KEN NET H

E.

I \T E R SON

A Programming
Language

JOHN WILEY AND SONS, INC.

NEW YORK LONDON SYDNEY

Copyright

1962 by John Wiley & Sons, Inc.

All Rights Reserved
This book or any part thereof
must not be reproduced in any fornl

without the written permission of the publisher.

FOUHTH PHINTING, 1\IAY,

lUG?

Library of Congress Catalog Card Nun1ber: 62-15180

Printed in the United States of America

To My MallY Teachers

PREFACE

Applied mathematics is largely concerned with the design and analysis of

explicit procedures for calculating the exact or approximate values of
various functions. Such explicit procedures are called algorithms or
programs. Because an effective notation for the description of programs
exhibits considerable syntactic structure, it is called a prograrnrning
language.
Much of applied mathematics, particularly the more recent conlputer
related areas which cut across the older disciplines, suffers from the lack
of an adequate programming language. It is the central thesis of this book
that the descriptive and analytic power of an adequate progranlnling
language amply repays the considerable effort required for its mastery.
This thesis is developed by first presenting the entire language and then
applying it in later chapters to several major topics.
The areas of application are chosen primarily for their intrinsic interest
and lack of previous treatment, but they are also designed to illustrate the
universality and other facets of the language. For example, the Inicro
programming of Chapter 2 illustrates the divisibility of the language, i.e.,
the ability to treat a restricted area using only a small portion of the
complete language. Chapter 6 (Sorting) shows its capacity to compass a
relatively complex and detailed topic in a short space. Chapter 7 (The
Logical Calculus) emphasizes the formal manipulability of the language
and its utility in theoretical work.
The material was developed largely in a graduate course given for
several years at Harvard and in a later course presented repeatedly at the
IBM Systems Research Institute in New York. It should prove suitable
for a two-semester course at the senior or graduate level. Although for
certain audiences an initial presentation of the entire language TIlay be
appropriate, I have found it helpful to motivate the developlnent by
presenting the minimum notation required for a given topic, proceeding
to its treatment (e.g., microprogramming), and then returning to further
notation. The 130-odd problems not only provide the necessary finger
exercises but also develop results of general interest.
vii

viii

Preface

Chapter I or sonle part of it is prerequisite to each of the remaining

"applications" chapters, but the applications chapters are virtually
independent of one another. A complete appreciation of search techniques
(Chapter 4) does, however, require a knowledge of methods of representa
tion (Chapter 3). The cross references which do occur in the applications
chapters are either nonessential or are specific to a given figure, table, or
program. The entire language presented in Chapter 1 is summarized for
reference at the end of the book.
I n any work spanning several years it is impossible to acknowledge
adequately the many contributions made by others. Two major acknowl
edgnlents are in order: the first to Professor Howard Aiken, Director
Emeritus of the Harvard Computation Laboratory, and the second to
Dr. F. P. Brooks, Jr. now of IBM.
It was Professor Aiken who first guided me into this work and who
provided support and encouragement in the early years when it mattered.
The unusually large contribution by Dr. Brooks arose as follows. Several
chapters of the present work were originally prepared for inclusion in a
joint work which eventually passed the bounds of a single book and
evolved into our joint Automatic Data Processing and the present volume.
Before the split, several drafts of these chapters had received careful
review at the hands of Dr. Brooks, reviews which contributed many
valuable ideas on organization, presentation, and direction of investiga
tion, as well as numerous specific suggestions.
The contributions of the 200-odd students who suffered through the
development of the material must perforce be acknowledged collectively,
as must the contributions of many of my colleagues at the Harvard
Computation Laboratory. To Professor G. A. Salton and Dr. W. L.
Eastman, I am indebted for careful reading of drafts of various sections
and for comments arising from their use of some of the material in courses.
Dr. Eastman, in particular, exorcised many subtle errors fronl the sorting
programs of Chapter 6. To Professor A. G. Oettinger and his students
r am indebted for many helpful discussions arising out of his early use of
the notation. My debt to Professor R. L. Ashenhurst, now of the Univer
sity of Chicago, is apparent from the references to his early (and un
fortunately unpublished) work in sorting.
Of my colleagues at the IBM Research Center, Messrs. L. R. Johnson
and A. D. Falkoff, and Dr. H. Hellerman have, through their own use of
the notation, contributed many helpful suggestions. I am particularly
indebted to L. R. Johnson for many fruitful discussions on the applications
of trees, and for his unfailing support.
On the technical side, r have enjoyed the assistance of unusually compe
tent typists and draughtsmen, chief among them being Mrs. Arthur

Preface

ix

Aulenback, Mrs. Philip J. Seaward, Jr., Mrs. Paul Bushek, Miss J. L.

Hegeman, and Messrs. William Minty and Robert Burns. Miss Jacquelin
Sanborn provided much early and continuing guidance in matters of style,
format, and typography. I am indebted to my wife for assistance in
preparing the final draft.
KENNETH

May,1962
Mount Kisco, New York

E.

IVERSON

CONTENTS

Chapter 1 THE LANGUAGE

1.1

I ntrod uction, 1

1.2

Programs, 2

1.3 Structure of the language, 7

Conventions, 7

Literals and variables, 9

Domain and range, 10

1.4 Elementary operations, 11

Arithmetic operations, 11

Logical operations, 11

Residues and congruence, 12

1.5 Structured operands, 13

Elementary operations, 13

Matrices, 14

Index systems, 14

1.6

Rotation, 14

1. 7

Special vectors, 15

1.8

Reduction, 16

1.9 Selection, 17

Compression, 17

Mesh, mask, and expansion, 19

1.10 Selection vectors, 22

1.11 The generalized matrix product, 23

1.12 Transpositions, 25

1.13 Special logical matrices, 26

xi

Contents

xii

1.14 Polynomials and positional number systems, 27

1.15 Set operations, 28

1.16 Ranking, 31

1.17 Mapping and permutations, 32

Reordering operations, 32

Permutations, 33

Function mapping, 34

Ordering vector, 36

1.18 Maximization, 36

1.19 Inverse functions, 37

1.20 Levels of structure, 39

1.21 Subroutines, 39

1.22 Files, 40

1.23 Ordered trees, 45

Directed graphs, 45

Ordered trees, 46

Right and left list matrices, 49

Well formation, 52

The index matrix as a function of the degree vector. 53

Tree, path, and level compression, 54

Extension of other operations to trees, 57

Homogeneous trees, 58

References, 62

Exercises, 62

Chapter 2 MICROPROGRAMMING
2.1 Instruction preparation, 72

Additive indexing, 73

Indirect addressing, 74

Dynamic relocation, 74

Branching, interruption, and trapping, 75

Complete instruction fetch, 76

2.2 Instruction execution, 77

Load and store, 78

Branch instructions, 80

71

Contents

xiii

Logical instructions, 83

Arithmetic instructions, 84

Shift instructions, 85

Convert instructions, 87

Input-output instructions, 88

2.3

Detailed logical design, 101

References, 103

Exercises, 103

Chapter 3 REPRESENTATION OF VARIABLES

3.1

105

Allocation and encoding, 105

3.2 Representation of structured operands, 106

The grid matrix, 106

Linear representations, 108

Nonlinear representations, 109

Chained representations, 110

Partitions, 115

Pools, 116

Sumnlary, 118

3.3

Representation of matrices, 119

3.4 Representation of trees, 121

Simplified list matrices, 122

The use of left lists, 122

Chain list matrices, 126

References, 128

Exercises, 129

Chapter 4 SEARCH TECHNIQUES

4.1 Scanning methods for ranking, 135

Directed scan, 135

Controlled scan, 141

4.2 Key transformations, 144

Nonunique key transformations, 147

133

Contents
4.3

xiv

Multiple keys, 155

References, 155

Exercises, 156

Chapter 5 METAPROGRAMS

5.1

Compound statements, 160

5.2

Lukasiewicz notation, 163

5.3

The minimax form of an Y-formula, 166

159

5.4 Translation from complete parenthesis to Lukasiewicz

notation, 169

5.5 Translation from Lukasiewicz to complete parenthesis

notation, 173

References, 174

Exercises, 175

Chapter 6 SORTING

6.1 Serial sorting methods, 177

Copy operations, 177

Simple classification and merge, 178

Classification and simple merge, 186

Partial pass methods, 191

6.2 Evaluation of serial sorting methods, 199

Simple classification and merge, 200

Classification and simple merge, 201

Partial pass methods, 203

6.3

Aids to serial sorting processes, 204

6.4 Internal sorting methods, 205

Simple classification and merge, 206

Classification and simple merge, 211

Special internal sorting methods, 212

6.5 Evaluation of internal sorting methods, 231

Expected number of transpositions, 232

Bubble sort, 233

176

Contents

xv

Ranking sort, 235

Odd-even transposition sort, 236

Repeated selection sort, 236

Sorting with replacement, 239

Comparison of internal sorting methods, 239

Appendix to Chapter 6, 239

References, 240

Exercises, 241

Chapter 7 THE LOGICAL CALCULUS

7.1

246

Elementary identities, 246

7.2 Canonical forms, 247

I ntrinsic vector, 247

Characteristic vectors, 249

7.3 Decomposition, 255

Disjunctive canonical form, 255

Other canonical forms, 260

References, 262

Exercises, 263

SUMMARY OF NOTATION

265

INDEX

277

ILLUSTRATIONS

Program
Program
Program
Program
Program
Program
Program

1.1
1.2
1.3
1.4
1.5
1.6

1.7

Table 1.8

Program 1.9

Program 1. 10

Table 1. I I

Program 1. 12

Progranl 1. I3

Program I. 14

Progranl I. I5

Figure I. I 6

Figure 1.17

Finite program
2

lnfinite program
Program for :r == n 2/3

Linear arrangement of Program 1.3

4

Matrix multiplication
4

Matrix lTIultiplication using leading decisions

6

A reformulation of Program 1.6, using an algebraic

statement of the branching

7

Typographic conventions for classes of operands

9
Selection on ban k ledger L (Example I. I)
19

I nterfiling program
20

The Cartesian prod uct A == Xl @ x 2 @ x 3

31

Mapping defined by a permutation vector j

35

Modes of subroutine reference

40

PrograITI for Exanlple 1.3

42

Program for Example 1.4

44

A general triply rooted tree with A(T) == 16, v(T) ==

(3, 3, 4, 3, 2), veT) == 5, fL(T) == (3, 7, 8, 5, 3), and

Il(T) == 26
45

A graphical representation of the directed graph

(n, U)

Program 1.18
Table 1.19
Progranl 1.20
Program 1.2 I

Figure 1.22
Progranl 1.23
Figure 1.24

47

Determination of i such that T i == X

49

Full list nlatrices of the tree of Figure 1. I 6

50

Deternlination of the index nlatrix I associated with a

54

right list nlatrix R

Determination of the index matrix I associated with a

55

left list matrix L

Compression of tree T of Figure 1.16 (with n == alpha-

bet)
56

Determination of the left list L == cx 2 /[(U/T)

57

Set selection and maximunl prefix and suffix operations

59

xvii

xviii

Illustrations

Figure 1.25

Homogeneous tree H and dispersion and 1110nlent

vectors
60

Table 2.1
PrograITI 2.2
Program 2.3
Program 2.4
Program 2.5
Program 2.6
Program 2.7
Program 2.8
Progran1 2.9
Program 2.10
Program 2.11
Progran1 2.12
Table 2.13
Table 2.14
Program 2.15
Program 2.16
Program 2.17
Program 2.18
Progranl 2.19
Program 2.20
Program 2.21
Figure 2.22

Central computer operands

72

Basic instruction fetch

73

Indirect addressing
74

Complete instruction fetch

77

Load and store instructions

79

Load and store index instructions

80

Skip type and other special branches

81

Basic transfer type branches

82

Logical operations
84

Add instructions
85

Shift instructions
86

Convert instructions
87

Channel operands
89

Input-output unit operands

90

Tape unit <D/

91

Channel i
93

I nstructions for special tape unit functions

94

Select unit and load channel instructions

96

Jnput-output branch operations

99

Trap control and store channel

100

Base two addition

101

Data paths
102

Figure 3.1 a
Figure 3.1 h

Representation of a vector x
107

Linear representation of r(x), with I == 34 + 4i + 2)

107

Linear representation of a matrix X

107

Deternlination of z == p(x k ) and z == x k fronl a linear

representation of the grid matrix r(x)

110

Determination of p(x k ) from a chained representation

of x
112

Progranl for x ~ vjx on a forward chained representa

tion of x and a backward chained stack of available

segments
114

Determination of p(xk ) from a chained representation

of x with ternlinal partitions at

116

Determination of the row vector Ai fronl a row-chai ned

representation of A
121

The conlpound logical statement .r 1\ (y V :)

122

Figure 3.2
Progranl 3.3
Program 3.4
Program 3.5

Progranl 3.6
Progranl 3.7
Figure 3.8

Illustrations

xix

Program 3.9

Partitioning of the left list of an n-tuply rooted tree

Figure 3.10
Program 3.11

Construction of a Huffnlan prefix code

124

Construction of the binary Huffman code T for charac

ters c with frequency f

125

Construction of the left list z of the binary Huffman

code for characters c with frequency f

126

Chain lists of the tree of Figure 1.16

127

Determination of the path p = T r from the right chain

list matrix P
128

123

Program 3.12
Table 3.13
Program 3.14

Table 4.1
Figure 4.2
Program 4.3
Program 4.4
Figure 4.5
Figure 4.6

Scan fraction of directed scans for uniform distribution

of arguments
136

Plot of cumulative probabilities of Table 4.1

137

Ranking of x in k by binary search

141

Generalized binary search

143

Tree traced by Progranl 4.4

143

Uniform tree and mapping vector for the set k = (200.

010, 120,001,022,202)
Figure 4.7
Program 4.8
Program 4.9

Program 5.1
Figure 5.2
Table 5.3
Program 5.4
Progranl 5.5

145

Contracted tree and associated leaf list matrix M for

the set k = (200, 010, 120,001,022,202)

146

Biunique transformation on key represented by x, using

the leaf list of the contracted tree of key set k

147

Programs and examples for methods of scanning

equivalence classes defined by a I-origin key trans

formation t
151

Analysis of the compound statenlent

z +- (x + y) X I' + (s - t)1l
161

Representations of a compound statement

162

Phrases in Y-system for symbolic logic based on

operators and, or, /lot

164

Evaluation of the fornlula z

165

Transformation of the fornlula z to mlnunax fOfln

168

Program 5.6
Progranl 5.7
Program 5.8

Translation from complete parenthesis to Lukasiewicz

notation
170

Translation from complete parenthesis to Lukasiewicz

notation with full checking of singularity

172

Translation fronl Lukasiewicz to complete parenthesi~

notation with complete test of singularity

174

xx

Illustrations

Table 6.1
Figure 6.2
Figure 6.3
Prograll1 6.4
Program 6.5
Program 6.6
Prograll1 6.7
Program 6.8
Table 6.9
Table 6.10
Prograll1 6.11
rrable 6.12
Prograll1 6.13
Table 6.14
Table 6.15
Progran1 6.16
Progran1 6.17
Progran1 6.18
Progran1 6.19
Progran1 6.20
Progran1 6.21
Table 6.22
PrograIn 6.23
Progran1 6.24
Table 6.25
Progranl 6.26
'fable 6.27
Program 6.28
Progran1 6.29
Figure 6.30
Figure 6.31
Prograln 6.32
Prograln 6.33

Prograln 6.34
Progran1 6.35
Tahle 6.36

176

Types of file operations

Copy operation
179

Simple classification and n1erge

180

1J7-way n1erge sort

182

177-Way single phase n1erge sort

]84
Single-phase nlerge without rewind
185

Two-phase column sort

187

Single-phase colunlll sort (output 011 h files)

188

The partial pass column sort

]92

Matrix M describing the partial pass colull1n sort of

Table 6.9
193

Method of partial passes for 171 files

194

AIl1phisbaenic sort on 3-digit base 3 key

196

An1phisbaenic sort
197

Cascade Inerge sort

199

J Ilternal aid to col un1n sort (177 = 4)

205

I nternal single-phase n7-way n1erge using 2117 fields

207

I nternal single-phase nlerge using two fields

208

Two-way internal single-phase 1l1erge

210

Pre-count colulnn sort

212

Radix exchange with v(h) < 2

215

Radix exchange with l'(p) = l{d) + 1

216

Bubble sort example

217

Bubble sort
218

Forward and backward bubble sort with string indicator

219

Odd-even transposition sort exan1ple

2I9
Odd-even transposition sort
220

Ranking sort example

221

Ranking sort
222

Ranking by insertion
222

The tree T representing the third-degree selection sort

for v(T) = (1, 2, 2, 4)

224

Contin uation of the third-degree selection sort

224

Repeated selection sort

225

Repeated selection sort of Progran1 6.32 executed on

the right list node vector r of a unifonn rooted h-way

tree
227

Repeated selection sort with replacelnent

228

Bubble sort with replacen1ent

230

234

Coefficients for ;:;(11) =-::_: ('(11)\ '7T17/2

1
'((/)

Illustrations

xxi

Table 6.37

Characteristics of internal sorting nlethods (for random

distribution of distinct keys)
238

Table 7.1
Table 7.2
Table 7.3
Table 7.4
Progranl 7.5
Table 7.6

The disjunctive canonical form

248
Intrinsic, characteristic, and specific vectors
250
Relations anlong characteristic vectors
254
Decornposition of the functionf(x) on u
256
Decomposition off on u
260
Decomposition in exclusive disjunctive form
262

chapter

THE LANGUAGE

1.1

INTRODUCTION

Applied nlathenlatics is concerned with the design and analysis of

algorithnls or prograrns. The systelnatic treatment of conlplex algorithms
requires a suitable prografnrning language for their description. and such a
programming language should be concise. precise. consistent over a wide
area of application. mnemonic. and economical of synlbols; it should
exhibit clearly the constraints on the sequence in which operations are
perfornled; and it should pernlit the descri ption of a process to be inde
pendent of the particular representation chosen for the data.
Existing languages prove unsuitable for a variety of reasons. Conlputer
coding specifies sequence constraints adeq uately and is also comprehensive.
since the logical functions provided by the branch instructions can, in
principle. be elnployed to synthesize any finite algorithm. However. the
set of basic operations provided is not. in general. directly suited to the
execution of commonly needed processes, and the numeric symbols used
for variables have little mnemonic value. Moreover. the description
provided by computer coding depends directly on the particular represen
tation chosen for the data. and it therefore cannot serve as a description of
the algorithm per see
Ordinary English lacks both precision and conciseness. The widely used
Goldstine-von Neumann (1947) flowcharting provides the conciseness
necessary to an over-all view of the process. only at the cost of suppressing
essential detail. The so-called pseudo-English used as a basis for certian
autonlatic programming systems suffers from the same defect. Moreover.
the potential mnemonic advantage in substituting familiar English words
and phrases for less familiar but nlore conlpact mathenlatical symbols fails
to materialize because of the obvious but un\vonted precision required in
their use.
Most of the concepts and operations needed in a progralnming language
have already been defined and developed in one or another branch of
mathematics. Therefore, much use can and will be made of existing
notations. However. since n10st notations are specialized to a narrow

l.t

The language

field of discourse~ a consistent unification must be provided. For exanlplc~

separate and conflicting notations have been developed for the treatment
of sets~ logical variables~ vectors~ matrices~ and trees~ all of which may~ in
the broad universe of discourse of data processing~ occur in a single
algorithnl.

t.2

PROGRAMS

A program statement is the specification of some quantity or quantities

in terms of some finite operation upon specified operands. Specification is
symbolized by an arrow directed toward the specified quantity. Thus ".II
is specified by sin x" is a statement denoted by

+-- SIn

x.

A set of statements together with a specified order of execution consti

tutes a program. The program is .finite if the number of executions is
finite. The results of the program are some subset of the quantities
specified by the program. The sequence or order of execution will be
defined by the order of listing and otherwise by arrows connecting any
statement to its successor. A cyclic sequence of statements is called a loop.

z
v

+-

x 3.1416

v~vxx

Program 1.1
program

Finite

+-

z~yxz

z+-2xz

Program 1.2
progran1

Infinite

Thus Program 1.1 is a program of two statements defining the result l' as
the (approximate) area of a circle of radius :r~ whereas Program 1.2 is an
infinite program in which the quantity z is specified as (2y)n on the nth
execution of the two-statement loop. Statements will be numbered on the
left for reference.
A number of similar programs may be subsumed under a single more
general progranl as follows. At certain branch points in the program a
finite number of alternative statements are specified as possible successors.
One of these successors is chosen according to criteria determined in the
statelnent or statements preceding the branch point. These criteria are
usually stated as a comparison or test of a specified relation between a
specified pair of quantities. A branch is denoted by a set of arrows leading
to each of the alternative successors~ with each arrow labeled by the

1.2

Progranls

comparison condition under which the corresponding successor is chosen.

The quantities compared are separated by a colon in the statenlent at the
branch point, and a labeled branch is followed if and only if the relation
indicated by the label holds when substituted for the colon. The conditions
on the branches of a properly defined program must be disjoint and
exhaustive.
Program 1.3 illustrates the use of a branch point. Statement ~5 is a
comparison which determines the branch to statements /1 I, () I, or }' I,
according as z > n, z == n, or z < n. The program represents a crude but
effective process for determining x == n'2:\ for any positive cube n.

, - - - - - - - - - - ~3

+-

y x y

~4

+-

f31
{J2

+--

k -7 2

~ ~5

>

y +-71 - k

z : n

x x

<

--)'1

k+-kx2

y2

y+--y+k

(51
Program 1.3

Progranl for.c = 1/;\

Program 1.4 shows the preceding program reorganized into a compact

linear array and introduces two further conventions on the labeling of
branch points. The listed successor of a branch statement is selected if
none of the labeled conditions is met. Thus statement 6 follows statement
5 if neither of the arrows (to exit or to statement 8) are followed, i.e., if
z < n. Moreover, any unlabeled arrow is always followed ~ e.g., statement
7 is invariably followed by statement 3, never by statement 8.
A program begins at a point indicated by an entry arrOH' (step I) and ends
at a point indicated by an exit arrOH' (step 5). There are two useful
consequences of confining a program to the form of a linear array: the
statenlents may be referred to by a unique serial index (statement number),
and unnecessarily complex organization of the program manifests itself in
crossing branch lines. The importance of the latter characteristic in
developing clear and comprehensible programs is not sufficiently appre
ciated.

1.2

The language
y

~O

~y

x y

z~yxx

>

z . n

k~kx2

y~y+k

k~k+2

y~y-k

Program 1.4

Linear arrangenlent of Progran1 1.3

+-

peA)

+-

v(B)

C/ ~O
C/ ~ Cji + Aki

k~k-l

>

>

11

>

v(A)

Bjk

0
~j

j:O
i

10

+-

k
j

Program 1.5

i-I
~

Matrix lTIultiplication

A process which is repeated a nunlber of times is said to be iterated, and

a process (such as Program ] .4) which includes one or more iterated
subprocesses is said to be iteratire. Program 1.5 shows an iterative process
for the matrix multiplication
C~AB

1.2

PrograJJlS

defined in the usual way as

v(A)

C/ = L

Akl

B/"

k=l

where the dimension of an fn

columns) is denoted by p(X)

X n

ri

= 1,2,

,p,(A),

lj

1,2,

, v(B),

rectangular nlatrix X (of 111 rows and n

veX).

Program 1.5. Steps 1-3 initialize the indices, and the loop 5-7 continues to
add successive products to the partial sum until k reaches zero. When this
occurs, the process continues through step 8 to decrement j and to repeat the
entire sunlnlation for the new value ofj, providing that it is not zero. I fj is zero,
the branch to step 10 decrements i and the entire process overj and k is repeated
fronl j = v(B), providing that i is not zero. If i is zero, the process is cOl1lplete,
as indicated by the exit arrow.

]n all exarTIples used in this chapter, emphasis will be placed on clarity

of description of the process, and considerations of efficient execution by a
computer or class of computers will be subordinated. These considerations
can often be introduced later by relatively routine modifications of the
program. For example, since the execution of a computer operation
involving an indexed variable is normally more costly than the corre
sponding operation upon a nonindexed variable, the substitution of a
variable s for the variable C/ specified by statement 5 of Progranl 1.5
would accelerate the execution of the loop. The variable s would be
initialized to zero before each entry to the loop and would be used to
specify C/ at each termination.
The practice of first setting an index to its maxirTIum value and then
decrementing it (e.g., the index k in Program 1.5) pernlits the ternlination
conlparison to be made with zero. Since zero often occurs in conlparisons,
it is convenient to omit it. Thus, if a variable stands alone at a branch
point, comparison with zero is impfie-d. Moreover, since a corrip~lrisoj'--olf
an index frequeniiy'occurs immedtatei"y after it is modified, a branch at the
point of modification will denote branching upon conlparison of the
indicated index with zero, the comparison occurring ({fier rTIodification.
Designing programs to execute decisions irTImediately after modification of
the controlling variable results in efficient execution as well as notational
elegance, since the variable must be present in a central register for both
operations.
Since the sequence of execution of statements is indicated by connecting
arrows as well as by the order of listing, the latter can be chosen arbitrarily.
This is illustrated by the functionally identical PrograrTIs 1.3 and 1.4.
Certain principles of ordering may yield advantages such as clarity or
sinlplicity of the pattern of connections. Even though the advantages of a

The language

t.2

particular organizing principle are not particularly nlarked, the uniformity

resulting fronl its consistent application will itself be a boon. The schenle
here adopted is called the Inethod o.lleading decisions: the decision on each
parameter is placed as early in the program as practicable, nornlally just
before the operations indexed by the paranleter. This arrangenlent groups
at the head of each iterative segment the initialization, ITIodiflcation, and
the ternlination test of the controlling paranleter. Moreover, it tends to
a void progranl flaws occasioned by un usual values of the arguments. For

i -(- fleA)

+1

i-(-i-l

.i ~- vCR) +
.i ~-.i - 1

C/ -(- 0

k-(-k-l

C/

Program 1.6

~--

v(A)

~-

C/ + Aki

B/.;

1\1atrix n1ultiplication using leading decisions

exanlple, Progranl 1.6 (which is such a reorganization of Program 1.5)

behaves properly for matrices of dinlension zero, whereas Progrc.uTI 1.5
treats every nlatrix as if it were of dinlension one or greater.
Although the labeled arrow representation of progranl branches
provides a complete and graphic description, it is deficient in the following
respects: (I) a routine translation to another language (such as a conlputer
code) would require the tracing of arrows, and (2) it does not pernlit
prograITInled nlodification of the branches.
The following alternative fornl of a branch stateITIent will therefore be
llsed as well:
:r : y, r -~ s.
This denotes a branch to statement nunlber Si of the program if the relation
The paranleters rand S may themselves be defined and re
defined in other parts of the prograITI. The null elenlent 0 will be used to
denote the relation which complenlents the renlaining relations r i ; in
particular, (0) -~ (s), or sinlply -~s, will denote an unconditional branch to

:rr,.IJ holds.

1.3

Structure

(~f

the language

statenlent s. Program 1.7 shows the use of these conventions in a refa fill u
lation of Prograrn 1.6. More generalIy~ two or rnore otherwise independent
progranls may interact through a statenlent in one progranl specifying a
branch in a second. The statenlent nunlber occurring in the branch rllust
then be augrnented by the name of the progranl in which the branch is
effected. Thus the statenlent (0) -~ Program 2.24 executed in Progranl I
causes a branch to step 24 to occur in Program 2.

+1

i +-p(A)

+-

i - I , (=ft, =)

+-

v( B)

j +-j - 1,

+-

v(A)

+-

k - 1, (/, =)

-*

(3, 9)

-*

(5, 2)

+1
(~f~~

=)

-*

(8, 4)

Program 1.7 A refornlulation of Progranl 1.6, using an algebraic

statenlent of the branching

One statement in a program can be modified by another statclllcnt

which changes certain of its parameters, usually indices. Morc gencral
changes in statements can be effected by considering the progranl itself as a
vector P whose components are the individ uaL serially n unlbered state
ments. All the operations to be defined on general vectors can then bc
applied to the statements themselves. For exanlple, the jth statenlent can
be respecified by the ith through the occurrence of the statement Pj ~ Pi.
The interchange of two quantities y and or (that is, .r specifies y and the
original value of y specifies .r) will be denoted by the statement y ~~ > .r.

1.3

STRUCTURE OF THE LANGUAGE

Conventions
The Sunlrnary of Notation at the end of the book sunlnlarizes the nota
tion developed in this chapter. Although intended prinlarily for reference,
it supplements the text in several ways. It freq uently provides a nl0rc
concise alternative definition of an operation discussed in the text, and it

1.3

The language

also contains inlportant but easily grasped extensions not treated explicitly
in the text. By grouping the operations into related classes it displays
their family relationships.
A concise progralnming language Inust incorporate falnilies of opera
tions whose nleInbers are related in a systenlatic manner. Each fanlily will
be denoted by a specific operation sYInbol~ and the particular nlenlber of the
family will be designated by an associated controlling paranleter (scalar~
vector. Inatrix~ or tree) which inlnlediately precedes the nlain operation
symbol. The operand is placed illlmediately after the nlain operation
synlbol. For exanl ple~ the operation k t x (left rotation of x by k places)
Inay be viewed as the kth melnber of the set of rotation operators denoted
by the symbol t.
Operations involving a single operand and no controlling paranleter
(such as l.r]' or rrl) will be denoted by a pair of operation syInbols which
enclose the operand. Operations involving two operands and a controlling
paralneter (such as the mask operation la~ U, bl) will be denoted by a pair
of operation symbols enclosing the entire set of variables~ and the con
trolling parameter will appear between the two operands. In these cases
the operation symbols themselves serve as grouping synlbols.
In interpreting a compound operation such as k t (j ~ x) it is inlportant to
recognize that the operation symbol and its associated controlling paraln
eter together represent an indivisible operation and must not be separated.
]t would~ for example, be incorrect to assunle that j t (k ~ x) were
equivalent to k t (j ~ x), although it can be shown that the conlplete opera
tions j ~ and k t do commute, that is, k (j ~ x) === j ~ (k x).
The need for parentheses will be reduced by assunling that conlpound
statenlents are, except for intervening parentheses~ executed fronl right to
left. Thus k t j ~ x is equivalent to k (j ~ x), not to (k j) ~ x.
Structured operands such as vectors and nlatrices, together with a
systenlatic component-by-conlponent generalization of elenlentary opera
tions, provide an important subordination of detail in the description of
algorithms. The use of structured operands will be facilitated by selection
operations for extracting a specified portion of an operand, reduction
operations for extending an operation (such as logical or arithInetic
multiplication) over all components~ and permutation operations for
reordering components. Operations defined on vectors are extended to
nlatrices: the extended operation is called a rOlf operation if the under
lying vector operation is applied to each row of the nlatrix and a colurnn
operation if it is applied to each column. A colunln operation is denoted
by doubling the symbol employed for the corresponding row (and vector)
operation.
A distinct typeface will be used for each class of operand as detailed in

1.3

Structure

(~ll he

language

Table 1.8. Special quantities (such as the prefix vectors a/ defined in Sec.
1. 7) will be denoted by Greek letters in the appropriate typeface. For
mneIllonic reasons, an operation closely related to such a special quantity
Rcpresentation

Typc of
Operand

Printed

Typed

Literal
Alphabetic
Nunleric

Ronlan, u.c. and I.c.

Standard nunleral

Circled u.c. and I.c. rOlllall.

Standard nunleral

Variable
Alphabetic
Nunlcric

Italic. u.c. and I.c.

Italic nunleral

Unnlarked
Underscore

Vector

Lc. boldfacc italic

Underscore

Matrix

u.c. boldface italic

Underscore

Tree

u.c. boldface rOinan

Wavy underscorc

Table 1.8 Typographic conventions for classes of operands

will be denoted by the )ame Greek letter. For exalnple, 'Y.ju denotes the
maxiIllum prefix (Sec. 1.10) of the logical vector u. Where a Greek letter
is indistinguishable from a Roman, sanserif characters will be used, e.g"
E and I for the capitals epsilon and iota.
Literals and variables

The power of any nlathematical notation rests largely on the use of

symbols to represent general quantities which, in given instances, are
further specified by other quantities. Thus PrograITI 1.4 represents a
general process which deternlines.r = 11'2:\ for any suitable value of 11. In a
specific case, say 11 = 27, the quantity .r is specified as the number 9.
Each operand occurring in a llleaningful process nlust be specified
ultimately in terms of commonly accepted concepts. The synlbols
representing such accepted concepts will be called literals. Exanlples of
literals are the integers, the characters of the various alphabets, punctua
tion nlarks, and ITIiscellaneous syITIbols such as S and ~~. The literals
occurring in Program 1.4 are 0, 1, and 2.
It is important to distinguish clearly between general synlbols and
literals. In ordinary algebra this presents little difficulty, since the only
literals occurring are the integers and the decinlal point, and each general
sylnbol employed includes an alphabetic character. In describing nlore
general processes, however, alphabetic literals (such as proper naITIes) also

10

1.3

The language

appear. Moreover, in a computer program, numeric symbols (register

addresses) are used to represent the variables.
In general, then, alphabetic literals, alphabetic variables, numeric
literals, and nunleric variables may all appear in a complex process and
must be clearly differentiated. The symbols used for literals will be roman
letters (enclosed in quotes when appearing in text) and standard numerals.
The synlbols used for variables will be italic letters, italic numerals, and
boldface letters as detailed in Table 1.8. Miscellaneous signs and symbols
when used as literals will be enclosed in quotes in both programs and text.
It is sonletimes desirable (e.g., for mnemonic reasons) to denote a
variable by a string of alphabetic or other symbols rather than by a single
symbol. The monolithic interpretation of such a string will be indicated
by the tie used in musical notation, thus: inL', inv, and INV may denote
'-'

----------

the variable "inventory," a vector of inventory values, and a matrix of

inventory values, respectively.
In the set of alphabetic characters, the space plays a special role. For
other sets a similar role is usually played by some one element, and this
element is given the special name of null elenlent. In the set of numeric
digits, the zero plays a dual role as both null element and numeric quantity.
The null element will be denoted by the degree symbol 0.
In any detenninate process, each operand Inust be specified ultinlately in
terms of literals. In Program 1.4, for example, the quantity k is specified
in terms of known arithmetic operations (multiplication and division)
involving the literals 1 and 2. The quantity n, on the other hand, is not
determined within the process and must presumably be specified within
some larger process which includes Program 1.4. Such a quantity is called
an arglllnent of the process.
Domain and range

The class of arguments and the class of results of a given operator are
called its domain and range, respectively. Thus the domain and range of
the magnitude operation (Ixl) are the real numbers and the nonnegative
real numbers, respectively.
A variable is classified according to the range of values it may assume:
it is logical, integral, or numerical, according as the range is the set of
logical variables (that is, 0 and 1), the set of integers, or the set of real
numbers. Each of the foregoing classes is clearly a subclass of each class
following it, and any operation defined on a class clearly applies to any of
its subclasses. A variable which is nonnumeric will be called arbitrary. In
the Summary of Notation, the range and domain of each of the operators
defined is specified in terms of the foregoing classes according to the
conventions shown in Sec. S.l.

Elenlentary operations

1.4
1.4

11

ELEMENTARY OPERATIONS

The elementary operations employed include the ordinary arithmetic

the elementary operations of the logical calculus~ and the
residue and related operations arising in elementary number theory. In
defining operations in the text, the symbol <=> will be used to denote
equivalence of the pair of statements between which it occurs.
operations~

Arithmetic operations
The ordinary arithmetic operations will be denoted by the ordinary
symbols +, -, x, and --:-- and defined as usual except that the domain and
range of multiplication will be extended slightly as follows. If one of the
factors is a logical variable (0 or I), the second may be arbitrary and the
product then assumes the value of the second factor or zero according as
the value of the first factor (the logical variable) is I or O. Thus if the
arbitrary factor is the literal "q~" then
Oxq=qxO=O
and

1 x q = q X 1 = q.

According to the usual custom in ordinary algebra, the multiplication

symbol may be elided.

Logical operations
The elementary logical operations and, or, and not will be denoted by A ,
V , and an overbar and are defined in the usual way as follows:
H' ~

H' ~ U
H' ~

v <=> vi' = 1 if and only if u = 1 and

V v<=>

l1

= I,

= 1

if and only if

u = 1 or v = I,

<=> H~ = 1

if and only if

u = O.

H'

If ~r and yare numerical quantities, then the expression x < !J implies

that the quantity x stands in the relation "less than" to the quantity y.
More generally, if (X and /1 are arbitrary entities and :!Il is any relation
defined on them~ the relational statement ('l)!Il;J) is a logical variable which
is true (eq ual to I) if and only if (X stands in the relation :!Il to /-J. For
example, if x is any real number, then the function

(x

>

0) - (x

<

0)

(commonly called the sign junction or sgn x) assumes the values I, O~ or

- 1 according as x is strictly positive, 0, or strictly negative. Moreover~
the magnitude function Ixl may be defined as Ixl = x X sgn x = x X
((~r
0) - (x <:: 0)).

12

1.4

The language

The relational statement is a useful generalization of the Kronecker

delta, that is, 6/ == (i == j). Moreover, it provides a convenient expression
for a number of fanliliar logical operations. The exclusice or, for example,
may be denoted by (u =1= c), and its negation (i.e., the equivalence function)
may be denoted by (u == r).

Residues and congruence

For each set of integers l1,j, and b, with b
of integers q and I' such that

n == bq

1',

>

j:S:: r < j

0, there exists a unique pair

b.

The quantity I' is called the j-residue o.l n n10dulo b and is denoted by b Ij n.
For example, 3109 == 0, 3119 == 3, and 310 10 == I. Moreover, if n 2 0,
then b 10 n is the remainder obtained in dividing n by band q is the integral
part of the quotient. A number n is said to be of ecen parity ifits O-residue
modulo 2 is zero and of odd parity if 210 n == I.
If two numbers n and In have the same j-residue Inodulo b, they differ
by an integral multiple of b and therefore have the saIne k-residue Inodulo
b for any k. If b /j n == b /j m, then In and n are said to be congruent rnod b.
Congruency is transitive and reflexive and is denoted by
In

==

n (nlod b).

In classical treatments, such as Wright (1939), only the O-residue is

considered. The use of I-origin indexing (cf. Sec. 1.5) accounts for the
interest of the I-residue .
.A number represented in a positional notation (e.g., in a base ten or a
base two nUlnber system) Inust, in practice, employ only a finite nunlber of
digits. It is therefore often desirable to approxilnate a nunlber or by an
integer. For this purpose two functions are defined:
I. the.fioor oj' or (or integral part of .r), denoted by l.rJ and defined as the
largest integer not exceeding .r,
2. the ceiling (~l,r, denoted by rorl and defined as the snlallest integer not
exceeded by .r.
Thus

f3. 141==4,

l3.14J==3,

l-3.14J == -4,

f3.001 == 3,

l3.00 J == 3,

l-3.00J == -3.

Clearly for == -l-rJ and lorJ

.r
frl Moreover, n == bln -:- bJ + b 10 n
for all integers n. Hence the integral quotient
bJ is equivalent to the
quantity q occurring in the defInition of the j-residue for the case j == O.

In -:-

Structured operands

1.5
1.5

13

STRUCTURED OPERANDS

Elementary operations
Any operation defined on a single operand can be generalized to apply
to each member of an array of related operands. Similarly, any binary
operation (defined on two operands) can be generalized to apply to pairs
of corresponding elenlents of two arrays. Since algorithrTIs conlnlonly
incorporate processes which are repeated on each member of an array of
operands, such generalization permits effective subordination of detail in
their description. For example, the accounting process defined on the
data of an individual bank account treats a number of distinct operands
within the account, such as account number, name, and balance. More
over, the over-all process is defined on a large number of sirTIilar accounts,
all represented in a conlmon format. Such structured arrays of variables
will be called structured operands, and extensive use will be made of three
types, called rector, rnatrix, and tree. As indicated in Sec. S.l of the
SunllTIary of Notation, a structured operand is further classified as logical,
integral, nUlnerical, or arbitrary, according to the type of elenlents it
contains.
A rector x is the ordered array of elements (Xl' X 2 , X 3 ' ,x,'(x). The
variable Xi is called the ith component of the vector x, and the number of
components, denoted by vex) (or simply v when the detennining vector is
clear from context), is called the dinlension of x. Vectors and their conl
ponents will be represented in lower case boldface italics. A nUlnerical
vector X may be rnultiplied by a numerical quantity k to produce the
scalar n1ldtiple k x X (or kx) defined as the vector z such that Zi == k X Xi.
All elementary operations defined on individual variables are extended
consistently to vectors as component-by-corTIponent operations. For
exarTIple,

w == (x

<

y)>w i ==

(Xi

<

yJ.

Thus if X == (1, 0, ], 1) and y == (0, 1, 1, 0) then

Y == (0, 0, 1, 0), and (x .::: y) == (0, 1, 0, 0).

X /\

+y

== (], 1, 2, 1),

14

1.5

The language

Matrices
A matrix III is the ordered two-dimensional array of variables

M 2 l,

MIl,
M

1 ,

Mf1(M)
1

2 ,

MJ~(M)
MI~(M)

M/I(M)

JI(M)

The vector (M/, M 2 i, . . . , M/) is called the ith rOH' rector of M and is
denoted by Mi. Its dimension v(M) is called the rOH' dimension of the
matrix. The vector (M/, M j 2, . . . , M/l) is called the jth colunul rector
of M and is denoted by M j Its dimension f-1(M) is called the colulnn
dinlension of the matrix.
The variable M/ is called the (i,j)th component or element of the matrix.
A matrix and its elements will be represented by upper case bol.dface
italics. Operations defined on each element of a matrix are generalized
component by component to the entire matrix. Thus, if 0 is any binary
operator,
P = M 0 N<:=>P/ = M/ 0 N/.

Index systems

The subscript appended to a vector to designate a single component is

called an index, and the indices are normally chosen as a set of successive
integers beginning at 1, that is, x = (Xl' X 2 , ,x~!). It is, however,
convenient to admit more general j-origin indexing in which the set of
successive integers employed as indices in any structured operand begin
with a specified integer j.
The two systems of greatest interest are the common I-origin systelTI,
\vhich will be employed almost exclusively in this chapter, and the O-origin
system. The latter system is particularly convenient whenever the index
itself must be represented in a positional nUlTIber system and will therefore
be employed exclusively in the treatment of COlTIputer organization in
Chapter 2.

1.6

ROTATION

The left rotation of a vector X is denoted by k t X and specifies the

vector obtained by a cyclical left shift of the components of x by k places.
Thus if a = (1,2,3,4,5,6), and b = (c, a, n, d, y), then 2 t a =
(3,4, 5, 6, 1,2), and 3 t b = 8 t b = (d, y, c, a, n). Fornlally, *
Z

= k t

X<:=>Zi

Xj'

where j

vlI(i

+ k).

* Restating the relation in ternlS of the O-residue will illustrate the convenience of the
I -residue used here.

1.7

Special rectors

Right rotation is denoted by

15

k ~ x and is defined analogously. Thus

where) == vll(i - k).

If k == 1, it may be elided. Thus t b == (a, n, d, y, c).

Left rotation is extended to matrices in two ways as follows:

~j

B~Ai

~ k

11

B<=>-C j == k j

==jl

Bi

tB

The first operation is an extension of the basic vector rotation to each row
of the matrix and is therefore called rOlt' rotation. The second operation is
the corresponding column operation and is ,therefore denoted by the
doubled operation synlbol 11. For example, if

k==(0,1,2),
and

C)

.~'

then

~)

and

Ii B

==

(~: ~

:).

.(

Right rotation is extended analogously.

1.7

SPECIAL VECTORS

Certain special vectors warrant special synlbols. In each of the following

definitions, the parallleter n will be Llsed to specify the dinlcnsion. The
interred rector lJ(n) is defined as the vector of integers beginning \vith j.
Thus l O( 4) == (0, 1, 2, 3), II ( 4) == (1, 2, 3, 4) , and l-' ( 5) == ( - 7, - 6, - 5,
-4, -3). Four types of logical vectors are defined as follo\vs. The jth
lInit rector J(n) has a one in the )th position, that is, (J(n))k == (k == j).
Thejul! rector (n) consists of all ones. The vector consisting of all zeros
is denoted both by and by "(n). The prefix rector (~llt'eight j is denoted
by a/(n) and possesses ones in the fIrst k positions, where k is the lesser ofj
and n. The sl~!fix rector wJ(n) is defIned analogoLlsly. Thus ~(3) ==
(0., I, 0),(4) == (I, 1, 1, 1),a:3(5) == (1,1, ],0,0),w:3(5) == (0,0, I, I, I),and
a'(5) == a 5 (5) == (I, 1, 1, 1, 1). Moreover, wJ(n) == j t aJ(n), and a J(Il) ::::::

j + wJ(n).

16

The language

1.7

A logical vector of the form all(n) 1\ u)i(n) is called an if?!ix rector. An

infix vector can also be specified in the fornl j t o.l''(n) , which displays its
weight and location more directly.
An operation such as x 1\ y is defined only for c0171patihle vectors x and
y, that is, for vectors of like dinlension. Since this cornpatibility
req uirement can be assunled to specify inlplicitly the dinlension of one of
the operands, elision of the parameter n nlay be pernlitted in the notation
for the special vectors. Thus, if y == (3,4, 5, 6, 7), the expressions E X Y
and E j X Y inlply that the dinlensions of E and E j are both 5. Moreover,
elision of j will be permitted for the interval vector lj(n) (or lj), and for the
residue operator Ij when j is the index origin in use.
It is, of course, necessary to specify the index origin in use at any given
time. For example, the unit vector E:~(5) is (0,0, 1,0,0) in a l-origin
system and (0, 0, 0, I, 0) in a O-origin systern, even though the definition
(that is, (Ej(n)h~ == (k == j)) remains unchanged. The prefix and suffix
vectors are, of course, independent of the index origin. Unless otherwise
specified, I-origin indexing will be assumed.
The vector E(O) is a vector of dimension zero and \vill be called the null
rector. It should not be confused with the special null elenlent
D.

1.8

REDUCTION

An operation (such as sumnlation) which is applied to allcornponents

of a vector to prod uce a result of a sirTI pier structure is called a reduction.
The C-reduction of a vector x is denoted by (~)x and defined as
z +- C)lx<:>z == ( ... ((Xl

x 2)

x:~)

... ) 0 X,,),

where 0 is any binary operator with a suitable dornain. Thus + Ix is the

surn, X Ix is the product, and V Ix is the logical sunl of the components of
a vector x. For example, x Il I (5) == I x 2 x 3 x 4 x 5, X Il1(n) == n~,
and + Il 1 (n) == n(n + 1)/2.
As a further example, De Morgan's law may be expressed as 1\ lu ==
V lu, where u is a logical vector of dimension two. Moreover, a sinlple
inductive argument (Exercise 1.10) shows that the foregoing expression is the
valid generalization of De Morgan's law for a logical vector u of arbitrary
dimension.
A relation "~ incorporated into a relational statement (..c~y) becomes,
in effect, an operator on the variables x and y. Consequently, the reduction
:!Jllx can be defined in a manner analogous to that of olx, that is,
"!Ii Ix == ( ... (( Xl ~3fx2) "~x:J 3f ... ) "!!lx ,,).

The parentheses now inlply relational staternents as well as grouping.

1.9

Selection

]7

The relational reductions of practical interest are ~/~/u, and =/u, the
exclusire-or and the equiralence reduction, respectively.
The inductive argument of Exercise 1.10 shows that ~F/U = 210 (+/u).
For example, if U = (1,0,1,1,0), then
-~ / II

= ((((I :1'= 0)
=(((1

1)~

= ((0 ::/= I)
=(1

I)

1)

0)

1)=1=0)

0)

0)=1,

and 2 10 (+ /u) = 2 10 3 = ]. Similarly, =

sequence,

/u

= 2 10 (+ Iii), and as a con

=I)U = =/ii,
a useful companion to De Morgan's law.
To complete the system it is essential to define the value of cJ/e(O), the
reduction of the null vector of dimension zero, as the identity elelnent of
the operator or relation
. Thus + /e(O) = V /e(O) = 0, and x /e(O) =
A /e(O) = 1.
A reduction operation is extended to matrices in two ways. A rOH'
reduction of a matrix X by an operator () is denoted by
Y~

,:)/X

and specifies a vector y of dimension fleX) such that Yi = c) Xi. A

COIUJ11n reduction of X is denoted by Z ~ all X and specifies a vector z of
dimension veX) such that Zj = c)/Xj
For example, if

u=

(~ ~ ~)
I

then +/U = (2,2,3), +//U = (2, L 3, I), A//U = (0,0, 1,0),

(0,0,1), =//U = (0, I, I, I), and +/(=//U) = 3.

1.9

/)U ==

SELECTION

Compression
The effective use of structured operands depends not only on generalized
operations but also on the ability to specify and select certain elenlents or
groups of elements. The selection of single elements can be indicated by
indices, as in the expressions Vi' Mi, M j , and M/. Since selection is a
binary operation (i.e., to select or not to select), nlore general selection is

18

1.9

The language

conveniently specified by a logical vector, each unit component indicating

selection of the corresponding component of the operand.
The selection operation defined on an arbitrary vector a and a compat
ible (i.e., equal in dimension) logical vector u is denoted by c +- uja and
is defined as follows: the vector c is obtained from a by suppressing from
a each cOlnponent a i for which U i === 0. The vector u is said to c(unpress
the vector a. Clearly v( c) === + ju. For example, if u === (1,0,0,0, I, I)
and a === (M, 0, n, d, a, y), then uja === (M, a, y). Moreover, if n is even
and v === (2) 10 ll(n) === (1,0, 1,0, I, ...), then vj l l(n) === (I, 3, 5, ... , n - I),
and +j(vj l l(n)) === (nj2)2.
Row compression of a matrix, denoted by uj A, compresses each row
vector Ai to form a matrix of dimension fleA) x +ju. Column compres
sion, denoted by ujj A, compresses each column vector A j to form a
matrix of dimension +ju x v(A). Compatibility conditions are v(u) ===
v(A) for row compression, and v(u) === fleA) for column compression.
For example, if A is an arbitrary 3 x 4 matrix, u === (0, 1,0, I) and
v === (I, 0, 1); the n

ul A ===

AI)
(A/
A
A:
2

A 23

A 43

(A A A A1)
A~3 A A A> '
41
(A
)
A
===
A:3 A
1

vilA ===

and

ulvll A === vllul A

It is clear that rOlf compression suppresses columns corresponding to

zeros of the logical vector and that column compression suppresses rOlfS.
This illustrates the type of confusion in nomenclature which is avoided by
the convention adopted in Sec. 1.3: an operation is called a rOH' operation
if the underlying operation from which it is generalized is applied to the
row vectors of the matrix, and a column operation if it is applied to columns.
Example 1.1. A bank makes a quarterly review of accounts to produce the
following four lists:

1. the name, account number, and balance for each account with a balance
less than two dollars.
2. the nan1e, account nun1ber, and balance for each account with a negative
balance exceeding one hundred dollars.
3. the nan1e and account nun1ber of each account with a balance exceeding
one thousand dollars.
4. all unassigned account nun1bers.

Selection

1.9

19

The ledger nlay be described by a Illatrix

I.)

L'I/

with colunln vectors L 1 , L'2' and L 3 representing nanles, account nUIllbers, and
balances, respectively, and with row vectors LI, L'2, ... , LI/I representing
individual accounts. An unassigned account nunlber is identified by the word
"'none" in the nanle position. The four output lists will be denoted by the
Illatrices P, Q, R, and S, respectively. They can be produced by PrograIll 1.9.

Program 1.9. Since L 3 is the vector of balances, and 2 is a conlpatible vector

each of whose conlponents equals two, the relational statenlent (L: 3
2) defIIles
a logical vector having unit conlponents corresponding to those accounts to be

P +-(L 3 < 2 )jjL

Q

R
S

+++-

kth account.

(L 3 < -100 )jjL

L 3k

(L 3 > 1000 )jja 2/L

(L 1 = none )j/2/L

Bank ledger.
Balance of kth account.
Account number of kth
account.

~
L1k

Name of kth account or

"none" if account number
L 2k unused.

Legend
Program 1.9

Selection on bank ledger L (Exanlple 1.1)

included in the list P. Consequently, the coluIlln conlpression of step 1 selects

the appropriate rows of L to define P. Step 2 is sinlilar, but step 3 incorporates
an additional row cOIllpression by the conlpatible prefix vector ex'2 = ( I, I, 0) to
select colunlns one and two of L. Step 4 represents the conlparison of the n~lIne
(in coluIlln L 1 ) with the literal ""none," the selection of each row which shows
agreenlent, and the suppression of all coluIllns but the second. The expression
"'none " occurring in step 4 illustrates the usc of the extended deflnitioll of
nlul t ipI ica t ion.

Mesh, mask, and expansion

A logical vector u and the two vectors a == til c and b == ul c, obtained
by conlpressing a vector c, collectively deternline the vector c. The
operation which specifies c as a function of a, b, and u is called a rnesh and

20

The language

1.9

is defined as follows: If a and b are arbitrary vectors and if u is a logical

vector such that +Iti == v(a) and +Iu == v(bL then the 111esh (~la and bOil
u is denoted by \\a, u, b\ and is defined as the vector c such that til c == a
and ul c == b. The nlesh operation is equivalent to choosing successive
components of c fronl a or b according as the successive conlponents of u
are 0 or I. If, for example, a == (s, e, k), b == (1, a), and u == (0, 1,0, I, OL
then \,a, u, b\., == (s, L e, a, k). As a further exarTIple, Progranl 1.1 Oa

+-

j+-O

k~k+l

3/ 1k

ck

10

ck

C ~

~O

\a,u,b\
(b)

v(a) +v(b)

~j
~

+ 1

a, b

Given vectors.

c = (al,bl,b2,a2,b3,b4,a3, ...).

Index of a.

bj

+ 1

-(- (Ii

Index of b.

j
k

Index of c.

u = (0,1,1,0,1,1,0, ...).

(a)

Legend

Program 1.10

Interfiling progran1

(which describes the merging of the vectors a and b, with the first and every
third component thereafter chosen from a) can be described alternatively
as shown in Program 1.1017. Since II == (I, 2, 3, 4, 5, 6, ... ), then
(3) 10 II == (1,2,0, 1,2,0, ... ), and consequently the vector u specified by
step 1 is of the form u == (0, 1, I, 0, 1. 1, 0, ... ).
Mesh operations on matrices are defined analogously, row mesh and
column mesh being denoted by single and double reverse virgules, respec
tively.
The catenation of vectors x, y, ... , z is denoted by x
y
z
and is defined by the relation
x

y EB ... EB z ==

(Xl'

x 2, ... ,

XJ'(X)'

YI' Y2' ... ,

ZI'(Z))'

Catenation is clearly associative and for two vectors x and y it is a special

case of the mesh \x, u, y\ in which u is a suffix vector.

21

1.9

In nunlerical vectors (for which addition of two vectors is defIned), the

effect of the general nlesh operation can be prod uced as the sunl of two
nleshes~ each involving one zero vector. Specifically~
\,x~ u~ y\\ == \x~ U~

== \O~

Ii,

0\ + \O~ U~ y\
x\ + \0, U, y\.

The operation \O~ U~ y\ proves very useful in numerical work and \vill be
called expansion of the vector y~ denoted by u\y. Compression of u\y by
II and by u clearly yield y and O~ respectively.Moreover~ any nunlerical
vector x can be deconlposed by a compatible vector u according to the
relation

x == u\ulx + u\ulx.
The two terms are vectors of the same dimension which have no nonzero
conlponents in common. Thus if u == (1,0, 1,0, 1), the decolnposition of
x appears as

Row expansion and column expansion of matrices are defined and

denoted analogously. The decomposition relations become

x == u\\ul X + u\ul X,
X == u\\uIIX + u\\uIIX.

and

The nlask operation is defined formally as follows:

C

+---Ia~ u~ bl<;~ulc ==

lila,

and

ulc

==

ulb.

