie (in the nineteenth and twentieth centuries) can be informatively restated with the use of this concept. Per haps the two most important results of modern logie are: (1) the discovery of a precise eymbolic language within which all statements and proofs of mathematics can be made and in terms of which a combinatorial symbolic criterion of valid proof can be given and (2) the demonstration that there is no universal algorithmic procedure for determining whether or not a state- ment of this symbolic language is true.} The first result emerged early in this century through work of Frege, Russell, and Whitchead (which in turn built on work of Boole, Peirce, and others in the nineteenth century). In Prin- cipia mathematica, Whitehead and Russell [1910] presented a considerable body of mathematics within such a precise symbolic language. (The study of this language and of similar mathematical languages is a principal concern of mathematical logic.) Mathematicians then began to look for a universal algorithmic “decision procedure” for determining the true statements in such # language, During the 1920's a number of mathematicians were active in this search. Partial progress was made, and suecess was belicved by some to be imminent. Then Godel showed in his epochal paper [1931] that no algorithm out of a rather broad ¢lass could serve as such a decision procedure.t Gddel’s work (the Gadel incompletenese theorems) preceded the } Both these concerns were foreshadowed in work of the seventeenth-century mathe- matician and philosopher Leibnits, who set forth two goals for science and philosophy: (1) the discovery of a universal precise symbolic language (characteristica universalis) within which all statements of science could be made and from which an especially clear insight into the meaning and validity of those statements could be obtained and (2) the discovery of « method (caleulue ratiocinater) for manipulating the statements of this language in a way that directly elucidated their meaning and interrelationshipa. {This conclusion is not explicit in Gédel’s work. But it was evident that Gadel's methods could be applied to demonstrate the inadequacy of any algorithm from 9 clam, thst inctuded all algorithms of the type then being pursued (see Herbrand {1931, 1931aj; see also Godel (1931, p. 197}}, The argument by which this was evident can be recon structed es follows. Gadel used a coding from formulas (ie., statements in the formaltam §2.5 Discussion 39 general study of recursiveness. Subsequent work on the formal characteri- zation of recursive function (§1.5) and on the Basic Result (§1.6) showed, in effect, that the broad class of algorithms to which the Godel method applies includes all algorithms (Church [1936]}, and hence that no universal decision procedure can exist. (As is indicated in work of Kleene (see ‘Theorem 1-VIID), the existence of recursively unsolvable problems is a direct and almost trivial consequence of the formal characterization. The Basic Result, together with Church’s Thesis, yields that any demonstration of recursive unsolvability is, in itself, an absolute demonstration that a uni- versal decision procedure for mathematics cannot exist.) As we shall see, recursive function theory provides valuable further insights into unsolvability and into the relation of unsolvability to mathe matical logic. This book will be mainly concerned with concepts and strac- tures that arise out of the unsolyability phenomenon. $2.5 DISCUSSION The results and examples of this chapter suggest that unsolvable prob- Jems can be classified according to the way and extent to which they are reducible to each other, How can the concept of reducibility be made precise? Several distinct reducibility concepts are defined, and a study of the resulting classification is begun, in Post [1944]. We shall present results of this and later work in Chapters 6 to 10. Do all instances of unsolvability in some sense embody the unsolvability of the halting problem (Theorem and Corollary 1-VII)? That is to say, is the halting problem reducible to every unsolvable problem? To put it another way, is it true that every unsolvable problem can be demonstrated to be unsolvable by means of a reduction from the halting problem? This question (for an interesting class of problems known as the recursively enu- of Whitehead and Russell) of number theory into integers. Let A be any set definable in number theory, ie, expressible ag {2|-- +--+}, where “s+ ass." ig a statment of number theory involving the variable “x”, Gédel presented s technique by whieh, given any such set A, a formula @ with code number 2, can be found whose clear snd immediate meaning is z, ZA. (This technique is described further in §11.6,) Now