The vectors C~ a~ u, and b are clearly of a common dimension and C i == a i

or b i according as u i ==
or u == 1. Moreover, the conlpress, expand,
mask, and mesh operations on vectors are related as follows:

la,
\a,

bl == \ula, u, ulb\,
u~ b\ == Iii \a, u, u\bl.

u,

Analogous relations hold for the row mask and row Inesh and for the
colulnn mask and colulnn rnesh.
Certain selection operations are controlled by logical nlatrices rather
than by logical vectors. The rOll' cOfnpressiol1 VIA selects clernents of A
corresponding to the nonzero elenlents of V. Since the nonzero elenlcnts
of V nlay occur in an arbitrary pattern, the result nl ust be construed as a
vector rather than a nlatrix. More precisely, VI A denotes the catenation
of the vectors V) Ai obtained by row-by-row compression of A by V.

22

1.9

The language

The colunln conlpressiol1 V II A denotes the catenation of the vectors VJ! A j.

Ie for example,

u ==

(~ ~ ~ ~)

then

VIA

==

(A 2\ A 4 1, A.")l, A 12, A 22, A 2a, A:3:3),

and

VilA

==

(A]2, A 21, A~2, A 2:3, Aaa, A4 1, A;)l).

Compression by the full matrix E (defined by E == 0) produces either a

r()\\, list (EI A) or a colurnn list (Ell A) of the matrix A. Moreover, a nunleri
cal nlatrix X can be represented jointly by the logical nlatrix V and the
row list VIX (or the column list VIIX), where V == (X
0). If the nlatrix
X is sparse (i.e., the components are predominantly zero), this provides a
conlpact representation which may reduce the computer storage req uired
for X.
The compression operations controlled by matrices also generate a group
of corresponding mesh and mask operations as shown in Sec. S.9.

1.10

SELECTION VECTORS

The logical vector u involved in selection operations lTIayitself arise in

various ways. It may be a prefix vector ai, a suffix Wi, or an infIX (i t a i );
the corresponding compressed vectors ailx, wilx, and (i t aJ')lx are called
a prefix, suffix, and infix of x, respectively.
Certain selection vectors arise as functions of other vectors, e.g., the
vector (x ~~ 0) can be used to select all nonnegative conlponents of x, and
(b
*) serves to select all components of b which are not equal to the
literal "*." Two further types are important: the selection of the longest
unbroken prefix (or suffix) of a given logical vector, and the selection of the
set of distinct components occurring in a vector. The first is useful in left
(or right) justification or in a corresponding compression intended to
eliminate leading or trailing "filler components" of a vector (such as left
zeros in a number or right spaces in a short name).
For any logical vector u, the Inaximum prefix of u is denoted by :xlu and
defined as follows:
v ~ :xlu<=-~v

==

ai,

where j is the maximum value for which 1\ I(ailu) == 1. The maXilTIUnl

suffix is denoted by (I)IU and is defined analogously. If, for example,
u == (1,1,1,0,1,1,0,0,1,1),
then
:xju == (1,1,1,0,0,0,0,0,0,0),
(I)/U == (0,0,0,0,0,0,0,0,1,1), +I:xju == 3, and +I(I)IU == 2.

The generalized n1atrix product

1.11

23

The leading zeros of a numerical vector x can clearly be renloved either

by compression:
y +-- ('X/(x == O))/x,
or by left j ustiflcation (nornlalization):
z

+--

(+I'X/(x

==

0))

x.

The extension of the maximum prefix operation to the rows of a logical

nlatrix V is denoted by 'XI V and defined as the compatible logical Illatrix
V, such that Vi == 'XI Vi. The corresponding maximuIn colunln prefix
operation is denoted by 'XII V. Right justification of a n uIllerical Illatrix X
is achieved by the rotation k t X, where k == +IU)/(X == 0), and top
just{jication is achieved by the rotation (+ 11'XII(X == 0)) 11 X (see Sec. S.6.)
A vector whose components are all distinct will be called an ordered set.
The.!,orlt'ard set selector on b is a logical vector denoted by alb and defined
as follows: the statement v +-- alb implies that v j == 1 if and only if b j
differs frOIn all preceding components of b. Hence vlb is a set which
contains all distinct conlponents of b, and + Iv/l is a Inininlunl. For
example, if c == (C, a, n, a, d, a), then (alc)/c == (C, a, n~ d) is a list of the
distinct letters in c in order of occurrence. Clearly (alb)lb == b if and only
if b is a set.
The backward set selector Tlb is defined analogously (e.g., (71 c)1 c ==
(C, n, d, a)). Forward and backward set selection are extended to Illatrices
by bot h ro ws (a IB, and TIB) and col unl ns (a II B, and Til B) in the es ta b
lished Inanner.
1.11

THE GENERALIZED MATRIX PRODUCT

The ordinary matrix product of nlatrices X and Y is conlnl0nly denoted

by ..Y Yand defined as follows:
v(x)

+--

XY <-~.:> Z/

==

.2 X

(i == 1,2, .. . /l(X)

Y/',

t.i ==

1.'=1

1, 2, . . . 1'( Y).

It can be defined alternatively as follows:

(XY)/

== +/(X'

Y).

This fonn ulation enlphasizes the fact that nla trix nl ultipl ication incorpor
ates two elCIllentary operations ( +, x) and suggests that they be displayed
explicitly. The ordinary Inatrix product will therefore be written as
X

Y.

24

1.11

The language

More generally, if 0 1 and O 2 are any two operators (whose dornains

include the relevant operands), then the generali:?ed Inatrix product
X (_ ~ Y is defined as follows:

ri

i
~ Y)/ = 0I/(X 0 2Yj),

(X

1, 2, ... , /1( X)
1, 2, ... , v( Y) .

tj =

For example, if

A=

then

(~

1 0

A~B=(l~ 14)

5 ,

20

'j

B =

!)
A ~

B=

(l :),

and

B=

and

(A

D'

T 0) ; B

(~

,6

The generalized matrix product and the selection operations together

provide an elegant formulation in several established areas of nlathe
matics. A few examples will be chosen from two such areas, synlbolic
logic and nlatrix algebra.
In synlbolic logic, De Morgan's laws (/\ lu
can be applied directly to show that

V = U:

ju

and

=ju = -/(ii)

v.

In matrix algebra, the notion of partitioning a matrix into subnlatrices of

contiguous rows and columns can be generalized to an arbitrary parti
tioning specified by a logical vector u. The following easily verifiable
identities are typical of the useful relations which result:
X ~ Y

eliI X) ~ eli II Y) + (til X) t (u II }T),

ul(X ~ y?) = X ~ (uIY),

ul/(X ~ Y) = (u/IX) ~ Y.

The first identity depends on the commutativity and associativity of the

operator + and can clearly be generalized to other associative conlmu
tative operators, such as /\, \j, and ~f=.
The generalized matrix product applies directly (as does the ordinary

1.12

Transpositions

matrix product X ~. Y) to vectors considered as row (that is, I

column matrices. Thus:
Z~ X
Z~Y
Z

~ Y<=>Zi ==

C'l!CX i

25

n) or as

02 y),

~ X<=>Zj == C)l!(Y CJ 2 X j ),

~ Y ;,~ x<=>z ==

0I/Cy

02 x).

The question of whether a vector enters a given operation as a row

vector or as a column vector is normally settled by the requirement of
conformability, and no special indication is required. Thus Y enters as
a column vector in the first of the preceding group of definitions and as
a row vector in the last two. The question remains, however, in the case
of the two vector operands, which may be considered with the pre-operand
either as a row (as in the scalar product Y '-: x) or as a column. The
latter case produces a matrix Z and will be denoted by

where Z/ == Yi C)2 x j , tt(Z) == v(y), and v(Z) == vex). * For example, if

each of the vectors indicated is of dimension three, then

~ Y

Y2'

(Yl'
==

Yl'

Y2'

Yl'

Y2'

1.12

:~)

Y ~

Y3

(Yl'
==

Yl'

Y2'

Y2'

Y:3'

Y:3'

::)

Y:3

TRANSPOSITIONS

Since the generalized matrix product is defined on columns of the

post-operand and rows of the pre-operand, convenient description of
corresponding operations on the rows of the post-operand and columns
of the pre-operand demands the ability to transpose a matrix B, that is, to
specify a matrix C such that C/ == B/. In ordinary matrix algebra this
type of transposition suffices, but in 1110re general work transpositions

* Since each "'vector" Yi () 2 X j is of dimension one, no scan operator 0

and the syn1bol may be interpreted as a "'null" scan.
0

is required,

26

1.12

The language

about either diagonal and about the horizontal and the vertical are also
useful. Each of these transpositions of a matrix B is denoted by a superior
arrow whose inclination indicates the axis of the transposition. Th us:

C/

C~B

---+

C~B

-..

==

B/

C/ == B~::;~;

i == 1,2,

,11(B)

C/

j == 1,2,

, v(B)

== Bj +l-i

For a vector x, either x or x will denote reversal of the order of the com
ponents.

For ordinary matrix transposition (that is, B), the commonly

used notation B will also be employed.

Since transpositions can effect anyone or more of three independent
alternatives (i.e., interchange of row and column indices or reversal of
order of row or of column indices), repeated transposition can produce
eight distinct configurations. There are therefore seven distinct transfornla
tions possible ~ all can be generated by any pair of transpositions having
nonperpendicular axes. *

1.13

SPECIAL LOGICAL MATRICES

Certain of the special logical vectors introduced in Sec. 1.7 have useful
analogs in logical matrices. Dimensions will again be indicated in paren
theses (with the column dimension first) and may be elided whenever the
dilllension is determined by context. If not otherwise specified, a matrix is
assumed to be sq uare.
Cases of obvious interest are the ju!1 matrix fern x n), defined by

fern x n) == 0, and the identity matrix I(m x n), defined by 1/ == (i == j).

More generally, superdiagonal matrices kl(m X n) are defined such that
kl/(m X n) == (j == i + k), for k :2: 0. Clearly 1 == I.
Moreover, for
sq uare matrices hi 1"1 == (hr-k)/.
Four triangular matrices will be defined, the geometrical Sylllbols
employed for each indicating the (right-angled isosceles) triangular area of

* These transpositions generate the rotation group of the square [cf. Birkhoff and
MacLane (1941) Chap. VI]. A pair of transpositions commute if and only if their axes
are perpendicular. Hence the pair +-- and t may be written unambiguously as-<t- .
Moreover, ~t- = X. The remaining two transformations can be denoted by 4 and~,
with the convention that the operator nearest the operand (i.e., the horizontal) is
executed first.

Po~vn()fl1ials and

1.14

the

111

X n

C +-- ...-J

C +--

~-=j

rectangular matrix which is occupied by ones. Thus

(111 X 11)

C ~- '~ (111
(111

C +---~ (111

27

positional nUlnher ,\ysteI11S

X 11)

<=> C/
<=> C: + 1 _

X 11) <--:?
X

C jl + 1

== (i +.i

In i 11

(111,

11 ) )

for i
andj

== 1, 2,
== 1,2,

, 111
,11.

J1) <=> C~: ll~ J

The use of the matrices E and I will be illustrated briefly. The relation
== 2 10 (u v) can be extended to logical matrices as follovvs:

u ;, v

U / V

== (2 E) 10 ( u ~

V) ~

the trace of a square numerical matrix X may be expressed as t ==

The triangular matrices are employed in the succeeding section.

+ /1/ X.

1.14 POLYNOMIALS AND POSITIONAL NlJMBER

SYSTEMS

Any positional representation of a number n in a base b number systcln

can be considered as a numerical vector x whose base b ralue is the quantity
n == w ~< x, where the lreighting rector w is defined by w == (hl'(X)},
b (X)--2, , b 2 , bI, 1). More generally, x may represent a number in a
mixed-radix system in which the successive radices (from high to low order)
are the successive components of a radix rector y.
The base y calue oj' x is a scalar denoted by y J__ x and defincd as the
scalar product y ~ x == w x, where w ==
y is the \veighting vector.
For example, if y == (7, 24, 60, 60) is the radix vector for the con1n10n
temporal system of units, and if x == (0, 2, 1, 18) represents elapsed tinlC in
days, hours, minutes, and seconds, then
11

n/

t == w ~ x == (86400, 3600, 60, 1) ~ (0, 2, 1, 18)

== 7278

is the elapsed time in seconds, and the weighting vector w is obtained as

the product
0

60

x /(24, 60, 60)

x /(60, 60)
x /(60)

60

" ; y ==

1
v

24

/(0)

86400
3600
60
1

If b is any integer, then the value of x in the fixed base b is denoted by

_L x. For example, (2e) _L a 2 (5) == 24. More generally, if y is any real

(be)

28

1.14

The language

number, then (YE) ~ x is clearly a polynomial in y with coefficients

that is,

Xl'

X 2 ' , Xl"

(YE) ~

==

xlyV(X)-l

Writing the definition of y

y

~ X

J_

+ ... + Xv-I?! + Xv.

in the form

==

(D

7y) :

exhibits the fact that the operation

is of the double operator type. Its
use in the generalized matrix product therefore requires no secondary scan
operator. This will be indicated by a null placed over the syrnbol _l. Thus

z +---- Xl

Y<=>Z/

==

Y~j.

Xi

For example, (YE) 1 X represents a set of polynomials in ?! with coeffi

cients Xl' X 2 , , Xl" and Y l X represents a set of eval uations of the
vector x in a set of bases yl, y2, ... , Y/l.

1.15

SET OPERATIONS

In conventional treatments, such as Jacobson (1951) or Birkhoff and

Mac Lane (1941), a set is defined as an unordered collection of distinct
elenlents. A calculus of sets is then based on such elementary relations as
set membership and on such elementary operations as set intersection and
set union, none of which imply or depend on an ordering anlong mernbers
of a set. In the present context it is more fruitful to develop a calculus of
ordered sets.
A vector whose components are all distinct has been called (Sec. 1.10)
an ordered set and (since no other types are to be considered) will hereafter
be called a set. In order to provide a closed system, all of the "set opera
tions" will, in fact, be defined on vectors. However, the operations will, in
the special case of sets, be analogous to classical set operations. The
following vectors, the first four of which are sets, will be used for illustra
tion throughout.
t == (t, e, a)

==
s ==

(a, t, e)
(s, a, t, e, d)

d == (d, u, s, k)
n

==
==

(n,

0,

n, s, e, t)

(r, e, d, u, n, d, a,

ll,

t)

A variable z is a Inell1ber of a vector x if ,: == Xi for sorne i. Menlbership

is denoted by,: E x. A vector X includes a vector y (denoted by either

Set operations

1.15

29

x ;2 Y or Y s; x) if each element Yi is a menlber of x. If both x 2 yand

x S; y, then x and yare said to be similar. Sinlilarity of x and y is denoted
by x == y.For example, t S; s, t s; r, t S; a, a S; t, t == a, and t
r. If
x S; y and x y, then x is strict~y included in y. Strict inclusion is denoted
by x c y.
The characteristic rector of x on y is a logical vector denoted by yX, and
defined as follows:
u

==

yx>v(u)

==

v(y), and u j

==

(Yj EX).

For example, / == (0, I, I, I, O),/ == (I, I, l)'sd == (1,0,0,0,1), d s ==

( I, 0, I, 0), and n r == (1, 0, I, 0, I, 1).
The intersection of y with x is denoted by y (\ x, and defined as follows:
y (\ x

==

yXjy.

For example, S (\ d == (s, d), d (\ S == (d, s), S (\ r == (a, t, e, d) and

r (\ s == (e, d, d, a, t). Clearly, x (\ y == Y (\ x, although x (\ y is not, in
general, equal to y (\ x, since the conlponents nlay occur in a different
order and may be repeated a differing number of tinles. The vector
x (\ y is said to be ordered on x. Thus a is ordered on s. If x and y contain
no common elements (that is, (x (\ y) == (O)), they are said to be di.yoint.
The set diJference of y and x is denoted by y ~ x and deflned as follows:
y ~ x

==

yXjy.

Hence y ~ x is obtained fronl y by suppressing those conlponents which

belong to x. For exalnple, e/ == (1,0,0,0, I) and s Ll t == (s, d). More
over, / == (0, 0, 0) and t ~ s == (O).
The union of y and x is denoted by y u x and defined as follows: *
y U x == y
(x ~ y). For exanlple, sud == (s, a, t, e, d, u, k), d u s ==
(d, u, s, k, a, t, e), sUa == s U t == s, and nut == (n, 0, n, s, e, t, a).
In general, x U y == y U x, and x == (x (\ y) u (x ~ y). If x and yare
disjoint, their union is equivalent to their catenation, that is, x (\ Y == (O)
implies that x U y == x
y.
In the foregoing developnlent, the concepts of inclusion and sinlilarity
are equivalent to the concepts of inclusion and equality in the conventional
treatrnent of (unordered) sets. The renlai ning deflnitions of intersect ion,
difference, and union differ fronl the usual fornlulation in that the result of
any of these operations on a pair of ordered sets is again an ordered set.
With respect to sinlilarit)', these operations satisfy the Sallle identities as do
the analogous conventional set operations on unordered sets \vith respect
to equality.
:1: The sY/11bols u and n
(and the operations they denote) arc conlnlonly called cup
and cap, respectively.

30

The language

1.15

The forward selection alb and the backward selection Tlb defined in
Sec. l.10 can both be used to reduce any vector b to a similar set, that is,

(alb)lb

==

(Tlb)lb

==

b.

Moreover, if f = (alx)lx, g = (aly)ly, and h = (alz)lz, then x = y n z

implies that f = g n h, and x = y u z implies that f = g u h.
The unit vector Ei(n) will be recognized as a special case of the charac
teristic vector E/ in which x consists of the single component j, and
y = l"(n), where h is the index origin in use. In fact, the notation
can be
used to make explicit the index origin h assumed for E i .
If z is any vector of dimension two such that Zl E x and Z2 E y, then z is
said to belong to the Cartesian product of x and y. Thus if x = (a, b, c)
and y = (0, I), the rows of the matrix

E:"

a
a

A=

b
b

c
c

0
1
0
1
0
1

are a complete list of the vectors z belonging to the product set of x and

y. The matrix A will be called the Cartesian product of x andy and will be
denoted by x @ y.
The foregoing definition by example will be formalized in a more
general way that admits the Cartesian product of several vectors (that is,
u ('9 v 09 ... 09 y) which need not be sets, and which specifies a unique
ordering of the rows of the resulting matrix. Consider a family of vectors
Xl, x 2 , .. , x' of dimensions d l , d 2 , . . . , do' Then

+- X

I?
0)

x-

(9 ... (9

A 1 :, d
x'"<=>.

(k -- )

0
= (1
Xk" Xh" . . . ,

xl" '
, )

for all vectors k such that I ::s:: k i ---: d i . Clearly v(A) = s, and ,u(A) =
x Id. As illustrated by Table 1.11, the rows of the Cartesian product A are
not distinct if anyone of the vectors Xi is not a set.
If the vectors Xi are all of the same dimension, they may be considered
as the columns of a matrix X, that is, Xi = Xi. The product Xl (>9 x 2
':'J' . .
x' = Xl @ X 2 09 ... 09 x, may then be denoted by (il X, or
alternatively by (2,IIY, where Y is the transpose of X. For example, if
X

l0(2)

E(3)

(0

0),

the n @IX is the matrix of arguments of the truth table for three variables.

1.16

Rank. in?:
a #
a #

31

*
1
b # 0

b # 1

=
x2 =
x:3 =
d =
Xl

(a, b, a)

(#, *)

(0,1)

(3, 2,2)

*
1

a # 0

*
0

a #
a *
a *
The Cartesian product A =

Table 1.11

1.16

1
Xl

x2

X3

RANKING

The rank or index of an elenlent C E b is called the b index (~l c and IS

defined as the smallest value of i such that c ~ bi. To establish a closed
system, the b index of any elelllent a b will be deflned as the null character
The b index of any elelllent (' will be denoted by b Ie; if necessary, the
index origin in use will be indicated by a sUbscript ap?ended to the
operator l. Thus, if b ~ (a, p, e), b 10 P == I, and b II P == 2.
The b index of a vector c is deflned as follo\vs:
0.

+-

==

c<-:>-k;

I Ct.

The extension to nlatrices 111ay be either row by row or (as indicated by a

doubled operator symbol ll) colulnn by column, as follows:
J
K

+-

+-

II

C<=->Ji
G'f<=;>K j

==
==

Bi

Ci,

Bj

C j .

Use of the ranking operator in a lnatrix prod uct reg uires no secondary
scan and is therefore indicated by a superior null synlbol. Moreover,
since the result lnust be linlited to a two-dinlensional array (nlatrix), either
the pre- or post-operand is required to be a vector. Hence
J
K

~- B~) C<:-~->Ji ~

+-

b{ C<~>Kj

==

Bi

I C,

b I Cj .

The fi rs t 0 f these ra nks the co III po nents 0 f c \v i th respect toeachoI' a set

of vectors B1, B'2, ... , B/I, whereas the second ranks each of the vectors
C I , (''2' . . . , C)O with respect to the fIxed vector b.

32

The language

1.16

The use of the ranking operation can be illustrated as follows. Consider

the vector b = (a, b, c, d, e) and the set of all 35 three-letter sequences
(vectors) formed from its components. If the set is ordered lexically, and
jf x is the ith lnember of the set (counting from zero), then

)=
For example, if x

1.17

-L

(v(b)e)

(b to x).

(c, a, b), then (b to x)

(2,0,1), and)

51.

MAPPING AND PERMUTATIONS

Reordering operations
The selection operations employed thus far do not permit convenient
reorderings of the components. This is provided by the mapping operation
defined as follows: *
For example, if a = (a, b, ... , z) and k = (6, 5,4), then C = (f, e, d).
The foregoing definition is meaningful only if the components of k each
lie in the range of the indices of a, and it will be extended by defining a j as
the null element if) does not belong to the index set of a. Formally,
C +--

am<=>c i

am.

if m i

ll(v(a))

if m i ll(v(a)).

The ability to specify an arbitrary index origin for the vector a being
mapped is provided by the following alternative notation for mapping:
am.
c+-mS;a<=?c;=,o
'
{

if m i

li(v(a))

if m i li(v(a)),

where i-origin indexing is assumed for the vector a. For example, if a

is the alphabet and m = (5,0,0,4,27,0,3), then C = m Jo a = (f, 0, 0, e,
0,0, d), and (c
oe)/c = (f, e, d). Moreover, mJ2a = (d, 0,0, c, Z, 0, b).
Elision of j is permitted.
If a ~ b, and m = b lj a, then clearly m.L b = a. If a $ b, then
m Jj b contains (in addition to certain nulls) those components common to
b and a, arranged in the order in which they occur in a. In other words,

(m

* oe)/(m.Lb) = a n b.

* For the purposes of describing algorithms, this notation is superior to the classical
"disjoint cycles" notation for permutations [cf. Birkhoff and MacLane, (1941)] because
(1) the direction of the transformation (from a to c) is unequivocally indicated, and (2)
the notation directly indicates a straightforward and efficient method for actual execu
tion, namely, indirect addressing.

1.17

MapjJing and jJermutations

33

Consequently, if p, q, ... , t are vectors, each contained in b, then each

can be represented jointly by the vector b and a mapping vector. If, for
example, b is a glossary and p, q, etc., are texts, the total storage required
for b and the mapping vectors might be considerably less than for the
entire set of texts.
Mapping may be shown to be associative, that is, m l (m 2.L a) =
(ml.L m 2)Jj a. Mapping is not, in general, commutative.
Mapping is extended to Inatrices as follows:

Ji

Mil'
J h Bi ,

MJJII B<=:>Cj =

MjJh B j .

A ~ M JrII B<=:>Ai

C~

Rowand column mappings are associative. A row mapping 1M and a

column mapping 2M do not, in general, commute, but do if all rows of
1M agree (that is, 1M = e p), and if all columns of 2M agree (that is,
2M = q ~ e). The generalized matrix product is defined for the cases
0

J A, and

M Ja.
The alternative notation (that is, c = an}), which does not incorporate
specification of the index origin, is particularly convenient for matrices and
is extended as follows:
A ~ B,n <=> Ai = Bn1,.,

A ~ B n1 <=:> Ai = B,ni'

Permutations
A vector k of dimension n is called a j-origin per/nutation rector if
k == lj(n). A permutation vector used to map any set of the same dinlen
sion produces a reordering of the set without either repetition or suppres
sion of elements, that is, k.L a == a for any set a of dimension v(k). For
example, if a = (f, 4, *, 6, z), and k = (4,2, 5, 1,3), then k Xl a = (6,4,
Z, f, *).
If p is an h-origin permutation vector and q is any j-origin permutation
vector of the same dimension, then q .L p is an h-origin permutation vector.
Since

the interval vector lj(n) will also be called the j-origin identity pernultation
vector. If p and q are two j-origin permutation vectors of the same
dimension n and if q .L p = lien), then p .L q = lien) also and p and q are
said to be inrerse permutations. If p is any j-origin permutation vector,
then q = p lj li is inverse to p.
The rotation operation k x is a special case of permutation.

1.17

The language

34

Function mapping
A function f which defines for each element b i of a set b a unique
correspondent ale in a set a is called a mapping fronl b to a. If f(b i ) = ale,
the element b i is said to map into the element ale. If the elements f(b i )
exhaust the set a, the functionfis said to map b onto a. If b maps onto a
and the elements j'(b i ) are all distinct, the mapping is said to be one-to-one
or biunique. In this case, v(a) = v(b), and there exists an inverse mapping
from a to b with the same correspondences.
A program for performing the mapping f from b to a must therefore
determine for any given element b E b, the correspondent a E a, such that
a = feb). Because of the convenience of operating upon integers (e.g.,
upon register addresses or other numeric synlbols) in the automatic
execution of programs, the mapping is frequently performed in three
successive phases, determining in turn the following quantities:
1. the index i = b l h,
2. the index k such that ale
3. the element ale.

= f(bJ,

The three phases are shown in detail in Program l.I2a. The ranking is
performed (steps 1-3) by scanning the set b in order and comparing each
element with the argument b. The second phase is a permutation of the
integers 1,2, ... , v(b), which may be described by a permutation vector
j, such that Ii = k. The selection of ji (step 4) then defines k, which, in
turn, determines the selection of ale on step 5.
Example 1.2.

If
b

= (apple, booty, dust, eye, night),

a = (Apfel, Auge, Beute, Nacht, Staub)

are, respectively, a set of English words and a set of German correspondents (both
in alphabetical order), and if the function required is the n1apping of a given
English word b into its German equivalent a according to the dictionary corre
spondences:
English:
apple booty dust
eye
night
German:
thenj

Apfel

(1,3,5,2,4). If b

Beute

Staub

"night," then i

Auge
5,ji

Nacht
4, and a

a4

Nacht.

If k is a permutation vector inverse to j, then Program I.I2b describes a

mapping inverse to that of Program I.I2a. If j = (1,3,5,2,4), then
k = (1,4,2,5,3). The inverse mapping can also be described in terms of
j, as is done in Program I.I2e. The selection of the ith component of the
permutation vector is then necessarily replaced by a scan of its components.
Programs 1. I2d and 1.12e show alternative formulations of Program 1.12a.

i+-O

i+-O
i~-i+l

+-

bi

k +-ii

k -<-ki

(0) b i

+- a7,~

aj

+-

b 1.:

(b) a j <=> b k

<=> a h

i+-blb
i -<-0

k -<-0

k+-k+l

6
7

+-

k -<-ii

-<-

ak

ai

(d) b i <=> aji

i : il;
h

+--

bk

+-

(b

= b)/i

a +-a k
(c) a; <=> bki

(e) b i <=> aji

Set of correspondents in
Programs (a, d, e) and set
of arguments in Programs

(b, c).
b

Set of arguments in Pro

grams (a, d, e) and set
of correspondents in Pro
grams (b, c).

j, k

Mutually inverse
permutation vectors.
Legend

Program 1.12

Mapping defined by a permutation vector j

35

36

1.17

The language

Ordering vector
If x is a numeric vector and k is a j-origin permutation vector such that
the components of Y = k.L x are in ascending order, then k is said to order
x. The vector k can be determined by an ordering operation defined as
follows:
k +- 8j jx
implies that k is a j-origin permutation vector, and that if Y = k Jj x, then
either Yi < Yi+1 or Yi = Yi+1 and k i < k i +1. The resulting vector k is
unique and preserves the original relative order among equal components.
For example, if x = (7, 3, 5, 3), then ()ljX = (2,4, 3, 1).
The ordering operation is extended to arbitrary vectors by treating all
nonnumeric quantities as equal and as greater than any numeric quantity.
For example, if a = (7,0,3, ,0 5,3), then ()lja = (3,6,5, 1,2,4), and if
b is any vector with no numerical components, then ()jjb = li(v(b)).
Ordering of a vector a with respect to a vector b is achieved by ordering
the b-index of a. For example, if a = (e, a, s, t, 4,7, t, h), and b is the
alphabet, then m = b II a = (5, 1, 19,20,0,0,20,8) and ()ljm = (2, 1,8,
3, 4, 7, 5, 6).
The ordering operation is extended to matrices by the usual convention.
If K = ())/A, then each column of the matrixB = KJ.LAisinascending
order.

1.18

MAXIMIZATION

In determining the maximum In over components of a numerical vector

x, it is often necessary to determine the indices of the maximum components

as well. The maximization operator is therefore defined so as to determine
a logical vector v such that vjx = m.
Maximization over the entire vector x is denoted by r x, and is defined as
follows: if v = r x, then there exists a quantity In such that vjx = In and
such that all components of vjx are strictly less than n1. The maximum is
assulned by a single com.ponent of x if and only if + jv = I. The actual
value of the maxin1um is given by the first (or any) component of vjx.
Moreover, the .i-origin indices of the maximum components are the
components of the vector vjt i .
More generally, the maximization operation v +- x will be defined so
as to determine the maximum over the subvector ujx only, but to express
the result v with respect to the entire vector x. More precisely,

ur

v +- urx<=> v = u\(r(ujx)).

The operation n1ay be visualized as follows-a horizontal plane punched

at points corresponding to the zeros of u is lowered over a plot of the

1.19

Inverse functions

37

components of x, and the positions at which the plane first touches them
are the positions of the unit components of v. For example, maximization
over the negative components of x is denoted by
v+---Cx

<

0)[ x

and if x = (2, -3,7, -5,4, -3,6), then (x < 0) = (0,1,0,1,0,1,0),

v = (0, 1,0,0,0, 1,0), v/x = (-3, -3), (V/X)l = -3, and. V/l 1 = (2,6).
Minimization is defined analogously and is denoted by ul x.
The extension of maximization and minimization to arbitrary vectors is
the same as for the ordering operation, i.e., all nonnumeric quantities are
treated as equal and as exceeding all numeric quantities. The extensions
to matrices are denoted and defined as follows:

v +--V

+---

V [X <=> Vi

rrX<=> V

\'";

Jl +--- U [ x<=> Vi
n

+--- U [

Vi rXi,

=
=

V j [Xj,

Vi[x,

X<=> V =
j

u [X j

As in the case of the ordering operation, nlaximization in a vector a with

respect to order in a set b is achieved by maximizing over the b-index of a.
Thus if
H

(d c h d h s h d c h c h d)
a6kq435k82j92

represents a hand of thirteen playing cards, and if

B=

then

c, d, h, s, 0, 0, 0, 0,
(

BloH= (

0, 0, 0, 0, 0)

2, 3, 4, 5, 6, 7, 8,9, 10, j, q, k, a
1, 0, 2, 1, 2, 3, 2, 1, 0, 2, 0, 2, 1)
12,4,11,10,2,1,3,11,6,0,9,7,

(4,13) 1_ (B i o H) = (25,4,37,23,28,40,29,24,6,26,9,33,13),
and
(e[((4, 13) ~ (B i O H)))/H = (s, 3)
is the highest ranking card in the hand.

1.19

INVERSE FUNCTIONS

To every biunique* function f there corresponds an illl'erse function g

such that g(f(x)) = ::c for each argunlent x in the donlain of the function f

* If the function f is many-to-one, the specification of a unique inverse g is achieved

by restricting the range of g to sonle set of "principar' values, as is donc, for cxanlplc,
for the inverse trigonometric functions.

38

The language

1.19

It is common practice either to introduce a distinct symbolism for the

inverse function, as for the inverse functions of logarithm (logo x) and
exponentiation (b 3), or to use a superscript -1, as in sin- 1 x or f-l(X).
The first alternative doubles the number of distinct operator symbols
required and obscures the relation between pairs of inverse functions;
the second raises other difficulties. The solution adopted here is that of
implicit specification; i.e., a statement is permitted to specify not only a
variable but also any function of that variable. Functions may therefore
appear on both sides of the specification arrow in a statement. For
example,
(2) 1- x +- z
specifies the variable x as the vector whose base two value is the number z.
Certain ambiguities remain in the foregoing statement. First, the
dimension of x is not specified. For example, if z = 12, x = (1, 1, 0, 0) is
an admissible solution, but so are (0, 1, 1, 0, 0) and (0, 0, 0, 1, 1, 0, 0).
This could be clarified by compatibility with a specified dimension of E.
Thus the statement
x+-z
(2(5))
specifies x unambiguously as (0, 1, 1, 0, 0). More generally, however, any
previously specified auxiliary variables will be listed to the right of the
main statement, with a semicolon serving as a separation symbol. The
current example could therefore be written as

vex)
(2)

+-

1- x +- z; vex).

The second ambiguity concerns the permissible range of the individual

components of x. For example, the base two value of x = (5, 2) is also
twelve. For certain functions it is therefore necessary to adopt some
obvious conventions concerning the range of the result. The assulnption
implicit in the preceding paragraph is that each component of x is limited
to the range of the residues modulo the corresponding radix. This con
vention will be adopted. Hence the pair of statements

y
Y

+-

(7, 24, 60, 60)

x+- 7278; y

determines x unan1biguously as the vector (0, 2, 1, 18).

It is also convenient, though not essential, to use selection operations on
the left of a statement. Thus the statement

u/b

+-

1.21

Subroutines

39

is understood to respecify only the selected components of b and to leave

all others unchanged. It is therefore equivalent to the statement

b ~ \ulb, u, a\.
Similarly,

ulb ~ ula
is equivalent to
b ~ Ib,

1.20

u, al.

LEVELS OF STRUCTURE

Vectors and matrices are arrays which exhibit one level and two levels of
structure, respectively. Although in certain fields, such as tensor analysis,
it is convenient to define more general arrays whose rank specifies the
number of levels of structure (i.e., zero for a scalar, one for a vector of
scalars, two for a vector of vectors (matrix), three for a vector of matrices,
etc.), the notation will here be lilnited to the two levels provided by the
matrix. * The present section will, however, indicate methods for ren10ving
this limitation.
The only essential particularization to two levels occurs in the provision
of single and double symbols (e.g., "I" and "II", "~" and "Jl") for row
and column operations, respectively, and in the use of superscripts and
subscripts for denoting rows and columns, respectively. In applications
requiring multiple levels, the former can be generalized by adjoining to the
single symbol an index which specifies the coordinate (e.g., "'II" and "/2'"
for row and for column compression, and, in general, "//'.) The latter can
be generalized by using a vector index subscript possessing one con1ponent
index for each coordinate.
The generalized notation can be made conlpatible with the present
notation for vectors and matrices by adopting the nam.e tensor and a
symbol class (such as capital italics) for the general array of arbitrary rank.

1.21

SUBROUTINES

Detail can be subordinated in a more general n1anner by the use of

subroutines. The nanle of one program appearing as a single statenlent in
a second prograln implies execution of the nanled progranl at that point;
the named program is called a subroutine of the second progran1. I f, for
example, "Cos" is the name of a progran1 which specifies z as the cosine of

* Further levels can, of course, be handled by considering a fanlily of Illatriccs 1M,

2M, ... , riM, or falnilies of faolilies /M.

40

The language

1.21

the angle between the vectors x and y, then Program l.13a uses the pro
gram "Cos" as a subroutine to determine r as the cosine of the angle
between the vectors p and q.

X+-P

Cos (p, q)

+-q

--1

r - Cos (p , q)

+- Z

Cos
r

+- Z

(a)

(b)

Program 1.13

(c)

Modes of subroutine reference

It is sometimes convenient to include the names of the arguments or

results or both in the name of the subroutine as dummy variables. Thus if
"Cos (x, y)" is the name of a subroutine which determines z as the cosine
of the angle between x and y, then Program l.13b uses Cos (x, y) as a
subroutine to determine r as the cosine of the angle between p and q.
Similarly, the program "z +- Cos (x, y)" can be used as in Progranl 1.13c
to produce the same result.

1.22

FILES

Many devices used for the storage of information impose certain restric
tions upon its insertion or withdra\val. The items recorded on a magnetic
tape, for example, may be read from the tape much more quickly in the
order in which they appear physically on the tape than in some other
prescribed order.
Certain storage devices are also self-indexing in the sense that the item
selected in the next read from the device will be determined by the current
state or position of the device. The next itenl read from a magnetic tape,
for example, is determined by the position in which the tape was left by the
last preceding read operation.
To allow the convenient description of algorithlTIs constrained by the
characteristics of storage devices, the following notation will be adopted.
A file is a representation of a vector x arranged as follows:

The null elements denote the "unused" portion of the file not employed in

Files

1.22

41

representing x. Each partition Pj determines a position (position j) in the

file. If a file (1) is in position j, then a forward read, denoted by
x, p

+- 0(1),

specifies x by the cOluponent Xj' the auxiliary variable p by the succeeding

partition Pi-i-l' and stops the file in the position j + 1.
The position of a file cD will be denoted by 7T(W). Thus the statenlent
j +- 7T(cD) specifies j as the position of (1), whereas 7T(I)) +- j positions the
file to j. I n particular, 7T(cD) +- 1 denotes the rewinding of the file and
7T(!)) +- v denotes lvinding, i.e., positioning to the extreme end of the file.
Any file for which the general positioning operation 7T(<l) +- j is to be
avoided as impossible or inefficient is called a s.erial or serial-access file.
Each terminal partition (that is, PI and p~,(P) assumes a single fixed
value denoted by}... Each nonterminal partition Pi may assume one of
several values denoted by AI' A2 , ,Av(A)' the partitions with larger
indices normally demarking larger subgroups of components within the
file. Thus if x were the row list of a matrix, the last conlponent might be
followed by the partition A3 , the last cOluponent of each of the preceding
rows by A2 , and the remaining components by AI. The auxiliary variable p
specified by the partition symbol during the read of a file is nornlally used
to control a subsequent branch.
A file may be produced by a sequence of jorH'ard record staternents:

where p is the partition synlbol recorded after the component Xj. As in

reading, each forward record operation increments the position of the file
by one. A file which is only recorded during a process is called an output
file of the process; a file which is only read is called an input file.
Different files occurring in a process will be distinguished by righthand
subscripts and superscripts, the latter being usually ernployed to denote
major classes of files, such as input and output.
Example 1.3. A set of 171 input files <I)il, i E LI (nl), each tcrn1inatcd by a partition
A2 , is to be copied to a single output file <1\ 2 as follows. Successive iten1s (C0I11
ponents) are chosen in turn fron1 files (D 1 1, (1)/, ... , <I) till, (1)/, (1)/, ... , always
on1itting from the sequence any exhausted file. A partition A 2 is to be recorded
with the last item recorded on <1\ 2 , and all files are to be rewound. The process is
described by Progran1 1.14.
Program 1.14. Step 8 cycles k through the values 1 to nl, and step 9 allows the
read on step 10 to occur only if Uk = O. The logical vector U is of di111cnsion 171
and designates the set of exhausted files. 1ts kth component is set to unity by
step 11 when file k is exhausted, as indicated by the occurrence of the partition A2
Each read is normally followed by step 13, which records on the output tl1e the

42

1.22

The language

U ~

e(m)
<I> .1

Input files for i E "l(m).

Each has terminal
partition A2

<1>2
1

Output file.

File <I> l is exhausted if

and only if u i = 1.

Itern to be recorded.

1,

2
7T(<I>1 )

i +-

n{<I>l)

i+-i-l

u:e

Uk

~O

~ 1

m
i=

ml 1(k

+-

Legend

+ 1)

10

b,p

11

Uk ~(p

12

u:e

13

<1>1

14

<D 12

+-

+-<I>k

= A2)

+-

b, Al

+-

b, A 2

Program 1.14

Program for Example 1.3

item read. However, when the last file becomes exhausted, step 14 is executed
instead to record the last item, together with the final partition A2
Steps 1-6 initialize the parameters u and k and rewind all files. After the last
item is recorded by step 14, the file rewinds are repeated before the final termina
tion on step 7.

It is sometimes convenient to suppress explicit reference to the partition

symbol read from a file by using a statement of the form
.Al

fh

I A2 .

~x+-o'-V~,

where the indicated branches depend on the value of the partItIon Pj+l
which terminates the read. Thus the left or the right branch is taken
according to whether Pj+I = Al or Pj+l = A2 Certain files (such as the
IBM 7090 tape files) permit only such "immediate" branching and do not
permit the partition symbol to be stored for use in later operations, as was
done in Program 1.14.
I n recording, the lowest level partition Al may be elided. Thus statement
13 of Program 1.14 may be written as
<D I 2

+-

b.

43

1.22

A file n1ay be read or recorded backward as well as forward. A backward

read is denoted by
x, P +-- I(I>,
and if <I) is initially in position j + I, then x == Xj' P == Pj' and the final
position becomes j. Backward recording is defined analogously. The
zero prescript may be omitted frOITI the syn1bol 0(1) for both fOf'Nard
reading and recording.
The conventions used for matrices can be applied in an obvious way to
an array of files (1)/. For example, the statement

7TDI)

+--

denotes the rewinding of the rOlV of Jiles (D/, j

7T( (I) J

ll(V(V)): the staten1cnt

+--

denotes the rewinding of the colunlll of Jiles <:1)/, i E llCUD)); and the

statement

ul(Di +-- ulx, ulp

denotes the recording of the vector con1ponent x j on file (D/ together \vith
partition Pj for all j such that u j == 1.
As for vectors and matrices, j-origin indexing D1ay be used and \vill
apply to the indexing of the file positions and the partition vector A as well
as to the array indices. I-Iowever, the prescripts (denoting direction of
read and record) are independent of index origin. O-origin indexing is
used in the following example.
Example 1.4. Files <Doo and <1\0 contain the vectors x andy, respectively, each
of dinlension n. In the first phase, the conlponents are to be nlerged in the order
x o, Yo, Xl' Y1' ... 'XlI-I' Yjl-l' and the first n conlponents of the resulting vector
are to be recorded on file (1)0 1 , and the last n on file <1\1. In other words, the
vectors Xl = anlz, andy 1 = wnlz are to be recorded on (1)0 1 and (1)/, respectively,
where z = \x, U, y\, and U = (0, 1, 0, 1, ... , 0, 1). In the next phase, the roles
of input and output files are reversed, and the sanle process is perfornled on Xl
and yl, that is, x 2 = an lexl, u, yl J, and y2 = w n lexl, U, yl J are recorded on
files <1)0 0 and <1)1, respectively. The process is to be continued through IJl phases.
Program 1.15. The program for Exanlple 1.4 begins with the rewind of the
entire 2 x 2 array of flIes. To obviate further rewinding, the second (and each
subsequent even-nunlbered) execution is perfornled by reading and recording all
files in the backward direction. Step 6 perfornls the essential read and record
operation under control of the logical vector u, whose conlponents u o, u l ' u 2
detennine, respectively, the subscript of the file to be read, the subscript of the
file to be recorded, and the direction of read and record. The 11le superscripts
(determining which classes serve as input and output in the current repetition) are
also determined by u 2 ' the input being u 2 and the output u 2 . The loop 6-8 copies

44 The language

1.22
O-origin indexing
E(2 x 2)

7T(<D)

+-

U +-

(3)

m+-m-1

U +-U

k +- n

8
9
10

>

File array of dimension

2 x 2; original input <Do;
original output <])1.

Control vector.

<
o

Column index of input

file.

u1

Column index of output

file.

u2

Row index of current

input file, and direction
of read and record.

Number of items per

file.

Required
merges.

<DU2
U

+-

o +-

<])

U2

<D Uu2

k+-k-l
u1
u1

+- U 1

u2

Program 1.15

i~

number of

Legend
Program for Example 1.4

n iten1s, alternating the input files through the negation of U o on step 7. When
the loop terminates, u 1 is negated to interchange the outputs, and the loop is
repeated unless u 1 = u 2 - Equality occurs and causes a branch to step 3 if and
only if all 2n items of the current phase have already been copied.
Step 3 decrements 111 and is followed by the negation of U on step 4. The con1
ponent u 2 must, of course, be negated to reverse direction, but the need to negate
U o and u 1 is not so evident. It arises because the copying order was prescribed for
the forward direction, beginning always with the operation

An equivalent backward copy must therefore begin with the operation

Not all computer files have the very general capabilities indicated by the
present notation. Some files, for example, can be read and recorded in the
forward direction only and, except for rewind, cannot be positioned
directly. Positioning to an arbitrary position k must then be performed by
a rewind and a succession of (k - 1) subsequent reads. In some files, re
cording can be performed in the forward direction only, and the positions
are defined only by the recorded data. Consequently, recording in posi
tion k makes unreliable the data in all subsequent positions, and recording
must always proceed through all successive positions until terminated.

Ordered trees

1.23
1.23

45

ORDERED TREES

Directed graphs

For many processes it is convenient to use a structured operand with the

treelike structure suggested by Fig. 1.16. It is helpful to begin with a more

114

223

1311

n16

2221

n21

n22

22231

n25

Figure 1.16 A general triply rooted tree with ).(T) = 16, 'J(T)
= 5, ~(T) = (3, 7, 8, 5, 3), and {leT) = 26

veT)

(3, 3, 4, 3, 2),

46

1.23

The language

general structure (such as Fig. 1.17) in which a unidirectional association

may be specified between any pair of its components.
A directed graph comprises a vector n and an arbitrary set of unilateral
associations specified between pairs of its components. The vector n is
called a node rector and its components are also called nodes. The associa
tions are conveniently specified by a (logical) connection matrix U of
dimension v(n) X v(n) with the following convention: there is an associa
tion, called a branch, fronl node i to node j if and only if Uj ; = 1.
A directed graph admits of a simple graphical interpretation, as
illustrated by Fig. 1.17. The nodes might, for example, represent places,
and the lines, connecting streets. A two-way street is then represented by
a pair of oppositely directed lines, as shown between nodes 3 and 4.
If k is any mapping vector such that
for i

2, 3, ... , v(k),

then the vector p = kJn is called a path rector of the graph (n, U). The
dimension of a path vector is also called its length. Nodes k 1 and k v are
called the initial and final nodes, respectively; both are also called
tenninal nodes. If j is any infix of k, then q = jJ n is also a path. It is
called a subpath of p and is said to be contained in p. If v(q) < v(p), then
q is a proper subpath of p. If k 1 = k" and p = kJn is a path of a length
exceeding one, p is called a circuit. For example, if k = (6, 1,7,7,2, 6,
1, 5), then p = (n 6, n 1, n 7, n 7, n 2, nfl' n 1, n 5) is a path vector of the graph
of Fig. 1.17, which contains the proper subpaths (n 7, n 2, n 6), (n 1, n 7,
n 7, n 2, n 6, n 1), and (n 7, n 7), the last two of which are circuits. Node j
is said to be reachable from node i if there exists a path from node i to
node j.

Ordered trees
A graph (such as Fig. 1.16) which contains no circuits and which has
at most one branch entering each node is called a tree. Since each node
is entered by at most one branch, a path existing between any two nodes
in a tree is unique, and the length of path is also unique. Moreover,
if any two paths have the same final node, one is a subpath of the
other.
Since a tree contains no circuits, the length of path in a finite tree is
bounded. There therefore exist maximal paths which are proper subpaths
of no longer paths. The initial and final nodes of a maximal path are
called a root and leaf of the tree, respectively. A root is said to lie on the
first leeel of the tree, and, in general, a node which lies at the end of a path
of length j from a root, lies in the jth level of the tree.

1.23

47

Ordered trees

n2

n5

u=

0
0

0
0

0
0

n4

Figure 1.17

A graphical representation of the directed graph (n, U).

A tree which contains n roots is said to be n-tup~v rooted. The sets of

nodes reachable from each of the several roots are disjoint, for if any
node is reachable by paths from each of two disjoint roots, one is a proper
subpath of the other and is therefore not maximal. Similarly, any node
of a tree defines a subtree of which it is the root, consisting of itself and
all nodes reachable from it, with the same associations as the parent
tree.
If for each level j, a simple ordering is assigned to each of the disjoint
sets of nodes reachable from each node of the preceding level, and if the
roots are also simply ordered, the tree is said to be ordered. Attention will
henceforth be restricted to ordered trees, which will be denoted by u pper
case boldface roman characters. The height of a tree T is defined as the
length of the longest path in T and is denoted by veT). The nunlber of
nodes on level j is called the moment of lel'el j and is denoted by fLlT). The
vector (L(T) is called the moment l'ector. The total number of nodes in T is
called the moment of T and is denoted by /leT). Clearly, v(fL(T)) = veT),
and +/fL(T) = /leT) = v(n). The number of roots is equal to fLI(T), and
the number of leaves will be denoted by ).(T).
The number of branches leaving a node is called its branchinK ratio or
degree, and the maximum degree occurring in a tree T is denoted by beT).
The dispersion rector of a tree T is denoted by veT) and is defined as
follows: vI(T) = (LI(T), and for j = 2,3, ... , veT), vlT) is equal to the
maximum over the branching ratios of the nodes on level j - I. For the
tree of Fig. 1.16, veT) = (3, 3, 4, 3, 2). The number of roots possessed by
a tree T (that is, vI(T)) is called its dispersion. A tree possessing unity
dispersion is called rooted or singular.

48

1.23

The language

Each node n i of a graph (and hence of a tree) may be identified by its

index i. Since a tree admits of more convenient index vectors, the under
lying index i will henceforth be referred to as the graph index.
In an ordered tree, any path of length k from a root can be uniquely
specified by an index vector i of dimension k, where i 1 specifies the partic
ular root, and the remaining components specify the (unique) path as
follows: the path node on level) is the ijth element of the set of nodes on
level) reachable from the path node on level )-1. The node at the end of
the path can therefore be designated uniquely by the index vector i. The
degree of node i will be denoted by b(i, T). The index vectors are shown
to the left of each node in Fig. 1.16.
The path from a root whose tenninal node is i will be denoted by T i .
In Fig. 1.16, for example, T i == (n 2 , ~8' n 13 , n 24 ) if i == (2, 2, 2, 3). A
vector i is said to be an index of T if it is the index of some node in T.
The subtree of T rooted in node i will be denoted by T i . Thus in Fig.
1.16, P == T(~.~,2) is a rooted subtree with v(P) == (1, 3, 2), and fJ.(P) ==
(1,3,3). A path in T i is denoted by (Ti)i. For example, if G is an
ascending genealogical tree* with the sword and distaff sides denoted by
the indices 1 and 2, respectively, then any individual x and the nearest
(Il - I) paternal male ancestors are represented by the path vector (Gi)(n),
where i is the index of x in G.
Example 1.5. Deternline the index i such that the path T i is equal to a given
argunlent x and is the "first" such path in T; that is, the function
(ex"(X) jv(T))

is a nlini111U111.

Program 1.18. The index vector i specifies the path currently under test. Its
last conlponent is incremented repeatedly by step 7 until the loop 6-8 is terminated.
If the path T i agrees with the corresponding prefix of the argument x, ternlination
occurs through the branch to step 9, which tests for conlpletion before step 10
augnlcnts i by a final zero component. Step 5 then respecifies d as the degree of
the penultilnate node of the set of d paths next to be tested by the loop. Termina
tion by a branch from step 6 to step 2 occurs if all d possible paths are exhausted
without finding agreement on step 8. I n this event, retraction by one level occurs
on step 2, and d is again respecified. Jf 1'( i) = I, the paths to be searched conl
prise the roots of the tree and d nlust therefore be specified as the nU111ber of
roots. This is achieved by executing step 3 and skipping step 5. Retraction to a
vector i of dinlension zero occurs only jf all roots have been exhausted, and final
ternlination from step 4 indicates that the tree possesses no path equal to the
argunlent x.

* Although such a genealogical tree is not necessarily a tree in the n1athen1atical

sense, it will be assumed so for present purposes.

1.23

Ordered trees
i

+-

e(2)

+-

wi/i

<

9
10

+i

Given tree.

Given path vector.

Path index vector to

be determined.

(5(;)1/i, T)

/; :

; +-

=I-

I-origin indexing

fJ.l(T)

v(i)
d

+-

49

+ wI

fJ.l(T)

Number of roots of T.

(5(i, T)

Degree of node i.

Legend

al'(i)/x

v(i) : vex)
i

+-

wI\i

Program 1.18

Deternlination of i such that Ti

If d is a vector of dinlension v(n) such that d i is the degree of node n i of

a tree T, then d is called the degree vector associated lrith n. In Fig. 1.16,
for example,
d = (3, 2, 4, 0, 0, 0, 2, ... , 1, 0, 0).
Moreover, if n is itself the alphabet (that is, n = (a, b, c, ... , z)), then
the vector n' of Table 1.19a is a permutation of n, and d' is the associated
degree vector. Table 1.19b shows another such pair, nil and d".
The degree vector provides certain useful infornlation lTIOSt directly.
For example, since each leaf is of degree zero, }.(T) = +/(d = 0). More
over, the number of roots is equal to the number of nodes less the total of
the degrees, that is, flol(T) = v(d) - + /d, and the 11laxinlunl degree
occurring in T is given by beT) = ((f d)/d)l. Finally, the degree vector
and the node vector together can, in certain permutations (those of Table
1.19), provide a complete and compact descri ption of the tree.

Right and left list matrices

If each one of the /-leT) index vectors i of a tree T is listed together with
its associated node (Ti)l'(i)' the list determines the tree conlpletely. Since
the index vectors are, in general, of different dimensions, it is convenient
to append null components* to extend each to the common lnaxinlunl
dimension veT). They may then be combined in an index /JIatrix of

* In the I-origin indexing system used here it would be possible to use the nUI11eric
zero to represent the null. In O-origin indexing, however, zeros occur as con1ponents of
index vectors and must be distinguishable from the nulls used.

50

1.23

The language

I~
3
4

0
0
0
0
0
1
2

0
0

c
f
d
r
e
z

11
i

P
q

111

0
0
1
0
0

t
w

0
0

j
1

ll

]11

nIl

b
k
h

]'

n'

y
v

1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3

1
1
1
1
1

1
2
3
4

1
2
3

3
2
2

4
5
6
7
8
9
10
11
12

4
0
1
0

2
3
3
3
3

1
1

1
2

1
2
2
2
2
2
2
2
2
2
2

1
2
2
2
2
2
2
2
3

15
16

21
22
23

24

25

2
2
2
3
3

1
2

2
3
1
2

0
0

1
1

0
0

3
2
2
2

17

111

18
19

0
0
0
0

26

2
1

0
0
0

s
x
t

0
0

2
2
2

1
2
2
2
2
2
2

2
2
2
3
3
2
2
2
2
2
2

Full left list 111atrix [T

Full right list 111atrix ]T

(a)

(b)

Table 1.19

P
q

1
2
3

20

1
f
d
r
e
i

13
14

0
0
0
0
0
0

b
g
c
z
11
k
h

2
2
3
3

1
2
1
2
3

1 4
1

2
3
1

:2
:2

:2

2
2

:2
3

1
3

1
:2
1

Full list 111atrices of the tree of Fig. 1.16

dimension /leT) X veT), which, together with the associated node vector,
completely describes the tree, If, for example, the node vector n is the
alphabet, the tree of Fig. 1.16 is described by the node vector n' and index
matrix I' of Table l.l9a or, alternatively, by n" and I" of Table 1.19b.
Because of the utility of the degree vector, it 'will be annexed to the array
of node vector and index matrix, as shown in Table 1.19a~ to form a full
list lnatrix of the tree. The degree vector and node vector together will be
called a list lnatrix. As remarked, the list Inatrix can~ in certain pernluta
tions, alone describe the tree.
Formally, the full list matrix M of a tree T is defined as follows: a2/ M
is an index matrix of the tree, M 1 is the associated degree vector, and M;!.

Ordered trees

1.23

51

is the associated node vector. Thus for each k E llCU(T)), M/;: = b(i, T),
and M/~ = (Ti).l'(i)' where i is the nonnull portion of -a,2/M\ that is, i =
((fi2/ Mk)
0)/fi 2/ Mk). The corresponding list matrix is a 2/ M.
Since a full list matrix provides a complete description of a tree regard
less of the order in which the nodes occur in the list, any column permu
tation MP (that is, any reordering among the rows) is also a full list matrix.
Two particular arrangements of the full list nlatrix are of prime interest
because each possesses the following properties: (1) the nodes are grouped
in useful ways, and (2) the list matrix (i.e., the degree vector and node
vector) alone describes the tree without reference to the associated index
matrix. They are called the full left list matrix and full r(r;lzt list matrix and
are denoted by [T and ]T, respectively. Table 1.19 shows the fuHleft and
full right lists of the tree of Fig. 1.16.
The left list index matrix I is left justified, * that is, the null elements are
appended at the right of each index. The rows Ij are arranged in increasing
order on their values as decimal (or rather (b(T) + 1)-ary) fractions with
the radix point at the left and the nulls replaced by zeros. More precisely,
the rows are arranged in increasing order on the function (1'(o))
(a to Ij), where a = (0, 1, 2, ... , ()(T)). t
The right list Inatrix is right justified and is ordered on the sanle function,
namely (v(a)) ~ (a Lo Ij). The rows are therefore ordered on their values
as integers, i.e., with the decimal point at the right. Fronl the exalnple of
Table I.I9b it is clear that the right list groups the nodes by levels, i.e.,
level j is represented by the infix (i J a l,)/ /(]T), where k = fJ-lT), and
i = + /aj-1/fJ-(T). In Table 1.19b, for exanlple, fJ-(T) = (3, 7, 8, 5, 3), and
if j = 3, then k = 8, i = 10, and level j is represented by rows i + 1 = II
to i + k = 18. The right list is therefore useful in executing processes
(such as the pth degree selection sort) \vhich require a scan of successive
levels of the tree.
The left list groups the nodes by subtrees, i.e., any node i is followed
inl111ediately by the remaining nodes of its subtree T i . Fornlally, if I =
cx 2/[T, and if i = (II,; :F o)/II,', then the tree T i is represented by the infix
((k - I) J a/I(Ti ))/ /[T. In Fig. 1.19a, for exa111ple, if k = 16, then i =
(2,2,2), {l(T;) = 7, and T i is represented by rows 16 to 22 of [T. The left
list is therefore useful in processes (such as the construction of a H uffnlan
code and the evaluation of a cOlnpound statenlent) which require a
treatnlent of successive subtrees.
The row index of a node in a right (left) list nlatrix is a graph index of
the node and will be called the right (left) list index.

* The term left list and the notation

[T are both intended to suggest left justification.

These statements hold only for I-origin indexing. ]n O-origin indexing, a =
0,1, ... ,()(T) -1).

-j(~,

52

The language

1.23

Well fornlation
A two-colulnn matrix which forms the right list of son1e tree is said to
be a It'ell fornled right list. Since the ordering of the nodes in a right list of
a given tree is unique, the right list of a given tree is unique. Conversely,
any well fonned right list specifies a unique tree according to the algorithm
of Progran1 1.20.
Identical relnarks apply to the left list, except that Progran1 1.20 is
replaced by PrograIn 1.21. Moreover, the necessary and sufficient
conditions for the well formation of a left list are identical with those for a
right list and are derived by virtually identical arguments. The case will
be stated for the right list only.
If R is a well fornled right list representing a tree T, then the dispersion
(i.e., the nUlnber of roots) vl(T) = v(R I ) - (+ IR I ) must be strictly
positive. Moreover, if S = CiJ II R is any suffix of R, then S is a right list of
the tree obtained by deleting from T the first j nodes of the original list.
For, such level-by-Ievel deletion always leaves a legitin1ate tree with the
degrees of the remaining nodes unchanged. Consequently, the nun1ber of
roots deternlined by every suffix of R l must also be strictly positive. In
other words, all components of the suffix dispersion rector s defined by

Inust be strictly positive. The condition is also sufficient.

Sufficiency is easily established by induction on the column din1ension
of R. The condition is clearly sufficient for v(R l ) = 1. Assun1e it sufficient
for din1ension v(R l ) - 1. If s, the suffix dispersion vector of R, is strictly
positive, then aIls, the suffix dispersion vector of filIIR, is also positive,
and by hypothesis alii R represents a tree G possessing S2 roots. Moreover,

implies that S2 :2:: R I \ and the number of roots possessed by G therefore

fulfills the nUlnber of branches required by the added node R 21. A
legitinlate tree corresponding to R can therefore be fornled by joining the
last R/ roots of G to the node R 2 1.
Tests for well formation can therefore be incorporated in any algo
rithm defined on a right or left list lnatrix M by computing the components
of the suffix dispersion vector s. The recursion Si-l = Si + 1 - MJ-l is
convenient in a backward scan of M, and the equivalent recursion
Si = Si-l 1 + Mf--l serves for a forward scan. The starting condition
for a forward scan is Sl = v(Ml ) - (+IM I ), and for a backward scan is
51' = 1 - MIll. Since the criteria of well fonnation are identical for right
and left lists, a nlatrix may be characterized sinlply as well or ill fonned.

1.23

Ordered trees

53

The purpose served by the degree vector d in the description of a tree is

sometimes served instead [e.g., Burks et a1. (1954)] by the vector g ==
- d. It is sonlewhat 1110re convenient in the analysis of well fOflnation,
since the expression for the suffix dispersion vector then sinl.plifies to
or

s == (I

+ 0);-1 g.

The index matrix as a function of the degree vector

The complete determination of the tree corresponding to a given list
111atrix Al is best described as the determination of the associated index
matrix I. For both left and right lists this can be achieved by a single
forward scan of the rows of M and of I.
For a right list R it is first necessary to deternline r, the nUlnber of roots.
The first r components of R are then the roots of the tree in order, the
next R 1l components of R are the second-level nodes reachable fronl the
first root, and so forth. Programs 1.20 and 1.21 describe the processes for a
right list and a left list, respectively.
Program 1.20. In each execution of the nlain loop 13--16, the ith row of the
right list R is exanlined to deternline the index vector of each node on the succeed
ing level which is directly reachable fronl it. The nunlber of such nodes is con
trolled by the paranleter d, initialized to the degree of the ith node by step 12.
The (right list) index of the nodes reachable fronl node i is deternlined byj, \vhich
is incremented on step 14 as the index vector of each node is detern1ined. The
index vectors of the successive nodes reachable fronl node i have the final con1
ponents 1, 2, 3, ... , and each nlust be prefixed by the index vcctor of node i.
This assignnlent is effected by the vector v, which is initialized by thc indcx vector
of node i rotated left by one (step 11), and which is incren1cnted by step 15 before
each assignment occurring on step 16. At the outset, v is set to zero and d is sct
to the nunlber of roots as deternlined by step 4.
Sincej is, at step 10, equal to the current nunlber of roots r augnlented by the
cumulative degrees of the first i - I nodes, then r = j - i + 1 and the exit on
step 10 therefore occurs always and only in the event of ill forn1ation. ;-\Itcrna
tively, the test can be viewed as an assurance that each row of the ll1atrix 1 is
specified before it is itself used in specification.
When step 5 is first reached, the index n1atrix I is cOlllplete but is cxprcssed in
I-origin indexing with zeros representing the null elen1cnts. Steps 5-7 translate
the matrix to the origin and nlask in the nccessary null elen1cnts.
Program 1.21. The index vectors lj are detennined in order under control of
the paranletcrj. The loop 5-18 traces a continuous path through the tree, deter
mining the index of each successive node of the path by rotating the index of the
preceding node (step 17) and adding one to the last con1ponent (step 13), and
maintaining in the connection vector c a record C i +1 of the indexj of the successor
of node i in the path traced. The path is interrupted by the occurrellce of a leaf
(that is, L1j = 0 on step 18), and the degree vector L 1 is then scanned by the loop

1.23

The language

54

I-origin indexing

i+-O
2

j+-O

v+-o

+-

~{Rl)

TJ

+-

(I = 0)

+-

~-II,

>

+-

+ (> -

I)E

Right index matrix of

T.

Index of row of R cur

ren tl y examined.

Right list index of node

reachable from node i.

i + 1

i :j

Current index vector.

>

Origin with respect to

which I is finally ex
pressed.

t Ii

11
12

d -(-R1i

13

d+-d-l

14

j +-j + 1

15

v -(- v +

16

Ii

+-

U,oEI

+-

Right list of T.

(+ IR])

i : peR)

10

Legend

<

WI

Program 1.20

Deternlination of the index n1atrix I associated

\vith a right list nlatrix R

(19-20) to deternline the index i of the last preceding node whose branches renlain
inconlpleted. Steps 22-23 then respecify v as the index vector of the node follow
ing node i in the path last traced, and step 21 decren1ents the conlponent L1i of
the degree vector. The branch fron1 step 19 to step 22 occurs at the con1pletion
of each rooted subtree. The test for well fonnation (step 12) is the sanlC as that
applied to the right list in Program 1.20, except that the notation for the relevant
paranleters differs. The concluding operations (6-9) include left justification on
step 7.

Tree, path, and level compression

The tree cOlnpressiol1

+--

U/T

specifics a tree P obtained fron1 T by suppressing those nodes corre

sponding to zeros of the logical tree U, and reconnecting so that for every
pair of nodes ::C, y of P, x belongs to the subtree of P rooted in y if and only
if x belongs to the subtree of T rooted in y. If, for exalnple, T is the tree of

Ordered trees

1.23

I-origin indexing

i~O

V~O

1 ~ ( + 1lJ.(U)

1 ~1

+--

10

~j

11

12

13

veL}) - (+IL 1 )

(1

Left index matrix of T.

Index of row of 1 being
determined.

Left list index of path

node preceding node j
in current path (Step
16), or index of last
previous node whose
branches remain un
exhausted (step 22).

C i +1

Index of node follow

ing node i in last path
traced from i.

Parameter for testing

well formation.

Current index vector.

0)

+ (1)

11,

- I)E

U, E1

+1
~r + L/
~-

+ w}

14
15

C i +I -(- j

16

i -(-j

Left list of T.

/1(L)
~

55

Origin with respect to

which 1 is expressed.
Legend

17

18

19

i -(- i - I

22

k ~- cHI

23

V ~-

II.'

Program 1.21 Deternlination of the index Inatrix 1 associated

with a left list nlatrix L

Fig. 1.16 with n as the alphabet, and U is the tree of Fig. 1.22a, then P
is the tree of Fig. 1.22b. The new indices are shown to the left of each node
of P. The set of nodes 221, 222, ... ,226, are all on the sanle level of P
although they have been shown with branches of different lengths to
permit easy identification with the nodes of the original tree T.

56

The language

1.23

111 f

131

21 k
2
b

U/T

(a)

(b)

Figure 1.22

Compression of tree T of Fig. 1.16 (with n

alphabet)

The compress operation is best executed on the left list because of the
grouping by subtrees. Program 1.23 gives a suitable algorithlTI which also
serves as a formal definition of the compress operation.
Program 1.23. The vector u is specified as the node vector of the left list of the
controlling logical tree U and controls the subsequent process. Step 4 deternlines
j as the index of the first zero conlponent of u. Steps 6 and 7 then delete the corre
sponding nodes of u and of the left list of T, but only after step 5 has deternlined d
as the change in degree which this deletion will occasion to the root of the sIllallest
subtree containing the deleted node. Steps 9-1 I perfornl a back\\'ard scan of the
degree vector to deternline j as the index of the root of the subtree, and step 12
effects the requisite change in its degree. The exit on step 9 occurs only if the node
deleted is a root of the original tree, in which event no change is produced in the
degree of any other node.

Two further compress operations controlled by logical vectors are

defined as follows. Path cornpressiol1 is denoted by

+-

u/T.

1.23

Ordered trees

~ a 2 j([T)

j~(+j'l.ju)+1

~-- L1j -

-o(-ejju

I-origin indexing

([U)2

u
E

r~l

~j

Left list of T.
Index of first zero of u
(Steps 4-8). Index of root
of smallest subtree contain
ing deleted node (Step 12).

Change of degree caused by

deletion of node j.

Number of roots indicated

by infix (a j ; \ -a.J')jjL, where
j is initial value and k + 1
is current value of j.

- 1

10
11

Left node vector of U.

7
8

57

>

12

Legend

Program 1.23

Determination of the left list L = a 2 j[(UjT)

P is obtained fronl T by suppressing every node on level j if u j == 0, and

reconnecting as in tree compression. Leref conlpressioll is denoted by

*-

ulIT,

and P is obtained from T by deleting each rooted subtree T i for which

U i == O.
Path compression by a unit vector j produces a tree of hcight onc.
Such a tree is, in effect, a vector and will be treated as one.
Two related special logical trees are defined: the path tree UE such that
u!UE == 0 and u!UE is the full tree E whose nodes are all unity, and the
level tree uE such that ulluE == 0, and ulluE == E.

Extension of other operations to trees

Two trees are cornpatible if they have the sanle structure. Elenlcntary
binary operations are extended node by node to conlpatible trees. For
example,

implies that node i of Z is the product of node i of X and node i of y for all
i. Similarly,

58

1.23

The language

specifies M as a tree (of the saIne structure as T) such that node i of M is

the j-origin b-index of node i of T.
The lnapping operation is extended to trees so as to pennute the rooted
subtrees of a tree. Forn1aIly
P ~ rn fJ T
implies that fLl(P) == vern), that Pi is a single null character if rn i ti(fLl(T)),
and otherwise Pi == T m.' where j-origin indexing is used for T.
Permutation of the s'ubtrees rooted in node i of T can be effected as
follows:
The notation ciiT will denote the application of the binary operator or
relation c to the nodes of T in right list order (i.e., dOl1'n successive levels)
and G /T will denote the san1e application in left list order (i.e., across
paths). If the operator is sYlnn1etric (i.e., its operands con1mute), then
oilT == OfT.
Maximization (UrT) and minin1ization (UlT) are extended to trees in
the obvious \vay.
The operations 'l./u, (!)/u, a/a, and T/a are each extended in two ways:
across paths and down levels. Examples of each appear in Fig. 1.24.
Operations extending down levels are denoted by double virgules and
represent an application of the corresponding vector operation to each
level of the tree considered as a vector. For example, the staten1ent
V~(JIIA

implies that each level of V is the forward set selection of the corresponding
level of A, that is, j /V == a/j/A. Operations extending across paths are
denoted by single virgules and are defined in tern1S of subtrees. Thus
V~'l./U

implies that V is obtained fron1 the logical tree U by setting to zero all
nodes of any subtree rooted in a zero node, and

(/)/U

implies that V is obtained fron1 U by setting to zero every node whose

subtree contains a zero node. The definitions of a/U and T/U are analo
gous.

Homogeneous trees
If, for all j, every node on level j of a tree T is either of degree zero or of
degree 'V j ;l(T), then the tree T is said to be unij()rn7. If all leaves of a

59

Ordered trees

1.23

uniform tree T lie in the same level (necessarily the top), then the tree is
said to be hOIJ1ogeneous. The structure of a homogeneous tree is completely
characterized by its dispersion vector veT). All maxin1al paths in a homo
geneous tree are clearly of the same length, nan1ely veT) == v(v(T)). Figure
1.25 shows a homogeneous tree and its associated dispersion vector.

I----~O

a/U

~-_....>@

~I---->"'@

~I-----~""CD

~I----"~CD

w/u

(J/U

~I---_->~@

a/IV

w/IV

O-//V

~f----""'>@
--~1

~--->-CD

~I---->~CD

r/U

Figure 1.24

r//U

Set selection and 111axin1unl prefix and suffix operations

60

1.23

The language

A tree T for which veT) = 111 is called an I1z-waJ' tree, and a tree for
which vI(T) = 1 and (iI/v(T) = m is called a singular I1z-lray tree.
The jth component of the mOlnent vector of
a homogeneous tree is clearly equal to the prod
uct of the first j components of the dispersion
n
vector, that is, fL(T) = (0 + I) / veT). The dis
persion vector is, in turn, uniquely deternlined
by the moment vector. The total nunlber of
nodes is given by [leT) = + /fL(T), and it can also
be shown that fl(T) = Y J_ y, where y is the
dispersion vector in reverse order.
Tree compression of a honl0geneous tree H
(that is, lJ/H) does not generally produce a
g
homogeneous tree, and, in fact, any tree P of
arbitrary structure can be represented by a pair
of homogeneous trees U and H such that P =
U/H. On the other hand, both path and level
q
compression of homogeneous trees produce
homogeneous trees. Moreover, ifP = u/H, then
r
v(P) = u/v(H), and if P = u//H, then v(P) =
v(H) - (+ /u)a].
Since the structure of a homogeneous tree is
completely specified by its dispersion vector k,
h
the structure of the special logical trees can be
specified in the forms E(k), UE(k), and uE(k).
In a homogeneous tree, the right list or left
list index of a node can be detennined as an
explicit function of its index vector. Conversely,
the index vector i can be determined directly
from the corresponding left list index, to be
denoted by I(i), or from the right list index rei).
U
In developing the relations between indices it'
will
be convenient to use O-origin indexing
II(H) = (2,3,2)
M(H) = (2,6,12)
throughout.
Figure 1.25 Homogen
The right list index is given by
eous tree H and disper
rei) = f(i) + g(i),
sion and moment vectors
f(i) = + /aJ/(i) -1 /fL(T)
where
is the number of nodes in the first 1'(i) - 1 levels, and
g(i)

(a '(i)/v(T))

1-

is the rank of node i in the v(i)th level. For exampIe, if i

(I, 0, 1) in the

Ordered trees

1.23

61

tree of Fig. 1.25, then fJ-(H) = (2, 6, 12), j(i) = +/(2, 6) = 8, and
g(i) = (2,3,2)
(1,0,1) = 7.
Since j(i) depends only on v(i), the index i may be detennined fron1 r by
first determining v(i) as the largest value for which j'(i) :s: r, and then
determining i such that

(cx'J(i)jv(l'))

_L i =

r - f(i).

In tracing a path through a tree, the kth node of the set reachable fron1
node i is the node j = i .2:) (k). It is therefore useful to express r(j) as a
function of rei). Clearly

f(j)

g(j)

= f(i) + (fJ-(T))I'(i)--l,
= g(i) X (v(T))IJ(i) + jl'-l.

In the special case of a singular hon10geneous In-way tree,

f(i)

+ 1n + n1 2 + ... + 1111'(i)-::' =

m 1!(i)-1

(111e)

l- e(v(i) -

1)

111 -

Hencej(j) = 1 + m X j(i), and g(j) = nl X g(i) + 11'-1' Recursion can

therefore be performed simply upon the single function rei) as follows:

r( j) =

111 X

r( i)

+ 1 + jl' - 1.

The left list index lei) is most conveniently expressed as a function of

v(i) and of the vector z(i) (zero extension of i), where z = cx '(i>Cv(T))\i.
Clearly v(z) = veT) and z is the index of the "earliest" leaf reachable froIn
node i. In Fig. 1.25, for example, z((I, 2)) = (1,2,0).
The zero extension has the obvious property that every node above the
path TZ(i) precedes node i in the left list, and every node belo\v the path
follows it. The number of nodes in the path which precede node i is
v(i) - 1.
The number of leaves above the path TZ(i) is veT) ~ z(i), and n10re
generally, the number of (j - 1)th level nodes above it is given by
(cxj/v(T)) ~ (cxj/z(i)). Consequently,
1

l'(T)

lei)

v(i) -

+ 2 (cxi/v(T)) l- (cxijz(i)).
j=l

For example, if i

lei)

v(i) -

(1,0) in Fig. 1.25, then z(i)

(1,0,0) and

+ (2) ~ (1) + (2, 3) ~ (1,0) + (2,3,2) l- (1,0,0) =

1 1.

The foregoing result may be written alternatively as

lei)

where W v = 1, and Wi-~ = 1

w = (10,3, 1), and w;- z(i)

v(i) - 1

+ (Wi
=

10.

w : z(i),

vieT)). In the foregoing exan1p1e,

This form is most convenient for

62

1.23

The language

determining i as a function of I, for since w z = I + 1 - v(i), then

zo(i) = II -7- woJ, zI(i) = l((wo II) - 1) -7- wIt etc. for all positive values
of the quotient, and all components thereafter are zero. The dimension
v(i) is then determined from the relation v(i) = I + 1 - w ~ z(i).

REFERENCES
Birkhoff, G., and S. MacLane (I941), A Survey of Modern A (fjebra, Macmillan, New
York.
Burks, A. W., D. W. Warren, and J. B. Wright (1954), "An Analysis of a Logical
Machine Using Parenthesis-free Notation," Mathel11atical Tables and Other Aids to
COl1zputation, vol. VIII, pp. 53-57.
Dickson, L. E. (1939), New First Course in the Theory of Equations, Wiley, New York.
Garner, Harvey L. (1959), "The Residue Number System," IRE Transactions, vol. EC-8,
pp. 140-147.
Goldstine, H. H., and J. von Neumann (1947), "'Planning and Coding of Problems for
an Electronic Computing Instrument," Report on the MathelJlatical and Logical
Aspects ofan Electronic COlllpUtil{[[ InstrUl11ent, Part II, vol. 1, Institute for Advanced
Study, Princeton.
Iverson, K. E. (1954), "Machine Solutions of Linear Differential Equations," Doctoral
Thesis, Harvard University.
Jacobson, N. (1951), Lectures in Ahstract A~fjebra, vol. 1, Van Nostrand, New York.
Kunz, K. S. (1957), Nurnerical Analysis, McGraw-Hill, New York.
Margenau, H., and G. M. Murphy (1943), The Mathel11atics of Physics and Chel11istry,
Van Nostrand, New York.
Phister, M. (1958), Logical Des({n of D{[{ital C0l11puters, Wiley, New York.
Richards, R. K. (1955), Arithl11etic Operations in Digital Computers, Van Nostrand,
New York.
Riordan, J. (1958), An Introduction to C0l11hinatorial Analysis, Wiley, New York.
Rutishauscr, H. (1959), "Zur Matrizeninversion nach Gauss-Jordan," Zeitschrift fiir
A/~{e~vandte Mathenzatik und Physik, vol. X, pp. 281-291.
Wright, H. N. (1939), First Course in Theory of Numbers, Wiley, New York.

EXERCISES
Organize each of the programs according to the method of leading decisions.
Except where otherwise indicated, use I-origin indexing. The conventions of
Sec. S.I of the Summary of Notation will be used in the statement of each
of the exercises.
1.1 Let d = (a, 2,3,4,5,6,7,8,9, 10,j, q, k), S = (c, d, h, s), u = (1,0, 1,0,1),
v = (0, I, I, 1, 0), x = (16, 8, 4, 2, I), and y = (2, 3, 4, 5, 6). Determine
(a) the dimensions v(d), v(s), and vex).
(b) the vectors x + y, x - y, x x y, x -:- y, and u + v.
(c) the logical vectors u /\ v, u V v, (u =/-= v), and (u = v).
(d) the reductions + lx, x IY, /\ lu, and V Iv.
(e) the base two values of u and of v, that is, +/(x x u), and +/(x x v).
(f) the rotated vectors 2 t d, 4 i s, and iy.
(g) the unit vector 1(5) in a I-origin system, and 3(5) in a O-origin system.
(h) the infixes (a 5 (7) /\ w 5 (7)) and 2 t a 3 (7).

Exercises
1.2 Show that
(a) x /L l(n) = 11! (Include the case n

63

0.)

(b) + /Lj(n) = n(n + 2j - 1) -:- 2.

(c) xj(k i x) = x/x.
(d) (k i x) + (k i y) = k i (x + y).

1.3 Write detailed (i.e., conlponent-by-conlponent) progran1s for the following

operations. Include tests for conlpatibility of the operands.
(a) w ~- U 1\ v.
(g) u ~- a)(k).
(b) W ~- V V V.
(h) u -(-- i 1 a)(k).
(c) b ~ - u/a.
(i) c ~- \a, u, b "
(d) B +-- u/A.
(j) c -(- la, u, b/.
(e) B ~- u/vj I A.
(k) c ~- u,a.
(f) x

+--

(x > O)/x.

1.4

Establish the identities

(a) la, u, b/ = \u/a, u, u/b;.
(b) \a, u, b, = /u\a, U, u'\b/.
1.5 The classic "rings-o-seven" puzzle can be posed as follows: an ordered
collection of n rings is to be placed on (renl0ved fronl) a bar under the following
constraints:
(i) ring n nlay be placed on or renloved at \vill.
(ii) ring k may be placed on or removed only if ring (k + 1) is on and all
succeeding rings are off.
The state of the rings can be described by a logical vector u, with Uk = 1 if ring k
is on. Write programs on U which describe the removal of the rings beginning
with
(a) U = [The successive values of u represent a r~flected Gray code; see
Phister (1958).]
(b) u arbitrary.

1.6 The ordered array of variables used to represent a variable x in S0111e coding
system 111ay be considered as a vector representation of ::r, denoted by p(~r). In the
8421 code for decimal digits, for exanlple, p(O) = (0,0,0,0), pel) = (0,0,0, 1),
and, in general, p(x) is defined by the relation + /[w x p(x)] = x, where w = (8,4,
2, 1). For each of the following coding systems, (see Richards, pp. 183-184 for
definitions), write a concise expression for p(::c):
(a) the excess-three code for decimal digits.
(b) any chosen two-out-of-five code.
(c) any chosen biquinary code.
(d) the semaphore code for the alphabet (see any good dictionary). Denote
each code by a two-component vector p(x) S; lO(8). Use a l:t', where
a = (a, b, c, ... , z).
1.7 Let X be a square sparse matrix represented by the logical nlatrix V =
(X
0) and either or both of the vectors r = V/X, and c = VIIX. Write pro
grams to determine the product Y = X ~- X, using the argunlents
(a) r, c, and U.
(b) rand V.
(c) c and U.

The language

64
1.8

Prove that

(a)

rrl

(b)

lla --:-

-l-.TJ.

bJ --:- cJ

la --:-

bcJ for all positive integers a, b, and c.

1.9 Let r = E/A, and c = EllA be the row list and colunln list, respectively, of
the nlatrix A, and let r ll , A/, and C k be corresponding elenlents of the three
representations of A.Deternline:
(a) h as a function of k, v(A), and fleA).
(b) k as a function of h, v(A), and fleA).
(c) the pernlutation vector h such that c = h r.

1.10

Show that

/ii (Use De Morgan's law for two variables and induction.)

(a)
/u
(b) ;/ /u

=/u
/u

210 + /u (Use
210 +/u.

=/ii.

(c)

(d)

(e) U'~ v = (2e)

10 (U

induction.)

~- v).

(f) U c~\ V = U v V.
(g) (t Xu) !\ (v Xw) = (t

Aw)

/\ (v

7\

u).

(a) Show that +/x = +/(u/x) + +/(u/x). (Include the case u = 0.)
(b) What properties are required of an operator 0 that it satisfy the relation
established for + in part (a)?
1.12 Show that
(a) X x Y = (ii/X)
(iiIIY) + (u/X) ~ (uIIY).
1.11

(b) u/(X

(c) ull(X

(d) (u

Y) = X ;, (u/Y).
Y) = (uIIX) ~ Y.
v)/a = (u/v)/(u/a).

Use the result of Exercise 1.Il(b) to extend the results of Exercise I. 12(a-c)
to logical operators.
1.14 Write progranls to determine:
(a) the value of the polynomial x at the point a, that is, to evaluate (I/e) __ L x
for.ll = a. Use no nlore than vex) multiplications.
(b) the derivative of the polynomial x, that is, the vector z such that
1.13

(IJe) ~ Z

=-

dy

((I/e)

x), and l'(Z)

lJ(X).

(c) the integral z of the polynonlial x satisfying the boundary condition

(d) the quotient

(ae) J_ Z = b.
q and remainder r obtained in dividing the polynonlial n by

the polynonlial d, for l'(d) ~ v(n}.

(e) the value of the polynonlial n at the point a by using part (d) with d

(I, -a).

d
(f) the value of - ((ye)
dy

_L n)

at the point a by t\\'O applications of part (e).

(g) an approxinlate real root of the equation (ye) 1_ x = 0, using parts (e) and
(f) and the Newton-Raphson formula [Kunz (1957)].

Exercises
1.15
Write
(a)
(b)

65

Let the conlponents of the vector r be the real roots of a polynolllial x.

a progralll to
deterllline the synlnletric functions of r. [Dickson (1939), Ch. X.]
deternline x as a function of r.

1.16 Write a progralll to deternline the polynolllial x consisting of the fIrst

terms of the exponential series 1 +,1/ + y2/2! -+- ....

11

1.17 Write a progranl to deternline the nloduli of all roots of the polynonlial x,
using the Graeffe method [Kunz (1957)]. Assunle that operations for the logar
ithm and exponential functions are available as subroutines.
1.18 List all the I-origin perlllutation vectors of dinlension four which are self
lnverse.
1.19 Using I-origin indexing, write progranls to derive
(a) the pernlutation k which is inverse to the pernlutation j.
(b) a pernlutation j which transfornls a given logical vector u to a prefix vector.
1.20 A square logical lllatrix U such that +/U = +// U = is sonletinlCs called
a per/11utation nlatrix, since prenlultiplication of a nUlnerical vector x deterlllines
a pernlutation of x. Write programs to deternline
(a) the permutation matrix U corresponding to the I-origin pernlutation vector
k, that is, deternline U such that U
x = k 1 x.
(b) the pernlutation k corresponding to a given pernlutation 11latrix U.
(c) the pernlutation V which is inverse to the pernlutation U.
1.21 Let p be the vector representation of a pernlutation and let c be the
standard representation in ternlS of disjoint cycles, including all cycles of one
[Jacobson (1951), p. 34.] Each cycle of c is enclosed in square brackets, each
half-bracket being considered as a component of c. For exanlple, if c = (L I, 3,
5,], L 2,4,], L 6,]), then p = (3,4,5,2, 1,6), v(c) = 12, and v(P) = 6, and, in
general, v( c) = v(P) + 2k where k is the nunlber of disjoint cycles in p. The
usual elision of cycles of one would give c = (L 1, 3, 5, ], L 2,4, ]), but this deter
Dlines a unique correspondent p only if the dimension of p is otherwise specified,
and inclusion of all cycles of one will therefore be aSSUDled. If each infix of
numerical cOlllponents in c is preceded by a left bracket and followed by a right
bracket, and if c determines a legitimate pernlutation vector p, then c is said to be
well./()rnlcd.
(a) Write a progranl to deternline p as a function of a well fornled pernlutation
c. Include determination of the dimension of p.
(b) Modify the program of part (a) to incorporate checks on the well fornlation
of c. If c is ill formed, the vector p is to be defined as the literal "ill
formed."
(c) Modify part (b) to process a sequence of vectors cl, c 2 , ,each pair being
separated by a single null element, and the end of the sequence being
indicated by a pair of successive null elenlents, i.e., to process z =
C,.
c1
(0)
c2
(0,0). Include checks on the well fornlation
of each pernlutation.
(d) Write a progranl to deternline the parity [Jacobson (1951), p. 36] of a
pernlutation vector p.

66
1.22

The language
Write detailed progranls for the following processes:

(a) k -(- 01/x

JI

(b) y -(~- nl
x
(c) V +- u r x

(i) nl-(--bl o a
(j) M -(- B ~o a
(k) u -(- b(l

(d) V-(ur X
(e) v -(- 'l..lu
(f) V-(wjjU
(g) v -( alb
(h) V -(- TjjB

c -(- b n a
(nl) c -(-- b L\, a
(n) c -(- b u a
(0) C -(- a (:) b
(I)

1.23 (a) Copy onto file (1)1 2 successive groups of itenls fronl the row of flies (1)1
in cyclic order, onlitting any exhausted files. The end of each group is
denlarked by a partition A2 , and the end of each file by a partition Aa .
(b) A file which is always recorded in the forward direction and read in the
backward direction functions as a stack. Using flle (1)2 2 as a stack,
nlodify the progranl of part (a) so as to reverse (in the output tile (1\2)
the order of the itenls within each group.
1.24 The acconlpanying node vector n and connection nlatrix C together specify
a directed graph (C/ = 1 indicates a branch fronl node i to node j) which is, in
fact, a tree.
n = (a, h, C, d, c, ,f, g)

C=

0
0

(a) Draw one of the possible ordered trees represented by nand C.

(b) For the tree T of part (a) show the full left list [T.
(c) Show the full right list ]T.

1.25

Write progranls which include tests on conlpatibility and which deternline

L = [T fronl R = ]T
(b) S = ](u/T) fronl fJ.(T), ]T, and u
(c) M = [(ujjT) fronl L = [T and u
(d) M = [(k II T) fronl L = [T and k
(a)

1.26 (a) Give an exalnple of a well fonned right list which denl0nstrates that a
prefix of a right list is not, in general, well fornled.
(b) Show clearly where the argunlent used in establishing the well fornla
tion of any suffix of a well fornled list breaks down when applied to a
prefix.

1.27 Give fornlal proofs for the facts that

(a) a left list groups nodes by subtrees.
(b) a right list groups nodes by levels.
(c) ~l(T) = lo(ll) - + Ill, where II is the degree vector of T.
1.28 Write progranls to deternline ~(T) as a function of
(a) the left list degree vector of T.
(b) the right list degree vector of T.
1.29

Trace Progranls 1.20 and 1.21 for the tree of Exercise 1.24.

1.30

Show that for a honlogeneous tree H, ,1f(H)

Y, where

y=

v(H).

1.31 If H is honlogeneous, v(H) = (3,2,3,4), and i = (1,0,2), deternline, in a

O-origin systenl
(a) the left list index I(i).
(b) the right list index rei).
(c) the index j of the nodc whose left list index is 27.
(a) If K = LO(Il) 1 ((11) ~ LO(I1)), show that K + K = n(E - n.
(b) If Y is any pernlutation of x and l'(x) = 11, show that x
K /', x
Y
K;~~ y.
1.33 Using the Euclidean algorithm, write progranls to deternline:
(a) d as the greatest conlnlon divisor of positive integers ,I' and /f.
(b) II as the g.c.d. of x and y where ct, x, and y represent polynonlials in ,:
(e.g., (: )
x).
1.32

1.34 To assure uniqueness, the nunlber of different digits (synlbols) used in a

base h nunlber systenl nlust not exceed h. The linlitation to the particular range
o ( l i ' h is, however, not essential. For exanlple, a base three systcnl can be
constructed using digits -1, 0, and 1, for which it is convenient to adopt the
synlbols -,0, and +, respectively. The positive nunlbers beginning at zero are
then represented by the sequence 0, +, + -, +0, + +, + - -, + -0, + - +,
+0 -, +00, etc. The negative nunlbers beginning at zero are 0, -, - +, -0,
- -, - + +, - +0, - + -, -0+, -00, etc.
(a) Construct addition and nlultiplication tables for this nunlber systenl and
calculate the sunl and the product of the nunlbers 0 - and - -. Usc the
decinlal systenl to check all results.
(b) Negative nunlbers are represented in this systenl without the attachnlcnt
of a special sign position. Spccial rules regarding sign are therefore
banished except that it is nece~sary to fornlulate a rule for changing the
sign of a nunlber, i.e., to nlultiply by nlinus one. Fornlulate such a
rule.
1.35 For any integer 11, let.('~ = 21n n, 1'3 = 31011'.('5 = sin n, and 1'7 = 71n 11.
As shown by Garner (1959), the ordered array(.('~, 1':3' .1'5' 1'7) provides a representa
tion of the integer n in a so-called residue nunlber systenl.
(a) Write the residue representations of the flrst ten nonnegative integers.
(b) For integers 11 in the range 0
n
(2 x 3 x 5 x 7) show:
(1) that the representation is unique.
(2) that an addition algorithnl nlay be defined which treats the several
<'

68

The language
colunlns independently, i.e., there are no carries. (The sunlS must
also lie within the specified range.)
(c) Discuss the choice of moduli for extending the range of the representa
tion.
(d) Show that the algorithm derived in part (b) is valid for all positive and
negative integers in the range -a12 :::;: n < al2 for a = 2 x 3 x 5 x 7.
(e) Derive an algorithm for obtaining -n fronl n.
(f) Derive an algorithm for multiplication.
(g) The sign of the number (i.e., its relation to zero) is not displayed
directly by this representation. Convince yourself that its deternli
nation is nontrivial.

1.36 Let x, y, and z be the positional representations of the nunlbers .r, .'I, and z
respectively. Using the floor and residue operations, write progranls to deter
mine z as a function of x andy, where z = J~ + y and the representation in use is
(a) base b.
(b) mixed base b.
(c) the +, -,0 base three system (of Exercise 1.34).
(d) the residue nunlber system (of Exercise 1.35).
1.37 Write programs for the multiplication z = x
Exercise 1.36.

X y

for each of the cases of

1.38 Write programs to convert in each direction between the following pairs of
number systems:
(a) base b l and base b 2
(b) base b l and base b 2
(c) base three and the +. -, 0 base three of Exercise 1.34.
(d) residue and base b (Exercise 1.35).
1.39

(a) Show that the superdiagonal matrices satisfy j/ 7,'/ = U ! 0/.

(b) A matrix of the form J = (xl + II) is called a Jordan box. Write the
expansion of the nth power of J.
(c) Show that X
y = Xl ~ y l + X 2 ~ y2 + ... + XI'(X) ~ yl'(X).
(d) Determine an explicit solution to the set of linear equations A .~ x = y,
where ulx = a and vlY = b are known and where +Iu + +Iv =
v(A) = II(A). State the conditions for the existence of a unique solu
tion.

1.40 Any nonsingular matrix A can be reduced to the identity / by a sequence of

row operations of the form Ai +-- xAi + yAi, or Ai +-----). Aj. The process which
accomplishes this (using row operations only) by reducing successive column
vectors to the successive unit vectors is called Jordan or c0l11plete elilnination.
If the same sequence of row operations is executed upon the identity nlatrix, it
A = I. The inverse of A can
will be transfornled to the matrix B such that B
therefore be obtained by performing Jordan elinlination on the matrix M =
A
/ so as to reduce the first l{A) colunlns to the identity. The last v(A)
columns are then the inverse of A.

(a) Write a program to detennine the inverse of A by Jordan elimination.

Exercises

69

ei

(b) The sequence of operations which reduce the ith colunln of A to is callec
the ith step of the process, and the ith diagonal elenlent at the beginning of
the ith step is called the ith pivot elelnent. Modify the progranl of part (a)
so that each step is preceded by a colunln pernlutation which yields the
largest (in absolute value) pivot element possible. This nlodification tends
to reduce the accunlulation of round-off errors.
(c) In the Jordan elinlination of part (a), it is unnecessary to store the identity
matrix explicitly, and, since the ith colunln is first affected at the ith step,
only one new colunln need be brought in at each step. Moreover, the ith
colunln of A may be discarded after its reduction to e i on the ith step, and
it is therefore necessary to store only a square nlatrix at all ti!nes. Show
that by shifting all columns to the left and by nloving the rows upward
cyclically, a very unifornl process results, with the pivot elenlent in the
leading position at every step [Iverson (1954) or Rutishauser (1959)].
Write a progranl for the process.
(d) Modify part (c) to allow the choice of pivot elenlents as in part (b). The
effects of the pernlutation on the not explicitly recorded identity cannot be
indicated directly, but the performance of the sanle set of pernlutations in
reverse order upon the rows of the resulting inverse produces the sanle
result. Verify this and prograrTI the process.

1.41 (a) Show that a group [Jacobson (1951)] can be represented by a square
matrix M such that each row and each colunln is a pernlutation vector.
(b) Show that M i = M; = II for sonle i.
(c) What are the necessary and sufficient conditions that the group repre
sented by M be Abelian?
(d) Write a program to deternline all cyclic SUbgroups of a group repre
sented by M.
1.42 If U is a logical nlatrix whose rows are each nonzero, nlutually disjoint, and
collectively exhaustive (that is, (+/U
e) = e, and +//U = e), then U defines
an 111-way partition of n, where 111 = li(U), and n = v(U). The partition is nlore
commonly represented by the vector p = +/U [Riordan (1958), p. 107]. Clearly
+ /p = n. Write a program to generate
(a) all partitions U of a given integer n.
(b) all distinct partitions of n, where U and V are considered equivalent if
p = +/U is a permutation of q = +/V.
1.43 Let x be a .'pace vector (i.e., of dinlension three), and let R(x) be the square
nlatrix l j (e !\ x). Sho\v that
(a) +/R(x x y) = (x x y) X e
(x
y) = x x y
(c) (+ / R ( x x y x (w x z) = (x ~- y) x (w ~- z)
(d) (x ~- y) x (x x y) = (x x y) x (x x y) + 2(lx

(b)

e;

x jy)

~ (lx

x jy).

1.44 Let x . y = (j x x ly) - (lx x jy) be the vector product of x and y for
vectors of dinlension three. Show that
(a) this agrees with the usual definition [Margenau and Murphy (1943)].
(b) x' y = -(y' x)

70

The language

(c) x y is perpendicular to x, that is, x

(x . y)
t x = 21 x for a vector of dinlension three.)
---

1.45

Let [x]

\/ (x

t x) be the length of. x, let x J! y

the angle between x andy, and let x a y =

Use the results of Exercises 1.43 and 1.44
(a) [x y] = [x] x [y] x (x a y). Note
parallelogram defined by x and y.
(b) (x . y) . z = (x
z) x y - (y
z)
(c) (x y) ; z = x

t
t (y . z).

o.

(Use the fact that

x +y
= [ ] x [

x y]

be the cosine of

'/1 - (x y y)2 be the sine of the angle.

to show that for space vectors
that [x . y] is the area enclosed by the
x x

chapter

MICROPROGRAMMING

The algorithms to be executed by an automatic computer must be

described in the restricted set of operations .(called instructions or C0111
mands) provided in a given computer, and an algorithm so described is
called a C0111puter progran1. Since computer instructions are relatively
complex, they may be described in turn by 111icroprogra111S employing
more elementary operations.
Microprograms may be used to define a computer instruction set for the
programmer, to define the detailed algorithms by \vhich the computer
circuits produce the operations of the instruction set, or for a variety of
other purposes. In the design and development ofa computer, for example,
it is important to maintain precise and complete communication between
the computer programmer, the computer (or system) designer, and the
logical circuit (or hardware) designer. The systeITI designer will, in fact,
ordinarily begin with a description at the programmer's level and proceed
through increasing detail to the hardware designer's level. Meanwhile, the
programmers concerned with evaluating potential performance and with
developing systems of nletaprograms (so-called autoITIatic programming
systems) should be enabled to follow and to influence the evolving de
sign.
The use of microprogramnling will be illustrated by a description of the
IBM 7090 computer (to be called the 7090) at a level approximately suited
to the programmer and the system designer. The final section treats
some problems in the extension to the hardware design level.
The programs together with the lists of operands constitute a self
contained description of the 7090 which, to readers already familiar with
computer organization and with the relevant sections (1.1 to 1.11 and 1.14)
of Chapter 1, should prove readable without reference to the text. The
text serves only to elucidate the microprograms and does not treat all of
the instructions described by them. Tables 2.1, 2.13, and 2.14 sumnlarize
the dimensions, format, and significance of the various operands and
should be consulted as each is first encountered. O-origin indexing will be
used throughout.
71

72

2.1

Microprogranlrning

2.1

INSTRUCTION PREPARATION

The operation of an automatic digital computer splits naturally into two

phases which normally alternate: the instruction fetch and preparation and
the instruction execution. The former involves the selection from sonle
information store C,ne/nory) of the next instruction to be executed, its
transfer to one or more control registers, and perhaps some modification
of the instruction introduced into the control registers through so-called
indexing, indirect addressing, or relocation. The execution phase begins
with the decoding of the operation code segment of the instruction in the
control registers to select the particular execution microprogranl to be
enlployed, and contin ues through the execution of the selected micro
progran1 upon variables in certain central registers and in certain menlory
registers determined by the address portion or portions of the control
registers.
The n1ain memory of the 7090 will be denoted by a logical matrix M of
dimension 215 x 36. Selection from M is limited to the selection of a row
Mi; each such row is called a It'ord, and Mi is called It'ord i or register i.
Dimension
Memory

2 15 x 36
3 x 15

I ndex accumulators

Sequence vector

s (instruction counter)

15

Command vector

36

u (s, q,p, 1,2, ... ,35)

Upper accumulator
Lower accumulator (Quotient register)

I (s, 1, 2, ... , 35)

Upper accumulator overflow

II

Lower accumulator overflow

Trapping n10de indicator

o.

f { 1:
2:

] nstruction fetch mode

norn1al
skip channel trap
skip trap and fetch phase

o.

no indexing
1: normal indexing (15 bit)
restricted indexing (9 bit)

I ndexing class

kO(c)

{ 2:

Indirect addressing class

k 1(

fo:

Console start signal (run)

Binary representation of z

Table 2.1

C)

no indirect addressing
ll: indirect addressing

r
p( z)

Central COlnputer operands

38

36

2.1

Instruction preparation

73

Each instruction is a full word selected from M, and the seq uence in
which instructions are selected is controlled by a control register called the
instruction counter or sequence register. This register represents a logical
vector s of dimension 15 whose base two value determines the word i to be
selected in the next instruction phase. The quantity*Ls is increnlented
after each instruction fetch and therefore selects instruction words in
natural ascending sequence. The value of s may, however, be respecified
by the execution of certain branch instructions.

+-

Program 2.2

M ls

Basic instruction fetch

The current instruction will be denoted by c. It is stored in a 36-bit

conunand register.
The basic instruction fetch involves only the variables M, s, and c, and
is described by Program 2.2. The second step shows that the increnlenta
tion of _Ls is reduced modulo 2 15 and that the selection of instructions
from the 2 15 word memory is therefore cyclic.
Additive indexing
It is convenient to the programlTIer to be able to add (or subtract) an
jndex quantity i to (fronl) the address portion of an instruction in the
comnland register c before its execution. This quantity is represented in
base two by a logical vector a and is stored in a special index register. In
the 7090, the data address portion of c is the fifteen-bit suffix w 1:>/ c and the
indexing is su btractive:
_lw};>/c

+--

21;> I CLw 1:>/c -

J_a).

The reduction lTIodulo 2 1:> again indicates cyclic treatnlent of addresses.

The 7090 contains three index registers or index accul11ulators which 11lay
be used independently or jointly. They will be denoted by the index
In:ltri I of dimension 3 x 15. One or nlore (or none) of the index regis
ters Ij are selected according to the value of the vector i == (18 t 0.:3)/ C, the
three-bit index tag portion of the conl11land, as follows:
~_ w 1.)/ c +-- 2 1;) I C~ _W I :> / C

J(( I8 t

3
0. ) /

c)

I) ).

The address in the conlmand register is clearly decrelnented by the base

* Since nurnber bases other than two will be used but rarely in the present chapter,
the elided form -.L x will be used instead of (2e) _~ x.

74

2.1

M icroprogranll11ing

two value of the vector obtained by oring together the selected rows of [.
The oring of the index accumulators permits simple circuits for their
selection. It is, however, of questionable value to the programnler, and
the 7090 index registers are normally used individually.
Indirect addressing
It is often convenient (as in the execution of a pernlutation) for a
programmer to specify a data address indirectly, i.e., to specify the address
of the word containing the desired data address. Such indirect addressing
could proceed through several levels, but in the 7090 it is limited to one.

C 12 /\ C 13

wI8jc -(_ w I8 jM Jw 15 jc

Program 2.3

I ndirect addressing

Only the last half of c is respecified by the corresponding portion of the

selected word, as described by Program 2.3. The occurrence of indirect
addressing is determined by components C 12 and C 13 of the operation
code.
Dynamic relocation
The correct execution of any computer program requires that each
instruction and each operand be stored in the register assigned in the
construction of the program. A program can, however, be relocated by an
integral amount n if each word originally assigned to address j is assigned
to address j + 11, and if each address in the program is also incremented by
11. The incrementation of program addresses can be performed explicitly
by an assembler or other metaprogram, or it can be performed dynanlically
by an additive index register containing the number 11. An index register
employed exclusively for this purpose is called a base address register.
More generally, the provision of a table of base addresses pernlits
independent dynamic relocation of different blocks of a program, where
each block is confined to a set of successive registers. This is equivalent to
one-level indirect addressing in which a certain portion of the address
(e.g., vj(W 15 j c)) selects from memory one of a table of base addresses to
respecify the same portion vj(wl.~jc) thus:
V;'(W15jC)

+---

Vj(W15jM~Vj(w15jC)).

If, for exalnple, v = a 7 , and the format is otherwise as in the 7090, then
columns 21-27 of registers 0 to (2 7 - 1) provide the base addresses for

2.1
successive blocks of 2 8 registers each.
relocation.

Instruction preparation

75

The 7090 provides no dyncunic

Branching, interruption, and trapping

The nornlal sequence of instructions (fetched from successive ll1enlory
registers) can be interrupted by respecifying the sequence register s. Such
respecification is perfornled in the execution phase of certain instructions,
prinlarily those called transjers, and skips. The sinlplest branch is the
TRA * (transfer), whose execution effects the following operation

The normal sequence can also be broken by the insertion of an instruc

tion in the cOlllllland register without disturbing the sequence register.
Unless the inserted instruction is itself a branch, the nornlal sequence is
resulned immediately.
If just before a branch (or insertion) the present value of s is stored in
sonle chosen memory register i, then the data in register i can be used in a
subsequent branch to reestablish the original sequence at the point reached
before the first branch. The storage of s and immediate branch are jointly
called an interruption. An interruption which is performed autonlatically
upon the occurrence of certain special conditions is called a trap. A trap
provides a convenient device for inserting in the nornlal progranl sequence
a subprograln delnanded by the occurrence of the special. conditions.
In the 7090, the so-called channel trap is controlled by an 8 x 3 logical
matrix T whose elements are determined by three different conditions
existing in each of the 8 input-output channels of the cOlnputer. A corre
sponding enable n1atrix E (also 8 x 3) and an enahle toggle e deternline
which elelnents of T are effective.
The channel trap is effected in the first phase of the instruction fetch
(Program 2.4) as described by steps 2-8. If the matrix eE /\ T is zero, the
branch on step 2 skips the trap operation and begins the nornlal fetch on
step 9. If not, step 3 determines j as the index of the first nonzero row,
step 4 stores s in a memory register detennined by j, and step 5 stores the
nonzero row (which indicates the particular condition causing the inter
ruption) in another portion of the same register. Step 6 resets the indi
cators which occasioned the trap. Step 7 resets the enable toggle e and
hence (as is clear from step 2) prevents the occurrence of further traps
until e is again set to one by the execution of a special enahle instruction
ENS. (The reset of e to zero prevents the uncontrolled interruption of
interruptions.) Step 8 performs the actual insertion by transferring to the

* 7090 instructions will be referred to by the mnemonic codes used in the IBM Alan[{a!
(1960).

76

Microprogramming

2.1

command register the content of a second memory register determined by

j. The seq uence register is undisturbed.

The 7090 can also be operated in a special trap nl0de which effectively
converts all transfer instructions (but not skip instructions) into inter
ruptions. Discussion of this type of trapping will be deferred since it is not
relevant to the instruction fetch phase.
Complete instruction fetch
The complete instruction fetch comprises three phases: channel
trapping, the fetch proper, and the instruction preparation by indirect
addressing and j ndexing. They are described by steps 2-8, 9-10, and 11-18,
respectively, of Program 2.4.
Certain of the three phases may be skipped according to the setting of
the fetch mode indicator f In the normal case (j' == 0) none are skipped.
Iff == I, the trap phase only is skipped. This case occurs after execution
of an instruction such as the RDS (read select), which must be followed by
a certain auxiliary instruction within a fixed time limit. If f == 2 (a case
which occurs only after execution of the XEC (execute) instruction), the
trap and the fetch proper are both skipped, and the comlnand already in c
is merely prepared by indexing and indirect addressing. In every case, j'
is reset to its normal zero value by step 11.
Not all instructions are subject to indexing. The indexability of a
command c is determined by a class function kO( c), which assumes the values
0, 1, or 2, according to whether c is subject to no indexing, nonnal indexing
(affecting all fifteen bits of the address), or restricted indexing (affecting
the last nine bits of the address), respectively. This behavior is determined
by the branch on step 13.
A second class function k 1 ( c) determines whether the instruction c is of
a type subject to indirect addressing. Actual indirect addressing of any
particular instruction of the appropriate type is initiated by the configura
tion C 12 == I and C 13 == 1. The function k 1 ( c) is itself independent of C 12
and C 13 .
The class functions kO( c) and k 1( c) will themselves be specified by
prefacing the mnemonic code of each instruction described by a pair of
digits. Thus,
11 CLA
indicates that clear and add is subject to both indexing and indirect
addressing, and
10 CHS
indicates that change sign is subject to indexing but not to indirect ad
dressing.

Instruction execution

2.2

>

.l:

1
: 0

eE /\ T

.i

. 5

(15

+ jaj

+-

/\ j(eE /\ T)

lO-+2j -<-- W

c -<-

(1.3)j M

15(36)\s
10 --j2j ~-

Mll+2j

-<-M~s

+ _ s)

]1

f'

~O

12

-<-

13

kO(e) : 1

14

~w15je -<- 215

15

Jw 9 je

16

a : 0

1R

(/

i = (1 8

(eE /\ T)j

J.s -<-2151 (I

10

77

-<--

k 1( e) /\ e 12 /\ e 13

>

I (J_w15 je

-<- 29 (J w 9 jc 1

L(i

X I))

J (i ::: w 9 jl)

t (1.3) Ic
Program 2.4

Complete instruction fetch

The phases of instruction preparation are performed in the following

order:
indexing (if indicated); indirect addressing (if indicated).
Moreover, if indirect addressing is performed, the new address is itself
re-indexed (if indicated). As shown by steps 12, 16, and 18, the indirect
addressing is limited to a single level.

2.2

INSTRUCTION EXECUTION

The execution phase begins with the "decoding" of the operation part
of the command c to select the appropriate microprogram to be executed.

78

A;ficroprogramn1ing

2.2

Except for the format of the operation code* (which in the comn10n case
occupies the prefix a 12/ c) the details of the decoding are, however, of no
interest to the programmer, and attention will be confined to the execution
microprogran1s. These may be grouped into a small number of falnilies;
for the 7090 they are load and store, branch, logical, arithnletic, shift,
concert, and input-output.
Certain of the arguments and results of the computer instructions are
represented by three central data registers to be denoted by I, u, and d.
The registers u and I serve as accumulators in the addition and other
arithmetic operations, and, since u and I jointly represent double precision
numbers (i.e., carries are in some operations propagated between the high
order end of I and the low order end of u), they will be called the upper and
lower accumulator, respectively. Since I receives the multiplier in a
multiplication and the quotient in division, it is called (in the 7090 manual)
the Multiplier-Quotient or MQ register, and the letter Q occurs in the
mnemonic code for instructions affecting it.
Signed numeric quantities are represented in base two with the sign in the
first component, i.e., register i represents the quantity y = (1 - 2 MOi) x
(~al/Mi). The lower accumulator I is, like each memory register, of
dimension 36, and the sign of a numeric quantity is represented by 10 ,
The upper accumulator is of dimension 38 and represents the number
(1 - 2u o) x (~a.l/U). The two extra components u 1 and u 2 are called
orerfioH' positions and are excluded from normal transfers of data fron1 u
to the memory. The component u 2 (called the p-bit) is, however, included
instead of the sign bit U o in certain logical instructions. The component
u 1 (called the q-bit) is made accessible only by certain shift operations.
The register d (distributor) serves only as intermediary in transfers
between main memory and the central data registers u and I and is not
accessible to the programmer.

Load and store

In each member of the family of basic load and store instructions
(Program 2.5), the memory word involved is selected by W 15/ c, the address
portion of the instruction. The instruction STA stores only the address
part of u, and STD stores the decrement part, so called because it is used
in certain instructions to specify the amount of decrement to be applied to
an index register. The STP stores the p-bit and the first two digits of the
magnitude part of u; that is, the three-bit prefix of the logical part of u
which enters into logical operations. The STO instruction stores the

* The operation code representing instruction x is a logical vector to be denoted by

p(x). Thus p(CLA) = (000101000000).

Instruction execution

2.2

11

Load Q

11

Store Q

11

Store left half Q

LDQ

I *- M.