for every algorithm (of the kind being studied by mathematicians seeking & uni- ‘versal decision procedure), it was evident (gia coding) that a set A could be found, defin- able in elementary number theory, such that for all 2, z € 1 if and only if 2 is @ code number of a formula asserted to be true by that algorithm. Applying Gédel's technique to get a corresponding © and zo, we see that either (i) the formula @ is true, in which case 2 & A, and the algorithm fails to tell us that @ is true, or (i) is false, in which case 29 & A, and the algorithm tells us thet this false formula is true. Kither way, the algorithm fails. Gédel’s original paper is explicitly concerned with matters of consistency and prova- bility in certain formal theories, We discuss these specific results further in Chapters 7 and Ll.40 -Unsolvable problems Theorem IE merable problemst) was one of Post’s [1944] main concerns. He answered it for certain kinds of redueibility. The answer for one of the most general and important redueibility concepts, however, remained unknown until the work of Friedberg [1957] and Muchnik [1956] in 1956 (see Chapter 10). §2.6 EXERCISESt §§.2 to 110 2-1. Show thot the function g defined in §1.3 is primitive recursive. (Hint: Let & = dalmaz (0,2 ~ A = daylmax 0, 2 — yl, f= de[min (21)], Show that A and j are primitive recursive, lise 4 to show h primitive recursive. Use A and f to get g) 2-2, Define f by 1, ife(z) +1; fa= {8 otherwise, Is f recursive? 2-3. Consider the list of primitive recursive derivations described in §1.4. Tet f. de the primitive recursive function determined by the (x + Ist derivation in this list =0,1,2,.... Define g ~ devif.(y)l. Is g recursive? Is g primitive recursive? 2-4. For sny k > 1, define the “halting problem for partial recursive functions of & variables.” Show that it is recursively unsolvable. (Hint: Reduction from ‘Theo- gem i-VIL ean be used.) 2-5. An instantaneous description (see §1.5) will be called finite if all but » finite number of cells on the tape are blank. Exch finite instantaneous deseription possesses unique minimal expression describing it, namely, that expression (see §1.5) which bas ‘no superfluous 5's. (a) Using these minimal expressions, describe a coding from finite instan- taneous descriptions into integers, : () Given any Turing machine end any finite instantaneous description, the machine can be started in that instantaneous description, sad 2 corresponding eomputa- tion will follow. (If the internal state of the instantaneous description dees not appear in the quadruples of the machine, the machine does nothing.) If and when a computs- tion halts, @ final finite instantaneous description is obtained. Hence, tia the fized cod- ing of (a), every machine yields a partial function of one variable. Let yz be the partial funetion yielded by machine Pin this wey. (Machine P, is the machine that ordinarily computes @.. To obtain y(y) for a given y, find the description with code number y; start P. in this deseription; and if and when it halts, take the code number of the finul deseription to be v.{y).) Ts y, partial recursive for all 2? _ ,., ©) Does the class ivey:, . . .) contain all partial recursive functions of one variable? (Hint: Show that for all «, domasn ¥, is either empty ot infinite.) {A problem is called recureively enumerable if it concerns a set or relstion of integers that can be effectively listed, i.c., is empty or is the range of some recursive function, where, for a relation to be the range of a function, a coding from K-tuples to integers must be used. All problems specifically discussed in §2.2 are recursively enumerable. ‘The problems of Theorem 1-VII snd Corollary 1-VII, and of ports (8), (c), (f), and () of Theorem I ate recursively enumerable. The problems of Theorem 1-VIII and of parts (a), (d), and (¢) of Theorem I are not. Recursive cnumerability will be discussed in Chapter 5. 4 Exercises are grouped according to related sections of the main text. Exercises on Chapter 1 ceeasionally use methods and concepts from Chapter 2. §2.6 Exercises a (2) Does a universal partial reeursive funetion exist for the class {Yoyes, . » 1? (Such a function might be called a generat universal function, since it can be used to duplicate the result of any Turing machine started in any Sinite instantaneous description. } (e) Show, by trivial reduction from Theorem 1-VII, that {|¥e(v) convergent} does not have a recursive characteristic function. (This gives the unsolva- bility of what might be called the general halting problem for Turing machines.) (f) Show that {eldomain y. = 9} has a reoursive characteristic function (in eontrast to the result of Exercise 2-17(5)). (Note: In 8 personal communication, Marvin Minsky and Hilary Putnam have announced that {2|¥z #8 lotel| does not have a recursive characteristic function (see also Exercise 2-9).) 