STQ

M.; w 15 jc

SLQ

jc

-0(

15

I
a 18 jl

--+

jc +-- ( ~ a,2)/U

--+

a18jM' w
15

15

jc

+-

Store

11

STO

M-

Store zero

11

STZ

Store address

11

STA

W15/M.:.w15jc +-- W 15 /U

Store decrement

]I

STD

(3 ~ ( 15)/Ml.w 15 jC

Store prefix

II

STP

a 3 /M

Store tag

JI

STT

(18 ~ ( 3)/M .Lw

Store instruction
location counter

11

STL

W 15 /M! w

Store logical \vord

11

SLW

Program 2.5

79

~ w 15 jc +--

.w

15

jc +--

15

-0(

(5 ~ ( 15)/u

(2 ~ ( 3)/u
15

jc

-0(

(20 ~ ( 3)/u

jc +-- S

w 15 jc -(- (j2/ U

Load and store instructions

normal numeric part of u (that is, all but the overflow bits), whereas SLW
(store logical word) stores the p-bit instead of the sign.
The instructions which load and store the index accumulators (ProgralTI
2.6) are of four types, as indicated by the leading letter of each of the
mnemonic codes, L for load index from memory, S for store index in
memory, A for load index from the address of the command register~ and
P for place the index in the upper accumulator or the upper accumulator
in the index. The portion of nlenl0ry, command register~ or upper
accumulator involved in each of the ten instructions which specify the
index is shown in steps 1-10. The last fi ve of these differ from the corre
sponding nlembers of the first five only by complenlentation on 215~ as
shown in step 11. Since the subtraction occurring in indexing (step 14 of
Program 2.4) is reduced modulo 21.\ the effect of complenlentation is to
add rather than subtract the quantity used to load the index aCCUlTIU
lator.
Step 12 shows that the index accumulators specified are selected by the
three-bit tag vector i == (18 1a. 3 )jc and that each receives the same specifying
quantity. Since the tag vector is used to select the index registers to be
specified, it cannot also be used to specify indexing of the instruction
itself, and, consequently, none of the load and store index instructions are
indexable. Neither do they permit indirect addressing.

80

2.2

M icroprogral1l111 ing

Load index from address

00

LXA

Load index from

decrement

00

LXO

2
15

Address to index true

00

AXT

+- W

Place address in index

00

PAX

+-

Place decrement in index

00

POX

Load complement of
address in index

00

LAC

Load complen1ent of
decrement in index

00

LOC

Address to index
complemented

00

AXC

Place address in index

complemented

00

PAC

Place decrement in
index complemented

00

POC

Store index in decrement

Place index in address
Place index in decrement

00
00
00

w 15 1u

(5 ~ a 15)lu

X +-

i III

00

_L x

Store index in address

/c

+- 215
+-

10

I ( - _l

x)

e( + Ii) ;, x

SXA
SXO

(3 ~ a 15 )IM ~_w15/C

PXO

u -(-tl -(-

w1f>/c --

w 15 (38)\(i
(5

12

i XI

W151M,

PXA

11

--

J3
i \ I

X 1)

~ a 15 (38))\(i

14
15

I)

]6

(i = (18 ~ a 3 )lc)

Program 2.6

Load and store index instructions

The last four steps show the storing of the index accun1ulators. The
quantity stored is the or function of the accumulators selected by the tag
(18 1 a:3)jc.
Branch instructions
The basic branch instructions are of two main types, the transfer
z.Jenoted by a leading Tin the mnenlonic code) and the skip. The behavior
of the skip instructions is shown in steps 1-10 of Progranl 2.7, and is
typified by the PST (p-bit test) of steps 1 and 10. If the p-bit of the upper

2.2

81

Instruction execution

accumulator is not zero, the sequence vector is increnlented so as to skip

the next instruction in the sequence; if the p-bit is zero, the instruction has
no effect. The various skip instructions differ in the particular tests nlade,
and the last two (CAS and LAS) differ also in providing three alternatives,

p-bit test

10 PBT

Low order bit test

10 LBT

s -(-- U 37

Storage zero test

11 ZET

S +-

Storage nonzero
test

11 NZT

S -(-

E: \~:c M _Lw 15 jc

Compare accum
ulator with storage

11 CAS

6c

2
M _~w15jc

4
5
6

Logical compare
accumulator with
storage

11 LAS

~-

_Ls

+-

+ (d
l(s + J_S)
0)

(d
2

1ii

0)

<}

10

Execute

It XEC

c +-1\"11- w 15 jc

11

f~-

12

Store location and

trap

00 STR

wF)jM+-s

.-is -(-

13

14

Leave
mode

trapping

10 LTM

t -(- 0

----+

15

Enter
mode

trapping

10 ETM

~-

--+

16

11 TTR

+-wl.'jc

Trap transfer
Console
reset

clear

or

Program 2.7

(1, f,

[(, I, h)

--+ 17
+-

~~18

Skip type and other special branches

skipping 2, 1, or 0 instructions according to whether the quantities COIll

pared stand in the relation <, =, or >, respectively.
When operating in the 110ntrapping 11lode (t = 0), the essential operation
of the transfer type of branch (Program 2.8) is the (conditional) respecifl
cation of the sequence vector s by the address portion of c. The HTR
(halt and transfer) also suspends operation of the conlputer until a run
signal is received from the console (steps 27-29).
J n the trapping 11lode, all transfer operations are converted to inter
ruptions; the sequence vector is first stored in register zero, and a

82

2.2

Microprogranlnling

(~(i

Transfer on index high

00 TXH

Transfer on index low

or equal

00 TXL

b -(-- ( ~ (i

XI)
XI)

Transfer on index

00 TIX

b ..;- (~(i

XI)

+-

Y -(- _L(i

XI)

J_ (3

t o.};)jc)
t o. )jC)

-i (3

~(3

- .l(3

15

o. 1 ;)jc)

t o. ;)jc

.Li/II -(- (215IY)( +ji)

Transfer on no index
Transfer with
incremented

index

XI) ~ ~ (3 t o. ;)jc)
.If -(- I (i XI) + ~ (3 t o. ;)jc
_L ilII -(.- (2 :Jly)( + ji)
J. il II -(- (2151( - _Ls( + ji)

b -(- (~(i

00 TNX
00 TXI

=1=

Transfer and set index

00 TSX

Transfer
Halt and transfer

II TRA)
]1 HTR

h -(-- 1

Transfer on zero

11 TZE

h +-(a. 1 ju)

LE

11

Transfer on nonzero

]1 TNZ

h ~- (a.Iju)

'!;:

12

10

13

Transfer on plus

11 TPL

Transfer on minus

11 TMI

14

15

Transfer on overflow

11 TOY

(h, u) -(- (u, 0)

Transfer on no over
flow

11 TNO

(b, u) -(-

Transfer on Q plus

11 TQP

Transfer on Quotient
overflow

11 TQO

(b, !) -(- (I, 0)

Transfer on low Q

11 TLQ

b..;-(l - 2IoLL~ljl< (l - 2Uo)~a1jU)

(ii,

17

b -(- b V I ~

-=1=

=1=

18

0.

1\ (u ~

19

20

t: ]

21

b : 0

22

S -(-W 15 jc

23

w 15 jMO -(- s

24

b : 0

25

-is +-

Program 2.8

16

0)

26

o. jc : p(HTR)

27

r -(- 0

28

r : 0

29

12

Basic transfer type branches

2.2

Instruction execution

83

(conditional) branch is made to register one. This behavior is useful in

trace programs. The conditions for the various transfers are indicated
in the setting of the logical variable b. Special indicators (such as the
overflow toggle u) are reset by the transfer instructions which they control.
Transfers based on the condition of the index accumulators (steps 1-9)
are combined with modification of the index registers. The quantity from
the index accumulators is again the or function of the accurTIulators
selected by (181 a.3)jc. This quantity is compared with the decrenlent part
of the command (that is, (31 0. 15 )/ c)to control the conditional branches and
is decremented or augmented by the decrement part to nlodify the selected
index accumulators.
The TSX (transfer and set index) inserts the complement (on 2 1;) of the
sequence vector into the selected index accurTIulators before effecting an
unconditional transfer. This instruction is convenient for incorporating
closed subroutines or other interruptions, since a subsequent TRA
(transfer) with a zero address and indexed by the same index register
restores the program sequence to the point of interruption.
As shown in Prograrl1 2.7, only the TTR (trap transfer) is exenlpt fronl
trapping. The trap indicator is set by the ETM (enter trap nl0de) and is
reset by the LTM (leave trap mode) as well as by a console clear or reset.
The XEC (execute) instruction performs no operation upon the central
data registers but inserts in the normal instruction sequence (without
breaking it) the instruction in the register specified by the data address
accompanying the XEC. This is effected by sinlply loading the specifIed
register into c and (by setting! == 2) skipping the trap and the fetch proper
of the instruction fetch phase.

Logical instructions
The logical operations (Progranl 2.9) concern the logical part of ll,
which differs from the numeric part by including the p-bit rather than the
sign, and hence comprises a,2/U. The first instruction of the family (ORS)
produces the logical or of the word selected fronl storage with the vector
a2/u and returns the result to the sanle location in storage. The instruction
ANS (and to storage) is similar. In the ORA (or to accumulator) the
15
result u is of dimension 38, and the second operand (that is, M w /C) is
expanded to this dimension by inserting two leading zeros before oring
it with u. The instruction ANA is similar. It is easily verified that ORA
leaves the extra bits unchanged and that ANA resets them. The ERA
(exclusive or to accumulator) is anomalous in that it resets the extra bits.
The ACL (add and carry logical) is normally used only in certain parity
check algorithms. It adds the selected word to the logical part of u,
treating both as 36-bit positive base two numbers, again adds any resulting

84

2.2

M icroprogral11111ing

15

w 15 /C

(a 2ju) V M

Or to storage

11

ORS

M_Lw

Or to accumulator

II

ORA

+-

And to storage

II

ANS

L w15(C +-

And to accumulator

11

ANA

+-

u 1\ (a2 \MLw 15 /C)

Exclusive or to accumulator

11

ERA

+-

(u

* (a2

a2ju

+-

+-

Cilju

/C

+-

u V (a2 \MI

10

COM

Cilju

Clear magnitude

10

CLM

a1ju +-0

Change sign

10

CHS

o +- U o

Set sign plus

10

SSP

o +- 0

Set sign minus

10

SSM

o +- 1

Store logical word

11

SLW

M_i_ w15 / C +-

Clear and add logical

11

CAL

Add and carry logical

1 ACL

+-

/C)

(a2 ju) 1\ M

15

w15/C

/C))

a 2ju

a 2\M_Lw 15 /c
_L(a2ju)

J_M_Lw

l-a 2ju -(- (2 36 jy)

Program 2.9

15

\Mlw

Complement magnitude

+-

+ (y

15

(c

236)

Logical operations

overflow to the low order end, and places the result (which will not exceed
1) in the logical part of u. The behavior of the remaining logical
instructions is evident from Program 2.9. As shown by the class functions
k O and k\ five of them do not permit indirect addressing. *
2~~(j

Arithmetic instructions
The description of arithmetic instructions will be illustrated by the family
of fixed pointt add instructions shown in Program 2. 10. The CLA (clear
and add) transfers the selected memory register to u, resetting the two

* Since each of these five instructions involves the accumulator only, the normal
address portion W 15 Ie does not represent an address, and its last three components are
used in representing the operation code itself. The possibility of indirect addressing
would suggest to the programmer the nonexistent possibility of specifying an arbitrary
indirect address in w 15 je.
t The 7090 incorporates three arithmetic functions: addition, multiplication, and
division, each of which nlay be performed in either a fixed (radix) point or floating point
nlode.

Instruction execution

2.2
Clear and
add

11

CLA

u +- (

f (i2)\M Jw 15 /c

Clear and
subtract

11

CLS

f ( 2)\(al

+-- (

15

:f

M _~w15/C)

85

----+

Add

11

ADD

d +- M--

Subtract

11

SUB

d +- (a l ~~ M

Add
magnitude

11

ADM

d *- (il 1\ M:

Subtract
magnitude

11

SBM

Round

11

RND

/c

3
w

15

15

/C)

/c

9
~al/u +-2 37
II
U

1(lzl)

*- U V x :j=

u 2) 1\ (U o = do))

o +- (z < 0) V (u o 1\ (z = 0))

Program 2.10

10
11
12

Add instructions

overflow positions. The CLS (clear and subtract) differs only in that the
sign is reversed in transfer.
The instructions ADD, SUB, ADM, and SBM each transfer the
selected word to d with an appropriate sign, add it to the nun1ber repre
sented by u (including the overflow positions), and place the sum reduced
modulo 237 in u. The sign of a zero result is (as indicated by step 12) the
sign of the number originally contained in u.
'The overflow indicator u is set only by a carry (but not a borrow) [ron1
u 3 to u 2 . This indicator controls, and is reset by, certain branch in
structions.
The RND (round) instruction is used to round up the n1agnitude of the
number represented jointly by the upper and lower accun1ulator by one
unit in the high order position of the lower accumulator. As shown in
Program 2.10, the content of the upper accumulator only is actually
affected.
Shift instructions
In describing computer instructions, the term left shift of a vector x by r
places refers either to the left rotation x +- r i x or to the left rotation

86

2.2

Microprogramming

combined with a reset to zero of the last r "vacated positions," that is,
x

+-

(r

x) 1\

wr

Both types of shift occur in the 7090 and concern various portions of the
entire accumulator u EB 1, as shown in Program 2.11. The portion affected
is determined by the mask vector m.

+-

X(mj(u

V (a r

mj(u EB I)

+-

(r

t (m/(u 8)

Accumulator left
shift

10 ALS

+-

(il(38)

(f)

(36)

Long left shift

10 LLS

+-

(il(38)

~E\

(il(36)

Logical left shift

10

LGL

+--

(il(38) ffi e(36)

Accumulator right
shift

10 ARS

~-

(il(38)

Long right shift

10 LRS

~-

(il(38) c) (il(36)

Logical right shift

10

+-

(il(38) (]) e(36)

LGR

m/(u 8) I)

+-

10 RQL

I+-rt l

Exchange accumu
lator and MQ

00

( ~ a2)/u ~-~ 1
(

(i2)/U

-0(-

I))) 1\ wl'

(36)

Rotate MQ

XCA

I))

(r

(m/(u CB I))) 1\ (il'

Exchange logical
accumulator and MQ 00- XCL

_lw 8 jc

Program 2.11

Shift instructions

The first three instructions are left shifts. Each sets the accumulator
overflow indicator if any nonzero bits are lost, i.e., if any of the first r
positions of the affected portion are nonzero. The next three are analogous
right shifts, which do not, however, set the overflow indicator. In the

Instruction execution

2.2

87

"long" shifts LLS and LRS, one sign position specifies the other, although
the sign positions are otherwise excluded from the shift by the mask m.
The LGR shifts all positions save the sign of u; RQL rotates MQ
without resetting any positions; and XCA, which "exchanges" the
accumulators, is effectively a rotation except that it resets the overflow
bits. The amount of shift r is in each case determined by the base two
value ofw 8 /c.
Convert instructions
Each convert instruction (Program2.12) selects a six-bit infix of one of the
accumulators, adds its base two value to a "base address" initially specified
by the address portion of the instruction, and selects the men10ry register

00
00

CAQ}
CRQ

3
d

+-

1+-6t 1

j +-j -

+- 215 / (_,_W 15 III

_L a. 6 II)

a.lOlc : p(CAQ)

a. 6 II

+-

_La.! lu
00

a. 6 Id

10

+- 2 37 /(

_~Cil/u

+ : d)

11
12

CVR
d

+-

13

<

j +-j 1
a

2 l5 j( lW 15 1d

+-

~w6Iu)

d +-Ma

Cil/u
6

+-

a. j"a lu

Program 2.12

14

15
16

(6

f (al/u)) /\

+-

(a. lCi lu) V a. 1d

6

Ci6
6

Convert instructions

17
18

88

Microprogralnlning

2.2

specified by the resulting address. Part of the selected register is used to

respecify the base address and part to modify one or other of the accumu
lators. The process is reapplied to successive six-bit infixes in cyclic order
a number of times determined by the base two value of (10 1( 8)/ c. If
c 20 = 1, the last fifteen bits of the last word selected in the operation are
transferred to index accumulator ]0.

Input-output instructions
Because the data transmission rates of input-output equipment serving
a computer are n1uch lower than those of the computer proper, computer
systems are normally designed to permit concurrent operation of the
computer and one or more input-output units. The units are therefore
more or less autonomous.
In the 7090, the autonomy of input-output equipment is carried further
by providing eight data channels which transmit data to, and are controlled
by, the computer proper, and which in turn control and accept data from
the individual input-output units. The entire input-output process there
fore con1prises three levels of microprograms: a semiautonomous input
output unit controlled by a semiautonomous data channel controlled by
the cOlnputer proper.
Attention will be restricted to the magnetic tape input-output units of
the 7090. Each unit is available to one specific data channel i (for i =
o - 7), and a particular unit can be characterized as the file $/. The unit
is completely autonomous only in certain simple operations, such as
reH'ind, write end offile, backspace, and continue to end of record. Except
for these special operations, a given data channel can control only one of
its files at a time. The eight data channels may, however, operate con
currently under control of the computer proper.
Each channel i behaves as a subcomputer with its own sequence vector
Si, command vector Ci, data register Di, and other miscellaneous operands,
as shown in Table 2.13. The instructions of the subcomputer (listed in the
matrix K) are called channel commands and differ from the instructions of
the computer proper in both format and function.
Tape Units. Each tape unit behaves as a file (Sec. 1.22); each recorded
component is an alphanumeric character represented in a seven-bit odd
parity error-detecting code, the lowest level partition Ao is represented by
the intercharacter space on the tape, the second level partition Al (called
an end of record gap) is a longer blank space on tape.
Each record gap is immediately preceded by a parity check character
which is appended to the normal data of the record to permit an even
parity "longitudinal" parity check on each of the seven-bit positions of the

Dimension
Channel data registers

8 x 36

Channel sequence vectors

8 x 15

Channel command vectors

8 x 36

Channel trap

r Toi

~I

I
.
L T 2 'l

: End of file A2
:

Parity check

Channel command

8 x 3
8 x 3

Channel trap enabled

Channel trap enabled

Tape position limits

Li: (Beginning, End)

Limit position on tape

(Determined by reflective marker)

Busy indicator

Write or read indicator

8 x 2

8
8

Tape unit index

o : Normal
Functions

read-\vrite

1 : Backspace record or write end of

file

2 : Backspace to file mark

3 : Rewind
Load channel waiting

Write record gap (AI) next

Current character

Current parity check

(reload)

8
8

Xi is the 7-bit representation

8 x 7
8 x 7

1 if character Xi is loaded

Interlock vector

Xi

Current character selector

Vi jDi is current character

8 x 36

End of file indicator

Qi : (Counter, Potential error)

8 x 2

Input-output indicator

c d

IOCD

TCH

Channel commands

r p

IORP

r t

IORT

c P IOCP
c t IOCT
s P
s t

Table 2.13

IOSP
lOST

Channel operands
89

90

2.2

M icroprogralnlning

preceding record. The check character is recorded automatically and when

read from the tape is used in the parity check but is not translnitted with
the data of the preceding record.
The third and highest level partition (called end offile) is represented by
a special recorded character A2 which has the seven-bit representation
p(A 2) = (0, 0, 0, ], ], ], 1). It is recorded together with the appropriate
check character (which, since the check is of even-parity is also A~) as a
separate record. The character A2 alone is not recognized as an end of file
partition; only the sequence AI' A2 , A2 , Al is so recognized. Tapes are
normally stopped only at a record gap so that, on restarting, the tape is
fully accelerated before the end of the record gap (and hence data) is
reached.
Character buffer
Partition buffer
Logical association (connection)
Busy indicator
Write-read status
Function status
End of file counter

Dinlension

Zl
PI

AI
B ~ 8 x

wi
F

# of units per channel

RIJ
AO intercharacter gap

File partitions

A Al inter-record gap

Table 2.14

A2 end-of-file symbol (000 1111)

Input-cutput unit operands

The tape unit parameters are listed in Table 2.14, and the operation of
tape unit (1)/ is described by Program 2.] 5. The unit idles at step 5 until its
busy indicator B/ is turned on by data channel i. After a starting delay of
about 650 microseconds reg uired to accelerate the tape and reach the
beginning of the record, one of four functions (listed under f in Table 2.13)
is perfornled as determined by the function indicator F/.
If F/ = 0, a normal read or write is perfornled under direct control of
the data channel as detailed in steps 18-37. If F/
0, one of the several
completely autononlOUS operations is initiated. If F/ is not zero and
W/ (write indicator) is unity, the autonOITIOUS function write end oj'file
is performed by steps 1-3, after which the busy indicator is turned off and
the unit returns to idle status. The end linlit indicator Lli is set by step 3
if the tape position exceeds a limit v set by a reflective marker attached to
the tape a short way before its extreme end.
If W/ = 0 and if F/ = ], 2, or 3, the unit backspaces to the next earlier
record gap, to the next earlier end of file position, or to position zero,
respectively. The last is called relt'ind. If, in backspacing, the tape becomes

<V/ ~A2' AO

<v j i

Lli ~ Lli

Bji

B/ : 0

starting delay

~ A2, Al

F/ : 0

Wji : 1

Rji

=
~

LOi ~ LOi

12

Z/,

=F

V (Fji:::j= 3)

pji : Ao
Fji :

R/ :

-d:::.

<D/

20
21

Pi ~P/

22

Xi

23

~p(Zji)

25
26

A2) x (R/ - 1)

Aji : 1

f-

Xi

<

29
30

p(Zji)

<D/
gi

yi

~ Z/, Al

~O

Aji :

Program 2.15

Tape unit (1) /

> v)

31
32

record gap delay

=ft'

28

<v j i ~ Zji, AO

gi : 0

27

Xi ~ 1

L1i ~ Lli V (7T(<V/)

em
N
N

24

pji : AO

~1<V/

Rji ~ (Zji

16

A/ : 0

p(Zji)

Rji : 0

>

17

P/

P/

record gap delay

11

15

19

~3

7T(<D ji) : 0

W/ : 1

Xi

10

14

18

Z/,

~O

13

V (7T(<v j i ) > v)

gi ~ gi V .11/

34

~
~
~.

35

33

;:::;:

36
37

~
~
~

s::
~.

;:::;:

'C
~

\0
N

T 2 i +- T 2 i V (r i J\ (kl

2 I,

==

t))

I k l : P

3 I I ~ r i J\ (k l
4 III

--

t) : 1

I (b i , A:) +- 0

B:i) :
I (B: FI W:

6 III

1
,

Xi) +- (l,fi' Wi, Wi)

>

IJi : 0

: III

9 II~ri+-O
=

10 I I

I r i : 0

11 I I

I (Pi EB

12 I I

I (g i EB r i EB yi)

A: i EB Qi)

+-

+-

(A o' 1, 3, 0)

13 I ~ Ci +-M.lS i

rl.-LS

14
15
16

eft

<-

15

36

(k o =1= r) J\ ((3

10 +

a ;C i : p(TCH)

=F c) J\ (p i

a:15)jCi ~ e) : 1

.-LSi)

= AI) : 0

Xi : 0

40

I II

Tli +- Tli V (=jXi)

II 43

(=jXi) J\ Eli : 1

II 44

Q oi

Q oi

+-

+-

Si +- w 15 jCi

+-

~"'"':
a

CT.3

S
S

Vq

1-111 45

(Xi ~ p(A 2 ))
i

Q I V (Qoi
=

(Qoi - 1)
I

III 46

> 0)

AI) : 0

r-i II 48

III 47

III 49

T 1 i +-T1 i V (QliJ\(Qoi =0))

=

~....:

:~

yi +- (yi =I- Xi)

TOi

38
39

(Ql i = 1) J\ (p i

37

Pi +-A o

QI

17 I ~ (~a:18/Ci =2 17) V(~a:18jCi = 6 x 215)

Xi +-0

delay

35

Vi +- a:6 (36)

(k

I (b i V

5 II I

Di +-0

Eli J\ Q1 J\ (Qoi

0) : 1

~II

50

II

51

C,()')

18 I

C 18i

~ WI5/M~w15/Ci

wI5jCi

19

20
21

K-,-a. /C

Qi ~

T1i

22

(3 ~ ( 15 )jCi : 0

23

gi

24

gi : 1

VijDi

Di

Vi

25

(k o

yi
=I=

r)

~Mlw15/Ci

26

~wl5jCi ~

27

Vi

28

Xi

29

Xi ~O

215 1(1

l-w I5 jCi)

55

56

~O

Q oi

ill ::

0
~

aIjXi

~6 ~

II 59

Vi

*/

Vi : a 6

(3

a 6 (36)

a l5 )jCi : 0

~(3 ~

54

T1i V (V jyi)

a l5 )jCi

I Xi

(:l-jVijDi) E8 VijDi

M --,-w 15 /C i

31

I yi

(yi :f- Xi)

~wl5jCi ~

32

~ Vi

+--

6.f Vi

33

V~

60

61
-1(3 ~ a l5 )jCi - 1

C 19i : 0

30

34

53

(3,0)

(V j yi) 1\ E 1 i

Wi:

52

Pi : AO

em
~
N

Di

215 1(1

~wI5/Ci)

62
63

64

I"":l

65

;::=

~
""":

s;:::

~.
~

~
~

I"":l

s;:::

~.

: a6

;::=

-~(3 ~

a l5 )jCi

~ l-(3 ~ ( 15 )jCi -

I
~

Program 2.16

Channel i

94

2.2

Microprogramming

rewound before the appropriate partition is found, the process is terminated

and the beginning limit indicator Loi is turned on (steps 10-11).
The file is read backward repeatedly by step 12. When a record gap
occurs, step 16 is executed, and if F/ = 1, the branch to step 4 returns the
unit to idle status. If F/ = 3, termination can occur only from step 11, at
which point the tape is rewound. The counter R ji is used to detect an end
of file partition. It is reduced by one (step 14) if the character read is A2 or
to zero if it is not. Since R/ is set to 3 after each record gap, step 17 is
reached with R / = 1 if and only if the end of file sequence AI, A2 , A2 , Al has
occurred.

20 WEF

k +-4

20 BSR

k +-1

20 BSF

20 REW

4
i

+-

bi

Ii

+-

Wi +-

Program 2.17

(-l(9

w 4 )jc)

k - 3 (k
(k

- 1

= 4)

= 4)

8
9

Instructions for special tape unit functions

Before completing the discussion of the remaining functions of the tape

unit, it may be helpful to follow through the entire process initiated by the
BSR (back space record) instruction. The channel idles (Program 2.16) on
step 5 with the busy indicator hi off. The BSR instruction (Program 2.17)
first determines the index i of the channel addressed, waits on step 6 until
the selected channel becomes free, sets the tape index t i = j to select the
particular unit Q) /, the function indicator ii to unity, the write indicator Wi
to zero, and the busy indicator hi to unity. This last action initiates oper
ation of channel i, which, as soon as unit <l>/ becomes free (B/ = 0), executes
steps 6-7 and then (since ii > 0) returns the channel immediately to idle
status. Step 6 transfers to the parameters of the selected unit the relevant
channel parameters which were themselves specified by the BSR instruction.

2.2

Instruction execution

95

Step 6 also makes the selected unit busy (B/ == 1), and hence starts it in
operation.
The normal read-write functions of the tape unit are described by steps
18-37. They are initiated by a branch from step 7 in the event that the
function indicator is zero. Over-all control by the channel is exercised
primarily through the association indicator A/, which is equal to unity if
unit j is logically associated with channel i, that is, if data are pennitted to
flow between them. If writing is in progress (Wji == 1) and A/ beconles
zero, the unit stops; if reading is in progress, the channel is freed ilnmedi
ately but the tape continues to a record gap.
Step 20 performs the read from tape, and if A / == 1, the partition read
is transferred to Pi' the seven-bit representation of the character read is
transferred to Xi (both for use by the channel), and the channel-unit
interlock Xi is set to unity to initiate appropriate disposition of the
character by the channel. If P/ is not a record gap reading continues, the
intercharacter delay (not shown) permitting time for the channel to
dispose of the character before the next is actually read. If P/ is a record
gap, the corresponding delay elapses before A/ is tested. If A/ == 0,
the branch to step 4 stops the unit. The tape stops only at a record
gap although transmission of data may be discontinued earlier by step
21.
The writing process begins at step 37 and nlay be discontinued before
any writing occurs (although the current record gap will be lengthened by
a few inches of blank tape). The main writing is perfornled by the loop
28-32, employing the channel interlock Xi. Step 31 sets the tape end lilllit
indicator. The loop terminates (step 32) when the HTite record gap
indicator gi is set to unity by the channel. Steps 33-36 then write the
longitudinal parity check character Yi supplied by the channel, together
with the inter-record gap partition AI. The write loop is then re-entered
unless A/ == O.
Channel operation. Operation of a channel is initiated either by one of the
special functions (WEF, BSR, BSF, REW) already described, or by a
WRS (write select), or an RDS (read select). The loading of the channel
command Ci required to control the two latter functions is, however,
controlled by a subsequent RCH (reset load channel), which transfers to
Si the address in memory of the desired channel conlnland.
The WRS (Program 2.18) selects the channel i specified by a portion of
its address, waits until the channel is free, sets its tape unit index t i as
specified by another portion of the address, sets the write indicator Wi to
unity and the function indicator ii to zero, and, finally, sets b i to start the
channel. The fetch mode indicator j' is also set to one so as to skip the

\0
0'\

5
e)

20 WRS}
20 RDS ~

i
bi

-L(9

t ( 4)jc

e)

- 1

8
i ~ -Lw jc

Wi

~ (a jc ~ p (WRS
3

Ii ~O

.f~

LCH

~ -L(1 ~

all)jc - (8e) .l (5, 4, 4)

+ Co

ri

10

ri

/\

bi

11

bi

Ci

~O

bi

11

12

~ 113
14
i +- -L(1 t all)jc - (8e) ~ (5, 4, 0) + Co I

h +-1
11

RCH

h +- h V
bi

o --+

Si

bi

Channel i, 9

~ w 15 jc

ri+-l
Program 2.18

J.s

I 15

16

I 17
h---J 18
~

19

Select unit and load channel instructions

em
N

2.2

Instruction execution

97

channel trap on the next instruction fetch. This prevents a trap frolll
intervening between the WRS and the following instruction (which is
nornlally an RCH). The RDS differs only in the setting of Wi on step 4.
If the channel is not busy (i.e., not selected), the RCH instruction
(Program 2.18) selects the channel specif1ed by a portion of the operation
code, sets the input-output indicator h, and copies the instruction address
to the channel sequence vector Si. If the channel is busy, the RCH
instruction sets the selected channel to its step 9, whereupon" the channel
waits on the interlock at step 10. Meanwhile, step 18 of the RCH sets
Si and step 19 sets the interlock 1'"i so that the channel Inay proceed.
Steps 13 and 14 of the channel operation load the channel conllnand
register Ci and increment the channel seq uence register Si. Ifthe comnland
is a TCH (Transfer in Channel), step 16 causes a branch to a new seq uence
of commands. If not, the ll'ord count, represented by (3 1 a.L/Ci, is
tested. Ifit is zero and if the current command is either an IOSP or IOCP,
the branch to step 13 immediately fetches the next COlnn1and in seq uence.
Otherwise, indirect addressing of the conlInand occurs (step 19) unless
C{K is zero.
Step 20 specifies k according to the class of the command being executed.
The cOlnnlands are listed in the nlatrix K of Table 2.13.
The first cOlnponent KOi assumes the value c, r, or s according as the
COl1lnland having code i is terminated by a word count test, a record gap,
or by either (signal). The second component K 1 i assunles the value d, p,
or t according as the channel discontinues operation, proceeds to the next
cOlnmand in sequence (as determined by SI), or transfers to an LCH
(Load Channel) instruction which may be awaiting execution by the
computer proper. Execution of the LCH (Progranl 2.18) is delayed at
step 1I and branches to step 18 (to respecify Si in the nlanner of the RCH)
only if the channel reaches step 3.
Channel operation continues on the right-hand segn1ent (steps 35-65) if
the operation is a read (w i == 0), and on the left (steps 22-34) if it is a
uTite. In the latter case, a zero word count causes in1nlediate ternlination
of the current command.
The norn1al tern1ination of a command in either read or write n10de
occasions a branch to step 1, where the tests for continuation begin. Step
1 sets the Channel Conllnand Trap indicator T./ if the current conlInand is
of the transfer type and an LCH (Load Channel) is not awaiting execution
in the con1puter proper. If the con1nland is of the proceed type, step 2
branches to step 13, where the next comn1and in sequence is fetched. If the
comnland is of the transfer type and an LCH is waiting (1'"i == 1), step 3
branches to step 9 to reset parameters and permit the channel to be reloaded.
1Il all other circulllstances step 4 is executed to disassociate the unit fro III the

98

M icroprografJun ing

2.2

channel and to return the channel to idle status. In read status, certain
abnornlal events-the occurrence of a parity error, or an end of file
partition-return the channel to idle status ilnmediately, regardless of the
type of command being executed.
The write operation (steps 22-34) is relatively simple. If the word
count in (3 1 0. 15 )/ Ci is zero, steps 23-24 ternlinate the current conlnland
but first initiate the writing of an end of record gap* if the comlnand is of
the "record" type (e.g., an IORP). If the word count is not zero, step 25
transfers to the channel data register Di the memory word selected by the
address portion of the command. The loop 28-33 transfers to the tape
unit successive six-bit infixes of Di and maintains the longitudinal parity
check Yi (originally reset on step 12). When all six have been transferred,
the branch to step 34 decrements the word count and, unless it beconles
zero, repeats the entire process from step 25.
The read operation (steps 35-65) begins by resetting D i to zero and the
infix selector Vi to a G Step 37 ternlinates the current conlnland if it is of
the count or signal type and the word count is zero. Steps 38-39 tenl1inate
the command if it is of the record or signal type and if the last file partition
read is a record gap. The partition indicator Pi is reset to Ao by step 39.
Thus a record gap present when termination is caused by a zero count is
still present on the first execution of the succeeding conlmand, whereas a
gap which itself causes termination is not present on the succeeding
command.
Steps 40-43 show the data interlock, the deternlination of the longi
tudinal parity check, and the setting of the parity error trap T 1 i in the
event of a parity error in the character. If the corresponding channel trap
is enabled, step 44 causes immediate ternlination in the event of a parity
error. Steps 45-48 detect an end of file configuration (using a counter,
Qoi in a manner similar to that used in Program 2.15), set the indicator
Q/ if a partition character A2 appears at the beginning of a word, and
cause termination (from step 49) \vith the end of file trap Toi set if an end
of file configuration occurs. If the character A2 occurring at the beginning
of a word is not part of an end of file configuration, step 50 sets the tape
error trap TIl, and step 51 causes termination if the corresponding channel
trap is enabled.
Steps 53-56 are executed only if Pi is a record gap. They reset the
counters Qi controlling the end of fIle test, test and reset the longitudinal
parity vector yi, and may cause termination in the event of an error. Step
57 causes the character transfer of step 58 to be ski pped if the character is

* Since the partition Al is represented by a gap, the writing of one gap inlnlcdiately
following anothcr, with no intervcning data, has the effect (whcn subsequently rcad) of a
single rccord gap.

Instruction execution

2.2

~s

Input-output test

10 lOT

*-

215 1(1

_Ls)

h : 0
h

*-

*-

Beginning of tape test 10 BTT

End of tape test

10 ETT

Transfer on channel 11 TEF

end of file

Transfer on channel 11 TRC

redundancy

Transfer on channel
in operation

11 TeO

Transfer on channel 11 TCN

not in operation

k *-1
i

~(1 ~

b *-(b i

"*-

all)jc - (8e) _L(6, 0)

k)

t :

b : 0

MO

w I5 (36)\s

b : 0

Program 2.19

Input-output branch operations

99

100

Microprogramn1ing

2.2

a potential end of file. Steps 62-65 decrement the word count and transfer
completed words to successive memory locations unless Cfn is zero. Step
61 suspends these operations when the word count reaches zero. Since
step 56 is followed by step 61, the occurrence of a record gap occasions the
(potential) transfer of a word to memory even though it is incomplete.
Because of the reset of Di on step 35, the incompleted part of the word is
zero.
Auxiliary channel instructions. Program 2.19 shows those branch instruc
tions which are controlled by indicators associated with the data channels.
Each indicator tested is also reset. The last four instructions shown are
subject to the trap mode.

11 SCH

i ~ 2 _1 (9 ~ a.:~) j C

bi

+ Co

e(36)

y~Ci

(3

a.,15)jy ~ Si

(Y18' Y20) ~ 0
M .lw 15 /C +-- y

11 ENB

y ~ MLw 15 /c

Eo

~w8jy

E 1 ~ (10 ~ a 8)jy
E2

~w8jy
-+

E~E

10 RCT
Program 2.20

Trap control and store channel

The channel indicators T may also cause interruptions as detailed in the

instruction fetch phase. They are controlled by the enable matrix E and
the enable trigger e which are set by the ENS (enable) and RCT (reset
traps) instruction of Program 2.20. The instruction SCH (Program 2.20)
pernlits storage of the channel registers.

2.3

Detailed logical des(f!;Jl

2.3

101

DETAILED LOGICAL DESIGN

Although a description couched at the progranllner~s level specifies

completely the functions of a computer~ it requires considerable extension
to provide a basis for the so-called logical design of circuits for realizing
the computer. The extensions include: (I) the specification of seq uence in
the lnicroprograms themselves; (2) further detailing of certain cOlnplex
functions; (3) reduction of the number of operands (registers) required;
and (4) economization in the underlying functions provided. The nature
of these extensions will be indicated briefly.
In principle~ the problem of seq uence control in the microprogranls does
not differ from the sequence control in computer progranls. However~ the
function served by the sequence vector s (a base two representation of the
address of the succeeding instruction) is freq uently served instead by a ring
or combination of rings. A ring is a logical vector r of weight one (that is,
+/r = 1) capable of rotation (i r or lr) and of resetting to one of several
initial positions Pi (that is, r +- e Pi .)
Certain steps of a microprogram, which at the progralnlner~s level nlay
be considered as monolithic, must themselves be realized in circuitry as
more detailed microprograms. The addition of two unsigned (positive)
nunlbers represented in base two by the vectors x and y might~ for exanlple,
be performed as in Program 2.21. The result x produced is correct only if
the sum is less than 2 '(x).
Economization in the underlying functions provided is achieved by
restricting the "data paths" provided between the various operands (i.e.,
registers) and by restricting the operands to which certain operations
apply. Restriction of data paths implies that it is not possible for each
operand to specify every other operand directly. For exanlple, nlemory
nlay be restricted to communicate only with the buffer register d so that
any transfer fronl memory such as
1

y:o
~y

t (x 1\

+-

(x ::/= z)

Program 2.21

z)

Base two addition

102

M icroprogranll11 ing

2.3
Address
decoder

r: E (36) , - - - - - r

Adder

Figure 2.22

Data paths

111ust in fact be perfonned in two steps:

d~Ml-.S

c~d.

An operation such as address decoding (i.e., conversion of the nonnal

base two representation of an address i into a one-out-of-n code of the
fornl i suitable for selecting word i from memory) is relatively costly and
is not normally provided for all relevant operands (such as sand w I :>/ c in
the 7090). Instead, decoding 111ay be provided on a single auxiliary
operand a; the selection of an instruction in the 7090 would then be
executed in two steps:
a~s

~ M~_a.

All microprograms specified at the programmer's level nlust, of course,

be translated into equivalent microprogranls which satisfy the path
constraints. Path restrictions are perhaps best displayed as a "block
diagram" showing the data paths provided between the various registers
and operations units. Figure 2.22 illustrates a convenient representation in
which the paths are shown as connecting lines with arrowheads indicating
the possible directions of transfer. If the indicated paths are further rc
stricted to selected components of the operands, this restriction 11lay bc
indicated by a pair of selection vectors separated by a colon. Thus the
notation
15
W (36) : (15)
on the path between d and s of Fig. 2.22 indicates that transfers occur
between w 1 ;)/d and s. The symbols rand c denote the selection of a rnatrix
row and a nlatrix column, respectively, as illustrated by the path betwccn
M and d. Pernlutations may be rcpresented in the fornl pJ. Thus if the

Exercises

103

vector d were to be transposed (reversed in order) in the transfer to c, the

path would be labeled with the expression

pS : ,
where p

35

Il- (36)1

(35,34, ... , 1,0).

REFERENCES
FalkofT, A., (1962) '"Algorithms for Parallel-Search Menlories," l.A.C.M. (to appear).
IBM Reference Manual, 7090 Data Processil{f{ Systeln, (1960), Form # A-22-6528,
International Business Machines Corporation.
Phister, M., (1958), Logical Design of D{fJital Conzputers, Wiley, New York.

EXERCISES
2.1 Write 7090 progranls for each of the following operations:
(a) M8 +-- M8 V M9
(b) M8

+-

M8

f~M6 +-- 236 1 (~M8 + ~M9)

M7 +_ ((~M8 + ~M9)
(c)
236 )
(d) M6 -(-IM7, a 6 , M 9 1
(e) M6 +--IM7, w 6 , M 9 1
(f) MG -(- IN1 7 , (8 1 a IO ), M 9 1
(g) MG +-IM7, M8, M 9 1 [Use lx, u, YI
(h) MG -(- U1W 1

Li

2.2

= (x

1\ u) V (y 1\ u)]

I n the nlagnetic core technology enlployed in the 7090, logical disjunction

(or) and negation are nluch easier to produce than conjunction (and). Linliting

the logical functions employed to disjunction and negation, write microprogranls

for the following 7090 instructions:
(a) ANS (Use De Morgan's law, Sec. 1.8)
(b) ERA
2.3 In the nlagnetic core technology used in the 7090, each transfer of a quantity
Y into a register x is actually an or with the present content of the register, i.e.,
x +- Y V x. A register lnay also be reset to zero. Subject to the foregoing
restriction, write microprograms for
(a) the operation I +-- Mi. (Use two steps.)
(b) the operations of Exercise 2.2(a).
2.4 Describe the nlain portion of the instruction fetch of the 7090 (steps 9-18
of Progranl 2.4) in an algorithnl which satisfies the data path constraints of Fig.
2.22.
2.5 Repeat Exercise 2.4 so as to satisfy the constraints of Exercises 2.2 and 2.3
as well.
2.6 A vector p which pernlits only the follo\ving types of operation:
(i) p -(- \p, wI, ,t\

.. {y -(- w1jp

(11)

l p

-11p

-(~W

Microprogranllning

104

is called a pushdown or stack vector. If the successive conlponents of a stack

vector p are represented by successive 111enl0ry registers, then the operations
affecting it can be controlled by a single address counter v, which is autoinatically
increnlented at each operation of type (i) (addition of a new final conlponent) and
is autolnatically decrenlented at each operation of type (ii) (reference to the final
conlponent acco111panied by its deletion).
(a) Using the 7090 registers and formats augnlcnted by an address counter l'
of dinlension 15, write a nlicroprogra111 for an operation LOS (load stack)
15
which transfers M w /c to the top of the stack [operation type (i)].
(b) Write a 111icroprogranl for STS (store stack) which transfers the top of the
stack to M_w 15 /c [operation type (ii)].
(c) Write a nlicroprogranl for an operation AND which produces the and
function of the top two components of the stack, deletes thenl, and appends
the result as a new final conlponent. [The net reduction in v(P) is olle].
(d) The AND of part (c) has no associated address. Show that aJl 7090
instructions (other than input-output) can be redefined so that only the
L OS and STS require associated addresses.
2.7 In the 7090, a decinlal digit ,r is represented in direct binary coding in a
six-bit logical vector x, (that is,
x = :c), and each register accomnl0dates six
decinlal digits. Use the cOllvert instructions (Progra111 2.12) in a 7090 progranl to
(a) convert froln binary to decinlal.
(b) convert fronl decill1al to binary.
(c) replace aJl leading zeros (i.e., all preceding the first significant digit) of a
number represented in deciInal.
2.8 Write a 7090 program to convert
(a) froll1 a 36-bit binary code to a 36-bit reflected Gray code [see Phister (1958)].
(b) fronl a reflected Gray code to binary.
2.9 A ll1enlory M is called a tqr: or associative nlemory if for any argunlent x
it yields a direct indication of the row or rows of M \vhich agree with x. If the
resulting indication is in the forll1 of a vector 5 such that the 111atrix 5//M contains
the indicated rows, then 5 = M (I, x. More generally, a logical JJlask vector m is
added to the system so that mlM is cOll1pared with the argunlent mix and sonlC
desired function of miMI; is represented by miMI; for each k. In the following
exercises M is assull1ed to be a logical 111atrix.
(a) Use De Morgan's laws (Secs. 1.8 and 1.11 or Sec. 7.1) to derive froIll the
relation 5 = M
x an expression for 5 which would be suited to a circuit
technology in which disjunction and negation are easier to perform than
conjunction.
(b) Write a detailed algorithnl using a row-by-row scan of M to deternline
5

= (mIM)

I~ (mix).

(c) Repeat part (b) using a colunln-by-column scan of M.

(d) Use a colunln-by-eolumn seanof Mtodetermine 5 such that 5//Mcontains
the rows of M of ll1aximull1 base two value [see Falkoff (1961)].

chapter

REPRESENTATION OF
VARIABLES
3.1

ALLOCATION AND ENCODING

Although the abstract description of a program may be presented in any

suitable language, its automatic execution must be perfonned on sonle
specified representation of the relevant operands. The specification of this
representation presents two distinct aspects-allocation and encoding.
An allocation specifies the correspondences between physical devices and
the variables represented thereby. An encoding specifies the correspond
ences between the distinct states of the physical devices and the literals
which they represent. If, for example, certain numerical data are to be
represented by a set of 50 two-state devices, the two-out-of-five coding
systenl of Exercise 1.6 might be chosen, and it would then remain to
specify the allocation. The two-digit quantity "hours worked" might be
allocated as follows: devices 31-35 represent components 1-5, respectively,
of the first digit, and devices 29, 16, 17,24, and 47 represent components
1, 2, 3, 4, 5, respectively, of the second digit.
The encoding of a variable will be specified by an encoding nlatrix C and
associated fornlat vector f such that the rows of flC list the representands
and the rows of flC list the corresponding representations. The encoding
is normally fixed and normally concerns the programmer only in the
translation of input or output data. Even this translation is usually
handled in a routine manner, and attention will therefore be restricted
primarily to the problem of allocation.
However, the encoding of numeric quantities warrants special comment.
It includes the representatIon of the sign and of the scale, as well as the
representation of the significant digits. Small numbers, such as indices, ad
nlit not only of the usual positional representation but also of the use of the
unit vector j to represent the number j (i.e., a one-out-of-n coding systenl),
or of the use of a logical vector of weight j (i.e., a base 1 nunlber systenl).
Allocation will be described in terms of the physical rector 1t, which
denotes the physical storage elements of the cOInputer. Each conlponent
of 1t corresponds to one of the v(1t) similar physical devices available, its
105

106

Representation oj'l"ariahles

3.1

range of values is the set of physical states achievable by each device, and
its index is the address of the device. Each component of 1t Inay corre
spond to a computer register, an individual character position in a register,
or an individual binary digit within a character, depending on the degree of
resolution appropriate to the allocation problem considered. The O-origin
indexing normally used for cOlnputer addresses will be used for the physical
vector, but I-origin indexing will, throughout this chapter, normally be
enlployed for all other structured operands.
An index of the physical vector will be called an address and will itself
be represented in the (perhaps mixed) radix appropriate to the given com
puter. The Univac, for example, employs base ten addressing for the
registers, and (because of the use of ]2-character words) a radix of twelve
for finer resolution. The address of the fourth character of register 675
nlight therefore be written as 675.3. In computers which have two or
nlore independent addressing systems (e.g., the independent addressing
systems for main menlory and for auxiliary storage in the IBM 705),
superscripts may be used to identify the several physical vectors 1tJ
In general, the representation of a quantity x is a vector (to be denoted
by p(x)) whose components are chosen from the physical vector 1t. Thus
p(x) == kJ1t, where k is a mapping vector associated with x. The dimension
of the representation (that is, v(p(x))) is called the dimension oj'::c in 1t. If,
for example, p(x) == (1t lO ' 1tn, 1t17 , 1t IH ), then k == (10, 9, 17, 18), and the
dilnension of x in 1t is four. If p(x) is an infix of1t, then the representation
of x is said to be solid. A solid representation can be characterized by two
parameters, its dimension d and its leading address f, that is, the index in 1t
of its first component. Then p(::c) == (j'1 a d )/1t.

3.2 REPRESENTATION OF STRUCTURED

OPERANDS
The grid matrix
If each component of a vector x has a solid representation, then the
representation of the entire vector is said to be solid and may be charac
terized by the grid matrix rex), of dimension vex) x 2, defined as follows:
rli(x) is the leading address of p(xJ, and 2 i (x) is the dimension of Xi in 1t.
If, for example, the vector x is represented. as shown in Fig. 3.1 a, then

rex)

==

17

19

27

23

32

C()')
~

x3

x4

X2

Xl

1t 22

Representand

X5

I 1t 23 I 1t 24

Physical vector
(a) Representation of a vector x

r 11

r 21

r 22

r 12

r 13

')

r 14

r 24

r 25

r 15

Representand
~

1t 40

1t 41

1t 42

1t 43

1t 44

1t 45

1t 46

1t 47

1t 48

1t 49

1t 50

1t 51

1t 52

1t 53

1t 54

1t 55

1t 56

1t 57

1t 58

fi

59

Physical
vector

~
~

~
~

I 0

(b) Linear representation of

rex) with I

= 34

Actual
value

~
~.
~

~
~
""':

:::::

+ 4i + 2).

Figure 3.1

':J

~
~

Xl
I

X2
1

Xl
2

X2
2

~
~

t::i
~

c::;-

Figure 3.2

Linear representation of a nlatrix X

l-'

.....,J

108

Representation of l'ariahles

3.2

Any structured operand can first be reduced to an equivalent vector, and

the grid 111atrix therefore suffices for describing the representation of any
construct, providing only that the representation of each of its elements is
solid. Thus a matrix X 111ay be represented by either the row-by-row list
r == E/ X or the column-by-column list c == Ell X, and a tree T 111ay be
represented by the left list matrix [T or the right list matrix ]T, either of
which may be represented, in turn, by a vector.
If a process involves only a sn1all number of variables, it is practical to
n1ake their allocation implicit in the algorithn1, i.e., to incorporate in the
algorithm the selection operations on the vector 1t necessary to extract the
appropriate variables. This is the procedure usually en1ployed, for
example, in simple con1puter progran1s. In processes involving numerous
variables, implicit allocation lTIay become too cumberSOlTIe and confusing,
and more systelTIatic procedures are needed.

Linear representations
The representation of a structured operand is said to be lincar if each
component is represented by an infix of the form (/1 a d )/1t, where I is a
linear function of the indices of the component. For exalTIple, the
representation of the matrix X indicated by Fig. 3.2 is linear, with d == 2
and I == - 11 + 5i + 8).
A linear representation is solid and can clearly be characterized by a
small number of paran1eters-the dilTIension d of each component and the
coefficients in the linear expression /. The representation of a vector x is
linear if and only if r 2 (x) == de and the difference () == r/(x) - r~ lex) is
constant for i == 2, 3, ... , vex).
If I == P + qi + 'i is the function defining a linear representation of a
matrix X and if a is the leading address of a given element, then the leading
address of the succeeding element in the row (or column) is sin1ply a + r
(or a + q). Freq uently, the succession must be cyclic, and the resulting
sum must be reduced modulo veX) x r (or p(X) x q). The inherent
convenience of linear representations is further enhanced by index registers,
which provide efficient incrementation and comparison of addresses.
Linear representation of a structured operand req uires that all COlTI
ponents be of the san1e dimension in 1t. This common din1ension n1ay,
however, be achieved by appending null elements to the shorter C0I11
ponents. The convenience of the linear representation n1ust then be
weighed against the waste occasioned by the null elen1ents. Moreover, if
several vectors or matrices are to be represented and if each is of unspecifIed
total dimension in 1t, it n1ay be in1possible to allot to each an infix suffi
ciently large to permit linear representation. Consequently, a linear
representation is not always practicable.

3.2

Representation oj'structured operands

109

Nonlinear representations
Since the use of the grid matrix imposes only the condition of solidity
for each component, it permits an allocation which is sufficiently general
for most purposes. The grid matrix serves in two distinct capacities: (1)
as a useful conceptual device for describing an allocation even when the
actual allocation is implicit in the program, and (2) as a paranleter
which enters directly into an algorithm and explicitly specifies the allo
cation.
If the grid matrix is used in a program as an explicit specification of the
allocation, then the grid matrix must itself be represented by the physical
vector. There remains, therefore, the problem of choosing a suitable
allocation for the grid matrix itself; a linear allocation is illustrated by
Fig.3.lb.
If the grid matrix rex) itself employs a linear representation, its use
offers advantages over the direct use of a linear representation of x only if
the total dimension of r in 7t is much less than the total dimension of x in
7t when linear representations are employed for both. This is frequently
the case, since each element of a grid matrix belongs to the index set of 7t
(that is, to lO(V(7t))), and the dimension of each element in 7t is therefore
both uniform and relatively small. Program 3.3 shows the use of the grid
matrix rex) and the encoding matrix C in determining the kth conlponent
of the vector x.
Program 3.3. A linear representation is assumed for rex), with elenlent r/(x)
represented by the infix ((p + qi + ~i) 1 aJJ)!Tt. Moreover, each eleI11ent of rex)
is assunled to be represented in a base b nunlber systenl. Step 1 detennines the
leading address of the representation of r lk(X). Step 2 specifies.1' as the base b
value of this representation, i.e., as the leading address of p(x k ). Steps 3 and 4
specify d as the dimension of x k in Tt, and step 5 therefore specifies z as the
representation of x k'
Steps 7-9 perfornl the decoding of z = p(x k ) to obtain z as the actual value of
xl." Since this process is normally perforI11ed by hUI11an or I11echanical I11eanS
(e.g., a printer) outside the purview of the programmer, it is here expressed directly
in terms of the encoding matrix C rather than in tenns of its representation. The
left-pointing exit on step 7 is followed only if z does not occur as an entry in the
encoding matrix.

The fornl chosen for the grid matrix is one of several possible. The two
columns could, for example, represent the leading and final addresses of
the corresponding representations or the dimensions and final addresses.
The present choice of leading address / and dimension d is, however, the
nlost convenient for use in conjunction with the notation adopted for
infixes; the logical vector (/ 1 ad) selects the appropriate infix.

3.2

Representation of variables

110

O-origin indexing for rt only

I
2

+-

+ qk + r x

p, q, r

.f +- b.l.((l t aY)jrt)
l+-l+r

+-

+- ( [

h ..;- /leC)

h+-h-l

z : ljCk

+-

b.l.((l
~

t ag)jrt)

ad)/rt

liCk

Constant, coefficient of row

index, and coefficient of column
index in the linear function for
the representation of rex).

Base used in representing ele

ments of rex).

lJ

Dimension in rt of each element

of rex).

.1'

Leading address of p(xJJ.

+1

d
Dimension of p(x k ) in rt.
=I
Z

p(x k ).

Encoding matrix for com

ponents of x.

Format vector for C.

Character encoded by x Ie'

-------+

Legend

Program 3.3 Detennination of Z

tion of the grid matrix rex)

p(x k ) and z

x k fronl a linear representa

Chained representations *

If a linear representation is used for a vector, then the deletion of a

component (as in a compress operation) necessitates the moving (i.e.,
respecification) of the representations of each of the subseq uent conl
ponents. Similarly, mesh operations (insertion) and permutations
necessitate extensive respecification. The use of a grid matrix
obviates such respecification in x, since appropriate changes can instead be
lnade in rex), where they may be much simpler to effect. If, for exanlple,
x is the vector represented as in Fig. 3.1 a, and z is a quantity of dimension
six in 7t, then the mesh operation

rex)

nlay be effected by specifying the physical infix (70

1 aY)/7t

by p(z) and by

* Chained representations have received extensive treatnlent, frequently under the

name "Iists." See, for example, Shaw et al. (1958) and Blaauw (1959).