2-6. Let {So5i, . . -] be an infinite alphabet of symbols available for use in Turing machines {in addition to B and 1). Modify the definition of Turing machine to sllow use of any of these symbols (any machine, of course, consists of a finite number of quad- ruples). Modify the definition of finite instantaneous description to allow for appear- ance of these symbols on the tape. (a) Describe coding for these new finite instantaneous deseriptions, () Show that the results of Exercise 2-5 all carry through for these new machines and configurations, where the numbering indicated by the notation “P,” is replaced by an appropriate numbering for the new machines. 2-7. Give a more formal proof for Theorem J-VIII analogous to the more formal proof for Theorem 1-VIL. 2-8. Prove Theorem 1-XI. (Hind: Use the unsolvability of the halting problem.) 2-9. Let A = {zlps is @ total function}. Prove that there is no recursive function f such that range { = A. (Hint: See proof of Theorem 1-VIIT.) 2010. Let ® be the class of all partial recursive functions of one variable. Let + be any map from N onto ®. + is called a aumbering for the partial reeursive functions of one variable. The standard Godel numbers of §1.8 provide such a numbering; call ite, A numbering + is acceptable if it is possibie to got back and forth effectively between rand mo, Le, if the following two conditions hold: Condition 1, There exists a recursive function f {not necessarily one-one} such that wof =~. (For every x number, we can find @ > number.) Condition 2. There exists a recursive function g (not necessarily one-one) such that wg =». (For every o number, we can find a z number.) (@ Show that any effective listing of all the Turing machines yields an acceptable numbering. (>) Show that condition 1 is # necessary and sufficient condition that any + have a partial recursive universal function, i.e, satisfy an appropriate version of Theorem 1-IV, the enumeration theorem. (c) Show thst condition 2 is a necessary and sufficient condition that any + have an s-m-n theorem; ie., satisfy an sppropriste version of Theorem 1-V. (Assume that the indexing for partial recursive functions of more than one variable remains the standard one of §1.8. This assumption can be dropped if the definition of numbering and condition 2 are both appropriately generalized to apply to partial recursive functions ‘of more than one variable.) (d) Show that condition 2 implies that x-y} is infinite for every partial recursive function y of one variable, i.e., that x satisfies an xppropriate version of Theorem LIL (int: Use result of Exercise 2-18(5).) Comment. The above results give invariant significance to the notion of acceptable numbering and to Theorems 1-IIT, 1-IV, and 1-V Gee discussion early in §1.8). In particular, say seceptable numbering has an enumeration theorem and an s-m-n theorem. Any numbering , by definition, includes all partial recursive functions of one variable.42 Unsotvable problems Condition 1 may be viewed as requiring that the numbering be “algorithmic,” i.c,, that every number yield an algorithm. Condition 2 may be viewed as requiring that the numbering be “complete,” i.e. “include all algorithms” (sco comment after Part III of Basic Result in §1.6, and see Exereiso 2-11). A211. Assume terminology of Exercise 2-10. Lot g be a recursive function giving 4 one-one mapping from N X N onto W. (Any effective listing of N X NV will supply such ag.) Define x, as follows. Given any y, find the #: and zz such that g(z,22) = y; then define ¥ = xi(y) to be the partial reoursive function antislying the following conditions: divergent, if 21 = 0; vio = { ree if a1 0; end v(2) = gale), allz #0. (a) Show that v, is s numbering, i.e., includes all partial recursive functions of one variable. @) Show thet, for m1, condition 1 holds but condition 2 does not hold. Comment. m1 is henee a numbering for all the partial recursive functions of one variable but does not “include ait algorithms.” ‘The enumeration theorem holds for it, but the s-m-n theorem does not, 2-12, Show that for every recursive f of one variable, there exists a recursive g such that f = dzlaylez,y) = UI) A219. Show that the partial recursive functions are not closed under p; ie., show that there exists a partial recursive y such thet X2(uy{H(2,y) = 1]]is not partial recursive, (int: Define zy) = Lify = Lorif both y = 0 and o,(2) convergent.) 2.14. Let a finite-state machine be a device with two tapes, each of which can move in only one direction, Input integers are presented ta the device on one tape; outpute are printed on the other. The device has a Snite number of internal states, and ite operation is determined by a finite set of ‘Turing-mnchine-like mechanical rules, Erasing operations are not used, (a) Is the following problem recursively solvable: the problem of deciding, given any finite-state machine and any input, whether or not the resulting computation alts? (8) Is the following problem recursively solvable: the problem of deciding, given any finite-state machine, whether or not that machine yields a total funetion? (©) The class of functions computable on such finite-state machines will depend upon the symbolism allowed for expressing input and output integers. Let Gq be the class of functions resulting when input and output integers are both written to the base n,n = 2, 3, . ‘Let @, be the class of functions resulting when inputs and outputs are both written as sequences of 1's, with each integer x represented as sequence of z+ 11's, Show that az[2s}isin 1 @. wal A(@) Show that az{z*] is not in U ay. aol (e) Give an example of nm > 2 such that @. ~ @a. Show that @1 x & for all a> 2. Show that © @, contains the nonnegative parts of all linear functions. pe (f) Characterize G1. (Answer: A funetion is in @) if and only if it is equal, at all but a finite number of places, to a function whieh can be expressed 2s the sum of & function of the form m[z/n} + g (a Hnear step function of nonnegative slope) and periodic function of period n.) Prove a unique-decomposition theorem. 215. Show that there exists a recursive h such that for all primitive recursive g, GxXV lz < y= gly) < My) §2.6 Exercises 8 A2-16, Can we conclude from the following assumptions, and using no other facts about x, that f is primitive recursive: (i) f is the junction mentioned, as f, in the first paragraph of §1.5; Gi) the funetion mentioned in (¢) in §1-1 is primitive recursive; and Gi) Fis total? (Hint: Show that thore is a real number p such that g = Azthe integer < 9 whose digit occura at the (2 + 1)st place in the decimal expansion of p] is primitive recursive, and such that f =, [the position of the left-hand digit of the first run of * or more 5's in the expansion of p] is total but not primitive recursive. Tako t, p, g from §1.10 (they may be assumed primitive recursive), and use the function A constructed in Exercise 2-15 to get this p so that: the decimal of p consists entirely of 0's and 5's; g is primitive recursive (via use of p, g, and Axwlé(m,x,y)}, where o, = h}; and (Wz)[A(z) < f(2)l. Experience with primitive recursive functions is helpful but not necessary.) $2 2-17. Show directly (e,, without the use of Rice's theorem, stated at the end of §2.1) that none of the following bas a recursive characteristic function: (a) icles Aas infinite domain}. (3) izle has emply domain}. (c) {|domain o. = domain gy}. 2-18. Show directly that for any given z,, neither of the following sets has a recursive characteristic function: (a) {z|domain y, = domein ox}. () tzlee = ea) (Hind for (a): Treat cases domain y,, = Band domain y., > 9 separately.) 2-19. Complete the proof of Theorem 2I(g} by showing that there oxists an zy such thet range yz, does not heve 8 recursive characteristic function, (#ini, See proof of Theorem 2-111.) 2-20, Show that there exist x» and 2 such that domain ys, has a recursive charac- teristic function but domain ¢-, does not. 2-21. Do parts (6) and (c) of Exercise 2-17 with “range” in place of “domain.” 2-22, Do part (a) of Exercise 2-18 with “range” in place of “domain.” 2623. Show that the problem of Theorem 1-VIT is directly zedueibie io the problem of Corollary 1-VIT. (Hint: Use a construction similar to that used in the proof of Theorem 2-1(e).) 2-24. Show that problem (¢} of Thoorem 21 is reducible to the halting problem. 42-25. Show that problems (6) and (j) of Theorem 21 are each reducible to the halting problem. 2-26. Show that none of problems (a), (d), or (¢) of Theorem 2-1 is reducible to the halting problem. A227. Let f and g be recursive functions such that for some set 4, A = range f and A = range g. Does A have a recursive characteristic function? Af-28. Modify the definition of Turing machine by allowing the special internal state g* in addition to 9o 91, . . . . Modify the method of associating partial functions with machines by stipulating that an output is to be used only when a machine stops in state q*. An input can hence fail to yield an output in either of two ways: (1) the computation does not terminate; (2) the computation terminates, but not in state ¢*. ‘We refer to the former situation as an infinite singularity and to the latter as a block singularity. (a) Is every partial recursive function obtained in this way? If 20, would ‘an effective listing of these modified machines yield an acceptable numbering in the sense of Exercise 2-107 (@) Is the following problem solvable: to decide, given any machine and input, whether or not an infinite singularity appeare?