3.2

Representation of structured operands

respecifying

111

rex) as follows:

rex) =

17

19

70

27

23

32

rex)

However, if the representation of

is itself linear, then insertions,
deletions, and permutations in x will occasion changes in all components of
whose indices are affected. The need for a linear representation of the
grid matrix (and hence for all linear representations) can be obviated by
the use of a chained representation defined as follows.
Consider a vector Y, each of whose components Yk has a solid represen
tation P(Yk) whose infixes (g 1 a.Y)/p(Yk) and a.g/P(Yk) are, respectively,
the dimension of p(Y,J in 1t and the leading address of the representation of
the (cyclically) succeeding component of Y (both in a base b system), and
whose suffix -a. 2g I p(Y,J is the representation of the kth component of SOlne
vector x. Then (the representation of) Y is called a chained representation
oj' x. In other words, the representation of Y incorporates its own grid
matrix (with the address column r1(y) rotated upward by one place) as
well as the representation of the vector x.
For example, if g = 2, b = 10, and x = (365,7,24), then

rex)

P(Yl)
P(Y2)

=
=

(1t 17 , 1t 18 , 1t 19 , 1t 20 , 1t 21 , 1t22 , 1t 23 )
(1t 68 , 1t69 , 1t 70 , 1t 71 , 1t 72 )

(6, 8, 0, 7, 3, 6, 5),

(2, 6, 0, 5, 7),

and

is a suitable chained representation of x.

The parameters required in executing an algorithm on a chained
representation yare g, the conlnl0n dimension in 1t of the elenlents of the
grid nlatrix r(y); b, the base of the number system employed in their
representation; and f and h, the leading address and index, respectively,
of the representation of some one component of y. The paranleters g and
b are usually common to the entire set of chained representations in use.
Program 3.4 illustrates the type of algorithnl required to deternline p(x/,.)
froln a given chained representation of x.
Program 3.4. The loop (1-3) is executed l{X) 10 (k - 11) times, with the result
that at step 4 the paranleter.{ is the leading address of p(y /.'). Step 4 therefore
specifies d as the dinlension of p(y k), that is, as the base b value of r /'(y). Step 5

112

Representation of variables

3.2
O-origin indexing for

vex)

11 (h +

1)

+--

----+

h : k

j'is the leading address of the

11th component of the chained
representation of x.

Base used for representation

of the elements of the grid
matrix.

Dimension in 1t of elements of
the grid matrix.

Dimension in 1t of kth com

ponent of the chained represen
tation of x.

kth component of the chained

representation of x.

bl-C(f + g) + a. )!1t
fJ

Z +- ( /

p(x k )

+--

+ a. d )/1t

+--

-a 2

(J

/z

only

h"r

.r +-- b l-((f + a. U)/1t)

+--

1t

-----+

Legend

Program 3.4

Determination of p(x k ) fron1 a chained representation of x

then specifies z as P(Yk). Step 6 deletes those con1ponents of z which represent

the elements of the grid I11atrix, leaving p(x/l)'
The parameters.{ and hare then1selves respecified in the execution of the
algorithm so that Iz becomes k and,f becomes , appropriately, the leading address
of p(Y,J. A subsequent execution then begins from this new initial condition.

The chained representation used thus far is cyclic and contains no

internal identification of the first or the last con1ponents. Such an identi
fication can be incorporated by adding a null component between the last
and first components of x. Alternatively the identification may be achieved
without augmenting the dimension but by sacrificing the end-around
chaining~ i.e., by replacing the last component of jr 1(y) by a null element.
Moreover, a chained representation may be entered (i.e., the scan Inay be
begun) at anyone of several points, provided only that the index 17 and
corresponding leading address f are known for each of the points.
The number of cOlnponents of a chained representation scanned (steps
1-3 of Program 3.4) in selecting the kth component of x is given by
1-{x) 10 (k - h), where 17 is the index of the component last selected. The
selection operation is therefore most efficient when the components are
selected in ascending order on the index. The chaining is effective in the
forward direction only, and the component (17 - 1) would be obtained
only by a complete cyclic forward scan of vex) - 1 components. The

3.2

Representation of structured operands

113

representation is therefore called a forlt'ard chain. A backlt'ard chain can

be formed by incorporating the vector lrI(y) instead of irI(Y)' and a
double chain results fron1 incorporating both.
A vector x which is respecified only by either deleting the final COIl1
ponent or by adding a new final component (i.e., by operations of the
form x +- WI/X, or x +- x EB (z)) behaves as a stack (cf. Exercise 2.6). A
backward-chained representation is clearly convenient for such a stack.
A simple example of the use of a chained stack occurs in representing
the available (i.e., unused) segments of the physical vector 1t. This will be
illustrated by a progranl for the vector cOln.pression

x +- v/x
executed on a forward-chained representation of x. The unused segments
representing the components of ii/x are returned to a backward-chained
stack or pool of available cOlnponents. A linear representation can usually
be used for logical control vectors such as v; in any case the problellls
involved in their representation are relatively trivial and will be subordi
nated by expressing each operation directly in ternlS of the logical vectors
and not in terms of the physical components representing them.
Program 3.5. In the major loop (6-23), k deternlines the index of the current
component VI,', and i andj determine the leading addresses of p(x /,) and p(x/; !1),
respectively. These three parameters are cycled through successive values by
steps 7, 8, and 12 and are initialized by steps 2,5, and 12. lf VI,: = 0, the infix
p(x/;) is returned to the pool by steps 21, 22, 23, and 6 so as to construct a back
ward chain.
The paranleter.r specifies the leading address of p(x 1 ) unless l'(X) = 0, in which
case J: is null. Step 1 ternlinates the process if ll(X) = 0, and otherwise step 4
respecifies.r as the null elenlent. If V = 0, this null value of.r renlains ~ if not, the
first nonzero component of v causes a branch to step 14. Since T = 0, step 15 is
executed to respecify :c as the leading address of p( (vi X)I). Step 16 then specifies 11,
the leading address of the last conlpleted conlponent of vlx. Step 15 is never
again executed.
Components of vjx other than the first nlust each be chained (in a forward
chain) to the preceding one. Hence the leading address i of a newly added conl
ponent nlust be inserted in the last preceding conlponent (whose leading address
is h). This is normally done by steps 18, 19, and 6 ~ step 20 respecifies h. If, how
ever, the component X k - 1 were also included, it would appear as the last conl
pleted component of vlx and would already be chained to the new C0111pOnent XI,'.
This situation is recognized by step 17 and occasions a branch to step 16. Step 16
then respecifies h and repeats the loop without executing steps 18, 19, and 6.
The process ternlinates when the cycle through the chained representation of X
is conlpleted, that is, when i returns to the original value of .r, preserved as t by
step 3. Step 10 is then executed, ternlinating the process directly if r(vlx) = 0.

114

3.2

Representation of variables

Otherwise, step 11 is executed to close the chain of v j x, that is, to insert f , the
leading address of p((vjX)l)' in the representation of the last con1ponent of vjx.
O-origin indexing for rt only

~c

i+-x

t+-x
+-

:.c =

~(;

k +-1

bl((r

k+-k+l

a 9 )jrt)

+- S

Index of v.
Leading address of p(x,J.

Leading address of P(X k +1 ).

Leading address of last

preceding C0111pOnent of

#
;

~)

11

b L((h

12

13

VI.:

14

+-

b _L((i

vjx.

+-

16

+-

i -

17

Vk - 1

18

r+-h

19

+-

20

+-

21

+-

22

s +-p

+-

Leading address of last

preceding component of
pool of available seg
ments.

Dimension in rt of ele
ments of grid matrices.

Base of representation of
elements of grid matrices.

a 9 )jrt)

15

23

a 9 )jrt)

+-

Logical vector.

9
10

i +-j

Leading address of p(x1 )

if vex) > 0; otherwise

Legend

Program 3.5 Program for x -(- vjx on a forward chained representation of x

and a backward chained stack of available segments

A chained representation can be generalized to allow the direct represen

tation of more complex constructs, such as trees, by incorporating the
address of each of the successor components associated with a given

3.2

Representation oJ'structured operands

115

component. This notion is fonnalized in the chain list nlatrix of Sec. 3.4.
The same scheme can also be employed to produce an efficient combined
representation of two or more vectors which share certain COInnlon
components. If, for example, x j = Z!" , and chained representations are
used for both x and z, then x may be represented in standard form except
that component x j incorporates a secondary address, which is the leading
address of Zk+l. Moreover Z has a standard representation except that
Zk-I is chained to Xj' with an indicator to show that the secondary address
of the succeeding conlponent is to be used. Deletion of any vector
conlponent in such a shared systenl must occasion only the corresponding
change in the address chain of the vector, the actual representation of the
conlponent being deleted only when no associated address renlains.
Partitions
If the set a is the range of the cOlllponents of the physical vector 1t, and
if SOIne element, say aI' is reserved as a partition sYlnhol and is excluded
from use in the normal representation of quantities, it can be inserted to
denlark the end (or beginning) of an infix of 1t. If the vector Y is repre
sented by a single infix of 1t such that the beginning of component Yj+l
follows imnlediately after the terminal partition of Yj' then the structure of
Y is completely represented by the partitions, and Y is called a partitioned
representation. A partitioned representation can be used for more cOlnplex
operands, such as matrices, if a set of two or nlore distinct partition
symbols are provided, one for each level of structure. The distinct
partition symbols can, of course, be represented by multiple occurrences of
a single symbol a l rather than by distinct Inenlbers of a.
A partitioned representation is sinlilar to a double-chained representa
tion without end-around chaining in the following particular: beginning
from component Yi, the component Yj can be reached only by scanning
all intervening components between i and j in increasing or decreasing
order according as i < j or i > j. The fIle notation introduced in Sec.
t .22 clearly provides the operations appropriate to a partitioned repre
sentation of a vector, with conventions which suppress all inessential
references to the partitions themselves.
The use of a partition to demark the end of an infix is particularly
convenient when the infix must be processed conlponent by conlponent
for other reasons, as in the use of nlagnetic tape or other serial storage.
The partition also appears to be more economical than the grid lnatrix,
which it replaces. This apparent economy is, however, somewhat illusory,
since the reservation of a special partition symbol reduces the infornlation
content of each nonpartition component by the factor log2 (v(a) - 1) -:
log2 v(a), where a is the range of the conlponents of 1t.

116

Representation of variables

3.2

Partitions can be employed in chained representations. For example,

the dimension in 7t of each component of a chained representation y can be
specified implicitly by terrninal partitions instead of explicitly by the vector
r 2(y) of the grid matrix. Thus if the elements of rI(y) are of dimension !S
in 7t, then w1/p(Yj) = aI' and (a 9 1\ w1)/p(Yj) = p(x j ), where x is the
vector represented by y. Program 3.6 shows the determination of p(xk )
from a chained representation y with terminal partitions a l .

O-origin indexing for

+--

v(x)

11 (h +

.l +-- b ~ (c'f t

h :k

40

-(-.l + g

4b

nJ

a1

4c

j +--j

4d

d +--j -.l
Z

p(x k) -(-

only

is the leading address of the

hth component of the chained

representation of x.
b

Base used for representation of

the elements of the grid matrix.

Dimension in n of the elements

of the grid matrix.

u1

Partition symbol.

kth component of the chained

(f t ad)jn

+--

h,.l

aJJ)jn)

=F

1t

1)

representation of x exclusive of
the terminal partition symbol.

aU jz
d

Dimension of z in n.
Legend

Program 3.6 Deternlination of p(xk ) from a chained representation of x with

terminal partitions a 1
Program 3.6. The program is similar to Progranl 3.4 and the step nunlbering
indicates the correspondences. The dimension d is so determined (steps 4a-d) as
to exclude the terminal partition itself from the quantity z specified by step 5.
Since only the first column of the grid matrix is incorporated in the partitioned
representation, step 6 excises a prefix of dimension g rather than 2g as in
Progranl 3.4.

Pools
Components of the physical vector 7t in use for the representation of one
quantity must not be allocated to the representation of sonle other quantity.
The construction of a chained representation therefore poses one problem
not encountered in its use, namely, the specification and observation of
restrictions on the availability of components of 7t. The restrictions can

3.2

Representation of structured operands

117

conveniently be specified as a pool, consisting of the available components

of 1t. Each allocation made must then be reflected in a corresponding
change in the pool. Moreover, as each piece of data is deleted, the com
ponents allocated to it are returned to the pool.
If, as in Program 3.5, a pool is treated as a stack, then the component
next taken from the pool is the component last added to it. The queue of
components in the pool thus obeys a so-called last in first out, or LiFO
discipline. The dimension in 1t of the last component of a pool will not, in
general, agree with the dimension required for the next quantity it is called
on to represent. If it exceeds the requirements, the extra segment may be
left in the pool, and the pool therefore tends to accrue more and Inore
components of smaller and smaller dimension. Hence it nlay be wise, or
even essential, to revise the pool occasionally so as to coalesce the segments
into the smallest possible number of infixes. This process can even be
extended to allow substitutions in other vectors in order to return to the
pool short segments which may unite existing segments of the pool. This,
however, will require a systematic scan of the chained vectors.
If the dimension of the last component (or perhaps of all components)
of the pool falls short of the requirements for representing a new quantity,
segments of the pool can be chained together. This requires the use of a
special partition symbol or other indication to distinguish two types of
links, one which marks the end of a given representation and one which
does not. More generally, it may be convenient to use multilevel partition
symbols to distinguish several levels of links, as was suggested for the
representation of a matrix.
Queue disciplines other than LIFO may be used. Three other types of
primary interest in allocation queues are the FI FO (first in first out), the
dirnensiol1-ordered, and the address-ordered disciplines. FI FO uses a
forward chain and may be preferred over LIFO because it uses the entire
original pool before using any returned (and usually shorter) segnlents.
The components of a dimension-ordered pool are nlaintained in
ascending (or descending) order on their dimensions in 1t. This arrange
ment is convenient in selecting a pool element according to the dimension
required. The components of an address-ordered pool are arranged in
ascending order on their leading addresses. This arrangement facilitates
the fusion of components which together form an infix of 1t.
If each of the available components of 1t is set to a special value which
is used for no other purpose, then the available conlponents can be
determined by a scan of 1t. Such a pool has no structure inlposed by
chaining and will be called a rnarked pool.
A marked pool requires little maintenance, since conlponents returned
to it are simply marked, but selection from it requires a scan of 1t and is

118

Representation of l'ariahles

3.2

therefore relatively slow. The use of marked and chained pools nlay also
be cOITIbined-all returned conlponents go to a marked pool which is left
undisturbed until the chained pool is exhausted, at which time the entire
nlarked pool is organized into a chained pool.
Summary

Since any structured operand can first be reduced to an equivalent

vector, the problems of representation can be discussed in terms of vectors
alone. The characteristics of the linear, chained, and partitioned repre
sentations of a vector may be sunlmarized as follows. A linear representa
tion permits the address of any conlponent to be computed directly as a
linear function of its indices and hence req uires no scanning of the vector.
However, the strict limitations which it imposes on allocation ITIay en
gender: (1) conflicts with allocations for other operands, (2) waste of storage
due to the imposition of a common dimension in 1t for all cOlnponents, or
(3) uneconomical execution due to the extensive reallocations occasioned
by the insertion or deletion of other than terminal conlponents.
The concept of the grid matrix is helpful even when the corresponding
allocation is ilTIplicit in the progranl. The explicit use of a grid ITIatrix
which is itself in a linear representation removes the restrictions on the
allocation of the vector itself while retaining the advantage of direct
address computation. The address computation differs frolTI the linear
case only in the addition of a single reference to the grid nlatrix and hence
requires no scanning. The difficulties enulTIerated for the direct linear
representation are not elinlinated but merely shifted to the linearly
represented grid matrix itself, where they nlay, however, prove much less
serious.
A chained representation allows virtually arbitrary allocation, relatively
simple operations for the insertion and deletion of components, the direct
representation of more complex structures such as trees, and econonlical
joint representations of vectors which have one or more COITIpOnents in
COmITIOn. However, a chained representation requires extra storage for
the grid matrix which it incorporates and occasions additional operations
for scanning when the components are selected in other than serial order.
The required scanning can be reduced by the retention of auxiliary
information which allows the chained representation to be entered at
several points.
A partitioned representation req uires the allocation of a single infix of
1t, and selection requires a fine scan, i.e., a component-by-component scan
of 1t to detect partition synlbols. Partitioning removes the need to
incorporate the grid matrix explicitly and does not impose a conlnlon
dimension in 1t for all components.

3.3

Representation oj'rnatrices

119

Mixed systems employing conlbinations of linear, chained, and parti

tioned representations are frequently advantageous. Block chaining, for
example, involves the chaining of blocks, each consisting of an infix of 1t
and each serving as a linear representation of some infix of the represented
vector. Alternatively, each chained block may be a partitioned represen
tation of some infix.

3.3

REPRESENTATION OF MATRICES

Structured operands other than vectors may be represented by first

reducing them to equivalent vectors which can, by employing the tech
niq ues of the preceding section, be represented, in turn, in the physical
vector 1t. In the case of a matrix A, t\VO alternative red uctions are of
interest, the row list' = Ej A = Al E8 A2
... E8 All and the colunln
list c = EllA. If 1"11' A/, and c k are corresponding elements of the three
alternative representations, then in a O-origin system:

+ j,
+ /-1j.

h = vi
k = i

Consequently,

= lh -:- vJ =
j = v 10 h = l k
i

and

/-1 10 k,

-:- /-1 J.

The dependence of h on k can be obtained directly by substituting the

foregoing expressions in the identity

lh -:- vJ + v 10 h
+ lk -:- IlJ.
x (v 10 h) + lh -:- vJ.

h= v x

to yield

Similarly,

k = /-1

x (/-110 k)

The pernlutation h which carries the row list r into the colunln list C
(that is, C = hJor) can be obtained directly from the foregoing expression
for h as follows:

The expression for the kth component of h is identical with the expression
for h above. Hence, if C = hJor, then ck = rhk = 'Ii as required.
If the row list (or column list) is itself represented linearly, then the
address of any component A / is obtained as a linear function of the indices
i and j. If either a file or a chained representation is to be used for the list
vector, then the components are processed most efficiently in serial order,
and the use of column list or row list is dictated by the particular processes
to be effected.

120

3.3

Representation of L1ariahles

If a large proportion of the elements of a matrix are null elements, it is

called a sparse matrix. Sparse matrices occur frequently in numerical
work (where zero serves as the null element), particularly in the treatment
of partial difference equations. A sparse matrix A can be represented
conlpactly by the row list r == Uj A, and the logical matrix U, where
U == (A 1= 0). The matrix A may then be obtained by expansion:
A == U\r.
Alternatively, the column list c == (A i= O)jjA may be used. The
transformation between the column list c and row list rmust, in general,
be performed as a sequential operation on the elenlents of U. Since it is
frequently necessary to scan a given nlatrix in both row and column order
(e.g., as either pre- or post-multiplier in a matrix multiplication), neither
the row list nor the column list alone is satisfactory. A chaining systelTI
can, however, be devised to provide both row and colunln scanning.
Let L be a matrix such that L 1 is a list of the nonzero elements of a
matrix A in arbitrary order, L/ is the column index in A of elenlent L 1 i,
and L 3 i is the row index in L of the next nonzero element following L 1 i in
its row of A. If L/ is the last nonzero element in its row, La' == 0. Letj~
be the row index in L of the first nonzero element of row AJ, and let
jj == if Ai == O. The following example shows corresponding values of
A, L, and j:

6
A

L==

7
5

f==

The matrix L will be called a row-chained representation of A and nlay be

used, together with the vector j, for the efficient scanning of any row A i as
illustrated by Program 3.7. The vector L 3 can be modified so as to give the
address in 1t directly rather than the row index in L of the next element in
the row, and Program 3.7 can then be easily re-expressed in terms of the
physical vector 1t.

Program 3.7. Step 2 yields the index in L of the first elel11ent of the ith row of A.
Step 4 deternlines its COlUI11n index j, and step 6 deternlines the index of the
succeeding conlponent. The process ternlinates at step 3 when the scan of the
row is c0111pleted.
If L 1 is chosen as a row list, the vector L:3 reduces to the form L:/, == k + I
or L:/ == c. Its function can then be served instead by incrementation of

3.4

Representation of trees
I-origin indexing

Ai +-0
2

k +-li

Ii

+-

121

Row index in L of first nonzero

element of row Ai.Ii = if Ai = O.
0

L 2k

Row index in L of next element.

L1

List of nonzero elements of A.

L 2k

Column index in A of L/".

L 3k

Row index in L of next nonzero

element following L 1 k in its row in
A. L 3 k =
if no such element
exists.

Aji

+-

L 1k

+-

Ll;

Legend

Program 3.7
tation of A

Determination of the row vector Ai fronl a row-chained represen

the index k and by the use of the logical vector u = (L 3 = O) for deter
mining the end of each row.
The construction of a column-chained representation is analogous to
that of a row-chained representation, and the two representations can be
combined in a single matrix L which gives both row and colunln chaining
employing but a single representation (that is, L 1 ) of the nonzero elelnents
of A.

3.4

REPRESENTATION OF TREES*

A tree T may be represented by a matrix and hence, in turn, by a vector

in a number of useful ways as follows:

1. by a full right list matrix ]T or by any column permutation thereof

(Sec. 1.23),
2. by a full left list matrix [T or by any column permutation thereof,
3. by a right list matrix cx 2 /]T,
4. by a left list matrix cx 2 /[T,
5. by various chain list matrices.
The full left and right lists seldom prove more convenient than the more
concise left and right lists. Except for the special case of a homogeneous

* Johnson (1962) provides a comprehensive treatment of the representations of trees

and discusses the suitability of each representation for a variety of search procedures.

122

Representation

(~r

3.4

variahles

tree~ both the right list and the left list are awkward to use for path tracing.
This function is better served by the chain list matrix, to be defIned as a
formalization of the chaining schenle suggested in Sec. 3.2.

Simplified list matrices

In certain important special cases, the various list representations of

trees may be sin1plified. If the degree of each node is a known function ()
of the value of the node, then for any list matrix M, M/ == b(M2t)~ and the
degree vector M l 111ay be elinlinated with
out loss. The node vector alone then repre
sents the tree and n1ay be referred to as a
right or left list rector as the case nlay be.
For example, in the tree of Fig. 3.8
(\vhich represents the compound logical
/\
statement x 1\ (y V z))~ a fixed degree is
associated with each of the logical operators
and, or~ and not (nalnely, 2~ 2~ and 1)~ and the
degree zero is associated with each of the vari
ables. The statement can therefore be repre_
Figure 3.8 The compound
sented unambiguously by the left list vector
logical staten1ent i' ,\ (If V z)
v==(/\,

,x, V~y,z).

This is the so-called Lukasielricz, Polish, or parenthesis-free form of the

compound statement [Lukasiewicz (1951) and Burks et al. (1954)].
Frequently, the only significant nodes of a tree T are its leaves (e.g.~ in
Example 3.2 and in a certain key transfornlation of Fig. 4.7) and all other
nodes may be considered as nulls. Hence if M is any list matrix~ the
significant portions of M l and M 2 are (M1 =F 0)/ M 1 and (M1 == 0)/ M2~
respectively. These significant portions may then be coalesced to form the
single vector
which, together with the logical vector (Ml == O)~ fonns a leaf'list nlatrix
that describes the tree. Moreover~ if the values of the leaves are distin
guishable from the components of the degree vector~ the logical vector
(Ml == 0) may also be dropped.

The use of left lists

The use of the right list nlatrix is illustrated by the repeated selection
sort treated in Sec. 6.4. The use of left lists will be illustrated here by two
examples, each of interest in its own right: the partitioning of the left list
of an n-tuply rooted tree to yield the left lists of the component singular

l~epresel1tation

3.4
p
2

+-

123

I-origin indexing

( O)

i+-O

Given left list of T.

i : {l(Z)

Row index of Z in ascending

scan.

Indicated number of roots of

current rooted subtree.

Moment of current rooted

subtree.

m +-0

r+-O

i : {l(Z)

i+-i+1

m+-m+l

(1 trees

+-

10

r : 1

11

+-

+1

Partition vector of Z, that is,

Pj = p:(T j ).

- Zli

Legend

<

p E8 (m)

Program 3.9

Partitioning of the left list of an n-tuply rooted tree

subtrees and the construction of a Hufflnan minimum-redundancy prefix

code.
Example 3.1. Partitioning of an n-tuply rooted tree. Progranl 3.9 shows a
schenle for partitioning a left list Z of a tree T into conlponent subtrees, i.e., for
deternlining the vector p such that Pj is the monlent of the singular subtree T j .
Thus v(p) = fJ.l(T), Pj = /l(T j ), and the infix ((p .~. a j - 1 ) t aPj)!!Z is the left list
of T j
The loop 6-10 scans successive components of the degree vector Zl (in ascend
ing order) and computes r, the indicated number of roots. The value of r in
creases by, at most, one per iteration, and when r becomes unity, the end of a
singly rooted tree has been reached. Its monlent In is then appended (step 11) as
a new final component of the partition vector p, the paranleters 111 and r are reset,
and the scan of the next rooted tree is begun. Nornlal ternlination occurs at
step 3; tennination at step 6 indicates ill fonnation of Z.
Example 3.2. Huffman minimum redundancy prefix code. If b is any set such
that ll(b) > 1, then any other finite set a can be encoded in b, that is, represented
by b. (The sets a and b may be called the "alphabet" and "basic alphabet,"
respectively.) If l{a) <; v(b), the encoding nlay be described by a nlapping vector
k such that p(a i ) = b k . Ifl'(a)., v(b), then each a i nlust be represented by a
vector Xi := b. For exa~lple, if a = lO( 10) and b = l (2), then the decinlal digits a
nlay be encoded in the so-called 8421 systenl:
(2

(4) ~

Xi

= a i'

124

3.4

RejJresentation of variables
0.380

(0)

(1 000)

(1001)

(101)

(1 1 0)

(111)

Figure 3.10

Construction of a H uffnlan prefix code

In so-called fixed length coding the vectors Xi have a conlmon dinlension d, and
the decoding of a nlessage m (consisting of the catenation of vectors Xi) involves
the selection of successive infixes of dinlension d. If the probability distribution
of the characters Q I occurring in messages is not unifornl, more compact encoding
may be achieved by using variable length codes and assigning the shorter codes to
the more frequent characters. Decoding of a nlessage in variable length coding
can be perfornled only if the boundaries between the successive Xl are indicated in
some way.
The boundaries between characters in a message in variable length code nlay
be demarked by special partition symbols (which is inefficient) or by using a
prefix code in which no legitimate code point Xi is the prefix of any other legitinlate
code point, including itself. The index vectors of the leaves of any tree possess
this property; conversely, any set of prefix codes can be arrayed as the leaves of
sonle tree. Hence if each character of the set to be encoded is assigned as the leaf
of a conl11l0n tree, and if each character is encoded by the associated index vector,
a so-called prefix code is attained. Figure 3.10 furnishes an exanlple of a binary
code (i.e., the branching ratios do not exceed two) constructed in this nlanner.
O-origin indexing is used. The discussion will be limited to binary trees.
] f ii is the freq uency of the ith character and 1i is the length of the assigned code
(i.e., the length of path to the root), then the nl0st efficient code is attained by
minimizing the scalar producti x 1. This nlay be achieved by the following con
struction, shown to be optimal by Huffman (1952). First, the characters to be
encoded are all considered as roots, and the two roots of lowest frequency are
rooted to an auxiliary node (shown as a null elenlent in Fig. 3.10), which is then
assigned their combined frequency. The process is repeated until only two roots
relnain. The tree of Fig. 3.10 is optilnal with respect to the frequencies shown to
the left of the leaves. The appropriate conlbined frequencies are shown to the left
of each of the nonleaves.
Programs 3.11 and 3.12 show the construction of the tree T representing a

3.4

Representation of trees
T

125

~c

fJ.l(T) : 2

(8/I)ST

+-

\\(0) EB a 2 //T, al,

+-

(8/I)Sf

-a 2//T\\

+ ja2 jj, al, a2 jI\

Program 3.11 Construction of the binary Huffman code T for characters c

with frequency I
H uffnlan code for a set of characters c i with frequencies Ii' the former in ternlS
of the tree itself and the latter in ternlS of its left list.

Program 3.11. The frequency vector I is permuted (step 5) to bring it to ascend

ing order, and the tree is subjected (step 3) to the sanle pernlutation. Step 4
replaces the first two rooted subtrees of T by the single subtree obtained by root
ing them in a null, and step 6 nlakes the corresponding alterations in the
frequency vector. The tree is initialized (step 1) as a one-level tree whose roots
are the given characters, and the process tenninates when the nunlber of roots of
T has been reduced to two.
Program 3.12. The tree T of Program 3.11 is represented by the left list node
vector z, in conjunction with the inlplicit degree vector d = 2 x (z = oe). The
algorithnl differs fronl Progranl 3.11 primarily in the reordering of the subtrees
(steps 6-9). Step 7 appends to x the left list of the ith subtree (of the reordered
tree) selected by the partition vector p according to the conventions of Progranl
3.9. Step 1a prefixes x by the new null root, and steps 11-12 redefine p
appropriately.
Program 1.21 can be applied to the left list produced by Progranl 3.12 to deter
nline the associated index nlatrix (in a a-origin systenl), and hence the actual
codes assigned.
It is not essential that the characters be assigned to leaves in precisely the order
specified by Programs 3.11 and 3.12, and it is sufficient that the dinlension of the
leaf index increase monotonically with decreasing frequency of the character. It
is therefore unnecessary to carry the characters themselves through the process ~
it suffices to deternline the structure of the tree, sort the corresponding index
matrix to right list order (which is ordered on dinlension of the index vectors),
and assign the characters (in decreasing order by frequency) to successive leaves.
Since the structure of such a tree (whose nodes have a conlnlon irrelevant value
and whose nonleaves all have a conlmon branching ratio equal to the nunlber of
roots) is sufficiently determined by the moment vector ~(T), the process of Pro
gram 3.12 can be simplified.

1--+

z+-c

+-

I-origin indexing

e(v(z))

v(f) : 2

+-

+-

e(O)

+-

(Olf)i

+-

x CD (((p

+-

8
9

3.4

RejJresentation of variables

126

:::;

t ex

I)

i : v(f)
z

+-

(0) ffi x

11

+-

(O/f)Sp

12

+-

(1

13

+-

(Olf )Sf

14- f

+- (

+ (p

~- ex 2 ))

-1- lex 2 /f) ffi

CD

-a 2 1f

Given character set.

Left list of Huffman

tree.

fi

Frequency of ith sub-

tree of z.

Partition vector Pi is

the moment of the ith

subtree of z.

Reordered left list

with subtrees in

ascending order on

frequency.

exPi)lz)

+1

10

-a2 /p

Legend

Program 3.12 Construction of the

left list z of the binary Huffnlan
code for characters c with fre
quency f

Chain list matrices

The full chain list matrix of a tree T is a matrix P of dimension peT) x
(b(T) + 2) defined as follows: P 2 is some node vector of T, PI is the
associated degree vector, Pj+2 is null if j exceeds the associated degree P Ii
and is otherwise the row index in P of the jth node emanating from node
P 2 i. Table 3.13 shows a full chain list matrix for the tree of Fig. 1.16. A
full chain list matrix is called a full right (left) chain list matrix if the nodes
occur in right (left) list order.
The full chain list matrix is a formalization of the scheme suggested in
the discussion of chained representations (Sec. 3.2). Its convenience in
forward path tracing is obvious. Since it does not identify the roots of the
tree, an auxiliary vector must be provided for this purpose. However, if
the ordering chosen for the nodes is that of a right list, the roots occur
first in the list, their number r = yep!) - (+/P I ) is specified by the degree
vector P v and the need for the auxiliary vector vanishes. Moreover, since
a right list groups all nodes emanating from a given node, each row of
0. 2/ P is simply a sequence of integers followed by null elements, and the
information necessary to path tracing is provided by the column P 3 alone.
The right chain list matrix of a tree T is therefore defined as a 3 / P, where

d'
n'
-n
I
1
2
2
g
0
3
u
0
t
4

3
2
0

8
9

5
6

a
b

18
16

26

24

9
20

p
d" nil
- - -
3
a
4
b
7
2
2
9
g
c
11
4
0
z
n
1
15

0
0

10

11

0
0

I
f
d

0
2

y
s

e
i

0
3
2

0
2
0

3
22
0

10

17

12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

127

Representation of trees

3.4

3
0
0
1
4
0
0

i
w
h
e

p
x
c
q
1

a
3
0
0
0

k
h

23

nl

0
0
0

p
q

0
0

14
19
6
2

14
25

t
a

16

2
0
0

n
g
u

1
2
0
3

n'

- -

19

d'
-

19

0
0

21

nl

20
1
17
15
21
0

23
12
26
7

18
16
0
0

24
8
0
0
0

0
0

0
0

4
0
,')

3
22

10

2
1

s
x

24
26

0
0

25

13
11

15

12

21

1
4
0
0

P
x

13
11

::;

::)

A full chain list Inatrix

The right chain

list matrix

Filial-heir chain list

(a)

(b)

(c)

Table 3.13

1
2
3

4
5
6
7
8
9

10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

Chain lists of the tree of Fig. 1.16

P is the full right chain list matrix of T. It is illustrated by Table 3.13b

Program 3.14 shows its use in path tracing. Although the degree vector PI
is redundant (that is, PI and P3 can be determined one from the other), it
provides a direct check (step 6) on the legitimacy of the index vector r
which would be difficult to obtain from P 3 alone.
For a search of the type described by Program 3.14, it is necessary to
scan down a level until agreement is reached and then across to the next
level. For this type of scan, the filial-heir chain list is compact and con
venient.

128

3.4

Representation oj'rariables
k +-1

}+-O

d +- v(PI )

.i

+ jP I

vCr)

j+-}+1

5
6

I-origin indexing

>

rj

i+-k+rj-l

Pj

+-

p 2 i

d +- pli

10

k +- p 3 i

Given index vector.

Right chain list matrix of T.

PI

Degree vector of T.

P2

Node vector of T.

P3

Chaining vector of T.

Path vector Tr.

Degree of current node.

Base address of the infix con

taining the current node.

Index of succeeding node in the

path Tr.

Current index of index vector

r.
Legend

Program 3.14

Determination of the path p

fronl the right chain list

matrix P

The set of (j + 1)th level nodes of the subtree T i are collectively called
thejthfilial vector of node i, and the first member of the first filial vector of
node i is called the heir of node i. (For brevity, the first filial vector of a
node will also be called its filial vector.) If each node is chained only to its
successor in the filial vector containing it and to its heir, the resulting
representation is called a filial-heir chain list. Fornlally, the filial-heir
representation of a tree T is a matrix F of dimension Il(T) x 4 such that
F 2 is a node vector of T, F 1 is the associated degree vector, Fa is a .filial
chain such that F:/ == j if node F 2 i is the successor of node F./ in the
smallest filial set containing it and F:/ == ifnode F 2 i has no such successor,
and F 4 is an heir chain such that F 4 i == h if node F 2 /1 is the heir of node F/
and F 4 i == 0 if F 2 i is a leaf. The filial-heir chain list is illustrated in Table
3.13c.
0

REFERENCES
Blaauw, G. A., (1959), '"Indexing and Control-Word Techniques," IBM Journal o(
Research alld Developl11ellt, vol. 3, pp. 288-301.
Brooks, F. P., Jr., and K. E. Iverson, (1962), (in press) Autol1wtic Data Proces.\il{f{,
Wiley, New York.
Burks, A. W., D. W. Warren, and J. B. Wright, (1954), "'An Analysis of a Logical
Machine Using Parenthesis-free Notation," Mathel11atical Tahles alld Other Aids
to COll1putatioll, vol. VIII, pp. 53-57.

Exercises

129

Dewey, Godfrey, (1923), Relatiu Frequency of E/~~/ish Speech Sounds, Cambridge

University Press, p. 185.
Huffman, D. A., (1952), ""A Method for the Construction of MiniInum Redundancy
Codes," Proc. IRE, vol. 40, pp. 1098-1101.
Iverson, K. E., (1955), "Report by the Staff of the Computation Laboratory to the
American Gas Association and Edison Electric Institute," Section III, Report
No.1, Harvard Computation Laboratory.
Johnson, L. R., (1962), ""On Operand Structure, Representation, Storage, and Search:'
Research Report # RC-603, IBM Corp.
Lukasiewicz, Jan, (1951), Aristotle's Syllogistic frolll the Standpoint o( lv/odern FOflnal
Logic, Clarendon Press, Oxford, England, p. 78.
Marimont, R. 8., (1959), '"A New Method of Checking the Consistency of Precedence
Matrices," J. ACM, vol. 6, pp. 164-171.
Ross, I. C., and F. Harary, (1960), "The Square of a Tree," Bell Systelll Tech. J., vol.
XXXIX, pp. 641-8.
Shaw, J. C., A. Newell, H. A. Simon, and T. O. Ellis, (1958), A Command structure
for complex information processing," Proc. Western Joint COlllputer Con(erence,
pp. 119-128.

EXERCISES
The symbols a and c will be used exclusively to denote lower case and capital
alphabets defined as follows:
a = (0, a, b, c,
, z, . , , , #, *, +).
c = (0, A, B, C,
, Z, . , , , #, *, +).
The expression 1t S; x will be used to specify the set x as the range of the conl
ponents of 1t.
3.1 For each of the following cases, specify a suitable encoding n1atrix and
fornlat vector and show the explicit value of the infix of 1t which (in a solid
representation) represents the given exanlple vector x:
(a) the decinlal digits d = LUCIO) in a ranked fixed-length code for 1t S; lO(2).
Exanlple: x = (6, 8, 9).
(b) the set a in a ranked fixed-length code for 1t S; lO(2).

Exan1ple: x = (c, a, t).

(c) the set a u c u lO( 10) in a fixed-length code for 1t S;; lO( 10).

Exanlple: x = (M, a, y, 0,3", 1,9,6,0, .).

(d) the set a u c in a two-case code (with single-character shift) for 1t S;; (I.
(See Brooks and Iverson, 1962.)
Exanlple: x = (T, r, 0, y, " N, . , Y, .).
(e) the set a in a Huffman prefix code for 1t S; lO(2). Assunle the frequency
distribution given in Dewey (1923).
Example: x = (t, r, e, e).

3.2

For each of the cases of Exercise 3.1 write a progranl which decodes the
infix (i 1 a.J)/1t, that is, which produces the vector z represented by the infix. The
auxiliary physical vector 1t l S;; S may be employed to represent the first colun1n of
the encoding matrix, where s is the set encoded. Perfornl a partial trace of each
program for the example value used in Exercise 3.1.

Representation oj'rariahle.\'

130

The ordered set of nlonths nl = (JANUARY, FEBRUARY, ... ,

DECEM BER) is to be represented by the physical vector 1t S;; C U l()( 10). For
each of the following types of representation, specify a particular representation
and show the values of the relevant conlponents of 1t:
(a) a linear representation (en1ploying null elenlents for tllling to a conlnl0n
dinlension in 1t).
(b) a solid representation for each elen1ent of 111 and an appropriate grid
nlatrix itself represented linearly.
(c) a chained r~presentation.
(d) a double chained representation.

3.3

3.4 (a) For each of the cases of Exercise 3.3, write a progran1 which selects
n10nth mIl'.
(b) Trace each progranl for the case k = 2.
(c) For case (d) of Exercise 3.3, write a progra111 \vhich selects 111 1,' by
forward chaining if k
l'(rn) -:- 2, and by backward chaining if
k . vern) -:- 2.
3.5 For each of the cases of Exercise 3.3, write a progran1 which "'prints out"
the set of 1110nths in a n1inin1un1 nun1ber of n-character lines, inserting a single
null between successive n10nths except where (i) further nulls n1ust be added to
prevent the continuation of a single word fron1 one line to the next, or (ii) no null
is needed between two successive \vords, the flrst of which is coternlinous with
the line. I n other words, produce a nlatrix Z of row di111ension n and of nlinin1unl
colulnn din1ension such that (Z
~E)/Z = (p(nl 1 )
p(rn 2)
p(nll~)'
and such that each ro\v Zi nlay be partitioned into one or n10re vectors of the
fornl p(rn/J
O , all but the last of \vhich nlust be of din1ension l{p(nl l.,)] + I.
3.6 Assunling a linear representation for each of the logical vectors involved,
and a forward-chained representation for each of the ren1aining operands, write
progranls for the following operations. Assunle in each case that the argunlents x
and y need not be retained, and aSSU111e the use of a backward-chained pool
where necessary.
(a) z ~
(b)

(c) z ~

(d) z ~

x, u, y

lx, u,y/
k
k

i x
1x

3.7 Repeat Exercise 3.6(a), using separate grid nlatrices for x, y, and z instead
of chained representations. Specify a suitable linear representation for each of
the grid n1atrices.
3.8 (a) If a chained representation is used for a vector x, then the selection of a
specitled con1ponent can be 111ade faster by providing a nUlnber of
alternative starting points for the required scan. State precisely the
quantities required in such a process and \\;Tite a progranl showing its
usc.
(b) If provision is nlade for starting the scan at any conlponent of x, the
chained representation nlay itself be sinlplified. Sho\v precisely what

Exercises

131

the Sill1plified forll1 is and identify the type of representation to which it

is equivalent.

3.9 Frequently a vector x kept in a partitioned representation (for efficient use

of storage) ll1USt be "unpacked" to a linear or other 1110re accessible fornl for
efficient processing. The converse operation of "packing" is also required. Let
the partitioned representation be a file (I) enlploying an interconlponent partition
AI' and a ternlinal partition A~, and write both packing and unpacking progranls
for each of the following cases. Assunle that the nlaxinlunl dill1ension in 1t of any
conlponent is n.
(a) A solid linear representation eInploying null fill.
(b) An allocation prescribed by a grid nlatrix G with G'2 = !lE.
3.10

Let 1t S; lO(2), let the set a be encoded in a five-bit code such that (2E)
and let each conlponent of the vector x be an (uncapitalized) English
word. Using a-origin indexing throughout, specify a suitable partitioned repre
sentation in 1t for the vector x, and repeat Exercises 3.9(a) and 3.9(b), using it in
lieu of the files.
p(ll/) = i,

3.11 For each of the following pool organizations, write a progranl to convert
a given ll1arked pool into a backward-chained pool:
(a) dinlension-ordered.
(b) address-ordered.
3.12 For each of the following queue disciplines, write progranls which take
fronl and return to the pool an infix of length n. Use secondary linking and
relegate to a Inarked pool any infix which is too short for linking. In each case
choose the type of chaining best suited to the particular queue discipline.
(a) LI FO (last-in-first-out).
(b) FI FO (first-in-first-out).
(c) Dinlension ordered.
(d) Address-ordered (utilize the possibility of fusing adjacent infixes).
3.13 Give a conlplete specification of a schenle for representing a tree T by a
full chain list matrix which is not in right list order. Write a prograll1 (expressed
in ternlS of the physical vector 1t) which deternlines the path vector T i for a given
index vector i.
3.14 Give a conlplete speciflcation of a schenle allowing joint representation of
those conlponents shared by two or nl0re of a fanlily of vectors Xl, x'2, ... , Xii as
suggested in Sec. 3.2. Write progranls to (i) select conlponent x/, and (ii) delete
conlponent x/.
3.15 Let 1t ;; II U l(lO), and let xl, x 2 , . . . , x n be a fanlily of vectors whose
cOll1ponents belong to the set o:l/[a U l(lO)]. Let the average and the nlaxin1unl
dinlensions of the vectors xi be a and Ill, respectively. AssunlC that the chaining
index is represented in decinlaI, with each digit represented by one con1ponent
of1t. Deternline (as a function of 17! and n) the value of a below which a chained
representation provides nlore con1pact storage than a linear representation with
null fill.

132

Representation oj' l'ariahle.~'

3.16 Write a program which uses the n1ininlization operation u..

determine the ordering pernlutation vector p -( - 01/( a 'I b).

I'

x to

Let U = (X
0) and r = UIX jointly represent the sparse l11atrix X.
(a) Write a program \vhich detern1ines (as a function of U and r) a suitable
row-chained and column-chained representation of X.

3.17

(b) Write a program defined on the representation produced in part (a) to

conlpute the product Y = X x X, itself represented in the fonn V =
(Y
0) and p = VIY.
(c) Write a progranl to deternline the trace (that is, +IIIX) of X fronl the
representation produced in part (a).
3.18 The unique assignment of Hufflnan codes produced by Progranl 3.12 is,
in general, only one of nlany equally efficient assignnlents, since the syn1bols to be
coded need only be assigned, in decreasing order on frequency, to the leaves of
the code tree in increasing order on their levels. Show that the structure of the
tree produced can be sufficiently described by its nloment vector alone, and write
a progran1 for the construction of a Huffman code based on this fact.
3.19 Following the notation and ternlinology used in Progranl 3.9 for the
analogous case of a left list write a progranl which detern1ines fron1 the right
list R of a tree T, the partition vector p which partitions it by levels.
3.20 Write a progranl which detern1ines the right list R = cx~/]T as a function
of the left list L = cx 2 /[T. Incorporate tests of well forn1ation.
3.21 Let [X8~]j) denote the pth power of the square nlatrix X with respect to the
,operators 0 1 and 02' that is, [X8~]1) = XCl~ X~<~ ... ~j~X to p factors.
(a) Show that ([C ]Ji)/ = 1 if and only If the-re is a ~path of length p fronl
node i to node in the graph (n, C).
(b) Show that [C ]1 = 0 for some p
I'(C) if and only if (n, C) contains no
circuits.
(c) If (n, C) contains no circuits, the connection nlatrix C is said to be "con
sistent. n The result of part (a) can be used to check consistency. Progran1
the alternative nlethod of Marin10nt (1959).
(d) If H = C
I, then ([H ::~ ]1
= 1 if and only if i =.i or there exists a path
fronl node i to node.i of length n
p + I. Show that for any connection
n1atrix C, [H /: ]/1 converges to a linlit.
1

1
)/

3.22
(a)
(b)
(c)
(d)

Devise progranls to determine

whether a given connection nlatrix C represents a tree.
the left list of the tree (n, C).
the right list of the tree (n, C).
a node list n and connection matrix C as a function of
(i) a left list I.J
(ii) a right list R.

3.23 Show that (n, C) and (n p ,

tion p.

CpP)

represent the sanle graph for any pern1Llta

3.24 lf (n, C) is a tree and if K = C ; C, then C can be detern1ined as a function

of K (see Ross and Harary, 1960). Write a progranl for deternlining G' fr0l11 K.

chapter

SEARCH TECHNIQUES

In classical applied 111athenlatics 1110st functions of interest ca n be

approxinlated by some algorithm which beconles, for practical purposes,
the deflnition of the function. In other areas, however, nlany functions of
practical, if not general, interest (such as the correspondence bet\veen
employee nalne and salary) can be specitled only by an exhaustive listing of
each a rgunl en t val ucan d its cor re spo ndin g fun ction val ue. Such a fu nction
will be called j()rtuitous.
The basic algorithm applicable to the eval uation of a fortuitous function
is a search of the list of argulnents, i.e., a cOlllparison of the givcn argu
111ent with the list of argunlents to detenlline the correspondent to bc
selected. In such an algorithlll it is convenient (as illustratcd by Progranl
1.12a) to distinguish three phases which successively detenllinc thc
following quantities:
( I) the index or rank r == k I k of the argunlcnt k in k.
(2) the index i == PI' of the correspondent in s.
(3) the correspondent s == ,,'I'
Step (2) is a perlllutation defined by the pernlutation vector p. Stcps (2)
and (3) arc silnple selections from structured opcrands (nortllally in linear
representations) and rcquire no further discussion. Step I is called
ranking, and the nlethods for acconlplishing it nlcrit dctailed treatlllcnt.
The argunlent k of a 1l1apping (and hence of a ranking) operation \vill bc
called a key. The Gernlan-English dictionary nla pping of F~xalll pic 1.2 is
typical of nlappings fronl key to correspondent. Ranking is itself a spccial
nlapping from the key set k onto its own index set ll( }'(k)).
If the represcntation used for sonlC or all of the data inlposes ccrtain
restrictions (such as serial access), there nlay be sonlC advantage in
coalescing the three phases of the Inapping operation so as not to dctcnnine
the rank explicitly. It will, however, be convcnient to lilnit the discussion
altnost exclusively to the problcln of ranking.
There arc two Inain types of ranking processes: scanning and kcy
transfonnations. A scanning process conlpares the key k sllccessively with
133

134

Search techniques

selected elements of the key set k to determine the rank r == k l k. A key

transjarlnation is any function or algorithITI t(k) which maps the set k into
some subset of the integers. The set of deriL'ed keys is defined as the set
d containing all derived keys arranged in ascending order. The set of all
keys which map into d j is called the jth equicalence class defined by the
transfonnation t.
If v(d) == v(k), the key transformation is biunique and the ranking
operation may therefore be completed by a permutation p such that
Pj == i for j == t(k;). If v(d) < v(k), then at least two distinct elements of k
map into the same element of d and the ranking process nlust be completed
by a scan of one of the equivalence classes defined by t.
If, for example, k == (n1, tu, w, th, f) is the ordered set of working days
encoded according to the encoding matrix
1

In

tu

1 0

th

1 0

G==

and fonnat vector f == (0, 1, 1, 1), then scanning can be accomplished by

cOInparing p(k), the encoded representation of the key k, with successive
rows of fiG to determine the rank r of the row on which agreeInent occurs.
Moreover, the key transformation
t(k) == (2)

_ p(k)

is uniq ue, but req uires an associated mapping vector m == (1, 2, 3, 4,

Had p(f) (that is, fiG:)) been chosen as (1,0, 1), the mapping vector
would not have been required. Finally, the key transfonnation
0,

0,5).

t'(k) == 1

p3(k)

has the range d == (1, 2), is not uniq ue, and req uires a subseq uent scan of
one or other of the equivalence classes e 1 == (tu, th), and e 2 == (m, w, f),
represented by (010, 100) and (00 L 0 II, Ill), respectively.
Although a strict ranking operation maps elenlent k; into the integer i,
any biunique mapping onto the index set ll(V(k)) will frequently serve as
well. For, if p is the permutation required to cOInplete the ranking process,
and if a subsequent permutation j is required (as in step 4 of Progranl
1.12a), the two permutations can be combined in the single pern1utation
q == jJp. Sinlilarly, the ranking of a set k may be considered as equivalent
to the ranking of any set obtained by pernluting k.

4.1

Scanning nlethods .fc)r rank ing

4.1

135

SCANNING METHODS FOR RANKING

The two nlain methods of scan are called directed and controlled. A
directed scan is begun in one of twopossiblc directions fronl a given initial
point i~ either ascending (that is~ i~ i
I~ i
2~ ... ) or descending. The

direction chosen may be determined by a conlparison bet\veen the given

argunlent and the set element k i at the initial point i. A controlled scan is
executed in a seq uence which is detennined by successivc cOlnparisons
between the argument and each element scanned. In an effective controlled
scan~ each comparison 11lust determine thc choice of the next clenlcnt for
conlparison so as to (approximately) lllinilllize the expected nunlber of
eleinents scanned. The directed scan is clcarly well suited to the use of a
file or chained representation~ which illlposes serial access to the eleinents,
whereas the controlled scan is 1l0t.
The scan Iength (i. e. ~ the n uIn be r 0 f key eIenl ents scann ed) \\/ iII be used as
a Ineasure in analyzing and evaluating scanning 11lethods. The sanle
nleasure and the same analysis apply also to the converse situation, \vhere
the rank of an element is given and the element itself must be obtained
from a chained or other representation which pennits only serial access.
An alternative related nleasure is the norlllalized scan length or scan
jj'action~ defined as the scan length divided by the nunlber of elenlents in
the set of keys.
A scan is said to be fooled if each exccution begins at the sanle point f.
A rooted scan may be advantageous \vhen the freq uency distribution of the
argunlents is nonunifonn and the nlost frequent keys can be grouped near
the root. A scan is called catenated if each execution is begun at the end
point of the preceding scan. A catenated scan Inay be elllployed in lIsing
a file when the intervals between successive scans are so short as to allo\\'
little or no time for return to a fixed root or \\;'hen the arguinents are
arranged in the saIne relative order as the items in the file.
Directed scan

A directed scan is called cyclic if elenlent k 1 follo\\/s k in an ascending

scan and if k follows k 1 in a descending scan. It is called nonc~l'clic if the
direction of scan is reversed whenever either of the ternlinal elenlents k 1 or
k is encountered. A cyclic scan is appropriate to a chained representation
with end-around chaining: a noncyclic scan is appropriate to a flle or to a
chained representation \vithout end-around chaining.
The initial direction of scan nlay be chosen in several \vays, the nlore
inlportant of \vhich are enuinerated and discussed belo\v. For independ
ently and unifornlly distributcd argunlcnts, the expected scan fractions
j

ll

W
0'\

V:l

~
~

~.

Type of Scan

pre!

Initial
Position

Initial
Direction

Cvclic or
Noncyclic

Any
Any
1
Catenated

l(I'Ud+ l)--;-2J

Catenated
Catenated
Catenated
Catenated
Catenated

Fixed. continue or reverse

To argument
Fixed
Fixed
To argument
To argument
To nearer terminal
To farther terminal
Continue previous
Reverse previous

Cyclic
Cyclic
Noncyclic
Noncyclic
Noncyclic
Noncyclic
Noncyclic
Noncyclic
Noncyclic
Noncyclic

No.1
1
2
3

7
8
9
10

Table 4.1

o :S

for Scan Fraction f

b

b
b
1

b
2b
b
-h'2j4 + h
2h
_b'2 + 2h
-b'2j2

b)

b
-h'2j4

b
b 3 j4 - h 2 j2
-b 3 j4 + b

+b

< 1~

+h

-h'2j4

-h 2 + 2h
b 2 j2 + 1
-b'2
2b b 3 j4 - b 2 j2
-b 3 j4 + b

+h

+ 3b-1

+b

1~

-b'2
b'2j2 b 3 j4

+.5

- 3b'2jt+ 3b - 1
-b 3 j4 + b 2 - b + 1

h'

Aver
age

I
-h'2j4 + h
1
1
1
-h'2j2 + 2h - 1
b 3 j4 - 3h'2j2 + 3b -

-h 3 j4

+ h2

Maxi
mum

~
~

1~

1
1

Scan fraction of directed scans for unifonn distribution of argunlents (from Iverson, 1955)

em
~

'"'"'"

4.1

Scanning 117ethods.fc)} ranking

137

are sUllllllarized in Table 4.1. Certain of the results are also plotted in
Fig. 4.2.
Initial direclionfixed (fixed scan). The ascending direction will beassulllcd.

If the scan is rooted and cyclic, the root may, without loss of generality, be
assulned to be one. The expected scan length for a set is then given by
"(h)

e ==

2:

j=

f(kJ

i == f

+ II

where j{k) is the normalized expected frequency of occurrence of the

argument k i and j' is the corresponding frequency vector deflncd by
ii == j{k j ).
The DlOSt efficient fixed rooted scan is therefore obtained by Llsing the
pernluted set a == (O/C - f))Jk such that the conlponents of a are arranged

1.0

...c

VI

D.. 0.5

0.5

1.0

1.5

2.0

Scan length b ~

Figure 4.2 Plot of cunlulative probabilities of Table 4.1 (Nul11bers refer to the
entries in Table 4.1)

138

4.1

Search techniques

in decreasing order on frequency. If the distribution of arguments is

uniform (that is,fi === llv(f)), then the expected scan length is (v(k)
1) -:-
2, and the expected scan fraction is approximately one half.
If a fixed scan is cyclic and catenated, the expected scan length depends
on the distribution f, but if the arguments are independently distributed,
then the expected scan length is independent of the ordering of the scanned
set k. This may be shown as follows. The expected length lr of a scan
rooted at r is given by Mr f, where M is the square matrix such that
Mr === (r - I) t LI . The probability of beginning a catenated scan at r is
the probability of ending the previous scan at r, that is, fro Consequently,

e === 1 ~ I === 1
Since, in general,
where N === M

M===

+ M.

t 1

1 t M~- 1 === 1 t M

f,

2e ===

then

(1 ~.

N ;~ 1),

If, for example, v(k) === 4, then

1 2

1 2

and

N===

It is easily shown that N is of the form

N === 2

v(k)1

and consequently Ng === N for any permutation p. But 1 +- N .~. f ===

(Ng); (1p ) in general, and since Ng === N, then 1 t N ~ 1 ===
.~ N -~ (1p ). Hence the expected scan e === (~) x 1 .~ N x 1 remains
the same for any permutation of1 or, equivalently, for any permutation of
the key set k.
If the arguments are not independently distributed, the analysis is, in
the general case, very complex. However, a sirnple but effective use of
correlation is made in the method of hatching. If a is a collection or batch
of uncorrelated arguments, each of which is to be ranked in the set k, then
the total expected scan time for a fixed catenated scan will be v(a) times the
expected scan time for a single item. If, however, the set a is ordered on k
(that is, a === k n a), then the entire set a may be ranked in a fixed cate
nated scan whose normalized length does not exceed one. * If a given

(1p )
(1p )

* The length of the scan will be determined by the maximum rank (in k) occurring in
the set a. The expected value of the scan fraction is approximately equal to the expected
value of the maximum occurring in a sample of size n = l'(a) chosen from the continuous
interval from zero to one. This value is known for various distributions [e.g., Cramer
(I 951) p. 370]; for a uniform distribution it is n -:- (n + 1).