“4 Unsolvable problems (c) Do part (6) with “infinite” replaced hy “block.” (a) Can we “replace” block singularities by infinite singulsrities; i.e, is there an effective procedure for going from any machine M to a new mschine M’ such that Bf and M” represent the same partial function and such that M’ has only infinite singularities? {e) Can we “replace” infinite singularities by block singularities? (Hint: Answers to (¢) and («) are “yes” and “no,” respectively.) $2.2 42-29. Consider the “metaproblem” of deciding, for any zp, whether or not problem (@) of Theorem 2-1 is recursively solvable. Show that this metaproblem is recursively unsolvable; ie., show directiy that {z\range e, haz a recursive characteristic function} docs not have a recursive characteristic function. $2.3 2-30. Let two partial recursive functions have nonempty sets A and B as their respective domains, and assume that A) B = 9. {a) Does there necessarily exist 2 partial recursive function ¥ such that ¥(A) = 10} and YB) = {1}? A(b) Does there necessarily exist a recursive funotion f such that f(A) ~ {0} and f(B) = 1}? Gini: Prove Theorem 2-11 (the extension theorem) for the case of functions and partial functions whose ranges are subsets of [0,1}.) 2031. Ts dzluylec(y) divergent)) partial recursive? A212, Let two partial recursive functions have A and B as their respective domains, Give a necessary and sufficient condition that A and B satisfy the following: there erists 4 partial recursive such that for ali z, {domain gs = A => yz) = 0] ond [domain gs = B = (2) = 1h 2-33. Consider the problem of deciding, for any z, whether or not g; is extendible to a recursive function. Is this problem recursively solvable? 43-34, Let A and B be sets. Consider the “problem” of deciding, for any given # EA, whether or not z GB. We say that this problem is relatively solvable if there exists 4 partial recursive function y such that [z €.4 B= dz) = iJand [gs EAA 5 = oz) = 0]. Investigate relative solvability of the following problems: (a) The problem of deciding, for any z such that w- is total, whether or not isa constant function, (in this ease, A = (alex is tofal} and B = (zp. is a constant Function}.) (5) The problem of deciding, for any z such that y, is tots}, whether or not os 15 one-one. (c) Tho problem of deciding, for any z, y, and 2 such that , is total, whether or not g-(y) = 2. (a) The problem of deciding, for any z such thet ¢, is totsl, whether or not range os is infinite. (Hert: For unsclvability, make reduetions from the halting problem (Corollary J-VI1) which proceed by associsting with each z @ total recursive function whose value for any argument y depends upon the result of performing y steps in the computation of P. applied to z.) Definition, If + is. partial function of two variables, define av = AslaviXayy) = O}}. 2-45. What functions occur in (| recursive gf = uo])? 2-36. (Uspenskii {1957].) If y is partial recursive, must av be partial recursive? A257. Characterize {f|f recursive & (Wz)(4 recursive g)lo: = Jugl|. (Hint: Consider functions which take eaeh integer as value infinitely often. For related. results, sce Markov [1947] and Kuznecoy [1950].) §2.6 Exercises 45 A2-38. Characterize {y{(@ recursive g)[~ = ug]}. (Hint: Consider thoso functions which, ag reistions, have a teoursive characteristic function, By Theorem 2-IIL, this class does not include all partial recursive functions. For related results, see Skolow [1944] and Post [1946).) Exercises 2-35 to 2-38 complete the structural theory begun in ‘Theorems 1-X and 2-111. 425 2-39. (2) Prove Rice’s theorem, stated at the end of §2.1. (Hint: Show that a recursive characteristic function for any other case could be used to solve the halting blem.) me {b) Deduce, as an immediate consequence of this theorem, that: {z\domain ex is infinife] cannot have a recursive characteristic function. (c) Deduce the results of Exercises 2-17, 2-18, 2-29, and 2.33.3B Purposes; Summary §3.1 Goals of Theory 46 $3.2 Emphasis of This Book 48 $3.3 Summery 48 §3.1 GOALS OF THEORY The material presented in Chapters 1 and 2 concerned major features of recursive function theory as developed before 1940, Included were the Basic Result, certain more technical results, such as the enumeration theo- rem (Theorem 1-IV) and the s-m-n theorem (Theorem 1-V), and the recur- sive unsolvability of specific problems embodying the halting problem. Development since 1940 has been varied and extensive. For purposes of summary, the goals and results of this more recent research can be grouped into six areas: (1) unsolvable problems; (2) unsolvability structures; (8) logic and foundations; (4) subrecursiveness; (5) recursive structure; (6) analogue structures. In our brief description this grouping is somewhat artificial and ignores important interrelations. 1. Unsolvable Problems This area was discussed in Chapter 2, It concerns results on the solva- bility or unsolvability of specifie problems. Since 1940, increasingly refined techniques for demonstrating unsolvability have been develeped, and, as a result, questions of unsolvability have been settled for increasingly broad classes of problems. A few typical results were stated in §2.2. Continuing effort has also gone into the study of solvable problems and into the identifi- cation of solvable subcases of more general problems known to be unsolvable. 2. Unsolvability Structures This area concerns concepts and structures that arise in a more general analysis of the unsolvability phenomenon; i.c., it concerns concepts which prove useful in classifying unsolvable problems (e.g, various exact versions of the idea of reductbility mentioned in Chapter 2), and it concerns structures which appear as @ result of such a clessification. Work in this area bas broad application to work in other areas. Its concepts, terminology, and results are widely used—especially in logic and foundations, where a number 46 §3.1 Goals of theory 47 of major resulis find their most natural expression in these terms. More abstract and general kinds of classification can also be studied, and certain generalized kinds of computability ean be defined and investigated. Area (2), sometimes called the area of recursive invariance,t includes the study of hierarchies, where clessifications by reducibility are related to classifica- tions by complexity of definition in certain formalized logical systems. 3. Logie and Foundations Applications of recursive function theory to logic have been extensive and profound, yielding new results and new insights. ‘As indicated in the historical discussion of §2.4, logic is, by its nature, concerned with algo- rithms. Many of the specific unsolvability results of area (1) have been obtained in logic (most often, such an unsolvability result will concern the problem of identifying the provable statements in some formalized system of deduction). A variety of syntectical and semantical notions (eg. oxiomatizability and incompletableness) can be usefully restated in recursive- tunction-theoretic form, and various more subtle distinctions can be drawn. t Results from area (2) can be used to study the relative power of various Iogical methods, systems, and means of expression. In modern work on foundations of mathematics, the notion of consiructivity has also played a basic role; recursive function theory has proved useful in reformulating and analyzing various notions of constructive proof and in isolating the construc- tive content of known classieal proofs. 4. Subrecursiveness The study of generalized forms of recursivencss was included in area (2). Area (4) concerns concepts of computability more restrictive than recursive ness, and it concerns the way in which such concepts ean be used to refine and stratify structures thet arise in the study of ordinary recursivencss. Natural and significant results in this direction have been difftcult to obtain. Considerable effort has been made in this area from the time of the earliest work on recursive function theory.§ (Primitive recursiveness is, of course, an, example of such 2 more restrictive notion.) 5. Recursive Structure This area concerns ways in which recursive structure can be imposed ‘upon existing and familiar mathematical objects to yield a new and richer t Recursive invariance will be defined in Chapter 4. { Syntaz is the study of formalized systems as pure formalisms, apart from intended meanings. Semantics (in logic) is the study of the relation between formalized systems and the mathematical objects (e.g., real numbers) about which the systems appear (or are intended) to speak. The distinction between syntax and semantics is intuitively useful but diffeult to make fully precise. 4A number of methods and results antedate the definition of recursiveness, and can be viewed as a part of the search for s satisfactory general definition of algorithm.48 Purposes; summary theory. The imposition of such structure is analogous to the imposition of topological structure on groups to yield the theory of topological groups. 