4.1

Scanning 117ethods fe)J' rank ing

139

argunlent set z is not ordered on k, it nlay first be permuted by sonle sorting

process to yield the set a == pJz which is ordered on k. The set a may then
be ranked in k and, if required, the set of ranks I11ay then be subjected to
the inverse pernlutation q == p (1 L1 to yield the ranks in the original set z.
The decrease in the total expected scan length nlay far outweigh the effect
of the additional permutations required.
If a fixed scan is noncyclic and rooted at one, the expected scan length is
the same as for the fixed cyclic rooted scan. The fixed noncyclic scan is
generally unsuited to any initial points r other than one (e.g., to a catenated
scan), since the first (I' - 1) elements of k are then reached only after a
reversal of direction and a rescan of the set fir Ik.
For the case of a uniform distribution, the behavior of the fixed scan is
sUI11marized in entries 1, 3, and 4 of Table 4.1. The derivation will be
illustrated in discussing entry 7.
Initial direction gil'ing shortest scan. For a fixed root r, the mlnlnlunl
expected scan is achieved if the items are disposed on either side of the root
so that the frequency fi is a monotone decreasing function of Ii - 1'1, the
scan length in a direct scan to the argument. In a cyclic scan, the position
of the root is inlmaterial; in a noncyclic scan it is best centered at the floor
(or ceiling) of (v(k) + 1) --:- 2. In a O-origin system this expression
becomes v(k) --:- 2.
For an arbitrary frequency function, the expected scan length is given by
the scalar productf >/ 1, where 1 == ILl - rEI. For a uniform distribution,
the results are given in entries 2 and 5 of Table 4.1. For a catenated scan,
the corresponding results appear in entries 2 and 6.
The possibility of choosing the initial direction so as to give the shortest
scan to the argument depends on the infonnation available. If the elements
of k are strictly ranked on some function g(k;), then the shortest direction
from root r to argument x can, in the noncyclic case, be deternlined by a
cOlnparison of g(x) and K(k r ). For the cyclic case this does not suffice, and
it is necessary* to know the index in k of the argument.r. This case is
therefore of interest primarily in selecting a specifled elenlent fronl a
serial-access representation and is of little interest in an actual ranking
operation. However, any double-chained representation or reversible tIle
can be used in a noncyclic as well as a cyclic manner and hence adnlits of a
choice of direction which is best in the noncyclic sense.
Initial direction to nearer (farther) terminal. If the value of the root I' is
known for each individual scan in a catenated scan, the direction to the
nearer terminal can be determined by c0I11paring r with the I11id point

* Approximating functions may, however, be used for estinlating the index and the
probable best direction.

140

4.1

Search techniques

(v(k) + 1) -:- 2. A noncyclic scan starting toward the nearer end is clearly
less efficient than one starting toward the argument, but it may be useful
when the most direct route to the argument cannot be determined. The
expected scan fraction is shown in entry 7 of Table 4.1; its analysis will
illustrate the method used in constructing the entire table.
It is assumed that the number of elements v(k) is sufficiently large that
the scan fraction / may be considered as a continuous variable. Let
pr (/ ::::;: b) be the probability that scan fraction f does not exceed b, and let
the function be represented in three parts such that pr
b) == prj (j' s:: h)
in the ith half-unit interval in b. Let x be the normalized initial position of
a given scan. Then 0 ::::;: x ::::;: 1, and, since the scan always begins toward
the nearer terminal, the fraction of the set covered in a scan of length b is
the same for the starting point (1 - x) as for x. Using this symmetry,
attentioncan be restricted to valuesof x in the range 0 to ~~' For the function
pr1 ( [ ::::;: b), the value of b is also restricted to the range 0 to J.
Consider fixed values of b and x with h ::::;: J. If 0 ::::;: x ::::;: b/2, the fraction
of the file covered by a scan of length b is given by b - x, for the scan
begins at x, proceeds a distance x to the nearer terminal, and returns to
the point b - x. If b/2 ::::;: x ::::;: b, the fraction covered is clearly x, for the
scan will reach the nearer terminal but will not return past x. If h ::::;: x ::::;: t,
the scan does not reach the nearer terminal, and the fraction scanned is
therefore b. Since x is uniformly distributed, the function pr1 (j' :s: b) is
obtained by integration as follows:

c.r : : ;:

pr 1 (f ::::;: b)

==

[l

b/2

(b - x) dx

(1)

+ ~".

x dx

/1/2

l~ i

J == =-17 + b.
2

17 dx

The factor of two arises from the syn1n1etry in x and the restriction of .r
to the interval 0 :=:;: x ::::;: -!. Similarly,
pr 2 (j'::::;: b)

==

[l

b/2

1
+ l~ixdxJ== -17 +-,
2

(b - ::r)dx

')/2

-b 2

+ 3b

and
pr:3 (f ::::;: b)

2[lb-l
dx +l~i (b o
b-l

.1:) dxJ

~.
4

Entry 8 shows the behavior of the scan starting toward the farther
terminal. Although the distribution differs n1arkedly fronl that obtained
for starting toward the nearer terminal, it has the same expected value of -~'
As may be expected, the function obtained for a fixed scan (entry 4) is
the average of the functions obtained for cases 7 and 8 and is linear in h.
Case 7 (toward nearer terminal) yields the smallest maximum scan length
of the three.

4.1

Scanning nlcthods .lor rank ing

141

Initial dircction rcccrsed (continued)fronl precious scan. In the absence of

any other basis of choice, the initial direction can be chosen as a reversal
or continuation of the direction which ternlinated the preceding scan. The
behavior is shown in entries 9 and 10 of Table 4.1. The relations between
the functions for fixed scan (F), continuation (C), and reversal (R) are

and

F -

F -

== - (F - C)
== -(F - C)

:?:: 0

for b

1,

for b 2 1.

Controlled scan
The seq uence followed in a controlled scan.is comnl0nly detennined by
a conlparison which determines the relative ranking of any pair of elements
x and y in the set k. It will therefore be assumed that comparison of the
argument x == k h with the element k j determines whether h < j, h == j, or

i~l

v(k)

j +-lCi

+ k)

-:-

2J
>

k +-j - 1
i +-j

Program 4.3

Ranking of;(' in k by binary search

::=:> j. The subsequent scan may then be limited to one or other of the
two subsets a/-- 1 jk and aJjk. Themaxinluln dimension of the subset renlain
ing to be scanned is therefore minimized by choosing j == l( v(k) + I) -:- 2j.
If each subsequent element for comparison is chosen so as to (approxi
mately) halve the dimension of the set renlaining to be scanned, the
process is called binary search. Program 4.3 shows the details of binary
search ~ i and k are the indices of the tenninal elenlents of the renlaining
subset, and j is the index of the element selected for cOlnparison.
If v(k) == 2k , then anyone of 2J -- 1 different argunlents Inay be isolated on
the jth comparison, for j E ll(k), and the one remaining argunlent will be
located on the (k + 1)th cOlnparison. Hence for a unifornl distribution of
arguments, the expected number of c0I11parisons required in a binary
search is gi ven by
e,/2 /,) == (1 .2 + 2 . 2 1 + 3 . 22 + ... + k .2 /,'-1 + (k + 1)) -:- 2 /,'.

142

4.1

Search techniques

It can be shown (e.g., by induction on k) that

e,/2 k )

==

[(k - 1)27\'

+ k + 2]

--:- 27,;.

The expected number of comparisons is therefore approximately (k - 1),

and for a general valueofv(k), the number is approximately rlog2 v(k)l - 1.
The expected value thus differs but slightly from the maximum value
flOg2 (v(k) + 1)1
If e..,(v(k)) is the expected number of comparisons required in a fixed
scan of k and if r is the ratio of the execution time for one step of binary
search to the execution time for one step of fixed scan, binary search is
(for a uniform distribution) the more or the less efficient according as e s
exceeds or is exceeded by relJ' Although the simplicity of Program 4.3
suggests that the ratio r is small, it will be large if the representation of the
elements of k permits serial access only.
The methods may be combined by using k steps of binary search to
select one of 2k subsets, which is then subjected to a fixed scan. If the
remaining subset contains m elements, the (approximate) reduction in the
expected number of comparisons achieved by one further step of binary
search is esCm) - es(/nI2), and binary search should therefore be discon
tinued when e.sCm) - esC/nI2) -s;: r. For a uniform distribution, this result
yields the following approximate expression for the optimum number of
steps of binary search:

The ranking type of comparison required in determining the sequence

in a controlled scan is always attainable for any arbitrary set k or for
some permutation thereof. For, if p(kJ is the representation of k i in 7t,
if 7t ~ lOeb), if t i == (be) ~ p(kJ, and if a == (8It)Jk, then the relative
ranking of any pair of elements of a can be determined by comparing the
base b values of their representations. If, for example, b == 10, and the
four successive elements of k are represented by (1, 0, 9), (0, 6,4), (7, I, 3)
and (5, 0, 6), then a is represented by (0, 6, 4), (1, 0, 9), (5, 0, 6), and
(7, I, 3), and relative ranking in a is determined by comparing elements as
decimal numbers.
In the execution of the binary search, the calculation of the index j (next
element for comparison), and the explicit determination of the terminal
indices i and k can be avoided by associating with each element k j a pair of
indices which indicate the two possible succeeding choices for j. More
precisely, if M is a matrix of dimension v(k) x 3, whose first column is the
set k, and whose second and third columns are vectors of indices (or nulls)
from the set ll(V(k)) U (0), then Program 4.4 describes a directed scan of k.

4.1

Scanning JJICfhodsj()r ranking

.i ~ l(p(M) + 1)
.i ~ M 2
.i M 3
+-

Program 4.4

143

2J
>

Generalized binary search

The elements of M 2 and M 3 can be so chosen as to execute a binary search

equivalent to that of Program 4.3. This is true~ for example~ for the
matrix M of Fig. 4.5.
The ordering of the elements of M 1 is clearly immaterial, i.e.~ if M 1 were
permuted~ then columns M 2 and M 3 could be respecified so as to yield the
original scanning order. One consequence of this is the fact that the rows
can be reordered so that the scan conveniently begins with the first row
rather than with row lCu(M) + 1) --:- 2]. A more important consequence
is the possibility of applying the method to the problem of multiple keys~
which will be raised in the treatment of key transformations.
As illustrated by Fig. 4.5~ the matrix M specifies a tree whose nodes are
the elements of Ml~ whose branching ratios are two, and whose paths are
traced by Program 4.4. The columns M 2 and M a can clearly be chosen to
specify a scan sequence other than that of binary search. In particular, the
element M/ selected for comparison may be chosen so as to equalize (as
far as possible) the total probability of the arguments in the two resulting

bee
2

cab

cat

dog

bee
cab

3
0

5M= egg 2 7
6

foe

had 6 8

hoe

nod

egg
foe
had

Figure 4.5

Tree traced by Progran1 4.4

144

Search techniques

4.1

subsets rather than to equalize the number of elements. This procedure

yields the most efficient controlled scan. ]f the probability distribution is
uniform, the method reduces to binary search. If the argunlents are
drawn froln two or more sets having distinct probability distributions, the
matrix M may be enlarged to include two index columns for each distinct
set. Each such pair of columns may then be designed to provide an
optimum scan for the associated distribution.

4.2

KEY TRANSFORMATIONS

Since a key transformation maps the set of keys k into a set of integers
(the set of derived keys d), any unique key transformation produces a
derived key which can be used to select the component of a mapping vector
directly and thus complete a ranking operation without the use of a
scan. If the transformation is not unique, it lnay still be used to partition
the original set k into v(d) subsets for scanning and so reduce the expected
scan time. Ideally a key transfornlation should be both simple and unique
and should produce a derived set d having a narrow spread; in practice,
compromises must be made.
Let k be the domain and d the range (in ascending order) of a key
transformation t(k i ) and let e J be the equivalence class in k which maps
into d j , that is, t(x) = d j for all x E eJ. The coalescence of t ill k is then
defined as the vector c such that c j = v(e J ), for j E ll(V(d)). Since the
equivalence classes are disjoint and collectively exhaust k, then +/C =
v(k). The spread of t ill k is defined as 1 + d" - d i . Thus if k is the set
(Sunday, Monday, ... , Saturday), and if t maps each day into the rank
(in the alphabet) of its leading letter, then d = (6, 13, 19, 20, 23), the
spread s = 18, c = (1,1,2,2,1), and +/c = v(k) = 7.
The key transformation is biunique if and only if C = . Moreover, if
the transformation t is biunique, the ranking operation (i.e., the deternlina
tion of the index of the argument in k) can be completed by a nlapping
vector whose components are selected by the index j = t(kJ - d l + 1,
and whose dimension is equal to the spread of t in k. The key transfornla
tion of the preceding example is biunique when restricted to the set x =
(Sunday, Monday, Tuesday), the set of derived keys is (13, 19, 20), and
the mapping vector m = (2, 0, 0, 0, 0, 0, 1, 3) of dimension eight serves
to cOlnplete the mapping if its cOlnponents are selected by the index
j = t(x;) - 12.
A key transformation is called j-or/gin if d l = j. Since the origin can be
changed by subtraction of a constant, attention will be restricted to
I-origin transformations. The spread of a I-origin transformation is
clearly d".

4.2

Key

tran.~/()r}}latiol1s

145

A biuniq ue key transfonnation is always

attainable since, as renlarked in the treat

o
ment of directed scan, the base h value of
the representation of the elements of the
domain k can be used. The spread of such
a transformation may, however, be im- 0
practicably large. Ie for example, x were
some small subset of the set of all ten
letter sequences, (e.g., all Ineaningful ten
letter words), then s would be 2610, and
the required dimension of the associated
mapping vector would be impracticably
large.
o
In general, if each element of k is of
dimension h in 1t and if the (used) range
of each element of 1t is the set lO(h), then the
Inapping vector required is of dimension
b h The use of the base b va~ue of the rep
resentation in selecting the cornponent of
the mapping vector is equivalent to select
ing a path through a uniform h-way tree as
illustrated (using O-origin indexing) in Fig.
4.6, for h == 3 and h == 3. The branch to
the jth level is selected according to the jth
o
component of the representation.
Sequentiallevel-b/-level selection in the
tree is less convenient than the direct use
of the base b value, except that the former
2
frequently allows the full tree to be greatly
contracted. If, for example, the tree of Fig.
4.6 is used for the set k == (200, 0 I 0, 120,
2
001,022,202) (as indicated by the numeric
leaves whose values are the O-origin ranks
in k), then the full tree can be contracted
to the nonhomogeneous tree of Fig. 4.7. Figure 4.6 Uniforl11 tree and
The contraction is defined formally as 111apping vector for the set k =
follows: if the subtree rooted in a given (200,010, 120, 001, 022, 202)
node contains no significant leaves, the
subtree is replaced by a single null leaf~ if the subtree contains exactly
one significant leaf, the subtree is replaced by that leaf. The contracted
tree can then be represented by a chain list Inatrix or, since all nodes save
the leaves are null, by a leaf list Inatrix M. For the exanlple of Fig. 4.6,

146

4.2

Search techniques

3
3

0
1
2
3
4
5
6
7
8
9
10
11

0
1
6 0
3 1
1 1
4 1
9 0
0
0
0 1
0
5 1
2

M=

2
4
5

10

0 0
9

10

5
7

11

2
8

Figure 4.7 Contracted tree and associated leaf list matrix M

for the set k = (200, 0 10, 120, 001, 022, 202)

M is given in Fig. 4.7. The sequential biunique key transfonnation on

the leaf list matrix of the contracted tree is described by Program 4.8.
Program 4.8. The components of the argument x are scanned by the index j,
and step 5 determines i as the index of the current node in the path. If node i is
not a leaf, then step 6 determines k as the index of the first node reachable fronl
node i. If node i is a leaf, then k is specified as the value of the leaf, and the right
pointing exit is followed on step 8 unless the exit at step 7 is taken first. This exit
occurs only if x is an illegitimate argument which leads to one of the null leaves
(such as the last leaf of Fig. 4.7) remaining in the contracted tree. The contraction
was performed in the specified manner so as to allow the incorporation of such a
test. I f it is not required, the tree can be further contracted by eliminating the
null leaves. The left-pointing exit on step 4 also indicates an illegitimate argu
nlent x, but one of insufficient dimension for the particular path specified.

The biunique key transformation provided by Program 4.8 is very

effective when applied to a set whose dimension is small compared to the
spread of the transformation prod uced by taking the base h val ue of the
representation as, for example, in a glossary of English words. * A dis
advantage of the process is the need to revise the entire leaf list matrix
when additions to or changes in the argument set occur. The process can

* See,

for example, Lamb and Jacobsen (1961).

Key transjorl11atioflS

4.2

O-origin indexing

j+- -1

j +-j

+-

147

+1

vex)

+-

+-

MOi

Mi
1

Argument.

Current node.
Current index of argument.

j
k

+ xl'

T-

Index of first node reachable from

node i and finally the rank of x.

Leaf list matrix.

M/i

M o

1 <=> h is a leaf.

Combined leaf and chaining vector.

Legend

Program 4.8 Biunique transfornlation on key represented by x, using the leaf

list of the contracted tree of key set k.

be modified to produce a simpler but nonunique transfornlation by

contracting the tree further so that some or all of the remaining leaves each
represent two or more significant leaves of the original tree.
Nonunique key transformations
Although it is frequently impossible to obtain a sufficiently siInple
biuniq ue transformation having a sufficiently sInall spread, it is always
possible to produce a simple key transfonnation of arbitrarily snlall spread
if the requirement of uniqueness is dropped. For example, the spread of
the key transformation

+-

l( (be)

p(.r)) -:- d

varies inversely with d, but the transformation is usually nonunique for

1.

If a key transformation is not unique, the ranking must be conlpletcd by

a scan of one of the equivalence classes which it defines. The scan of each
of the equivalence classes e J may, in general, be either directed or con
trolled, and the individual subsets Inay be ordered by frequency of occur
rence, by the base h value of their representations, or by sonle externally
imposed (e.g., chronological) order. If a chained representation or fIle is
used for each subset, a directed scan is normally used.
The expected length of a directed scan of each of the equivalence classcs
e j may be computed and weighted by the relative frequency of the class to
yield an expected over-all scan length. If the distribution of argunlcnts is

148

4.2

Search techniques

uniform, the expected scan length for eJ is given by

(v(eJ)

1) -:-- 2 = (c j

I) -:-- 2,

where c is the coalescence of the transformation. Moreover, the relative

frequency of arguments from eJ is c j -:-- (+ / c). Consequently, the over-all
expected scan length I is given by

[(c

+ e) ~ cJ =
2(+/c)

[1 + C~ CJ -:-- 2.
C

~ e

For a fixed dimension of the derived set d (and hence of c), and for a
necessarily fixed value of +/C = v(k), the value of I is clearly minimized
if the coalescence vector c is uniform, i.e., if the components of c are all
equal. Hence the expected scan length is minimized by a key transforma
tion whose equivalence classes are of equal dimension.
A given key transformation is frequently employed to rank a variety of
subsets of its domain k rather than k itself. For example, if k is the set of
English words in a given dictionary, then one of the subsets to be ranked
may be the set of distinct words in a particular sample of English text. If
a particular subset of k is specified, then the coalescence of the key trans
formation in the specified subset x can be determined, and the transforma
tion can be chosen accordingly. More generally (as in the case of samples
of English text), the active domain x may be only partially specified. The
transformation should then be chosen so that its coalescence is nearly
uniform for the expected active domains. If, for example, k is the set of
all five-letter sequences, and if each active domain is the set of five-letter
sequences beginning with a specified letter, then the key transformation
used should depend only on the last four letters. If the set of derived keys
produced by a key transformation has a spread s and a random uniform
distribution within that spread, then the expected length of a scan (of the
equivalence classes) can be shown * to be 1 + v(k) -:-- 2s.

Scanning of the equiralence classes. If an element x is to be ranked in k by

a scan of k itself, no auxiliary information is required since the rank of
component k j is simply its index j. If some permutation of k is used
instead, then an auxiliary ranking vector r (i.e., a permutation vector)
nlust provide the rank in the given set k. Specifically, if y = pJk is used
for ranking, then r is the permutation vector inverse to p, and the rank of
element Yj is r j. Finally, if the vector Y is itself to be scanned in some
prescribed order other than simple cyclic order, the order may be repre
sented by a chaining vector q.
The vectors y and r or y, r, and q can be combined into a two-column

* See Johnson (1961) and Exercise 4.4.

Key tran,sjonnations

4.2

149

matrix S or a three-column nlatrix C which contains all infonnation

requisite to the ranking operation. More generally~ a collection of Inat
rices F~ V, etc., can be used, each representing some subset of the given key
set k.
The method of scanning the equivalence classes defined by a given key
transformation is largely determined by the type of Inatrix or Inatrices used
to represent the key set k. The five major methods of practical interest are
enunlerated below. Each is illustrated in Fig. 4.9. A I-origin key trans
formation t is assumed.
(a) OrerfioH'. A two-column matrix F of column dilnension s = d"
represents the first elements of the equivalence classes (and their ranks), as
follows:
]? Id J = e 1 j ,j E II ( v( d) ),

Fl i

0,

i d.

All remaining elements of the sets (i.e., aI/e') are represented in arbitrary
order in a two-column ~~overftow" matrix V.
The scan procedure is given by the Program of Fig. 4.9a. If the given
argument x is not equal to F~(r\ then the overflow matrix V is scanned in
ascending order. The left-pointing exit indicates that x k.
For a uniform distribution, the expected scan length is clearly given by
I = 1

+ (Cu( V) +

I) x (l( V) -:- 2v(k),

where (l( V) = v(k) - v(d). The expected scan length is therefore large
unless the average dimension of the equivalence classes (that is, lICk) -:- v(d))
is close to unity. For a known nonuniform distribution, the expected scan
can be reduced by placing the most freq uent element of each eq uivalence
class in F and ordering the elements in V according to their freq uency.
(b) Orerfiow with chaining. The two-column matrices F and V used in
method (a) can each be augmented by a third column chaining vector
which chains each equivalence class. Thus F~k is the row index in V of
element e/" if it exists, and is otherwise null. Similarly~ if V/I = e)i, then
V 3 h is the row index in V of e,~ :-1 if it exists, and is otherwise null. The
program is given in Fig. 4.9h. The expected scan length for a uniform
distribution can, as shown earlier, be expressed in terms of the coalescence
vector c as follows:

(c) Single table lrith chaining. In the overflow methods [(a) and (b)],
certain rows of the matrix F go unused, and a saving in storage can be

k = (Sunday, Monday, Tuesday, Wednesday, Thursday,

Friday, Saturday)
t(k i ) = 1

+ (610 n i ), where

n = (19, 13,20,23,20,6, 19)

1
2
3
4
5

F=

Friday
Sunday
Tuesday

~I

Wednesday

(n i is the rank in the alphabet of the first letter of ki)

z = (2, 2, 3, 6, 3, 1, 2), where

Zi

1
2
3

t(k i ),

V=

j +-

"*

x : Fli

4 J

j +- F 2 i

Vj

;+-0

e,..;,

i : {leV)

i
x

Overflow

F=

6
1 1
3 2
0

Wednesday

2
3

Friday
Sunday
Tuesday

I Thursday
Monday
Saturday

-4

; +- lex)

Ox:
:I

F1i

j +- F 2 i

5
7

1
2
3

4
5
6
7

T=

Friday
Sunday
Tuesday
Monday
Thursday
Wednesday
Saturday

6 0
1 4
3 5
2 7
5
4
7 :

; +- Fai

+-

~
~

~
~.

+1

: VIi

.#

~
V)

j +- V 2 i

(a)

1
2
3
4
5

Ul

Monday
Thursday
Saturday

d = (1, 2, 3,6), and s = 6.

Data of examples

t(x)

:J

i +- lex)

x: T 1'

; +- Tai

j +- T 2 i

x : VIi

Single table with chaining

i +- Vai

Overflow with chaining

(b)

j +- V 2i

(c)

em
".

em
~

1
2
3
4
5
6
7

T=

Sunday
Monday
Tuesday
Wednesday
Thursday
Friday
Saturday

1 2

2 7
3

i+-l(x)

1
2

i+-m i

3
4
5

5
6

x .
m = (6, 1, 3,

c,

0,

4)

+-

(d)

1
2

3
5
4
7

T 3i

+-

leX)
Tli

Tli

Tli

j+---TZi

Single table with chaining and

mapping vector

T=

Friday
Sunday
Monday
Tuesday
Thursday
Wednesday
Saturday

+---

/l(T)

h(i + 1)

Tli +-.r

I~

TZi

Open addressing system

construction and use of
table

+---k
+---

i}

T 2i

(e)

\~

s::;

::::;

Figure 4.9

Progran1s and exan1ples for n1ethods of scanning equivalence classes defined by a I-origin key transforn1ation t

~
~.

-go
S5

....

til

152

4.2

Search techniques

effected by combining the three-column matrices F and V in a single table

T of column dimension
fl(T) = max (s, v(k)),

where s

ll

fl(F).

Let u be a logical vector such that U/l 1(fl(F)) = d. Then ii/IF constitutes
the unused and u//F the used rows of F. Let v be a vector of dimension
f1( T) obtained by appending a zero suffix to u. The first two columns of
T are then defined as follows:

v//(a 2/T) = u//(a 2/F),

v//(a 2/T) = a2/(VP),
where p ;2 llCU( J7)) and v(p) = + Iv. (The vector p permits an arbitrary
reordering of the rows of V). The third column of T is a chaining vector
which chains each of the equivalence classes e i
The appropriate scan program (Fig. 4.9c) is sinlilar to that of method
(b), and the expected scan length is identical. The serious disadvantage of
the method lies in the construction of the matrix T -all of the rows
(specified by v) required for the leading elements of the equivalence classes
must be known before any of the nonleading elements can be allocated.
The table T is therefore usually constructed in two passes over the given
key set. Moreover, any addition to, or change in, the active key set k
which introduces a new equivalence class may occasion reallocation of
some row of T.
(d) Single table with chaining and mapping l'ector. The main deficiency
remarked in method (c) is occasioned by the fixed vector v and the fixed
order of the rows of v// T, both imposed by the given key transformation I.
The difficulty can be shifted from the matrix T to a mapping vector m
which is used (as indicated in the program of Fig. 4.9d) to effect a further
transformation of the index i = I(X). The rows of T may then be arranged
in any desired order, provided only that for each h E ll(V(d)), m i = j,
where e/L = kk = T/, and I(e/L) = i.Moreover, if T 1 = k, then the
ranking vector T 2 may be omitted.
Except for the extra step occasioned by the operation j ~ rn i , the
expected scan length is again the same as for method (b). However, the
requirement that fl( T) ~ max (s, v(k)) may now be relaxed to the form
!l(T) ~ v(k), whereas vern) must equal or exceed s. Since the squared
length of the coalescence vector (that is, etc) can, in generaL be reduced
by increasing the spread s of the transformation I, the expected scan length
can now be reduced at the cost of increasing the dimension of the nlapping
vector tn rather than at the (usually much higher) cost of increasing the

4.2

Key transformations

153

column dimension of T. A similar advantage can be gained by employing

the mapping vector m in methods (a), (b), and (c).
(e) Open addressing system. * The open addressing system employs a
single table T but requires neither the chaining vector T 3 nor the mapping
vector m. As shown by the program of Fig. 4.ge, each argument ;r is
obtained by a forwars ~~an of T 1 , beginning at component T~(.r). Since
the scan is cyclic, it is necessarily successful. It can also be made fairly
efficient by constructing T as follows. The matrix is first specified as a
matrix of nulls. The elements of k are then assigned in order, element k j
being assigned to the first unassigned row following row t(kJ - I.
The program of Fig. 4.ge describes both the construction and the use of
the table T. The branch on step 3 can occur only if the element x has not
yet been entered in the matrix T, and steps 5 and 6 then complete its entry
and the specification of the corresponding component of the ranking
vector T 2 The use of T can, in fact, proceed concurrently with its con
struction, i.e., each argument x presented to the program defines a new
entry in T if it has not previously occurred, the k index of x being deter
mined by some algorithm independent of T.
If the active argument set k is not fixed, it may be desired either to add
new elements or to respecify the rank of some element already defined in
T. Respecification may be incorporated by allowing the scan used in
defining an entry in T to terminate on encountering either the null element
or the argument itself. Although respecificati~n of an entry may be
allowed, deletion of an entry and its replacement by the null element cannot,
for the occurrence of such an inserted null element between the beginning
point i = t(k j ) and the point at which k j is entered in T would later cause
an erroneous indication that k j was not defined in T. Replacement of a
deleted entry by a special "'deletion character" distinct from the null
element could, however, be used.
The expected scan length in the open addressing system exceeds that for
Jnethod (d), since the expected length of scan of each equivalence class is
increased by the potential interleaving of elements from different equiva
lence classes. Thus, in the example of Fig. 4.ge, the expected scan lengths
for each of the equivalence classes (Sunday, Monday, Saturday), (Tuesday,
Thursday), (Wednesday), and (Friday) are (1 + 2 + 6)/3, (2 + 3)/2, I,
and I, respectively, yielding an over-all expected scan length of 16/7. The
corresponding scan lengths for a chained scan (e.g., method Cd are
(1 + 2 + 3)/3, (1 + 2)/2, 1 and I, with an over-all expected scan length of
11/7. However, since it uses a fixed scan, the open addressing system is
better suited to a serial store than is the chained system.

* The open addressing system appears to have been first used by A. L. Samuel, G. M.
Amdahl, and E. Boehm in constructing address tables for an assembly program.

154

4.2

Search techniques

If the derived keys are uniformly distributed in the range 1 to /-leT)

then, as shown by Schay and Spruth (1961), the expected scan length is
1 + p -:- 2(1 - p), where p = v(k) -:- /-leT). For a nonuniform distribu
tion, the expected scan length can be reduced by allocating the most
frequent elements first, i.e., by defining T from the set k reordered in
descending order on frequency.

Bucket files. In certain files the locations divide naturally into blocks or
buckets of n successive locations each, such that the entire contents of any
bucket can be scanned in virtually the same time required to scan anyone
location in the bucket. Such a file is called a bucket file (Peterson, 1957).
In a magnetic disc file, for example, each track forms a bucket. Each of the
foregoing methods of scanning equivalence classes can be adapted to suit
the characteristics of a bucket file. The equivalence classes can be
grouped in buckets, with chaining provided only from bucket to bucket.
Clustering. * The active argument sets of interest may be relatively small
subsets of the complete set k. Moreover, their elements commonly share
some characteristic so that a key transformation which gives uniform
coalescence and uniform spacing of the derived keys with respect to k
may yield highly nonuniform coalescence or nonuniform spacing, or both,
with respect to a given active domain x. This effect is called' clustering. If,
for example, each element of k is represented by a vector of decimal digits
of dimension ten, then the key transformation
t(x)

l((lOe) ~ p(x)) -:- I07J

yields a mapping onto the range lO(l03) which has both uniform coalescence
and uniform spacing. On the active domain x, whose elements are all
represented by vectors p(x) such that a 3 jp(x) = (2,4, 7), however, all
elements "cluster" in the single derived key 247.
The deleterious effects of such correlations among elements of the
active domain can be reduced by employing key transformations which
depend on all components of the representation and do so in a manner
which shows no systematic relationship to the structure of the representa
tion. The mid-square method, for example, consists in squaring the given
key and extracting the middle digits of the resulting product. A commonly
used transformation is the taking of residues modulo some number m
such that m ~ vex) and is either prime or contains few factors.

* Note added in proof: M. Hanan and F. P. Palermo offer an important solution

to clustering by the application of Bose-Chaudhuri codes. R. T. Chien and C. v.
Freiman have remarked a similar application of Fire codes (private communications).

4.3

Multiple keys

4.3

155

MULTIPLE KEYS*

If in some mapping operation the access to both the key set and the set
of correspondents is serial (or partially serial), considerable advantage can
be gained by replacing the vector T 2 of ranks by the suitably reordered set
of correspondents, that is, T/ becomes the correspondent of the key T/.
For, the ranking operation on the argument k j which gives access to the
element T 1 i = k j also gives immediate access to the correspondent T 2 i in
the same row Ti. This is equivalent to eliminating the permutation
operation (through reordering of the set of correspondents) and coalescing
the ranking and selection phases so that together they require a single
access to the (partially) serial memory.
For a single functional correspondence, the coalescing of the ranking
and selection phases can (by a suitable ordering of T) be accomplished by
the single-table process (Fig. 4.9c) without introducing the mapping
vector m of process (d). Frequently, however, a number of related func
tional correspondences must be provided between pairs of a family of
vectors OJ so ordered that 0/' corresponds to 0/ for all i, j, and k. In an
accounting system, for example, 0 1 , O2 , 0 3 , and 0 4 might be, respectively,
the vector of account numbers, names, addresses, and balances in a given
ledger. Those vectors which may occur as arguments in a mapping process
are called key vectors ~ those which never occur as arguments are called
satellite rectors.
may be reordered (and augmented by a suitable chaining vector) so as
to permit the use of the program of Fig. 4.9c for some selected key set Oi.
However, for any other key set OJ' the order will, in general, be unsuitable.
The program of Fig. 4.9d may, however, be used together with an appro
priate mapping vector m j and chaining vector qj. For the sake of uni
formity and the advantage of allowing an arbitrary ordering for 0, the
distinguished key set Oi may also be provided with a mapping vector m i
and treated like the rest.
The generalized binary search of Program 4.4 can be applied to the case
of multiple keys by providing a pair of chaining vectors (M2 and M 3 ) for
each key. The open addressing system is clearly unsuited to Inultiple keys.

REFERENCES
Burks, A. W., D. W. Warren, and J. B. Wright, (1954) ""An Analysis of a Logical
Machine Using Parenthesis-free Notation," Mathenzatical Tahles and Other Aids
to COlnputation, vol. VIII, pp. 53-57.
Cramer, Harald, (1951), Mathenzatical Methods o.lStatistics, Princeton University Press.

* See Johnson

(1961).

156

Search techniques

Johnson, L. R., (1961), "An Indirect Chaining Method for Addressing on Secondary
Keys," Conurlunicatiofls of the Association for ConlputiJ~!j MachinelT, vol. 4, No.5,
pp. 218-222.
Lanlb, S. M. and W. H. Jacobsen, Jr., (1961), "A High-Speed Large Capacity Dictionary
System," Mechanical Translation, vol. 6, pp. 76-107.
Lukasiewicz, Jan, (1951), Aristotle's Syllogistic fronl the Standpoint of Modern FOrlnal
Logic, Clarendon Press, Oxford, England, p. 78.
Peterson, W. W., (1957), ""Addressing for Randonl-Access Storage," IBM Journal of
Research and Developrnent, vol. I, pp. 130-146.
Schay, G., Jr., and W. G. Spruth, (1961) ""Analysis of a File Addressing Method,"
Technical Menlorandum 17-051, Advanced Systenls DevelopnlentDivision, IBM
Corporation, New York.

EXERCISES
4.1 Give a formal proof of the fact that the binary search of Program 4.3 will
not work properly if steps 5 and 6 are replaced by the statenlents k +- j and i +- j,
respecti vely.
4.2 (a) The argument k i in the set of keys k = (kI , k 2 , , kl') occurs with the
relative unnormalized frequency i. For the case v(k) = 10, design the
nlatrix M which will minimize the expected scan length when applying
Program 4.4.
(b) Show how additional keys may be incorporated in the system of part (a)
without revising the entire matrix M. Discuss the effects on the expected
scan length.

4.3

Consider the ledger L defined as

3 1 2 5 AD A M S, S. H. 0 0
B A K E R, J. C. 0 0
7 9 2 6 FOX, R. L. 0 0 0 0
3 4 2 0 FOX, R. L. 0 0 0 0
L= 1 9 2 5 H ILL, K. 0 0 0 0 0
2 4 8 6 JON E S, J. C. 0 0
9 1 2 7 JON E S, J. C. 0 0
6 1 3 5 KIN G, K. M. 0 0 0

o1 6 8

435 o ASHooooooo
76 o ELMoooooooo
435 o LAURELoooo
435 o LAURELoooo
118 o LINDENoooo
61 oMAPLEooooo 0
736 o LINDENoooo
76oELMooooooo 0

and the argument domains k I , k 2 , and k 3 , consisting of a1l4-digit decimal numbers

(account numbers), all 12-letter names (in capitals, with null fill), and all 14
character addresses (alphanumeric with null fill), respectively, and let the rows of
'a 4 jL, (4 t a I2 )jL, and w I4 jL represent the corresponding active domains xl, x 2 ,
and x 3 .
(a) Specify a simple key transformation on the set k I whose range lies in the
set lI(p(L), which, when applied to the active key set xl, yields a derived
key set of dimension three or greater.
(b) Reorder the ledger L and add a chaining vector to chain the equivalence
classes so that the resulting matrix M may be used with the key trans
formation of part (a) and a program of the type of Progranl 4.9c. Show

Exercises

157

the resulting nlatrix M and the specific progranl used for selecting the row
Mi deternlined by an argument a E Xl.
(c) Specify simple key transfornlations on each of the sets k'!. and k 3 which
yield derived sets of dimension not less than three when applied to the
active donlains x'!. and x 3 , respectively.
(d) Augnlent the nlatrix M of part (b) by a pernlutation vector and a chaining
vector suited to each of the key transfornlations of part (c).
(e) Write a program which selects, as a function of a and j, the row of M
corresponding to the argunlent a E x j , for j = I, 2, or 3.
4.4 Let t i = f(k i), where f is a key transfornlation such that t S; LI ( t'(k)). The
vector t takes on n ft distinct values (where n = l'(k)), which are assunled to be
equiprobable. For n = 2, the cases are (I, I), (I, 2), (2, I), and (2, 2), with
expected scan lengths 1, -}, I, and }.
(a) Show that the over-all expected scan length is 1.25 for n = 2.
(b) Calculate the expected scan lengths for n = 3 and for n = 4.
(c) Generalize the result of part (b) to show that the expected scan length
rapidly approaches 1.5. [See Johnson (1961) for an alternative derivation.]

4.5 Design an open addressing system for the ledger L of Exercise 4.3 and the
key transformation of part (a).
4.6 Program and discuss the extension of binary search to Ill-way search.
4.7 Let I i be the time required to scan over the ith elenlent of a vector x which
is represented in a chained (or other serially-scanned) representation, and let Ii
be the frequency of occurrence of the argunlent Xi. Discuss the role of the
"standardized frequency" s = I -:- I in deternlining the optinlUI11 ordering of the
vector x.
4.8 The ne{[{hbors problenl requires that the near neighbors of an object in
n-dimensional space be found. The technique used depends on the dinlensional
ity and the particular criteria of adjacency.
(a) The position of a vehicle on a turnpike is represented by the distance in
nliles of the vehicle from the south end, and Pi is the coordinate of the ith
patrol car. Write programs to:
(i) identify the patrol car nearest an accident at position a,
(ii) identify the two cars nearest to each other.
(b) A three-column nlatrix V specifies the locations and radio call-signs of a
fleet of nlerchant vessels on a flat lake, where VIi is the call-sign of the ith
vessel, V 2 i is its distance in miles fronl the meridian tangent to the lake on
the west, V 3 i is its distance in miles fronl the perpendicular tangent to the
lake on the south. Write a progranl to deternline the call-sign of the neigh
bor nearest to a distressed vessel whose call-sign c is given.
(c) The matrix of part (b) is used to specify call-signs and locations at til11e f
of a fleet of b0l11bers over a flat territory. When each b0l11b is released,
neighboring planes must be at a safe distance. Construct a progranl which
will find the call-signs of all pairs of planes within r nliles of each other at
time f.
(d) In a certain hydrodynamic calculation, the nl0tion of each elel11entary

158

Search techniques
volUlne of fluid is traced. The ith elementary volume is described by row Li
of a matrix: (4 t ( 3 ) 1L represents the three space coordinates (in a recti
linear system), and a7 1L represents the remaining parameters. At each
time step the parameters of volume i are redetermined by those of the four
elements nearest it. Write a program to determine a 4 1Li as the set of
indices of the four nearest neighbors of element i. Hint. Attach an explicit
index vector before sorting.

chapter

lVIETAPROGRAMS

It is frequently necessary to treat a program as an argument of sonle

process, as in the systematic analysis of a faulty program or in the trans
lation of a program expressed in one language to an equivalent prograrn
expressed in a second language. Such a process defined on prograrTIs may
itself be formalized as a nletaprogram.
Formally, a n1etaprogranl is defined as a program whose domain is a set
of programs, each element of the domain being called an argunlent
program. If the range of a nletaprogram is also a set of prograrTIs, the
lTIetaprogram is called a translator. An element of the range of a translator
is called a function program; i.e., a translator operates on argurTIcnt
programs to produce function programs. A metaprogram whose range is
not a set of programs is called an ana(vzer. Th us an analyzer prod uces, not
a function program, but data useful in the analysis or application of the
argument program. If, for example, the instructions of a corTIputer
program are sorted on the data address and listed, the list brings together
all data references to each register used and therefore facilitates analysis of
the (possibly conflicting) uses of registers in the program. A metaprogranl
which schedules and directs the execution of other progranls (and nlcta
programs) is called a director or superl'isor.
Four main types of translator are distinguished: compilers, assenlblers,
generators, and interpreters. A compiler accepts programs expressed in a
given language (the argument language) and produces corresponding
programs expressed in a second language (the junction language).
An assembler is a special case of a compiler which is limited as follows:
(1) the statements of the argument program are virtually independent and
may therefore be treated one at a time, and (2) the statements of the
argument program are simple (not compound) and need not be analyzed
into component statements. There usually exists a fixed correspondence
between the operations of the argument program and those of the function
progranl; the translation thus consists essentially of a substitution of
symbols for the operations and/or operands.
A generator produces, by specialization, anyone of a fanlily of function
159

160

5.1

Metaprograms

programs. Thus a single generator might, for each specified value of n,

produce a specialized program for evaluating the function :r n . The
argument of a generator is usually viewed as two distinct parts; the
skeleton prograrn, \vhich determines the family of potential function
programs, and the specification, which determines the particular melnber
of the family produced. Generators are frequently incorporated in
compilers.
A translator is called an interpreter if it (1) executes the segment of
function program corresponding to a statement of the argument program
immediately after it is produced, and (2) selects the statements of the
argument program in a seq uence determined by the execution of the
function program. The function program itself is normally treated as an
intermediate result, and only the outputs of the executed argument program
are retained. The execution of an interpreter can therefore be viewed as
follows: each statement of the argument program is first "interpreted"
in the function language and is then executed.
The trace progranl and the utility progranl are special cases of the inter
preter. The former executes the argument prograrrl without modification
but produces, in addition to the normal outputs of the argument program,
a trace of the argument program listing each instruction executed and the
intermediate results produced. In a narrow sense, a utility program is an
interpreter whose argument program is supplied directly by a computer
operator via the control console. More broadly, the term is used to denote
any program frequently used by the computer operator.
The present discussion of metaprograms will be limited to one important
aspect of compilers, the analysis of cOlnpound statements occurring in the
argument program. The treatment embraces the translation between the
common parenthesis notation and the Lukasiewicz (1951) notation which
proves most convenient in the analysis of compound statements.

5.1

COMPOUND STATEMENTS

Each statement of a program can be considered as an operator which

maps the given operands into the specified result. If the operator corre
sponding to a given statement belongs to a given set of operators p, the
statement is said to be elementary in p. A finite program whose operators
all belong to a set p is called a program in p. A statement which is not
elementary in p but which can be expressed as a finite program in p is said
to be compound in p. The analysis in p of a compound statement is the
specification of a corresponding finite program in p.
For exanlple, the statement

z +- (x

+ y)

+ (s

- t)H

161

CO/l1pound staten1ents

5.1
is compound in the operator set

p = (addition, subtraction, multiplication, exponentiation),

and Program 5.1 a shows one possible analysis in p. Program 5.1 b shows a
similar analysis in the set

(addition, subtraction, nlultiplication, branch).

A metaprogram that translates all statements which are eleInentary in a

set of operations p can be extended to translate statements which are

-)

a+-x+y

a+-x+y

b+-axr

b+-axr

+-

s - t

+-

en

d+-1

z+-b+d

+-

s - t

i+-n

<

i+-i-1

z+-b+d

Program 5.1

Analysis in p

Analysis in q

(a)

(b)

Analysis of the conlpound statenlcnt ,: -(- (.r

+ y)

x r

+ (s

f)"

compound in p by the addition of a metaprograIn for analyzing cOIn pound

statements. The conventions adopted for the representation of conlpound
statements must, of course, be complete and precise so that interpretation
is unequivocal. These conventions should be fanliliar and convenient to
the progranlmer and should also permit easy analysis by a nletaprograIn.
The comInon parenthesis notation of elenlcntary algebra is congenial to
programmers, whereas statements in Lukasiewicz notation are easier to
translate and evaluate, easier to transfonn to an optinlunl fornl which
minimizes the amount of intermediate data storage and execution tinlC
required in their evaluation or translation, and possess the sinlple criterion
for well formation developed for the left list matrix of a tree in Sec. 1.23.
The analysis of conlpound statements will therefore be discussed in ternlS
of Lukasiewicz notation, and algorithnls for translating between the
parenthesis and Lukasiewicz notations will be presented.

162

5.1

Mefaprogralns

Tree

(a)

2
2

/\

/\

/\

(/\, V, p, 1\, q, r, V,

/\

Left list vector or Lukasiewicz form

(d)

S, -,

((q /\ r) V p) /\ (s V 7)
Parenthesis form

Full left list

matrix
(b)

Figure 5.2

Left list
matrix

(e)

(c)

Representations of a conlpound statenlent

t)

5.2

Lukasiewicz notation

163

Figure 5.2 shows several alternative representations of a compound

statement. The tree representation (5.2a) is perhaps the most graphic, and
the other forms may be viewed as alternative representations of the tree.
The common parenthesis form of Fig. 5.2e, for exalnple, specifies the
requisite structure primarily by grouping the nodes of each subtree within
parentheses with the root in the middle of the group. As shown in Sec.
3.4, the left list nlatrix of Fig. 5.2c can be supplanted by the simpler left
list vector of Fig. 5.2d, providing that the degree of each operator p is a
known function 6(p).
The left list vector notation for a COIn pound statelnent is also called
Lukasiewicz, * Polish, or parenthesis~ff'ee notation. The Lukasiewicz and
the parenthesis notations will hereafter be referred to asY:;-notation and
Y-notation, respectively.

5.2

LUKASIEWICZ NOTATION

Although !f-notation can be viewed as the left list vector of a tree

representing a compound statement, it is helpful to develop an alternative
equivalent formulation as follows. Let I and p be two disjoint sets whose
elements are literals and operators, respectively, and whose union v ==
1
P is called a l'ocabularJ'. A strictly positive integral degree function
6(p) is defined on each element of p, and each operator p of degree d
accepts d elements of 1 as operands to define a result or value in 1. In
synlbolic logic, for exanlple, 1 == (0, 1), P == (l\, V, -), 6( A ) == r)( V) ==
2, and ()( -) == 1. Consistent with these notions. the degree of each literal
is defined to be zero. t
Each operator p of degree v( q) defines a function (i.e., a nlapping) fronl
each vector q s; 1 into an elenlent y E 1. This function is denoted by the
vector f == (p) ffi q. The vector f is called an !//-phrase (~l/ength v(I), and
the elenlent y is called its ra/ue. Table 5.3 shows, for exanlple, theY)
phrases in the system for symbolic logic based on and, or, and not. The
vector f is clearly contained in the vocabulary v, that is, f s; v. Where its
omission raises no ambiguity, the prefix will be dropped fronl the tenn
"Y-phrase" and from similar terms to be defined.
A vector z s; v is called an !~rornlu/a (~llength l'(z). In particular,
every phrase is a formula. The degree rector of a fornlula z will be denoted
by 6(z) and defined by the relation (r)(z))i == ()(Zi).

First proposed by Lukasiewicz (1951) and first analyzed by Burks et al. (1954).

t The system is extended to include variables as well as literals by considering the

vocabulary v = x
1
p, where x is the set of variables, and x n (l
p) = e(O).
The degree of each variable is, like that of a literal, defined as zero. The domain of the
operators is still confined to the set I, and in any legitimate algorithm each variable is
specified as a literal by some statement before it enters as an argument.

164

5.2

A1 etaprogranls
Phrase

Value

(V, 0, 0)
(V,O, 1)
(V, 1,0)
(V, 1,1)
(/\ , 0, 0)
(/\,0,1)
(/\,1,0)
(/\,1,1)
(-,0)
(-,1)

1
1
1

Table 5.3 Phrases in~-systeln for syn1bolic

logic based on operators and, or, not

If sonle infix of z is a phrase, and if the infix is replaced by its val ue, then
the resulting vector y is called an2-reduction of z. If y is any Y-reduction
of z which cannot be further reduced, it is called a conlplete reduction of z
or, since it can be shown to be unique, the complete reduction of z. Com
plete reduction of z will be denoted by 2(z). A formula z is said to be
singular* if its complete reduction is a single literal, i.e., if l'(Y(z)) = 1 and
2(z) E 1. Thus q = (/\,1, V, 0,1) and r = (1) are singular, but s =
(/\ , 1, 1, 0) and t = (/\ ) are not.
For exanlple, complete reduction of the singular formula z = (/\ , ,1,
/\ , 0, I, V, 1, -, 1) may be performed as follows:

Zl

= (/\, V, 1, /\, 0, 1, v, 1, -, 1)
= (1\, V, 1, /\, 0, 1, v, 1, 0)

Z2

Z3

Z4

Z5

(/\,

V, 1, /\,0,1,1)

( /\, V,

1, 0, 1)

------

(/\,1,1)

Program 5.4 shows the complete reduction of a formula z, including tests

of singularity.
Program 5.4. The components of the given forn1ula Z are examined in reverse
order and assembled into a stack vector y = (Zi, Zi+l, . . . , z,,), where Zi is the

* The term well fOr/ned used by Burks et al. (1954) and others is avoided here because
singularity implies not only that the formula represents a well formed tree but also that
the tree is singuldr.

Lukasiewicz notation

5.2
+-

+-

s+-O

+-

e(O)

+-

v(z)

+1

i+-i-l

(Zi) CD y

(5(y}) : 0

I-origin indexing

(v(y) = 1)

+-

+-

j :

bey})

Given formula.

Reduced suffix of z.

Index of

Dimension of phrase to be
reduced.

Singularity indicator.

+1

~v -- ~ ( a. j / y)

11

y +-(x) CB aj/y

in descending scan.

Legend

>

v(y)

10

165

Program 5.4

Evaluation of the formula z

conlponent currently exanlined. When an operator (i.e., a node of nonzero de

gree) is first encountered, the prefix of y fornls a phrase of dimensionj = ()(Yl) +
I, which is inlmediately reduced (i.e., evaluated) on step 10 and is then replaced
by its reduced value on step II. Singularity of the vector z is indicated by a non
zero value of s, which is set to one only if the exit occurs fronl step I with v(y) = 1.
The case ~'(y)
1 can occur if the fornlula represents a well fornled but nonsingu
lar tree, i.e., if the formula contains two or more singular formulas. The exit
fronl step 9 occurs if the indicated dimension of any phrase exceeds the current
dimension of y and leaves the indicator s at its initial zero value.

The singular fornlulas are clearly the meaningful compound statements

in the system. Moreover, if L 2 is a singular fonnula and if L 1 === ()(L'!.),
then L is the left list of a singular tree. The singularity of a given formula
z can therefore be determined froln its associated degree vector d === r5(z).
The necessary and sufficient condition for singularity of the associated tree
is simply v(d) - (+ jd) === I. As shown in Sec. 1.23, the necessary and
sufficient condition for well formation is that all components of the suffix
dispersion vector s defined by*

s === (I

~) ~ ( -

d)

must be strictly positive. The nlaxinlum over the components of s will be

called the fnaximunl suffix dispersion 0.[ z.

* The suffix dispersion vector describes the dispersion (nunlber of roots) of all suffIxes
of

Z,

as may be more easily seen from the alternative expression

5

= ~'(aj /d) -

(+ /aJ)d).

166

5.3

Metaprogran1s

5.3

THE MINIMAX FORM OF AN 2-FORMULA

Two formulas are said to be equiL'alent if they have the same value for
each possible specification of their variables. If z is any formula whose
operators are all symmetric (i.e., whose operands commute), then any
reordering of the component singular formulas of z which leaves the span
of each operator unchanged leads to an equivalent formula. For example,
since the operators /\ and V are symmetric, the formulas
z = (/\, V, /\, V,

and

= (/\ ,p,

V,

}i', /\,

q, /\, r, V,

u, v, r, q, p)
}i', /\,

u, v)

are equivalent, as may be easily verified. In the tree representation this

reordering appears as a reordering of the group of subtrees rooted in a
common node.
A formula whose maximum suffix dispersion is minimal with respect to
the set of all equivalent formulas is said to be in minimax form.
The dimension of the stack vector y employed in the evaluation of a
formula z (cf. Program 5.4) takes on successive values equal to the number
of roots in the tree represented by the suffix of z currently scanned. It
therefore assumes the values (in reverse order) of the components of the
associated suffix dispersion vector s. The maximum dimension of the
stack vector determines, in turn, the amount of auxiliary storage required
in the evaluation of the formula or in the compilation of a function
program for its evaluation. It also detern1ines, in part, the number of
transfers of intermediate results to and from storage in evaluating the
formula in a computer having a limited number of central registers. A
formula in minimax form minimizes the maximum dimension of the
stack vector and is therefore to be preferred.
The transformation of a singular formula z to equivalent minimax form
is based on the following theorem: if each of the O(Zl) component singular
formulas of (iljz is in minimax form, then the entire formula can be
brought to minimax form by arranging the component singular formulas
in ascending order on their maximum suffix dispersion.
For example, if z = (1\, V, /\, a, b, /\, C, d, V, e,!), then <5(Zl) = 2,
and (iljz contains two singular formulas, yl = (V, 1\, a, b, 1\, C, d), and
y2 = (v, e, f), each in minimax form and possessing maximum suffix
dispersions of 3 and 2, respectively. Moreover,

q = (Zl) ffi y 2 ffiyl =

(1\, V,

e,j,

V, 1\,

a, b,

1\, C, d)

is an equivalent formula in minimax form, with a maximum suffix dis

persion of 3 as compared with a value of 4 for the same function of z.

5.3

The Hlininlax .!()nn

(~l an YJ~f()rnlula

167

To establish the theorenl, let z be any singular fornlula, let r(z) be its
dispersion (that is, the number of roots), let s(z) be the Inaxinlunl suffix
dispersion of z, let d == b(Zl), and let

be the unique (cf. Program 3.9) partitioning of (iljz into its conlponent
singular formulas. Then yi
yi-tl
yi I,' -1 represents a k-tuply
rooted tree and r(y! E8 yi~l
1) == k. Moreover,

== max

[S(yl)

+d-

1,

S(y2
yff)J

== Inax [s(y])

d -

jJ

jJ.

j=l

==

max [s(yJ) j=l

Since the component formulas are in minimax fonn, the s(yJ) are individ
ually minimal, and the maximum over S(yi) - j is clearly mininlized by
arranging the s(yj) in ascending order. This concludes the proof.
To ensure that each component fornlula is itself optinlal, it suffices to
apply this reordering procedure in turn to the successive singular fonnulas
encountered in scanning the given formula from right to left, as shown in
Program 5.5.
Program 5.5. The vector y is the suffix al'!z pernluted to optinlal [orI11, p is its
partition vector* (that is, ((p; a j - 1 ) 1aPj)/y is thejth singular formula of y), and
g is its maximum suffix dispersion vector (that is, gj is the nlaxinlunl suffix dis
persion of thejth singular formula of y). The nlain control parameter i is decre
mented on step 11, and, if it is not zero, the degree d = ()(z i) of the next
component to be added to y is examined. If Z; is not an operator, the branch
to step 8 occurs with h = I. The component Zi is then a fornlula of length 1
and steps 8-10 add it to y and nlake the appropriate changes in p and g.
If Zi is an operator (of degree d), the loop 15-22 scans the vector g and reorders
the first d component formulas of y so that their nlaximunl suffix dispersions are
brought to ascending order. This is accol11plished by the sinlple, but not necessar
ily efficient, sorting process of conlparing successive pairs of adjacent conlponents
of g and interchanging the corresponding C0l11pOnent fornlulas of y (by rotation

* The conventions used for p are those established in the subtree partitioning of
Program 3.9.