6. Analogue Structures Unlike area (6), where recursive function theory is used further to analyze existing mathematical objects and results, area (8) concerns the definition and study of new mathematics] objects which are, in some sense, recursive analogues to more familiar (and nonrecursive) mathematical objects. The nonrecursive objects may use nondenumerable sets or have other nonconstructive features. In the recursive analogues, all basic sets are (at most) denumerable, all basic entities are given integer code numbers, and mappings are taken to be admissible only if they are partial recursive on code numbers. Most work done on analogue structures has been in the realm of set theory. Cardinal numbers, ordinal numbers, topological spaces, and real-variable analysis have al] been studied in recursive-analogue form. Work on recursive analogues has been of continuing interest, both in the results and insights it affords and in the further questions it raises. §3.2 EMPHASIS OF THIS BOOK The chief concerns of this book will be area (2) (unsolvability structures) and (to a lesser extent) area (3) (logic and foundations). Many of the proofs will rely on the semiformal methods deseribed and justified in Chap- ter 1. We shall omit area (1) (unsolvable problems) insofar as it concerns specific problems in other parts of mathematics. More formal methods necessary for treating such problems can be found in Davis [1958]. As was done in Chapter 2, we use Church’s Thesis to give a semiformat treatment of unsolvable problems oceurring within recursive function theory itself. Several illustrations from area (6) (analogue structures) will be presented. Some indication of work in areas (4) and (5) (subrecursivencss and recursive structure) will be given. §3.3 SUMMARY Chapters 1, 2, and 3 present the concept of recursiveness. The Basic Result is given, and particular instances of unsolvability and reducibility are described. The scope and usefulness of semiformal techniques are indicated. In the remaining chapters (4 to 16), we define and consider certain gen- eral concepts that yield the examples and ideas of Chapter 2 as special cases, and we go on to treat further concepts that grow, in tur, out of these. Use and occurrence of these ideas in logic is also discussed. Chapters 4 $3.3 Summary 49 to 16 thus fall in area (2) (unsolvability structures) with some material from area (3) (logic and foundations). Occasional examples from areas (4), (5), and (6) are given, More specifically, Chapters 4 to 16 cover the following. Chapter 4 uses the concept of recursixe invariance to characterize more accurately the sub- ject matter of area (2). Chapter 5 introduces the basic concepts of recursive and recursively enumerable set. Chapters 6 to 9 define and investigate the basic reducibility concepts. Chapter 10 discusses a central probiem about reducibilities left unsolved by Post (see §2.5). Chapter 11 presents the recursion theorem. In expositions of recursive function theory, this basic and useful tool is often presented at the beginning. The theorem is not necessary, however, for the material that precedes it below. We believe that its position in this book gives the student a better perspective and deeper understanding of its use. Chapter 11 includes a variety of applica tions of the recursion theorem and concludes with an introduction to ordinal notations (an analogue structure from area (6)}. Chapters 12 and 13 give further resulis on the besie structures of Chapters 5 to 9; Chapter 12 con- siders the recursively enumerable sets as a lattice of sets, while Chapter 13 considers degrees of unsolvabiity (a degree is an equivalence class of sets under an equivalence relation of interreducibility) under partial ordering by reducibility. Chapter 14 introduces the arithmetical hierarchy (a classi- fication of sets of integers according to minimum logical complexity of definition) and relates this structure to the degrees of unsolvability studied in Chapter 13. Chapter 15 extends the arithmetical hierarchy from 2 classification of sets of integers to a classification of sets of sets of integers. Chapter 15 also treats the notion of recursive functional. Chapter 16 intro- duces the analylical hierarchy, a classification similar to, but more inclusive than, the arithmetical hierarchy. Chapter 16 also considers generaliza- tions of the notion of effective computability. Applications to logie occur throughout the book. Bibliographical references are incomplete. We make acknowledgement when we know the correct attribution. References for more advanced study are occasionally suggested, although the choice of such references is somewhat arbitrary. For examples of work in area (1), see Davis [1958]; Tarski, Mostowski, and Robinson [1953]; Rabin and Scott [1959]; Minsky [1961]; and Dreben [1962]; in area (4), see Grzegorezyk [1953]; Fischer [1962]; Ritchie [1963]; Axt [1959]; and Feferman [1961]; in area (5), see Rabin [1960]; in area (6), see Dekker and Myhill [1960]; Moschovakis [1963]; and Crossley [1968].