168

5.3

M etaprogranls

I-origin indexing

s+-O

+-

E(O)

+-

(0)

+-

(0)

+-

v(z)

+1

+-

ad /g

+-

+-

(h) ED ad/g

+-

((p tad)

10

+-

(zJ ED y

;;j=

Partition vector of y;
P j is the dimension of
the jth singular form
ula of y.

gj

Maximum suffix dis

persion of the jth
singular formula of y.

Degree of Zi'

Singularity indicator.

L1(d)

+ ((r1)/1)1
+ 1) 8) -ad/p

l+-i-l

12

+-

(v(p)

13

+-

O(Zi)

14

d : v(p)

15

+-

Legend

= 1)
>

h+-h+l

16
17

The transformed suffix

-ai/z.

11

Given formula.

>

18

gh

19

+-

(p

20

+-

\ti/y,

:::;

gh+l

21

gh~gh+l

22

Ph~Ph+l

Program 5.5

a h 1) ,}
U,

aPh +Ph+l

Ph t (u/y)\

Transfornlation of the fornlula

to nlininlax fornl

of the infix representing the pair) if an interchange is required. Steps 21 and 22

effect the corresponding interchanges in the vectors p and g. The loop is ternli
nated by the branch fronl step 17 to step 6, the first d fornlulas of y (fornling the
prefix a)/ /y, where n = ( + lad /p)) are then in ascending order on their maxinlunl
suffix dispersions, and the new fornlula (Zi)
all/y is therefore in optinlal fornl.
Its nlaximuIll suffix dispersion is conlputed by steps 6 and 7 and replaces the
prefix ad jg (step 8) so that g beconles the nlaximuIll suffix dispersion vector of

5.4

Translation fron1 c0l11plete parenthesis

169

y. The partition vector p is respecified by step 9. The one conlponent of g

remaining at the conclusion is the maxinlum suffix dispersion of the optinlized
statement y.
(zJ

The minimax form of a formula is, in general, not unique. It can be

made unique, however, by using some assigned orderings of the operators,
literals, and variables (e.g., the ordering specified by the vocabulary
v= x
I ffi p) as a minor category in the reordering of cOlnponent
formulas. Such a unique form is helpful in detecting the occurrence of
equivalent compound statements within a formula, with the ainl of
obviating repeated segments in a corresponding function progranl.

5.4 TRANSLATION FROM COMPLETE

PARENTHESIS TO LUKASIEWICZ
NOTATION
Ordinary parenthesis notation is complicated by the occasional or
consistent use of certain conventions for eliding parentheses. For example,
the expression
(x + (y x z))
may also be written as
(x

+y

z)

by the convention that nlultiplication takes precedence over addition, or as

x

+y

x z,

with the understanding that the entire expression need not be enclosed in
parentheses.
The problem posed by the use of such conventions can be segregated by
considering a cOfnplete parenthesis notation in which all inlplied paren
theses are included, i.e., in which each operator and its associated operands
are enclosed in parentheses. The analysis of a statelnent in parenthesis
notation can therefore be performed in two steps, a translation to conl
plete parenthesis notation according to the prescribed conventions, fol
lowed by the analysis of the resulting statement. The present discussion
will be limited to expressions in conlplete parenthesis form.
The conlplete parenthesis notation will be referred to as ;~-notation and
the terminology adopted for Y/-notation will be extended analogously.
Thus, z = ([, [, x, +, y, ], x, r,]) is a ;~-formula more conlnlonly
denoted by (x + y) X r. To avoid confusion with the nonnal use of
parentheses (e.g., in enclosing vectors), brackets will be Llsed (as in the
foregoing vector z) to represent the cfJ-notation.

170

5.4

M etaprogranls

The discussion will be limited to a system of unary and binary operations

only (i.e., ()(p) = I or 2) and will again be illustrated by the systelTI of
logical operations. Assuming a vocabulary of the form v = x ffi 1 ffi p c)
([, ]), the rules of composition for :YJ-notation may be formulated as
follows:

1. If rEX (B 1, then v is a singular &-fornlula.

2. If z is any singular formula, if U E p, and if ()(u) = 1, then (D (B (u) 8j
z ffi (]) is a singular :3'-form ula.
3. If Y and z are both singular formulas, if b E p, and if ()(b) = 2,
then (D
y ffi (h) 8) z
(]) is a singular &-formula.
In particular, ([, y, /\, z, ]) is singular but (y, /\, z) is not; (y) is singular
but ([, .If, ]) is not, and ([, -, y, ]) is singular but (-, y) is not.

7T(W) +-

I-origin indexing

A2

:c +- <!)1

<D 2 Output in eSe-notation (reversed in order).

:c

<1>3

Auxiliary stack file.

Set of operators.

<1>1

<I)2+(x E p) +-

<1>2 +-1<1>3

Program 5.6

:c

Input in .9'-notation (terminated by 0, A2).

Legend

Translation fron1 con1plete parenthesis to Lukasiewicz notation

The translation frolll &- to 2-notation can be performed with the aid
of one auxiliary file or stack vector. Program 5.6 shows a suitable process
which will correctly translate any singular formula, but which includes no
tests for singularity. It is noteworthy that all left parentheses are simply
ignored. A sinlilar process can, of course, be designed to use only the left
parentheses and to ignore all right parentheses. * Any translation which
tests for singularity clearly must use all parentheses. If the !f- and &
notations enlploy different sets of operator symbols (e.g., /\, V,
, and
x, +, r--;), the appropriate translation can easily be incorporated in the
progranl.
Program 5.6. The original staten1cnt is assun1ed to be recorded on a file <1>1'
with partitions Al following each syn1bol and with a null itenl and partition A2 at

* Oettinger
plete.

(1960) analyzes three types of parenthesis notation: left,

r(~ht,

and

COfll

5.4

Translation fr0l11 c0l11plete parenthesis

171

the end. File <1>2 receives the resulting.!f-formula in reverse order, i.e., fronl right
to left. The stack file <1>3 receives each operator synlbol as it is read from <1>1 and
transfers thenl one at a time in reverse order (i.e., by a backward read) to <1>2 at
each occurrence of a right parenthesis. A trace of the progranl shows, for
example, that the :3J-formula ([, [, x, V, y, ], !\ ,[, - , z, ], ]) translates correctly
into the ~-formula (:r, y, V, z, - , 1\ ) reversed fronl normal order.
A partial test of singularity can be provided by testing each component
for compatibility with its predecessor, the ordered pair being declared
compatible if and only if it can occur in some singular formula. For
example, an operator may be followed by either a left parenthesis or a
variable, but not by a right parenthesis or another operator. These first
order compatibility constraints can be expressed, in terms of the following
classes: left parenthesis, unary operator, binary operator, variable or
literal, and right parenthesis. These classes will be denoted by [, u, b, l\ and
], or alternatively by 1, 2, 3, 4, and 5. The constraints are sumnlarized in
the nlatrix M of Program 5.7.
The test of singularity provided by the first-order constraints is not
complete, * but can be conlpleted by the following expedient.-j- The
auxiliary file which receives the operators (file <1)3 of Program 5.7) also
receives the left parentheses in their turn. The following tests are then
added:
1. Each operator is accepted and replaces the previous entry in the
auxiliary file if and only if the previous entry is a left parenthesis.
2. The transfer of one component from the auxiliary file to the output
file normally occasioned by the appearance of a right parenthesis is
accepted by the right parenthesis if and only if the component trans
ferred is an operator.
3. The possible exhaustion of the auxiliary file is tested each time it is
read.
The first test prevents the acceptance of two successive operators without
an intervening left parenthesis. At each application of the test, the corre
sponding left parenthesis is removed from the file. Since the auxiliary fIle
may now contain left parentheses as well as operators, the second test is
required to prevent their acceptance as operators. The complete testing
and translation process is described by Program 5.7.
Program 5.7. The current conlponent .r is read fronl the input file on step 8 and
its class k is deternlined before the repetition of the nlain loop at step 5. Step 5

* The tests provided in conlpilers have frequently been lirnited to essentially this type.
See, for example, Carr (1959).
t This procedure is due to Oettinger (1960).

172

5.4

M etaprogralns
I-origin indexing

7T{<1

s+-O

<1>3

j+-2

[1

Mi.
k . 0

u2

<1> 2+(k<4) +-

j+-k

+-

<1>1

+-

<1>3

10

+-

11

+-

12

+-

[ u b v ]
1 2 345

0, Al

13

Al

14
15
16

A
i=

+-

o(x)

x : ]

20
21

+1

P : [

18

]5

Al

P +-1<1>3

k +-4

b3
v4

17

19

1 101 0
10010
10010
00101
00101

First-order
compatibility
constraints.

= 1.

Mlci

j accepts k:,>Mki

<1>3

Auxiliary stack file.

<1>2

Output in se -notation
(reversed order).

<1>1

Input in Y'-notation

(terminated by 0, AI)'

Class of previous component.

Class of current component.

Set of operations.

b(X)

Degree of operator x.

Singularity indicator.

Legend

+-

+-1<1>3

x:p

Program 5.7 Translation fronl conlplete parenthesis to Lukasiewicz notation

with full checking of singularity
determines the first-order compatibility of k with its preceding value j. (The
singularity indicator s is set to unity only at the exit on step 10.) Each conlponent
occasions the recording (step 6) of one itenl on one file-the auxiliary file (1)3 if
k
4, or the output file (1)2 if k
4. The item recorded is the current conlponent
unless it is a right parenthesis. I n the latter event, the variable .r is first respecified
(step 20) by a backward read from the stack file. The test on step 21 assures that
the itenl read is an operator.
] f the current component is an operator, the previous item recorded on the
auxiliary file nlust first be read, conlpared with ""[," and discarded. This occurs

5.5

Translation .lioln Lukasiewic2

173

on steps 15 and 16. The exits on steps 15 and 20 indicate nonsingularity due to
the exhaustion of the stack file. Step 9 provides a final test to ensure that the
stack file is exhausted when the input file beconles exhausted. Since the first
component of any singular fornlula must be either a variable, literal, or left
parenthesis, the initial setting ofj to 2 on step 4 provides the appropriate initial
conlpatibility condition.
Each of the translation programs considered produces the resulting
2-fornlula in reverse order. This is the order in which it is nlost easily
evaluated and, consequently, the order in which the synthesis of a corre
sponding function program is most easily performed. Synthesis nlay
therefore proceed concurrently with analysis. The analysis Inay, on the
other hand, be completed first, and the output file (1)2 rewound before
beginning the synthesis. The latter alternative allows the use of separate
metaprogranls for analysis and synthesis, and hence makes lesser denlands
for nletaprogram storage. It also allows the application of an intervening
transformation of the ,2-formula to some preferred equivalent fornl.
However, as shown by Program 5.4, the transfornlation to nlininlax I'ornl
also treats the Y-formula in reverse order. It can therefore be perforIned
concurrently with the translation from parenthesis notation.

5.5 TRANSLATION FROM LUKASIEWICZ TO

COMPLETE PARENTHESIS NOTATION
The inverse translation fronl Lukasiewicz to conlplete parenthesis
notation is, unlike the evaluation of the Lukasiewicz forn1ula, best
performed by a forward scan. The suffix dispersion criterion of singularity
must then be applied in the following way. The dispersion of the entire
stateIl1ent is assun1ed to be one, and the dispersions of successively shorter
suffixes are obtained by subtracting (1 - ()(.r)) for each succeeding con1
ponent .r. The suffix dispersion thus COIl1puted n1ust reach zero when and
only when the remaining suffix is null ~ if not, the staten1ent is nonsingular.
The translation of Program 5.8 provides cOlnplete checking of singularity.
Program 5.8. The resulting jP-fornlula is produced on file (I):! in reverse order.
Each operator encountered is recorded in the auxiliary file together with a preced
ing left parenthesis, and it also causes a right parenthesis to be recorded in the
output (steps 9-11). Each variable encountered is recorded (step 12) in the output
file and initiates a transfer fronl the auxiliary file to the output file which terrni
nates (step 19) only when an operator of degree two is encountered or (step 13)
when the file beconles exhausted. I n the latter event, steps 14 and 15 are executed
as a final check on singularity-exhaustion of the stack file, exhaustion of
the input file, and the first zero value of the suffix dispersion III nlust occur
together.

174

Metaprograms
rr((I)

+-- E

<1>3 -(- (),

s+--O

4
5

nl

A2

+--

Al

tll +--

m - (1 - (5(x))

;{'

(D

11

cI>2 +- ]

12

cI>2 +-- X

+--1<1>3

Al

15

+--

Input in Y-notation
(terminated by 0, A 2).
Output in :~-notation
(reversed order).

q)3

Auxiliary stack file.

Set of operators.

Suffix dispersion assuming

total weight is unity.

Singularity indicator.

+- [

10

13

<1>2

:p

(})3 +-- X

14

(D!

X +-- (D!

I-origin indexing

5.5

Legend

Al

cD 1

s+--l

16

17

<D 2

18

19

(5(p)

+--

P
-=F

Program 5.8

Translation from Lukasiewicz to complete parenthesis notation

with complete test of singularity

REFERENCES
Burks, A. W., D. W. Warren, and J. B. Wright, (1954), "An Analysis of a Logical
Machine Using Parenthesis-free Notation," Mathematical Tables and Other Aids to
COll1putation, vol. VIII, pp. 53-57.
Carr, J. W., III, (1959), "Digital Computer Programming," Chapter 2 of Grabbe, Ramo,
and Wooldridge (Eds.), Handbook 0fAutonlation, Conlputation, and Control, vol. 2,
Wiley, New York.
Lukasiewicz, Jan, (1951), Aristotle's Syllogistic from the Standpoint of Modern Formal
Logic, Clarendon Press, Oxford, England, p. 78.
Oettinger, A. G., (1961), "Automatic Syntactic Analysis and the Pushdown Store,"
Proc. Twe((th SYfnpositlfn in Appl. Math., April 1960, published by American
Mathematical Society.

Exercises

175

EXERCISES
5.1 For each of the follo\ving logical functions, exhibit an equivalent (i) tree,
(ii) .5I'-fornlula, and (iii) Y-fornlula:
(a) ,l(~r, y, z) = :r /\ (y V z).
(b) g(w, x, y, z) = (w V (y
z /\ (x V (w = y.
(c) the function of part (b), limiting the operators elnployed to and, or, and nol.

5.2 Let,l(:r, y, z) be a logical function of three variables, let q be an equivalent

formula in :JP-notation, and let r be an equivalent fornlula in YJ-notation. Write
progranls to determine the intrinsic vector i(t') (cf. Sec. 7.2) as a function of
(a) the :1'-formula q.
(b) the 2"-fornlula r.
5.3 Leta=(/\,-,O,V,I,O)
b = ( , 0, V, 1, 0)
c = (/\, V,O, V~ 1,0)
d=(/\, V,q,f, -,I)
e = ( V, q, r,
, I)

q = ([, [, q, \/, 1',

r = ([, q, V, f, ],
Trace
(a)
(b)
(c)
(d)
(e)

], /\ , [, - ,

/\ ,[, - ,

I, ], ])

I, ], ])

the operation of
Progranl 5.4 for each of the cases z = a, Z = b, and z = c.
Program 5.5 for z = d.
Progranl 5.6 for (1\ containing q.
Program 5.7 for (}\ containing q and for q)l containing r.
Program 5.8 for (1\ containing d and for q\ containing e.

5.4 Write a program for translating from :5I'-notation to Y-notation which is

analogous to Program 5.6 except that it ignores right rather than left parentheses.

5.5 Write a program to extend the mininlax transformation of Progranl 5.5 to

the case of an operator set of the form p = pI
p2, where pI and p:!. are a set of
symmetric and asymmetric operators, respectively.
5.6 Write a program which extends the nlinimax transfornlation of PrograITI 5.5
to include ordering on the variables and operators so as to bring the fornlula to
unique canonical form, as suggested in Sec. 5.3.
5.7 Write a progranl which will recognize all identical singular subfornlulas
occurring in a singular ,~-fornlula z and which will produce a record of the
associations in some convenient forn1.

chapter

SORTING

The order in which a set of items is arranged in a large-capacity store

often has a marked effect on the simplicity and speed of execution of
algorithms defined on them, and it therefore becomes necessary to sort or
rearrange groups of items.
The problem of sorting may be described as follows: given a vector a,
determine the ordering vector p = 6j(k(a)) and the permuted vector
c = pJa, where k(a,J is a numeric function defined on the components of
a, and k(a) is the vector defined by (k(a))j = k(aJ. The function k is
called the key of the sorting process, and k(a) is called the key vector
associated with a. The key function is frequently an index in some set b,
that is, k(a) = b l a. The components of a will also be called items; since
the vector a is commonly represented in a file, it will also be called a file.
Most sorting processes determine the ordered vector c= pJa without
explicitly determining the permutation vector p. A sorting process which
explicitly determines and uses the permutation p is called an address table
sort.
Sorting processes fall into two major classes, called serial- or random
access, according to whether the files used to represent the original and the
intermediate vectors produced are serial-access or not. Random-access
processes are also called internal, for they are normally performed in the
"internal" storage of a computer. Combinations of serial processes and
internal processes are used, but the two types can and will be described,
analyzed, and evaluated independently.

Input

Output

Name of Process

Single
Single
Multiple
Multiple

Single
Multiple
Single
Multiple

Duplication
Classification
Merging (or Merge)
Revision

Table 6.1

Types of file operations

176

6.1

Serial sorting fnethods

177

Four types of operations on files are distinguished, according to whether

one file or several files are used in input and in output. They are shown in
Table 6.1. A classification (merge) involving nl output (input) files is called
an nl-way classification (merge).

6.1

SERIAL SORTING METHODS

Copy operations

A serial sorting process is executed as a sequence of copy operations.

A copy operation is defined as follows: all items from a given set of input
files are transferred to a given set of output files, and each item read froIn
any input file must be transferred to some output file before a further itenl
is read from the same input.
Input
Files

Item
Positions

Figure 6.2

Output

Files

Copy operation

Figure 6.2 provides a graphic description of the copy operation. The

PI' P2' ... 'Pm denote storage for nl items, the kth input
file may be read to the kth item position only, and the occupant of any
item position may be transferred to anyone of the output files. Once an
item is so transferred, the evacuated position may be refilled by the next
item from the corresponding input file. Any copy operation can clearly be
performed by serial input files and serial output files, with each itenl being
read and recorded but once, and with no auxiliary repositioning of the files.
A copy operation is also called rank-preserl'ing, since it satisfies the
m components

178

6.1

Sorting

following definition of that term. [The concept of rank preservation and

its use in the analysis of sorting processes were first introduced by
Ashenhurst (1953).] Any item may be specified by a pair of coordinates
(f, r), the file coordinate.f designating the file to which it belongs, and the
rank coordinate r designating its rank (i.e., its index) in the file. Two items
with initial coordinates (11' r 1 ) and (j~, r 2 ) and with final coordinates (11" r/)
and (/2" 1'2'), respectively, are said to be relatable if and only if j~ = .f2 and
1/ = 12' An operation is called rank-preserving if precedence relations
are maintained for all relatable items, i.e.,
1'1

<

r2 >r/

< r2'

Henceforth the terms nlerge and classification will, unless otherwise

specified, refer to rank-preserving merge and rank-p1'eserL'ing classification,
respectively. A merge in which each input file forms an infix of the output
file is called a simple merge, and a classification in which each output file is
formed from an infix of the input is called a simple classification.
An m-way classification and a subsequent m-way merge together effect a
rearrangement from a single file to a single file. If the classification and
merge are both rank-preserving, the possible rearrangement effected is
restricted. However, a sequence of such orderings using alternate classi
fication and merge can effect an arbitrary reordering. In particular, the
following two important subclasses of such orderings will each be shown
to suffice:
1. simple classification and merge,
2. classification and simple merge.
A seq uence of copy operations of the first type used to effect complete
ordering on some key is also referred to as a merge sort. A sequence of the
second type is called a colunln sort.

Simple classification and merge

An infix in a file vector for which the key is a monotone increasing
(decreasing) function of the rank is called an increasing (decreasing) string.
The length of a string is the number of itenls it contains, and a maximal
string is a string contained in no longer string. A file containing a single
maximal string is ordered on the key.
For example, the sequence of keys
1, 3, 5, 8, 4, 7,9
contains several increasing strings including 1, 3; 1, 3, 5, 8; and 7, 9, but
it contains only two maximal increasing strings, 1, 3, 5, 8 and 4, 7, 9 and
six maximal decreasing strings. Henceforth the term string will normally
refer to a maximal string.

6.1

Serial sorting 111ethods

179

Two flies, each contaInIng one string, may be nlerged to produce a

single string by selecting at each step the item with the slllaller key of the
two next available from the input flies. More generally, if the inputs each
contain 11 strings, and if the foregoing process is generalized to produce the
longest possible output strings, each output string will contain precisely
one string from each input. If the inputs contain III and 11'2 strings, re
spectively, then J1 == nlax (111' 11;?) strings are produced in the output. A
subsequent simple two-way classification which assigns rl1 --:- 21 strings to
the first file and the remaining strings to the second, yields the greatest
possible reduction in the maximum n unlber of strings in anyone file.
Repetition of the merge and classification phases eventually produces an
ordered file.
The generalization of the process to an H1-way Inerge and H1-way classi
fication is inlmediate, * the optimum nUlnber of strings assigned to each
output file by the classification process being linlited to
In1. Referring
to Fig. 6.2, the 111-way merge Inay be described as follows: those itenl
positions containing keys which eq ual or exceed the key last recorded on
the single output file are said to be el(f!;ihle, and the next itenl chosen for
recording is the eligible item with the mininlum key. When no eligible
items remain, all positions are again Inade eligible and the process
continues, initiating another string in the output file. The number of
output strings produced is clearly the maxilnum of the nUlllber occurring
in an input file. Figure 6.3 illustrates the process for 111 == 3. The vertical
strokes in the figure indicate the division into maxilnal strings and do not
denote infonnation represented directly within the files.
In any sorting procedure, the smallest subprocess which treats the entire
set of items once is called a phase. The sinai lest subprocess which by
simple iteration prod uces the sorting process is called a stage. A stage
may comprise one or nlore phases. In the Inerge sort described above, for
example, the classification phase and the subsequent Inerge phase together
constitute a stage which is iterated until order is achieved ~ the process
is therefore called a two-phase rnerge. The use of a revision operation
(Table 6.1) permits the classification and the Inerge to be coalesced into a
single phase, and the resulting process is called a sil1gle-phase l11crgc. The
single-phase merge requires rn input files and 111 output tIles, whereas the
two-phase merge requires only (111 + I) files-one input and In outputs
in the classification phase, and 111 inputs and one output in the Illerge
phase.
The followillg fonnat will be assullled for the original files in all programs
in this chapter: the terminal item is a d ulllnlY (null) which is not to be

rn --:-

* This method, comnlonly credited to Goldstine and von Neunlann (t (48), was
presented by J. W. Mauchly in July 1946 in the Moore School lectures (t94H).

~
~

47,92, 38, 79, 20, 59, 98

I34, 36, 70, I 67, 96, '24

I
I15, 76, I 22, 43, 63 I

00

I47, 92, I 38, 79, I 20, 59, 98, I 34, 36, 70, I 67, 96,I 24, I 15, 76,I 22, 43, 63 I
V)

~.

15,34,36,47,70,76,92

I 38, 43, 63 67, 79, 96' I15, 34, 36, 47, 70, 76, 92, I 22, 38, 43, 63, 67, 79, 96, I 20, 24, 59, 98 I
122,
20, 24, 59, 98

'15,20,22,24,34,36,38,43,47,59,63,67,70,76,79,92, 96, 98'

Figure 6.3

Simple classification and nlerge

CD')

?'
~

6.1

Serial sorting nlethods

181

sorted and which is accompanied by a terminal partition AFt-I, where j is

the index origin in use; all other items are separated by the partition Aj
The final dummy item facilitates the use of the immediate branching
convention introduced in Sec. 1.22.

Two-phase merge. The (m + 1) files used will be labeled <1>11, <1>21, ... , <l> m l ,
and <1\2, the last serving as initial input and as input during each classi
fication phase. Program 6.4 describes the entire merge.
Program 6.4. The merge phase is shown on the left and the classification phase
on the right. The heart of the former is the loop 7-17. Its operation is controlled
by the logical vectors v and z (of dimension nl), which specify the set of exhausted
files and the set of ineligible items, respectively. An eligible iten1 is one which has
been read to one of the nl item positions Pi, is not a dumn1y, (i.e., null), and
possesses a key which equals or exceeds r, the key of the item last recorded.
The merge phase terminates on step 7, when all files are exhausted. Step 8
initializes the vector z (ineligible item positions) to the value v (exhausted files),
and step 9 increments the output string count s. Each execution of the subloop
10-14 records on cD 1 2 the item Pj having the smallest key of all eligible items.
Steps 10-12 select and record the item Pi and preserve its key as the variable r.
Step 13 reads the next item fron1 file j. If this exhausts the file, the branch to
step 15 is followed. Step 15 adds j to the set of exhausted files and step 16 adds it
(since the new item Pi is a final dummy) to the set of ineligible items. Step 17
then repeats the subloop if any eligible items remain and otherwise re-enters the
major loop at step 7. If the files are all exhausted, step 7 branches to step 18 to
begin the classification. If not, the production of a new maxin1al string is begun
on step 8.
If step 13 does not exhaust the file, it is followed by the decision on step 14,
which repeats the subIoop if r does not exceed the new key, and otherwise adds}
to the set of ineligible items on step 16.
The necessary initialization is perforn1ed by steps 3-6. Step 3 rewinds all input
and output files. Steps 4 and 5 perform the initial read from each unexhausted
file to the corresponding item position and respecify v to indicate any file which
may be exhausted by the initial read itself. The vector v is itself specified external
to the process, so that the initial set of input files may be restricted at will. On
subsequent repetitions of the merge, all files are n1ade available by step 2.
The classification phase begins by terminating the output file with a dun1n1Y
item and final partition A2 , and rewinding it. The /11 input files are also rewound
to serve as output files in the subsequent Ill-way classification, and the process is
terminated by step 20 if the output string count is equal to one. Step 21 redefines s
as the maximum number of strings to be allotted to the output files in turn, and
step 22 reads the first item from the input file (I\~.
The output files are selected in ascending order by the index i. The variable j,
which counts the strings, is initialized on step 25 and decren1ented on step 30 each
time a break is indicated by the preceding comparison on step 29. When s strings

182

6.1

Sorting

eD 1

A2

+- O ,

eD 1 2

v+-O

17(<1))

vip, vjq

+-

+-

f
+-

+-

17(eD)
vl<D 1

v V (q = A2 )

rs -:- ml

21

22

i -(- 0

Z -(- v

~-

10

+-

((zlk(p))/Llh

11

r -(- k(pj)

13

<D 1
A2

Pj

+-

:S;

-F

k(pj)

15

vj

-(--

16

Zj

+-

17

23

+1

24

+-

+-

25

+-

k(p)

26

Pj
:S;

14

<D 1

+-

([)i l +-

(D/

+-

20

S +-

12

19

s -(-0

+1

18

+-

0, A2

+- (1)1 2

27
A2

k(p)

j -(-j - 1

28
29

30

>
I-origin indexing
-:I

Legend

(I> i]

Input to merge; output of

classification, i E L](nl).

(1)1 2

Output of merge ~ input to

classification.

Item positions

File partitions

Exhausted files

Ineligible items

= 111.

String count.

Key of last recorded item.

Final dummy item.

Index of output file.

j
Program 6.4

J)

Index of input file.

111-\Vay 111erge sort

6.1

Serial ,,,'orting nlethods

183

have been recorded, the branch from step 30 to step 24 is followed to repeat the
process for the next output file.
The classification phase is terminated by the occurrence of the partition A2 on
step 28. Step 1 then records a dummy and a final partition on each output tile
(including any which may have received no items), and step 2 resets the vector v
to zero.

The merge phase with which Program 6.4 begins is essential, even though
the original data are contained in a single file, since it also serves to
determine the string count s needed in the subsequent classification. The
need for the string count can be avoided by using a classification process
which merely assigns successive strings to successive files in cyclic sequence.
This process does not satisfy the definition of a simple classification, and
it will be given the name string classification. String classification is
frequently more convenient to use than silTIple classification, particularly
in processes such as the single-phase merge to be described next. Two
successive strings assigned to a given file in string classification can coa
lesce to form a single string, but the probability of such an occurrence is
small, especially in later stages.
Single-phase merge. The two phases of the merge sort can be coalesced in
a single revision operation employing nl input and r1'1 output files. The
two rows of files (1)1 and <D 2 serve alternately as input and output on
successive stages.
Program 6.5. The Dlain subprocess (15-22) differs fronl the corresponding seg
ment (10-17) of Progranl 6.4 only in the control of file selection, the alternation
of input and output being controlled by the alternator G, which alternates between
1 and 2 (step 2) on successive stages. The classification is controlled by the variable
i, which selects the particular output file on step 17 and which is itself cycled
through the integers I to /11 by step 14. When all files are exhausted, the branch
to step 2 is followed, resetting the vector v to zero, ternlinating the output files,
and rewinding all files. The final output is contained in file <1\([.

Elimination ojjile rewind. Each of the sorting processes described req uires
a rewind of all files between successive stages. If the files enlployed are
capable of backward read, the processes can be ITIodified so as to elinlinate
the need for rewind. Since each file is alternately read and recorded in
successive stages, each will always be read in a fixed direction and recorded
in the other. Since the space needed for recording is not known in advance,
it is necessary to do all recording in the forward direction and therefore
to read in the backward direction. The changes required will be illustrated
by modifying the single-phase merge of Program 6.5.
Program 6.6. Since the alternate forward record and backward read effectively
reverse the order of all files on successive stages, alternate stages nlust assenlble

184

Sorting

6.1

(a, s)

+-

a+-2

V+-O

<1>a

7T(<1

(1, 0)

11 (a +

1)
I-origin indexing

A2E

+- E,
+-

<1>1
<1>2

Input to odd stage

Output of odd stag
I tern positions

P
q

File partitions

Exhausted files

s : 1

s+-O

vjp, vjq

v+-VV(q=A 2E)

10

i+-O

Key of last recorded item.

11

v :

Alternator (1, 2).

12

+-

13

+-

+1

14

+-

11 (i +

15

+-

(zlk(p))j l l)l

16

+-

kePi)

vj<1>a
Z

= nz.

Ineligible items
.String counL

Legend

1)

<1>~-a +- Pj

17
18

+-

A2

Pi

<1>ja

+-

19

r : k(pj)

20

Vj

+-

21

Zj

+-

22

Z :

Program 6.5

>

=F
m-way single phase merge sort

ascending strings and descending strings, respectively. This is achieved by re

versing the algebraic sign of the key in statements 20 and 24 (by use of the alter
nator a) on even-numbered stages.
Except when the entire process is terminated by failure to follow the branch
from step 8 to step 11, the output files are never terminated by a partition, nor
rewound, but are simply read backward as in statements 13 and 23. To ensure
that the backward read of a file terminates properly, each is provided with an
initial dummy item (step 11), and the branch on step 23 occurs on either the

(c, a)

7T(eD)

v+-O

c : 1

++-

(0, 1)

E
I-origin indexing

i=

eD l

Input to odd
stage

eD 2

Output of odd
stage

Item positions

Partitions

Exhausted files

Ineligible items

+-

185

Serial sorting nlethods

6.1

s : a-I

9
10

= m.

String count.

11

Key of last recorded item.

12

Alternator (1, 2).

Direction of read.

13

vip, vlq

14

15

i+-O

16

v :

17

Z +-

18

19

i+-ml l (i+l)

20

+-

((zl(3 - 2a)k(p))j l l)l

21

+-

k(pj)

Legend

+1

eD~-a +- Pj

22
23

vi ceDa

v V (q = A2E)

+-

+-

+-

).,A 2

Pi

ceD ja

+-

24

(3 - 2a)r

25

V j +-

26

Zj +-

27

(3 - 2a)k(pj)

>

Program 6.6

=f:.

Single-phase merge without rewind

186

6.1

Sorting

terminal partition A2 or the permanent initial partition J.. The entire process is
ternlinated by equality of s and a-Ion step 8, which can occur only if s = 1
and a = 2. This pair of conditions ensures not only that the nunlber of strings is
unity but also that an odd number of stages has been executed, and hence that the
final output (on (1)1 2 ) is in ascending order.
Since the initial input file must normally be read forward, an exception is made
through the agency of the variable c. On the first pass only, c = 0, and the
reads on steps 13 and 23 are therefore forward. The forward read on the first
pass necessitates a subsequent rewind of the input files, which is provided by
step 6.

A commonly used variant of the m-way merge sort (called string

doubling from its behavior in the case nl = 2) treats the initial input as if
the maximal strings contained were each of length one, and therefore
prod uces strings of uniform length nl in the output. These uniform strings
may not be maximal but are treated on the next stage as if they were, i.e.,
output strings of length m 2 are produced. In general, the kth stage pro
d uces strings of uniform length ml.', and rlogm n1 stages are required to
order n items. The number of stages does not depend on the initial number
of maximal strings, and no use is made of possible inherent order in the
original array. The comparison operations may, however, be somewhat
simplified, since the need to test eligibility is replaced by counts of the
items read from each file or by recording partitions between successive
strings. If 11 is not an integral power of m, some of the strings will be
shorter than the normal length. They can be expanded by dummy items,
although the use of partitions renders this unnecessary.

Classification and simple merge

The classification and simple merge sort is also referred to as a column
or digital sort, for the successive classification phases are controlled by
successive columns (digit positions) of a positional representation of the
key. The behavior of the process is not so obvious as the behavior of the
merge of the previous section, and a formal proof of its ability to produce
ordering will be given at the end of this section. Since the process is based
on a positional representation of the key, it will be convenient to use
O-origin indexing for all operands.
Let k be the (nonnegative) sorting key, let d be its digital representation
in a base b number system (that is, (bE) ~L d = k), and let q = v(d),
where b q exceeds the largest existing key. The complete column sort
comprises q stages, each stage consisting of a b-way classification
followed by a b-way simple merge. The classification on stage j is based on
d l ' j (the jth digit of the key counting from the low-order end), and each
item is assigned to file (J)i' where i = d j. The simple merge is defined
l

6.1

187

Serial sorting Inethods

such that the output file has the form <Po, (PI' . . .
the b output files of the preceding classification.

, <Db-I'

where the

(Pi

are

Two-phase COlUn1l1 sort. Like the merge sort, the base b column sort may
either be two-phase (classification followed by simple merge and using
b + 1 files) or single-phase (using 2b files); the two nlethods are described
by Programs 6.7 and 6.8, respectively.

j +-q
1T(<D)

+-

j +-j - 1

(b(q)) l-d

<D~.3

<D 1 +-

1T(<D)

i+-O

+-

+-

+-

k(p); b

11

<1>00 +-

<DO
0

Input to classification; output

of merge.

<1>1

Output of classification; input

to merge (v(<D 1 ) = b).

Positional representation of
key (v(d) = q).

Base of representation.

A1

Al

<1>l

Stage = q - j.

Current input file~

Legend

i+-i+l

12

14

Al

O ,

10

13

<D oo

+-

+-

O-origin indexing

<

<

i : b
<D oo

+-

0, Al

Program 6.7

Two-phase column sort

Program 6.7. Steps 1-7 constitute the classification and steps 8-14 the n1erge
phase. File <Do 0 is the single input to the classification and the single output of the
simple merge. The remaining b files are denoted by <D /, i E l O(b).
The component of d which controls the current classification (step 6) is selected
by j, which scans d from right to left. In the (implicit) specification of d on step 5,
b serves as an auxiliary variable (cf. Sec. 1.19). When the classification phase ends
with the exhaustion of <1>0 0 (step 4), the output files are terminated, all files are
rewound, and the simple merge is performed (steps 10-13) to re-collect on <D oo the
results of the preceding classification. This file is then terminated by step 14, and
the entire process is repeated from step 2 for the next lower value of j. The
dimension of d is specified (by compatibility with e(q) on step 5) as q, and q is,
of course, specified externally.

188

Sorting

6.1

Program 6.8. The alternator a again determines the roles of the two sets of files
(input or output) on successive stages, but because of the use of O-origin indexing,
it alternates between the values 0 and 1. The classification process (steps 7-11)
differs from that of Program 6.7 only in the control exercised by the alternator a,
and in the specification of a sequence of input files <1>oa, <1>l a, , <1>~_I instead of
the single file <1>0.

+-

a+-l

7T(<1

+-

O-origin indexing

j+-j-l

i+-O

+-

+-

<1>

+-

k(p); b

+-p

i+-i+l

10
11

Al

<1>i a

(bE(q)) ~d
ii

<

ii

dj

<1>0

<

12

<1>1

Input to odd stage }

Output of odd stage v

b.

Positional representation of
key (v(d) = q).

Base of representation.

= q - j.

Stage

Current input file.

Alternator (0, 1).

Legend

i : b
<I>ii +- oE, AlE

Program 6.8

Single-phase column sort (output on b files)

The program is deficient in two respects: it requires that the original input be
on a full set of b files, and it produces a final output on b files rather than one.
The first defect may be remedied by the use of an externally specified logical
vector v to designate the input files which are unused in the original stage. The
second may be remedied by a final simple merge following the qth stage. This
could be added as a separate program, but it can be effected more simply by
replacing q with q + 1 to ensure that the high order digit do is identically zero and
so restricts the final output to a single file.

Validity of the column sort. * If d is the q-digit, base b representation of the

key k of an item p, then the (q - j)th pass of the base b column sort assigns
the item to file <D i a, where i = d j The subsequent simple merge collects
the files in the order <D oa, <D I a, ,<Dba_I. Consider any two items PI and P2
with distinct keys k l and k 2 represented by d l and d 2 , respectively. It may

* This

proof is due to Ashenhurst (1953).

6.1

Serial sorting lnethods

189

be assumed, without loss of generality, that k i < k 2 Since the keys are
distinct, there exists an integer r such that d ri < d r2 and that arid l =
a r ld 2 Thus r is the (O-origin) index of the highest order column in which
the keys differ.
In the (n - r)th pass, the items PI and P2 are therefore assigned to files
<D h a and <D i a such that h < i. Hence at the conclusion of the subsequent
merge, item PI precedes item P2. In each subsequent stage the two items
are always assigned to the same file (since d .,! = d s2, S < r), and the two
items are therefore always relatable with respect to these operations.
Since each stage is a copy operation, and therefore rank-preserving, the
relative ordering of all relatable items is preserved, and PI thus precedes P2
at the conclusion of the process. Thus for any pair of items PI and P2 with
keys k i and k 2, such that k i < k 2, item PI precedes P2 in the final arrange
ment.

Vector keys and categories. The column sort is actually based on the
components of the vector d and only indirectly on the numerical key it
represents. The process can therefore be generalized to any numerical
vector key y defined on each item a i . It can be further extended to an
arbitrary key vector b belonging to the prod uct set

c=

@ c i @ ...

CO

Cl/(b)

as follows: on the classification stage based on component b j , the itenl is

assigned to file c j l b j An obvious generalization of the foregoing proof
of validity shows that the resulting ordering is that of the product set C.
The component set c j is called the jth category, CO is called the nlajor
category, cV(b)--l the minor category, and the ordering is said to be defined
on category cv(b) ~l within c (b)-2 within co. For example, a ninecolumn employee number b = (b o' b I , b 2 , b a) may be based on four
categories, the first component b o representing the eInployee's one-bit
payrOll classification (hourly or salaried), the second his two-decimal
digit department number, the third his two-alphabetic-character job code,
and the fourth his four-decimal-digit identification number. The colunll1
sort on the base b representation of a numerical key is clearly a special case
of a vector key in which each of the categories is the set lO(b).
It is frequently necessary to order a set of items on certain subsets of the
given categories, and on different rankings of the categories. In general, if
b is a vector key and m is any compatible mapping vector, then a related
ordering may be defined on the vector d = b m . If-continuing the previous
example-it is required to produce a list ordered by employee's department
within job classification, then m = (2, 1) and d = (b 2 , bI).
Any infix of b defines a related ordering which is actually achieved at
V

190

Sorting

6.1

some stage of the column sort on b, and the ordering defined by it is there
fore said to be contained in the ordering b. If the orderings dO, d 1, . . , d r 1
are all contained in b, they can all be achieved at some stage in the ordering
b, and the total number of passes required may be less than that required
to achieve the r different orders independently. For example, dO = (b 2 , bI)'
d 1 = (b o, b 2 , bI)' and d 2 = (b 2 , b I , b o' b 3 , b 4 ) are all contained in 1 =
(b o' b 2 , b I , b o' b 3 , b 4 ) and can be achieved jointly by sorting on the six
components of 1 rather than on the ten components of dO, d 1 , and d 2
The usefulness of this result is further extended [Ashenhurst (1953)] by
the following fact: if a given component recurs in an ordering vector, its
later occurrences may be ignored. For example, the second occurrence of
b o in the vector 1 (as 13) may be ignored, and 1 therefore contains the
ordering d 3 = (b o, b 2 , b I , b 3 , b 4 ) as well.
The propriety of suppressing later recurrences of a component of an
ordering vector is easily established. Let b be a given ordering vector and
let d = b m and c = b n be two related orderings. Moreover, let d be
obtainable fronl c by suppressing all later recurrences of components of b.
This implies that the mapping vector rn is obtained from the mapping
vector n by the same process, i.e., rn = (ajn)jn. If PI and P2 are two
items whose (distinct) keys c i and c 2 agree in all components up to but not
including the jth, then their relative order is determined by the component
c j = b n .. In the corresponding keys d 1 and d 2 it is clear that the first
compol{ent in which the items differ is again b n and that the same relative
order is therefore determined by d. Since a third ordering v = b p is also
equivalent to the ordering c = b n if P is also reducible to rn, the result
concerning eq uivalent orderings can be extended as follows: two orderings
y = zm and w = Zn are equivalent if (ajrn)jrn = (ajn)jn.
Choice oj'number base. In the merge process of Program 6.4, the value of
111 = n 1.
In the column sort, on the other hand, the number of files required is
determined directly by b, the base of the number system representing the
key. If the choice of b is otherwise arbitrary, it can be chosen as n - 1 for
any n > 2. However, explicit execution of the base conversion indicated
by the statement
(be) ~ d +- k(P)

m may be chosen to suit the number n offiles available, that is,

is usually avoided by using the base in which the key is represented in the
original files. If a base conversion is required, it can be performed once on
the first stage and the resulting vector d incorporated with the item for use
on subsequent stages.
Base conversion may sometimes be inconvenient or impossible as, for
example, in the case of special purpose sorting equipment devoid of

6.1

Serial sorting Inethods

191

arithmetic units. In this event the stage required for each digit (b-way
classification and subsequent merge) may be executed as a series of copy
operations each utilizing fewer than (b + I) files. If, for instance, the
digits are represented in a ranked binary n-bit code, a series of n two-way
classifications and two-way simple merges will effect the desired ordering
on one digit of the key. More generally, a group of p successive binary
digits can be treated at each pass, providing that 2 P + 1 does flot exceed
the number of files available.
Repeated block sort. If a set of items is classified on the high-order
column of an associated vector key, each of the resulting subvectors can be
separately ordered and then nlerged in a final sinlple merge. Thus each of
the subvectors forms an infix or block in the final arrangement. Such
block sorting can be used to distribute the labor among a number of
independent sorters.
Block sorting can also be repeated by further classifying each block on
the next lower-order column of the key. For a key vector of dimension q,
q repetitions of the block sort yield a conl plete classification, and ordering
can then be achieved by a subsequent sinlple merge. However, since the
number of blocks produced is (except for duplicate keys) equal to the
original number of items, the use of repeated block sorting is unattractive
unless simplified by special schemes for keeping record of and controlling
the numerous blocks. Two such schemes will be discussed. The radix
exchange sort (Sec. 6.4) is appropriate to random-access storage only: the
amphishaenic sort is appropriate to serial files.

Partial pass methods

Each of the sorting schemes discussed thus far is constrained to treat the
entire collection of items at each stage. Partial pass nlethods obtained by
relaxing this requirement normally achieve a reduction in the total nUlnber
of items handled, but at the cost of some increase in conlplexity. The
partial pass methods gain their advantage by largely obviating explicit
merge phases.
The basic column sort gives rise to two partial pass nlethods of interest,
the an1phisbaenic sort and the partial pass colurnn sort. The cascade sort
arises from the use of partial passes in a merge sort.
Partial pass column sort. * This method achieves the effect of one stage of
a column sort on a base b key with fewer than (h + I) files by using a
sequence of partial passes. The method will be illustrated by an example
[taken from Ashenhurst (1953)] involving four files and a decinlal key.

* Presented by John Mauchly in the Moore School lectures (1948) and treated morc
fully by Ashenhurst (1953).

192

Sortin,')a

6.1

Table 6.9. This table describes the partial pass column sort for reordering on a
single column of the decimal key (thejth). The parenthetical expression follow
ing each file (1) i indicates that it contains all itenls whosejth key digit equals one
of the enclosed integers. A second pair of parentheses indicates a second set of
itenls in the file following, and grouped separately from, the first set. Thus the
original input file (})o is described by (1)0(0, 1, 2, 3, ... , 9) or by any pernlutation
of the decimal digits enclosed in a single pair of parentheses.
Output Files
Step

Input File
Remaining from previous steps

Pass
Fraction

<I>.~ :

(0,2,4,7)
(l, 5, 6)
0,8,9)

1.0

<j>~ :

<1>1: (0,2; 4, 7)

<I>.): (l, 5, 6)
<I);: (3,8,9)
(V o :
'

(2,7)
(4)
(0)

0.4

<!):.!: (l, 5, 6)(2, 7)

(1)3: (3, 8, 9)(4)

(1)0: (0)
(1)1:

(5)
(1)(2)
(6)(7)

0.5

(j)o: (0)(1)(2)
(})1: (6)(7)

(3)(4)(5)

0.5

(j):!:

(9)

<1>0: (0, 1,2,3,4,5,6,7,8,9)

Copied

(J)3:

(3, 8, 9)(4)(5)

<1>1: (6)(7)(8)

<1>1:

<})o:
(1)2:

(0)(1)(2)(3)(4)(5)
(9)

(8)

(6)(7)(8)

0.3

(9)

0.1

(1)3 :
<1>0: (0)( I )(2)(3)(4)(5)(6)(7)(8)

<I):.!: (9)

(1)3 :

(1)1:

Total

Table 6.9

2.8

The partial pass column sort

The first step copies each item whosejth key digit is (0, 2,4,7) to file <1\, itenls
(1, 5, 6) to file (1)2, and items (3, 8, 9) to file (T)~3. After each step, only the previous
input file and the next input file are rewound, in this case <I>o and <I>1. In step two,
itenls (2,7) are copied to (1)2, and therefore.follow the group (1, 5, 6), as indicated
by the separate parentheses. Sinlilarly, items (4) and (0) are copied to (1)3 and <Do,
respectively.
Step three is preceded by a rewind of (1)2 and (1\. Since items (I) and (2) occur
in separate groups in the new input file $2' they can be copied to <1)0 in separate
groups (I) (2) as indicated. Similar remarks apply to items (6) and (7) copied to
(}\. The three subsequent steps complete the required ordering, producing in the
original input file (1)0 the ordered array (0) (1) ... (9). The rightnl0st colunln
shows that fraction of the original file (assuming a unifornl distribution of the key
digits) copied on each pass. The total at the bottonl indicates that the expected
execution tinle is equivalent to 2.8 full passes.

The partial pass process of Table 6.9 is described more concisely by the
O-origin matrix M of Table 6.10. Element M/ specifies the file to which

Serial sorting I1zethods

6.1
Digit

Digit

0
0

Digit

1 2

Digit

Digit

1 3

Digit

Digit

Digit

1 2

Digit

1 0

Digit

1 2

Input

Table 6.10

193

1 2

Matrix M describing the partial pass COIUlllll sort of Table 6.9

items with key digit s are to be assigned in the rth step of the process, for
r E lO(6) and s E lO(10). The eleventh and last row of M specifies the input
files, i.e., M/l is the index of the input file in the rth step. An algorithlll
based upon the matrix M is described by Program 6.11.
Program 6.11. The subloop 14-18 perfornls the classifIcation according to the
element M/, using file i (specified by M/) on step 9) as input. When the file is
exhausted, the branch to step 6 increnlents r and repeats the process unless
r = v(M). Equality indicates conlpletion of the jth colullln and causes a branch
to step 4 to decrenlent j and reset r. The conlparison on j is deferred to line 13
so as to follow the ternlination and rewind of the new input file. The branch on
step 10 prevents the recording of a dunlnlY itenl on the original input file. The
previous input file is rewound by step 8.

Program 6.11 includes the two-phase column sort as a special case, for
if (b + 1) files <Do, <D 1 , , <Db are available, the nlatrix shown on p. 194
specifies a process essentially identical with that of Progranl 6.7. File (1)/1
corresponds to file <D oo, the first column of M deternlines the b-way
classification, and each of the b succeeding colunlns specifies the copying
of one file in the b-way simple nlerge onto file <1)/).
The method of partial passes is frequently used in the sorting of alpha
betic data. Variants include the 1-} pass-per-character nlethod of sorting

194

6.1

Sorting
0
1

h
b

I
M== .

j +-q
O-origin indexing

1+-0

7T(D)

j +-j - 1

+-

I' +- -

1'+-1'+1

I'

l{M)

7TD i )

1 +- M/)

-+-

Files (i

As in Table 6.10.

Positional representation of

key (r(d) = q).

Base of representation.

0 for initial data).

Column stage = q

I'

Partial pass stage.

Current input file.

Legend

q - 1

10

j - r

11

<[); +-

C),

12

7T( <I);)

+-

13

14

15

(b(q))
~d +- kCp); b

16

+-

17

+-

18

<!)k +-

Al

<

0
+-

<D i

(1)

Al

dj
M/i

Program 6.11

Method of partial passes for

III

files

-.i.

6.1

Serial sorting nlethods

195

punched cards * and the proced ure described by McCracken et al. (1959),
p.312.
Anlphishaenic sort. The amphisbaenic sort (Nagler, 1959) is a particular
arrangement of the repeated block sort employing partial passes. For a
base b key it requires b + 1 files and proceeds as a sequence of classi
fications, with a simple merge of the last b subblocks produced occurring
immediately after each classification on the low-order colulnn. Each
classification allots digits to the available output files according to the rank
of the file index, e.g., if file $0 is the input, the digits 0, 1,2, ... , h - 1 are
allotted to files 1, 2, 3, ... , h, and if file (P:3 is the in put, they are allotted
to files 0, 1, 2, 4, 5, ... ,b. The files are recorded forward and (except for
the initial input) are read backward without intervening re\vind. The
subblocks are designated by the key digits on which they have been
classified, e.g., by 0, 1,2, ... , b - 1, 00,01,02, etc. The block chosen
for classification at each step is the one with the smallest designation
among those not yet n1erged, the designations being ranked according to
their values as decimal (h-ary) fractions. Thus block 213 precedes block
214, which precedes block 22.
Table 6.12 shows the steps of an amphisbaenic sort on a three-digit,
base three key using files <Po, <D 1, <D 2, and <1):3' with the initial and final data
on file <Do. The input files are designated by asterisks. Thus file (Do is the
input to step 1, and blocks 0, 1, 2, are classified to files (P t , <D:!, and (f):3 as
indicated. Step 2 classifies the smallest block (0) from file <PI to files <1)0' <D:!,
and <D 3 . Step 3 classifies blocks 000, 00 I, and 002 to files <D t , <])2' and (P:~.
The next step merges these blocks to file (1\), and the following step begins
classification of the next smallest block (01) on file <D:!. It is clear that if
the files are read backward, the next block to be classified is always
immediately available. The general process for a q-digit, base b key is
described by Program 6.13.
Program 6.13. O-origin indexing is used throughout. Each classification is con
trolled by the ""current" vector c (of dinlension b), whose conlponents are the
successive indices of the available output files. Thus c k is the index of the file to
which digit k is classified (steps 16-19). The current vector is deternlined by
step 10 so as to omit the index i of the current input file. The selection of the
block to be classified is determined by the vector h (of dimension q), the next
block to be classified on digitj being determined by the prefix a j +1 jh. Because the
classifications proceed for increasing values of j and the blocks just produced
appear last on the files, the selection can be determined by the last component of
the prefix alone, i.e., by h j This is done on step 7, where the index i of the input

Described in I Bl'Yt Form 22-3177-2 Sorter A1allllal, p. 12.

'0
0'\

Step

<Do
<1>1

4 (merge)
000-002

00

<D"

(0,0,0)

1
2

000
1,01,001
2,02,002
(0,0,0)

s
c

(0,0,0)
(1,2,3)

(0,1,0)
(0,2,3)

2
(0, 1,0)
(1,2, 3)

1,01 *
2,02*
(0, 1,0)
I
I

Step

12

13 (merge)

14

15 (merge)

<D;
j

1,01
2,02
(0,0,0)
I

<Do 000-102, 110 000-112 000-1 12, 120 000-122

$1
<1>2
<1>3
h

i
j
s
c

*
111
2,12,112
(1, 1,0)
1
2
(0, 2, 1)
(0,2,3)

*
2,12*
(l, 2,0)
2
1

121
122
2*
(1,2,0)
3
2
(0,2,3)
(0,1,2)

*
*
2
(2,0,0)

6 (merge)

8 (merge)

9 _ _1

000-002,010
OIl
I*
2,02,012
(0, 1,0)
2
2
(0,1,2)
(0,1,3)

000-012

16

17

18 (merge)

1-9

20 (merge)

000-122,20
21
22
*
(2,0,0)
3
1
(0,3,3)
(0, 1,2)

000-122*
21,200
22, 201
202
(2,0,0)

000-202
21*
22*
*
(2, 1,0)
3
1

000-202,210
*
22,21 I
212
(2, 1,0)
I
2
(0,3,1)
(0,2,3)

000-212

2,02*
(0,2,0)
1
1

00-012,020
021
1,022
2*
(0,2,0)
3
2

(0,1,3)
(0,1,2)

2
(0,3,0)
(l, 2,3)

000-022
1*
2

(1,0,0)

000-022, 10
11

III

~
g-.
~

2, 12*
(I, 1,0)
2
1

2
(0,2,0)
(l, 2, 3)

(merge)

000-102
11*

000-022*
11,100
101
2, 12, 102
(1,0,0)

2, 12
(1,0,0)
2
1
(0,2,3)
(0, 1,3)

22*
*
(2,2,0)
3
1

10

21

122 (merge)

000-212,220
221
*
222
(2,2,0)
2
2
(0,3,2)
(0, 1,3)

000-222
*
*
(0,0,0)
-I

I
Table 6.12

Amphisbaenic sort on 3-digit base 3 key

em
?'
~

O-origin indexing

Files (i = 0 for initial

and final data);
i E loeb + 1).
Base of representation.

dj

Jndex of current input.

.f

Controls forward read

and rewind of initial
input.

<l>i

controls current
classification.

Indices of current out

put files. v(c) = b.

aj+Ijh specifies input

block to classification
on digit}.

S j is index of input
file to classification on
digit).

Positional representa
tion of key.

=q

Legend
1T(<I +- e(b)

[+-0

h +- (q)

~l

[:0

20

[+-1

21

i+-O

1T(<I>O) +- 0

22

j+-O

P +-l<1>i

23

i +- (eijlO(b))h j

) : q - 1

24

j+-j+l

k +-(c o

Sj

+- (q)

+- i

c +- eijlO(b)

11

k +- (j

12

<l>Ck

+- 0, Al

13

<D Ck

+- 0, Ao

15

= q - 1) 1\ (co = 0)

<

Al

k : b

27

+-l<D Ck

28

k+-k+l

29
30

k : b

hj

Al

26

<1>0 +- P

<

)+-)-1

k+-k+l

<

25

0)

P +-1cI> Ck

10

14

+-

i +-

b 10 (h j

+ 1)

31
32

Sj

33

(bE(q)) ~ d +- k(p) ; b

<1>0 +- 0, Al

34

18

k +-d j

1T(<I>O) +- 0

35

19

<l>Ck

16

P +- f<l>i

17

+- P
Program 6.13

Amphisbaenic sort
197

198

Sortina
1-'"

6.1

to the preceding stage deternlines the vector .ijlO(h) of output indices, fronl which
the new input index is selected as the h}h conlponent. The succes~ive values of h,
i, j, S, and c listed in the example of Table 6.12 I11ay be helpful in tracing the
operation of the prograI1l.
When classification on the last digit is conlpleted, the last subblocks are nlerged
(steps 25-30) onto the output file (1)0. The variable j nlust then be decrenlented
and the component h j incremented I1lodulo b (steps 31-32). When h j conlpletes
a cycle (becomes zero), the corresponding subblocks are exhausted and j nlust be
decrenlented repeatedly until the corresponding h j does not beconle zero. The
nlain process through increasing values ofj is then repeated.
Although h j deternlines the input subblock for the classification on digit j, it
deternlines the index of the input file only indirectly (step 7) through the vector
of output files, itself determined by the input file i used in the classification on the
preceding digit. When j is increasing, the value of i is sinlply the value fronl the
preceding stage. However, when j is decrenlented (steps 31-32), it is necessary to
deternline the input i used in an earlier classification. A record of the value of i
corresponding to each classificationj is therefore kept (step 9) in the vector s, and
is used to redefine i on step 33.
Each file to be used as output (except possibly file (1)0) is first closed with a
partition (steps] 2-] 5) to demark the beginning of each of the subblocks to be
recorded. Since back\\-'ard read is to be used, two dunlnlY itenls are provided so
that the branch on Al coincides with the read of a dunlnlY itenl. An extra read
then disposes of the extra dunlnlY. Any partition recorded on the final output
file (1)0 before a fully classified subblock (when j = q - ]) would renlain in the
final output. Step]] prevents this by initializing the index k to ] ifj = q - ] and
the index of the first output file (that is, co) is zero. In a sinlilar manner, step 25
prevents the attenlpt to copy file (1)0 to itself during the nlerge phase (steps 25-30).
Step 28 reads the extra dumnly item recorded in the partitioning operation of
steps ]2-]5.
All files are read backward except the initial input in the first stage. This
behavior is controlled by the logical variable.r (steps] 6 and 20-2]). The branch
on step 20 fails the first tinle only, allowing/'to be respecified as 1 and rewinding
file <1>0. On subsequent stages, the branch to step 23 causes the read of the extra
dunlnlY partition.

Cascade sort. * The cascade sort is a partial pass merge sort, with each
stage proceeding as follows. The strings are initially distributed (unequally)
among In of the 111 + 1 available files; an m-way merge to the empty tile is
perfonned until sonle input is exhausted; an In - 1 way merge to the
newly emptied file is then performed from the relnaining inputs, and so on
to a final two-way merge. The effectiveness of the process depends on a
particular initial distribution of the input strings.
Table 6.14 illustrates the process for 190 strings distributed anlongfive
of six available files as shown in the first row. Succeeding rows show the

Due to Betz and Carter (1959). See also Gilstad (1960) and Carter (1962).

Eraluarion

6.2

(~l serial

sorting Incthods

Distribution of strings

Stage
0
1

IS

29

3
4

1
0

Table 6.14

41
12
3
1
0

(~ascade

50
14
4

55
15
5

I
I

199

nlerge sort

distribution of strings at the end of the succeeding stages. * The process

requires but four passes~ only the last of which is a conlplete pass (e.g., the
last five strings need not be copied in stage 1).
The power of a nlerge process ITIay be defined as the (average) factor by
which the nunlber of strings decreases per pass, i.e., as \jl-;, \vhere s is the
nun1 ber of strings whose ordering can be con1 pleted in p passes. For the
given exan1ple, the power is approxinlately 3.7. This surpasses the po\ver
of 3.0 attainable in a three-way single phase Inerge sort using the saine
nun1ber of files.
Gilstad (1960) has proposed a variant of the cascade sort (called
pO~l'phase) in \vhich every phase is an In-way n1erge, i.e.~ each newly
recorded output enters ilnmediately as input in the follo'Ning phase. Its
power is slightly greater than that of the cascade sort.

6.2

EVALUATION OF SERIAL SORTING METHODS

Three Inajor factors enter the eval uation of a serial sorting process:
the amount of program storage required~ the nun1ber of serial-access flies
used, and the execution time. The first two factors require little analysis~
and attention will be limited prin1arily to the third.
Because the execution tinle of a serial sorting process is nOrInally
determined almost cOITIpletely by the tilTIe required to transfer inforInation
to and from the serial files~ the execution time is assun1ed to be directly
proportional to the number of passes of the files required. Each phase
corresponds to a pass or (as in partial pass Inethods) to SOITIe fraction of a
pass, and the number of passes per stage is detern1ined by sun1n1ation over
the component phases.
The constant of proportionality relating actual execution tin1e and
number of passes depends on such factors as the average length of the
items, the reading and recording rate of the serial files~ and (in processes

* The jth column of the table refers not to a specific file but to that tIle which
ranks jth in nunlber of strings.

200

6.2

Sorting

requiring rewinding after each phase) the rewinding speed. Since these
factors are specific to particular eq uipn1ent and particular tasks, and since
the nature of the dependence is obvious, the present analysis is lin1ited to
consideration of the nun1ber of passes.
Consideration is also given to related orderings (cf. vector keys and
categories), which may, in the use of colurnn sorting, be achieved more
efficiently jointly than separately.

Simple classification and merge

The number of stages required in the tn-way merge sorting methods
depends on So' the number of maximal strings in the original file, and on
tn. For, if Sj is the nun1ber of (maximal) strings at the conclusion of the
jth stage, then Sj rI ==' rSj -;- m 1. Since the eventual string length must be
one, the number of stages required is given by

==

pogm sol,

for which the approximation r -~ log;n So will be used. The number of

passes is then 2r for a two-phase merge and r for a single-phase merge. If
the "cost" of the process is assumed to be of the form
c

==

(a

+ m)r == (a + nl) 10glll So,

for some constant a, then the optimum choice of

solution of the equation
loge m == 1 + a -;- /11.

/11

is obtained as the

The two-phase and single-phase methods may be compared for a fixed

number of files as follows. Let 11 == 2k be the number of files, * let S be the
number of strings, and let t I and 12 be the execution tin1es for the single
phase and two-phase methods, respectively. Then
t2
tI

2 logm 2

2 log(2k-I)

logml s

logk

2 logs k

logs (2k -

2 10gIo k

I)

loglo (2k -

1)

log('(2k -

1)

This ratio increases monotonically fron1 1.26 at k == 2 to an asymptotic

value of two. Since it exceeds unity, the single-phase process is to be
preferred.

Expected nurnber ofstrings. Since the original file is frequently specified in

terms of the nun1ber of items n rather than the number of maximal strings

* An odd number n would prejudice the result against the single-phase method, since
one file would necessarily be left idle.

6.2

Era/ltation oj'seria/ sorting nlcthods

s, it is desirable to determine the relation between

of s for a random distribution of the keys.

17

201

and the expected value

Let k be a vector of dimension 17 + 1 whose first n conlponents are the

successive keys of the file and whose last cOIllponent is infinite, and let
f = (k < lk), d = (k = lk), and b = (k ~.~ lk).
If, for example,
k=793558214x,
then
lk = x 7 9 3 5 5 8 2 1 4

f=

1010001100

d= 0000100000
b= 0101010011.

Each unit component of f marks the beginning of each nlaxinlal

ascending string in the forward direction (left to right), each unit com
ponent of t b marks the beginning of each maxinlal string in the backward
direction, and a unit component of d marks each key which duplicates its
predecessor. Consequently, + If, + Ib, and + Id are the nunlber of
forward strings, backward strings, and duplicates, respectively. Since the
relations <:, =, and > are exhaustive and ITIutually disjoint the three
logical vectors are exhaustive and mutually disjoint, i.e., j' V d V b = E.
and f 1\ d = f 1\ b = d 1\ b = O.
Consequently
(+/f) + (+Ib) + (+Id) = v(k) = 17 + 1.
Denoting the expected value of ~r by e(x), it follows that

e(+/f)

+ e(+lb) +

e(+ld) = n

1.

Symmetry and the assumed random distribution together iInply that

e( + If) = e( + Ib). Consequently, the expected number of strings e(s) is
given by
e(s) = e( + If) = (n + 1 - e(+/d) -:- 2.
lf there are no duplicates, the expected number of strings is approxi
mately n12, and the expected string length is therefore approximately two.
Assuming a uniform distribution of keys in a range of g values, the proba
bility that d i = I is clearly llg for all i E l2(n - I). Therefore e( + jd) =
(n - 1) -:- g.

Classification and simple merge

The number of stages required in the colunln sort is eq ual to q, the
number of significant digits in the representation of the key. To facilitate

202

6.2

Sorting

comparison between different nunlber bases, it is convenient to use the

variable g denoting the range of the key. Thus in any base b, the number
of significant digits q required in the key is given by q == f10glJ gl, or
approxinlately by q == log,) g. For any pair of bases b i and b2 , the corre
sponding number of stages qi and q2 are related as follows:

CfJ.
q2

==

10g'Jl g
Jog,)~ g

==

logy h2
logr, hI

==

10g'Jl h2
log'q hI

==

10g

b.).

b
1

A lTIOre practical form for calculation is

qi

loglo h 2

q'2

loglo b i

The foregoing expressions are identical in form to those obtained for the
number of stages required in the merge sort, but with the range g replacing
the initial number of strings SO' and with b replacing m. Moreover, the
nunlber of files required depends on b in the same way that the corre
sponding merge processes depend on In. This holds for both the two
phase column sort (b
1 files) and the single-phase colunln sort (2b files).
The analysis concerning the optimal value of m therefore carries over
directly to the choice of the base b, the only additional consideration being
the possible need for base conversions on the key. The comparison
between two-phase and single-phase processes also applies directly to the
column sort, with the conclusion that the single-phase method is superior.
Of the methods discussed, the column sort is the only one which shows
significant advantages in the joint treatment of two or more related
orderings. If x == zp and y == Zq are two vector keys, and if wijp == aijq,
then the ordering defined by the key zr' for r == wijp will, when applied
to the set of items ordered on y, suffice to produce ordering on x. The
total number of columns sorted to achieve the two orderings jointly is then
red uced by i.
More generally, if wijp agrees with a selected subvector of the prefix
ajjq and if the remaining elements of the prefix occur inwijp (in any order),
then sorting on the columns corresponding to w1)p may again be elided.
More precisely, if there exist integers i and j, and a logical vector u such
that

w1)p
and

(wl)p)

== u/(a j /q),

;2

Ciij(ajjq)),

then the ordering x can be achieved by applying the ordering wijp to the
ordering y, and the total number of columns sorted is reduced by i. It is
assumed that neither p nor q contains any repeated components, for if
they do, each of the later occurrences nlay be suppressed.

Era/uation oj'seria/ sorting 111ethods

6.2

203

Since the ordering on x can be performed before rather than after the
ordering on y, the roles of p and q may also be reversed, and the case
showing the larger reduction may be chosen. If three or Inore orderings
are prescribed~ the foregoing method may be applied to evaluate each of
the possible sequences of ordering.

Partial pass methods

In the absence of a general method for designing a partial pass colunln
sort, its efficiency will be indicated only by an analysis of the four-fIle
decimal key example of Table 6.9. If g is the range of the key, the number
of passes is given by PI = 2.8q = 2.8 logio g. This may be compared with
the value Pa = 2 loga g obtained for the straightforward base three column
sort, which can also be performed with four files. The ratio

!!.l

= (2.8 logio g) -:- (2 log:3 g) =

1.4 loglo 3 = 0.668,

P:3

indicates the superior efficiency of the partial pass column sort for this case.
The four files can also be used for a single-phase column sort in a base
two number system, yielding the value P2 = log2 g for the number of
passes. Hence PI!P2 = 0.843, and the method of partial passes is again the
more efficient.
In the amphisbaenic sort on n items with a q-digit base b key, h + I flies
are required, and the total number of items handled in the classification
phases is nq. In the merge phases, however, each item is handled at most*
once, and the total number of passes is therefore less than (q + I). This
may be compared with the 2q passes required in a two-phase column sort
using the same number of files. Alternatively, conversion of the key to a
base l(h + 1) -:- 2 J representation could permit the use of a more efficient
single-phase merge requiring approximately q log(h, 1)/~ h passes.
The disadvantages of the amphisbaenic sort reside in the more complex
program required and in the need for frequent reversal of the direction of
the files, i.e., from forward record to backward read. The time lost in such
reversal may be considerable for certain files.
The power of the cascade sort is, as indicated by the example of Table 6.14,
somewhat greater than that of the corresponding merge sort. Its behavior
is most conveniently analyzed [in the manner developed by Carter (1962)]
in terms of the difference equation satisfied by the number of strings
occurring in successive states. The formulation of these equations is
indicated in Exercise 6.19. Carter provides asymptotic solutions for cases
of practical interest.

* If the final output file (I)., is among the set of output files in the classifIcation on the
low-order digit, the subblock assigned to it need not be recopied in the nlcrge phase.

204

Sorting

6.2

In addition to its greater complexity, the cascade sort suffers from the
need for a particular initial distribution of the strings, and from the
dispersal of the file rewinds (which can be performed concurrently by n10st
computers) throughout the process. Moreover, in the event of a computer
error, a rerun froln the last correct input files is much Inore difficult to
program than is a corresponding rerun for a straightforward merge.

6.3

AIDS TO SERIAL SORTING PROCESSES

Internal sorting normally enjoys a much higher basic execution rate than
does serial sorting, but for large volulne files the limited size of internal
storage nlay make serial sorting necessary. The anlount of serial sorting
may, however, be reduced by some use of internal sorting. For example, a
preliminary internal sort can produce maximal strings whose lengths are
limited only by the size of the internal store and thus reduce the number of
strings presented for subsequent serial merge sorting.
The present section is devoted to methods of reducing serial sorting by
the auxiliary use of internal sorting. For this discussion, the only knowl
edge assumed concerning the internal sorting process is its capacity to
order a specified number of items.
Two classes of processes arise, one for aiding merge sorting and one for
aiding colUInn sorting. The aid to Inerge sorting is the simpler, since it
consists merely in assembling long strings by internal sorting before
beginning the serial merge sort. A serial column sort, on the other hand,
may be aided by a final internal sort performed after the column sort.
If k = k(x) is a key vector associated with x, and if In is any positive
integer, then a serial column sort performed on the key vector

k l = lk -:- n1J
produces the vector XO CB Xl
x P , where the vector Xi contains all
items such that k/ = i. If the infix vectors Xi are then copied in turn from
the file, individually reordered on the key

k 2 = (nl)!ok
and recorded, the resulting file will be ordered on k. Table 6.15 shows an
exanlple for m = 4.
If internal storage allows the internal ordering of as ITIany as n iteIns,
then the reordering of the infixes Xi can be accoInplished by an internal
sort provided that m is so chosen that v(x ::::~ n for all i. If the original
keys are all distinct, In may be chosen equal to n.
]f the sort on k l is perfornled as a base b serial colunln sort, the number
of stages required is reduced by approximately 10gb m fronl the number
1
)

Internal sorting IHethods

6.4
Original Order
x k kl k 2
Xl

1 3

x2 9 2 1
x3 2 0 2
x4 11 2 3
Xs

x6 4
x7 3

1 2
1 0
0 3

Ordered on k i
X
k kl k 2

x3 2 0
x7 3 0
Xs

7
6

x6

Xl

x2 9
x4 11

Table 6.15

2
3
1 3
1 2
1 0
2 1
2 3

Sets of Duplicates in k l

XO

I)
I
j

Xl

x2

205

Final Order
x k kl k 2

x3

x7
x6

Xs

0 2
0 3
1 0

6
7

1
J

x2 9
x.t 11

2
2

Xl

2
3

3
1
3

Internal aid to column sort (nl = 4)

required for a corresponding sort on the original key. The subsequent

internal sort on k 2 therefore serves as an aid to the serial colunln sort.
The arithmetic operations indicated in the definition of keys k l and k'2
may be simplified if nl is chosen as an integral power of the base b of the
original key. For, if the vector d is the q-digit base b representation of k i ,
and if n1 = h', then wIld, and wild are the base b representations of k/ and
k i 2 , respectively. The keys are therefore obtained from k by extracting the
specified columns, and the serial sorting is reduced by exactly t stages.

6.4

INTERNAL SORTING METHODS

Since the range of practical sorting nlethods is clearly broadened by the

use of random-access storage, internal sorting methods include all of the
serial processes treated in Sec. 6.1. However, since the use of randonl
access storage introduces certain new problems in the execution of these
processes, they will be reconsidered before proceeding to nlethods suited to
random-access storage only.
If the available random-access storage is divided into a number of areas
or fields, these fields can be used in lieu of the serial tIles. The serial
sorting methods then carryover unchanged except that the autolnatic
self-indexing property of the serial files must be replaced by an explicitly
programmed indexing of the corresponding fields.
The efficacy of an internal sorting process depends not only on the speed
of execution but also on the nUlnber n of items which can be ordered, using
a given storage capacity c (measured in nUlnber of items). Since n is nearly
linear in c, this property is measured in tenns of the storage ratio r = c -7- n.
In using random-access storage, the effect of a sinlple merge or a silnple
classification can be achieved rather easily through address nlodiflcation.
Hence there is little advantage in splitting either the nlerge or the column

206

6.4

Sorting

sort into two separate phases, and attention will be limited to the single
phase processes. The two-phase processes can prove superior if efficient
block transfer of data is available in the execution of the program, but
their behavior should, in any event, be clear from the treatment of the
analogous single-phase processes.
Simple classification and merge
An internal single-phase merge analogous to the serial single-phase
merge of Program 6.5 can be based on the assignment of two matrices 1 X
and 2 X to correspond to the sets of input and output files. The rows Xi of
the matrices a X correspond to the files <D i a of the serial process, each item
being represented by a single matrix element. Items are read sequentially
from and to fields; for each input field Xi, an index r j indicates the
elenlent axr ) to be read next; for each output field (3- a )Xi, an index 5 i
indicates th~ element next to be specified. The items in each field occupy
the leading positions in the field, and the index of the first unoccupied
element in the input field a Xj is indicated by the parameter t j The process
is described by Program 6.16.
I(

I(

Program 6.16. As in Program 6.5, the vectors v and z specify exhausted fields
(files) and ineligible fields (positions), respectively. The paranleters a and tare
initially specified external to the process. At the beginning of each stage the
vector v (exhausted fields) is specified (step 3) according to the unit conlponents
of t. At each stage except the first, t is respecified (step 2) by the final value of s
from the previous stage. When only one string renlains, each conlponent of s,
save the first, will remain at its initial value of unity, and this condition is used to
terminate the process at step 1. The vector k represents the keys of the current
itenls in the 111 input fields and is initiaJ)y specified (step 7) by the keys of the
colunln vector aXI of initial items of the input fields. The renlainder of the
process is closely analogous to Progranl 6.5.

If the total number of items to be sorted is n, then each of the 2m fields

111uSt accommodate n items, and the total storage allocated nlust be
2nm. The storage ratio is therefore 2m. It can be reduced to two by
putting the output in a single field and keeping a record of the beginning
location of each successive set of s' = s -:- In 1 strings, where s is the
maximum number of strings in anyone input field.
Let the input be represented by the single field xa, let b be a vector of
dimension m + 1, whose kth component specifies the beginning location
of the kth set of maximal strings, for k = 1, 2, ... ,j, thejth set n um bering
possibly less than s', and let b j \ I be the location of the first unused position
of the field. Obviously j ~ m, and if j < m, the remaining undefined
components of b are immaterial. If the j sets are to serve as inputs to an
m-way merge, then the prefix ant / b serves to initialize the vector r of
aXj

6.4

Internal sorting Hlethods

al/S

t-+-s

v -+- (t

s -+-

r -+-

a -+-21 l (a

I-origin indexing

aXkh

3-aX7/~

+ 1)

k(v/ a Xl)

v/k -+-

i-+-O

10

-+- v

11

i -+-

12

.i -+- ((zlk)/ll)l

13

r ~kJ

14

3-a Xi-+- a X

15

+-

rj

16

Si

-+-

Si

17

rj

19
20

>

21
22

lJl

11 (i + 1)

si

18

+1
+1

Vj

!i =

lJl.

v= n.

I tenl k of output
field h.
Index of current
input from field h.

Sh

Index of next output

to field h.

tit

Index of first unused

location of field h.

kit

Key of current item

from input field h.

v=

111.

Exhausted fields.

Ineligible fields.

Alternator (1, 2).

Current output field.

.i

Current input field .

Total number of items.

tj

Legend

kJ

-+- 1

Zj ~

-::j:.

rJ

I tern k of input field

h.

rh

k J -+- k(a X r /)
r

207

Program 6.16

Internal single-phase In-way nlerge using

21J1

fields

current input indices, and the suffix wmjb defines the ternlinating
locations t.
The strings produced by the merge from the sets of 111axinlal strings in
x a can be transferred without classification to the single output field
X 3 - a If the vector b is redefined by the beginning locations in the output
field of successive sets of
n11 items, it can be used to define the input
fields in a subsequent merge from the field X:l-a. Progranl 6.17 shows a
convenient arrangement of the process.

rs' --:-

I-origin indexing

3
Input field.
4

Output field.
b i+ I

+-

Beginning of
subfield d in
output.

"
\v = n
f
}

v = 111

+-

aftljb

+-

wmjb

I ndex of current in
put for subfield d.

(V/lI)'jl

Index of first loca

tion following sub
field d.

10

+-

kg

11

k(XruQ)

+-

12

Exhausted subfields.

Ineligible subfields.

s - 1

I ndex of next output.

14

15

i.( 0

-0(

16

11 +-

17

18

h -(- s

19

(/1

i -(- i

20

bi

11

21

Z +-

-0(--

+ 1)

V= In

Alternator (1, 2).

Output string counter.

Index of current output sub

field.

Index of current input sub

field.

Total number of iten1s.

Legend

Program 6.17

Internal single-phase
merge using two fields

22
r

23

+-

kj

24
~j

25

rj

26

1I+-11+1

+-

27
28

>

vj

+-

31

Zj +-

32

30

Key of current item

fron1 subfield d.

rs -:- n11

13

+-

kd

+ 1

208

6.4

Internal sorting

rnethod~,'

209

Program 6.17. As in Progranl 6.16, the vector r determines the current set of
items being examined, and t determines the terminal value of r. The first II/
components of b initialize r (step 7), and the last 111 components initialize t
(step 8). The index u determines the component of the output vector X3~a, which
is next to be specified (step 24).
The remainder of the process differs fronl Progranl 6.16 primarily in the deter
Inination of b. The counter Iz allows i and b i to be respecified (steps 19-20) at
the end of each group of s strings. The last specified component of b is deter
mined separately by step 5. Step 6 redeternlines v for the next stage.

Since some of the initial values so' vo, and bO may be unknown, the
initialization of the process (steps 1-3) merits some attention. If So alone is
known, VO and bO may be chosen as follows: VO == (il; b I O = 1, b 20 = veX).
The effect is to perform the first merge from a single input area. Con
sequently, the first stage performs no rearrangement but does determine
the vectors v and b.
If So is unknown, S may be determined by a preliminary string count.
Alternatively, it may be set to any integer val ue y =2 So. The process
remains valid, but the required number of stages is increased by [Iogm

yl

pogm Sol
Since So cannot exceed the number of items, the initial value s = veX) is
always valid, and for an assumed random distribution it exceeds the
expected value by a factor of two only. If greater initial order is expected.
it may be desirable to modify Program 6.17 to allow a small initial choice
of s, accompanied by automatic respecification in the event that it proves
too small. The modification may be simply made by inserting a branch on
eq uality of i and m, following step 18. The branch would succeed only in
the event that the initial specification of s were too small and should lead
to a process for increasing (e.g., doubling) s and repeating the process.
The case m = 2 is of especial interest, partly because it leads to signi
ficant simplification in the storage allocation, and partly because the
advantages which large values of m enjoy in serial sorting largely disappear
in internal sorting. These advantages arise fronl the reduction in the
number of stages with increasing m, with a nearly constant time per stage
due to the fixed reading and recording time for the serial files. In internal
sorting, the time required for key comparisons and address calculations in
choosing the minimum key becomes relatively more inlportant and, since
the key comparison time is an increasing function (frequently linear) of nl.
the advantage may lie with small values of m.
The sinlplification of storage allocation arises in the following way. A
two-way string classification on n items may be used in conjunction \vith a
single output field with a total capacity of n items by assigning the odd
numbered strings from the beginning of the field in normal order and the

210

6.4

Sorting

S2

veX)

a+-2 II (a

r -(- (1, veX))

S +-

r
xa

+-

+--00

>

> k( X r2 a))

+-

Sk

Index of output subfield,

k E (1, 2).

Index of current input

subfield.

Jndex of current output

subfield .

Key of last recorded

itenl.

Alternator (1, 2).

r : k(Xrja)

.i ~- 2 11 (j +

1)

13

+-21 1

1)

14

+-

15

x:3-a

+- X

16

r +-r

Qj

17

S +- S

Qi

(i

Output field.
Index of kth input sub
field, k E (1, 2).

11 (j + 1)

12

Input field.

rk

r2

r : k(Xria)

10
11

r1

X3-a

.i +- 1 + (k(Xr1a)

8
9

I-origin indexing

1)

k(Xria)

S i

ri

Program 6.18

Legend

Two-way internal single-phase merge

even-numbered strings from the end of the field in reverse order. Thus if
sj is the jth string of 2k strings so classified, the output field would contain
the array
Sl

EB S3 EB S5 EB ... EB

S2k-l

EB S2k EB ... EB S6 EB S4 EB S2,

-0(----------

-----------~

where the arrows indicate the increasing directions of the associated

strings. The restriction to an even number of strings in the foregoing
example is clearly not essential. The corresponding two-way internal
single-phase merge is described by Program 6.18.
Program 6.18. Since the current index vectors r (for input) and S (for output)
may always be initialized as shown in steps 3 and 4, and since termination of a
phase occurs when r 1 exceeds r 2 (step 7), explicit use of the vectors band t of
Program 6.17 is obviated. The only added complication lies in the different
treatnlent of indices r 1 and Sl' which must be increnlenfed whenever used, and of

6.4

Internal sorting rnethods

211

indices r 2 and S2' \vhich nlust be decrenzenfed. This treatnlent is effected by

addition (steps 16-17) of the rows Qi and Qi of the nlatrix

It is interesting to note that the use of indexed variables to allow conlparisons

with the larger (or snlaller) of two keys (steps 9 and II) reduces the requisite
number of comparisons from four to three.

The method of string-doubling also permits some simplification In

storage allocation and address calculations.

Classification and simple merge

As in the case of the internal merge sort, the internal column sort differs
from the corresponding serial sort primarily in the problem of storage
allocation. Again, the straightforward solution lies in the allocation of 2h
fields of n item positions each, and the use of b-dimensional indexing
vectors f, S, and t to control the input and output fields. The sorting
process used is identical with that of Program 6.8 (serial single-phase
column sort), and the indexing problems are analogous to those of Pro
gram 6.16 (internal single-phase merge using 2m fields).
As in the corresponding merge sort using 2m fields, the foregoing process
has a high storage ratio which can be reduced to two by a two-field process.
Unlike the corresponding case for the internal merge of Program 6.17,
the explicit classification process cannot be avoided. Consequently, it is
necessary to determine in advance the size of field required for each of the
b classes corresponding to digits 0, I, ... , (h - I ). This leads to the
so-called pre-count column sort of Program 6.19, in which each stage
incorporates an examination of the next higher order position of the key
and a count of each of the digits occurring.
Program 6.19. O-origin indexing is used, and the vectors XO and X(( (for a = 0
or I) serve as input and output fields. The classification on the key digit d i is
performed so that all itenls which agree in thejth colunln of the key fornl an infix
in the output X((, and so that the value of d j associated with successive infixes
increases monotonically. The output indexing is deternlined by the vector s,
which is, in turn, initialized by the vector b. The value of b for the succeeding
stage is deternlined by steps 13 and 14 *, according to the value of the next higher
key digit dj~l. The initial value of b is assunled to be defined externally. It I1lust

* Statement 14 is, for most computers, an inefficient procedure for deternlining b.

Normally it is preferable to make a simple count of each of the digits and to sunl th~
counts to determine s at the beginning of the next stage.

212

6.4

Sorting

1~

j+-q+l

2~

j+-j-l

s+-b

+-

+-

8~

(be(q

O-origin indexing
~

ii

~-

Xa

Input field.

Xii

Output field.

Sh

Index of output
subfield h.

bk

Beginning of out
put subfield k for
next stage,

e(b)

+-

-1

+1

veX)

1O

11

X Sh a

12

Sli -- Sit

~-

d
1))~

+-

k(xia); b

dj

Xa
l

+-

+1

13

~-

d j

14

~-

+ Cik+l

E lO

v=b

(b).

Positional re- }
presentation
v = q
of key.

+1

Base of representation.

j
i

Current colunln of key.

Current output subfield.

Alternator (0, 1).

rndex of currrent input.

Legend
Program 6.19

Pre-count colunln sort

be deternlined by a prelinlinary count, perhaps perfornled when reading the

items into the initial field. The use of q + 1 instead of q in steps 1 and 9 ensures
that the pre-count quantity d j _ 1 is properly defined even for the final stage.

Special internal sorting methods

The present section is devoted to internal sorting methods which are
specifically unsuited to serial files. The storage ratio provides a lnajor
categorization of methods; a method either possesses unity storage ratio
or it does not.
Unity storage ratio is achieved by methods which proceed by the inter
change of item pairs. The type of interchange may be limited to the
transposition of adjacent items, to "insertion" of an item acconlpanied by
a movement of all intervening itenls toward the evacuated position, or to
the exchange of an arbitrary pair. The corresponding methods are
characterized as transposition, insertion, and exchange methods, respec
tively.
Exchange methods include the radix exchange sort. This is an arrange
ment of the repeated block sort for a base two key, for which the operation

6.4

Internal sorting tnethods

213

count (number of elementary operations required to order n items) is of the

order of n log2 n.
Transposition methods include the bubble sort, odd-eren transposition,
and the ranking sort. They are characterized by relatively simple progranls,
an operation count of the order of n 2 for random initial order, and the
capacity to utilize existing order to reduce the operation count.
The only insertion method treated is ranking by insertion. The operation
count (counting comparisons only and not counting the individual iteITI
transfers of the block movements associated with each insertion) is of the
order of n log2 n for random order, and is reduced by existing order. It is
n10st attractive in a computer providing efficient block movement of
items.
Methods having a storage ratio greater than unity include the merge and
column sorts previously discussed. One additional method of this type is
treated-the ptll-degree repeated selection sort. The operation count is of
the order pn", where q == (p + I) -7- p.
Any internal sorting method can be broken into two distinct phases, the
first utilizing only the keys to determine the permutation required on
the items, and the second effecting the perrTIutation of the items. Since the
permutation vector is, in effect, a table of addresses of the iteITIS, the
process is called an address table sort. Address table sorting is particularly
advantageous if the volume of data in the item is large cOInpared to the
data in its key.
Any sorting method in which each stage isolates the item with the
smallest key (among the items remaining from previous stages) can be
modified to produce longer strings by the use of one auxiliary serial input
file and one auxiliary serial output file. The modification is called sorting
lrith replacement. It consists in recording the selected nlinimulTI itenl in the
output file and reading from the input file a replacement iteITI which enters
in the subsequent stages only if it is eligible for continuation of the string
already recorded.
The internal methods are evaluated and conlpared in Sec. 6.5, and the
results are summarized in Table 6.37.
Radix exchange. Radix exchange is a form of the repeated block sort for a
base two key. The high-order column of the key is scanned first. The fIrst
zero itenl (item with key digit zero) is exchanged with the last unit itenl, the
second zero item is exchanged with the second last unit itelTI, and so on,
until the first stage of the block sort is completed. The zero itenls no\\'
form a prefix in the vector of iteITIS, and the unit itenls fornl a suffix. The
process is then repeated on the next colUlnn of the key, first on the prefIX
obtained from the first stage, and then independently on the suffix. Each

214

Sorting

6.4

column of the key is treated in turn, the exchange process being performed
independently on each of the disjoint infix vectors defined by the preceding
stage. O-origin indexing will be used throughout the discussion.
Primary control of the process is exercised by the vector b, whose
successive components specify the beginning index of successive subblocks
(infixes). In particular, bo = 0, and bl'~l = vex), and at the beginning of
the kth stage, v(b) = 21\'~1 + 1. The storage required for b is therefore
significant but may be reduced by either of two expedients. The vector b
may be replaced by a logical vector u such that U/lo = b. Determination
of the next component of b then requires a scan of the components of u.
The use of U is thus normally unattractive except in a computer equipped
for the efficient determination of the location of the leading nonzero
component of a logical vector. The second expedient leads to a more
complex but more practicable process. The straightforward use of b will
be treated first.
Program 6.20. Steps 10~21 perfornl the exchange on the subblock with indices
k in the interval b i :::; k ,,;: b if1 . The indices Yo and Y1 designate the pair last
exchanged, k is the index of the current item exanlined,.i is the current colunln
of the key, and a is an alternator which is zero during the forward scan of the zero
section and unity during the backward scan of the unit section. The alternator a
determines which of the indices Yo and Y 1 will initialize k (step 13), the direction
of the scan (step 14), the type of key digit (0 or 1) sought in the next itenl to be
exchanged (step 18), and which of the indices Yo and Y 1 is to be redefined by k
(step 19) when the search ternlinates through failure of the branch at step 18.
If a does not beconle zero from negation on step 20, the process is repeated fronl
step 13 with a = 1, producing a backward search for a unit digit in the key. If a
beconles zero, both the forward and backward scans have been conlpleted, and
the required item exchange is perfornled on step 21. The final exit fronl the entire
loop 13~21 occurs when k = Y ii , that is, when k = Yo on a backward scan or
k = Y 1 on a forward scan. In either event, the final value of k is the beginning
index of the new subblock defined by the exchange, and it is used inlnlediately to
specify C I: on step 16. The vector c is eventually meshed with b (step 6) to re
specify b as (b o, co, b 1 , c 1 , . . . ). The initial specification of b on step 1 ensures
that the first subblock treated is the entire vector x.

The number of subblocks which must be distinguished at anyone titne

can be reduced to the dimension of the key by a method due to Hilde
brandt and Isbitz (1959). The process is controlled by a partition vector p
of dimension q + 1, whose successive nonzero components specify the
beginning indices of the subblocks of x remaining to be ordered. At each
stage, the first remaining subblock is exchanged on the appropriate key
digit d j , i.e., for j increased by one over the value used in generating the
particular subblock. When the exchange occurs for j = q - I, the order
ing of the two leading subblocks is complete, and they are rernoved froln

Internal sorting Inethods

6.4
b

(0, vex))

+-

j+-O

j +-j

j : q

+-

(2e) 10 LO(2 J

+-

\b, v, c\

i--i+l

bi

+-

O-origin indexing

11

r1

a -- 0

13

+-

r a

14

~-

k - 2a

15

k : r(l

16

ci

17

(2e(q l-d

18

d j : a

21

v(b) =

b i+ 1

+-

20

Beginning index
block i,

+-

+-

+-

+-

k(x k )

of sub

+ 2j

ei

Beginning index of ith new

subblock,
vee) = 2).

~esh vector
(0, 1, 0, 1, ... , 0),
l{V) = 2)+1 + 1.

Base 2 representation of
key
(ll(d) = q).

Current column of d.

Indices of last pair

exchanged
(v(r) = 2).

Alternator for scan of zeros

or ones.

12

ra

Vector to be ordered.

bi

vex)

r o -- b i -

=F

+ 1)

-1

10

19

215

Legend

ii

x To --) x
r1

Program 6.20

Radix exchange with l{b)

2"(1)

further consideration by respecifying Po by Pil and resetting Pil and PI' to

zero, where Pil and PI' are the first nonzero components of 1 /p. The end
of the new leading subblock is now determined by PU" the new leading
nonzero component of 1 /p, and the exchange is executed on the appro
priate column,j.
Record of the value of j appropriate to a given subblock is kept by
recording its terminal partition as Pflj. This is achieved, first, by recording
each new partition generated by exchange on column j - I in component
Pr/-j, and, second, by advancing the cOIllponent Pw (determined in the

6.4

216

Sorting

+-

e(q

+-

vex)

+ 1)

j+-O

j+-j-l

Po

j+-j-l

Pq-j : 0

+-

+-

O-origin indexing
Vector to be ordered.
Partition vector determin
ing beginning of subblocks

PI

<

(v(p) = q
d

+ 1).

Base two representation of

key (v(d) = q).

Pq_j

Current column of d.

Indices of last pair ex

changed (v(r) = 2).

Alternator for
zeros or ones.

Pq_j

j : q - 1

+-

+1

10

j +-j

11

r o +- Po - 1

12

Pq_j

+-

13

rl

14

a +-0

15

k +-r a

16

17

k : r{i

18

(2e(q))l-d

19

dj : a

20

ra

21

22

x ro

+-

+-

scan

of

Legend

k - 2a

+-

+-

+1

+-

k(x,,)

ii

x r1

Program 6.21

Radix exchange with

~{p)

= J{d) + 1

prefix removal phase) to Pw-I and resetting P/(j to zero. Incidentally, this
procedure ensures that P It and Pv always occur as PI and P2 In practice,
these two components need not be reset if the scan for Pw is begun with P3'
Program 6.21. The reader nlay find it helpful to trace the progranl (i.e., record
successive values of all paralneters in tabular fornl) for a sinlple case of a few keys

6.4

217

Internal sorting rnethoc!\'

of snlall dinlension. The exchange phase (15-22) is identical with that of Prograrl1
C i is onlitted. If j ,- q - I at the conclusion
of the exchange, j is incremented (step 10), r o is reset to Po - 1, and the ne\\'
partition k specifies both P(l-J and r 1 . If) = q - I, the prefix renl0val is executed
by steps 4-8. Step 4 respecifies Po, and the scan of steps 5-6 (which begins with
P3) locates a nonzero component PQ-J which is advanced to P'l~ (J; 1) (indirectly by
steps 7, 10, and 12) and is reset to zero by step 8. Steps 0- I provide the initial
specification of Po and (via step 12) of Prj'
6.20 except that the specification of

Bubble sort. The basic operation of the bubble sort

possible transposition of a pair of adjacent itenls so
of the two keys earlier in the sequence. The first
operations performed in sequence on the item pairs

is the conlparison and

as to place the smaller
stage consists of such
(X t , l' X,.), (X" ~,X'I I)'
I

Stage
2

Stage
3

Stage
4

Final
Itenls

6
II
9

8
6
11
9
4

4
8
6
1I
9

3
4
I- 6
8
9

x5
x6

Original
Keys

Stage

Xl

x:!
x3
x4
x5
x6

Original
I Itenls

3
4

Table 6.22

11

4
6

x~

8
9
11

x4
x:3

Xl

I
I

Bubble sort example

The result is to bubble each item upward in the sequence

until it encounters an item with a smaller (or equal) key and then to leave
it and continue bubbling the ne\v smaller item. In particular, the smallest
item is bubbled to the top. Successive stages repeat the process, but since
the jth stage brings the jth smallest item to the jth position, the () + I )th
stage need not treat the first j positions. It is clear that v(x) - 1 stages
suffice, but it is advantageous to allow termination at the end of the first
stage during which no transpositions occur.
Table 6.22 shows the arrangements prevailing at the end of each stage
of a bubble sort. The items above the staircase line are not re-examined.
Although the items are in correct order at the end of stage three, there is
no available indication of the fact until stage four is executed without the
occurrence of a transposition.
. . . , (Xl' X 2 ).

Program 6.23. The detailed behavior of the bubble sort process described by
this progranl should be clear from the foregoing discussion. It may, however, be
renlarked that at most I' = }I(x) - I stages are executed, even though the final
order is achieved only at the rth stage.

218

6.4

Sorting
i

~-

0
0

t+-O

i+-i+1

j+-j-1

+-

Vector to be ordered.

Stage number.

Termination indicator.

Xj~
+-

Legend

k(x j +1 )

k(x j )

vex)

10

I-origin indexing

x j +1

1
Program 6.23

Bubble sort

If a given set of items is completely ordered except for a single item

which is displaced upward from its proper position by j places, j stages of
the bubble sort will be required to complete the ordering. On the other
hand, a single stage of the bubble sort performed in the alternate direction
(i.e., scanning from Xl to XlI and bubbling the large items downward)
would suffice. In general, there is some advantage in perfornling successive
stages of the bubble sort in alternate directions.
If on a backward scan (from XlI to Xl) no transposition occurs between
items x j and Xj-l, then x j and x j + l are in correct relative order.
Consequently, if x j and (the possibly new) x j - l are not transposed on the
succeeding forward scan, then x j and x j + l will suffer no transposition. This
result may be extended to strings of items which suffer no transpositions,
and a record of this existing order can be used to obviate the corresponding
comparIsons.
More precisely, if S is a logical vector such that Sj = 1 if and only if no
transposition occurred between items x j and x j - l in a backward scan, then
no transposition between items x j - 1 and x j on the succeeding forward scan
will also imply no transposition (i.e., no need for comparison) between x j
and Xj-t-l if Sj = 1. More generally, if a = 1 is an alternator such that
a = 1 on the forward scan, then Sj may be defined as unity if and only if
no transposition occurs between x j and X j +W Program 6.24 shows the
entire process. This variant of the bubble sort is attractive only for
conlputers in which the indicated scan of the logical vector can be per
formed efficiently.

6.4

Internal sorting lnethods

219

Program 6.24. Multiplication of each key by the alternator a provides the

required alternation in the behavior of the branch on step 10. The behavior is
best appreciated by tracing a sinlple case.

I-origin indexing
3

s +- a 1(v(x)) V w 1(v(x))

Vector to be ordered.

(a,j) +-(-1,1)

String indicator
(v(s) = vex)).

Scan index.

a +- -a

Alternator ( 1).

j +--j

+a

(a #- 1)

Program 6.24

+ (a =

Legend

l}v(x)

Forward and backward bubble sort with string indicator

Odd-ecen transposition sort. Like the bubble sort, the transposition sort
has unity storage ratio and involves the comparison and possible trans
position of adjacent items. Each stage consists of two phases or half~stages.
In the first half-stage, each item with an odd index (except the last if it is
odd) is compared with its successor; in the second, each item with an odd
index (except the first) is compared with its predecessor. Table 6.25
provides an example.
Original
I tenlS

Original
Keys

Xl

8
6
11
9
3
4

x2
x3
x4
Xs

x6

Table 6.25

Stage

Stage
1

6
8
9

11
3
4

6
8
9
3
11
4

6
8
3
9
4
11

6
3
8
4
9
11

Stage
3
3
6
4
8
9
11

3
4
6
8
9
11

Stage
4
3
4
6
8
9
11

Odd-even transposition sort example

Final
Items
Xs

x6
x~

Xl

x4
x3

SortinO"

220

6.4

/)

Program 6.26. The subloop 6-10 perfornls the first or the second half-stage of
the transposition sort according as the alternator a is 0 or 1. Final ternlination
occurs on step 3 as soon as one half-stage is conlpleted without a transposition
occurring, except that a mininlunl of two half-stages nlust be executed. The
nlininlunl of two half-stages is assured by the initialization of t on step 1.

1~

t ..(- 2

a +-1

I-origin indexing

+-

t - 1

+-

ii

j +-a

j+-j+2

>

k(x j _ l )

x j - I +--)0

Vector to be ordered.

Termination indicator.

Alternator (0, 1).

Scan index.

vex)

10

<

+-

k(x j )

Legend

Xj

Program 6.26

Odd-even transposition sort

The validity of the termination conditions of Prograrn 6.26 follows from

the following proposition: if any half-stage except the first effects no
transpositions, the items are conlpletely ordered. If the half-stage is odd,
then

Since no transpositions occur, the conditions of the previous stage (which

exists by hypothesis) also remain valid, i.e.,

The two sets of inequalities together inlply ordering. * A similar argument

applies for the case of an even half-stage.
1n Sec. 6.5 the transposition sort is shown to be less efficient than the
bubble sort. However, it enjoys the unique advantage that all cornparisons

* This result may be combined with the fact (established in Sec. 6.5) that the nunlbcr
of transpositions required is finite to establish convergence of the transposition sort.
For, if the set is not ordered, each half-stage must effect at least one transposition.

221

Internal sorting Inetlzods

6.4

and transpositions in a given half-stage are independent and nlay therefore

be executed in parallel.

Ranking sort. If one new iteITI is added to a vector of itenls already ordered
on a given key, the resulting vector can be ordered by ranking the ne\V
iteITI, i.e., by comparing it with the iteITIS of the original vector in succession
until the appropriate rank of the added item is deternlined. Moreover. 11
repetitions of this process, which draw the new iteITIS fronl a given vector
of n itenls, will order the entire vector. Table 6.27 shows an exaITIple in

I Original Original I Initial

Stage

Stage

Itenls

Keys

Set

Xl

8
6
11

6
8

6
8
11

Stage I Stage

Stage
3

~I~

---

x~

x3
X,l

Xs

x6

9
3
4

I
Table 6.27

6
8

Final
Itcnls
---

xi)
x(}

6
R

11

6
8

11

x.I

11

x: 3

x~

Xl

Ranking sort example

which the individual ranking operations are each perfornled by conlparing

the added item with the ranked itenls in turn (starting with the largest),
moving forward by one place each item whose key exceeds that of the
added item.
Program 6.28. The index i controls the selection of successive itenls to define
the itenl z (step 4) which is to be added to the ranked set by steps 6-9. The indexj
controls the selection of successive itenls of the ranked set for conlparison with ::,
and each execution of the subloop ternlinates if either j beC0I11eS zero or if
k(z)
k(x j ). In either event, step 9 inserts the new itenl ;:, into the position x j l'
which was last evacuated. Fronl the initialization of the index i it is clear that the
process is actually c0I11pleted in vex) - 1 rather than l'(x) stages.

Ranking by insertion. Since each stage of the ranking sort ranks one ne\V'
item in an already ranked set, the determination of its position can be
accomplished by a binary search. This sharply reduces the required
number of comparisons. However, once the new position is deternlined.
each of the succeeding items in the ranked set must be moved down one
place to make way for the new item to be inserted. This ITIethod is par
ticularly good where (due either to suitable facilities in the COITIputer or to
the use of a chained representation for the vector of items x) such block
transfer is easy to perform.

222

6.4

Sorting

i-+-i+l

I-origin indexing

vex)

Xi

Z -+-

j -+-j - 1

k(z)

x j +l

-+-

x j +l

-+- Z

-+-

Vector to be ordered.

Index of current itenl to be ranked.

Scan index.

Current item to be ranked.

k(x j )

Legend

xj

Program 6.28

-+-

Ranking sort

1
ai(v(x)) 1\ aj(v(x))

ujx

i : vex)

-+-

+-- ~

I-origin indexing

(ujx)
X

ujx

Vector to be ordered.

Seglnent to be rotated
to insert (ujX)II'

lndex of item to be
ranked.

k(x i ) : k(x j )

k(x i) : k(Xl)

10

-+-

11

-+-

13

j -+-l(p

12

<

Lower partition in
binary search.

Upper partition in
binary search.

Scan index in binary

search.

ai(v(x)

-+-

+ q)

-:-

2J

Legend

k(x i ) : k(x j )

14

j : p

15

p-+-j
Program 6.29

Ranking by insertion

6.4

Internal sorting Inethods

223

Program 6.29. The binary search is perfornled by the loop II-IS, which so
determines j that the new itenl x i is to be inserted after x j Since the floor opera
tion is used on step 12, j will eventually reach the value p, and step 14 therefore
ternlinates the loop in a finite nunlber of steps. Step 3 perfonns the required
insertion by a right rotation of the infix of x bounded by x j and Xi' The cases
where Xi lies outside the previously ranked set are treated by the con1parisons on
steps 7 and 9. Incidentally, step 7 takes full advantage of any initial order in the
itenls, e.g., if the set is initially ordered, steps 8-15 are never executed.

Repeated selection sort. The process of scanning all items of a vector for
the smallest key and transferring the selected item to a separate output
area (in a serial or random access file) will be called selection. Repeated
selection on the successive remaining (unselected) items will serve to
transfer the items in the order determined by the specified key. This
method of ordering is called a (first-degree) selection sort.
If the given vector of n items is divided into m subvectors of at n10st
items each, then a selection from each subvector will produce a
vector of m items, thejth of which is the smallest of the jth subvector. A
selection performed on the m items of the resulting vector will then select
the smallest item of the entire original set. I f the selected item canlC
from the kth subvector, it is then replaced by a further selection on the
remaining items of the kth subvector. Repetition of the process n tilnes
serves to order the items. Because selection is performed on two levels,
the process is called second-degree selection.
In general, the smallest (first-level) item may be selected from a set of v?
second-Ieeel items, each of which is selected as the smallest of V: 3 third-Ieeel
itenlS. The process can clearly be extended to any desired number of levels.
If p levels of selection are used, the process is termed pth-degree selection or
repeated selection. It may be represented as a singular homogeneous tree
T of height p + I, as illustrated by Fig. 6.30.
Figure 6.30 shows the initial filling of the lower levels in a third-degree
selection sort performed on the sixteen items at the top of the tree T, with
veT) == (1,2,2,4). The keys are indicated in parentheses. The positions
of the third level are the nodes (I, I, I), (I, I, 2), (I, 2, I), and (I, 2, 2).
They are first filled by items x3 (6), x6(2), xn( I), and X 14 (4), respectively, each
selected as the smallest among the second-level nodes of the corresponding
subtrees TO,I,U' T O ,1.2)' T O ,2,1)' and T O ,2,2)' respectively. The first position
of level two is then filled by x 6 (2), selected as the smallest among the second
level nodes in its subtree, and so forth. Figure 6.3 I shows the continuation
of the process through the selection of the first two output iten1s.
If e is some value which exceeds the absolute value of all keys, then the
selection process may be made more uniform by assuming either that the
entire top level of the tree is occupied by the items to be sorted or that

rn -:- ml

224

6.4

Sorting

Level

Level
2

Level

Level

xl(9)

x3(6)

~::)
x4(16)

x6(2)

~, xs(3)

~:::
x7(7)

x8(l3)

~lt

~ X12(5)

~XlO(ll)
Xll (8)

xg(l)

~X13(l5)
~~M+

X14(4)

xls(12)

X16(lO)

Figure 6.30
(1,2,2,4)

The tree T representing the third-degree selection sort for v(T)

~Xl(9)

x2(14)

x3(6)

x3(6)

x4(16)

x6(2), xs(3)

~X5(3)
x6(2)

xs(3), x7(7)

x7(7)
xs(13)

xg(1), x6(2)

~X9(1)
X12(5)

xlQ(ll)

Xu (8)
X12(5)

Xg(l), X14(4)

Figure 6.31

Continuation of the third-degree selection sort

224

6.4

Internal sorting methods

225

unoccupied positions are filled with items having the key value e, and by
replacing each item selected from the top level by an item with a similar
key value. Termination of the process may then be determined by the
appearance of such a key value at the output. Moreover, the initial filling
may be simplified by filling all lower levels with dummy items having key
values of (-e). These may be recognized and discarded at the output.
The normal process will remove all the dummy items first, leaving all
levels in the state which would be produced by a special initial fill.
Program 6.32. The top level of the tree T initially contains the items to be
sorted (completed if necessary by dummy items having the key value of e), and
the remaining levels contain items with the key value (-e). The index vector s
scans the second-level nodes of the subtree T w1/s (steps 14-18) to determine the
index m of the node having the minimum key z. If z
e, step 10 replaces the
root of the subtree by the selected node value and step 11 respecifies s to begin
the scan of the subtree rooted in m. If z = e, then all second-level nodes contain
dummy items with "infinite keys," and step 9 branches to step 1 to replace the
root of the subtree by a dummy item as well. The branch from step 12 occurs
when the scan of the top level has been completed; it also results in the insertion
of a dummy item.
Since each complete scan (over all levels) begins with s = (1, 1) (steps 3, 11, 15),
the resulting minimum item is brought to the root of the tree. Step 2 specifies ::
as its key, and steps 4 and 7 determine its disposal. If z = e, all legitimate itenls
have been flushed fronl the tree, and the branch to step 5 terminates the output
file, rewinds it, and ends the process. If z = -e, the item is a dummy initial fill
and is discarded by skipping the recording of the output file on step 8.

Since the selection process proceeds by levels in the tree, a corresponding

computer program can best be based on a right list-specifically, on the
right list node vector e 2 /]T and the dispersion vector veT). The computa
tion of the list index res) required in the path tracing is described (for
O-origin indexing) by the recursion on the functions f and g developed in
Sec. 1.23. This recursion yields a relatively simple computer program for
a general homogeneous tree. It will be shown, however, that a b-way
rooted tree (i.e., a rooted tree with a common branching ratio b) is the case
of greatest practical interest, and in this case the simpler recursion
res)

= b

r(w1/s)

+ 1 + s"

(also developed in Sec. 1.23) can be used. Program 6.33 shows the repeated
selection sort of Program 6.32 executed on the right node list vector r of a
b-way rooted tree.
Program 6.33. The initial conditions are as assumed for Progranl 6.32, and the
steps of the two programs correspond very closely. The simple index modifica
tion required from stage to stage is shown in step 11.

226

Sorting

6.4

kT(;ilJS)I)
z

z : e

<1>21

7T(<1>l)

z : -e

<1>2 1

z : e

10

+-

I-origin indexing

(1)
=i

+-

+-

Leaves of T contain items

to be sorted or dummy
items with key = e; other
levels contain dummy
items with key = -e.

-e < k < e for all legiti

mate keys k.

Tl, Al

+-

Scan index.

Index of minimum in

Minimum key.

cD 2l

Output file.

(T m )1

e2jT'Wl/s'

m CD (0)

11

12

v(s) : veT)

13

+-

14

Sv

15

S +-- S

16

z : kT s )l)

17

18

+--

+- o ,A 2

(T wl/ m )1
+-

k(Tl)

+-

+-

>

Legend

vll(s)(T)

+ WI
~

kT s )l)

+-S

Program 6.32

Repeated selection sort

The superiority of a common branching ratio is demonstrated as follows.

The number of items scanned per item selected (ignoring initial fill and
termination) is clearly the sum of the branching ratios, whereas the maxi
mum number of items accommodated in the top level is equal to their
product. It is therefore easily shown (e.g., by induction or by the use of
Lagrange multipliers*) that for a fixed number of items, optimum execution
is furnished by a common branching ratio.
Sorting H'ith replacement. Certain of the internal sorting processes
discussed (bubble and repeated selection) proceed in a succession of
stages, each of which results in the selection of the smallest remaining

* See, for example, Margenau and Murphy (1943), p. 205.

6.4

Internal sorting methods

227

item. Since this item can be transferred immediately to an output area or

serial output file, the evacuated position can be refilled by an item from a
serial input file. Each output item can therefore be replaced by a new item
from a serial input, and the resulting process is called sorting H'ith replace
ment.
If the key of the newly introduced item exceeds or equals the key of the
last output item of a group, the new item may be treated as a member of

k(r p ) +- e
2

z +- k(r 1 )

m +-1

z : e

<1>2 1 +- 0, A2

I-origin indexing

Right list node vector

(]T)2 of tree of Program
6.32.

7T(eD l) +- 1

-e < k < e for all legit

imate keys k.

Scan index.

q)l +- r 1

171

Index of mininlunl node.

z : e

-e

10

r p +- r m

lOa

p +- 171

11

s +- b x (nl -

12

s : vCr)

13

z +- e

13a

} +-0

14

} : b

15

}+-}+1

15a

s +-s + 1

16

z : rs

17

z +- k(rs)

18

m +-s

z Mininlum key.

eD 21 Output file.

1) + 1
~

I ndex of previous node

(to be replaced).

Branching ratio of the

uniform rooted tree T.
Legend

Program 6.33 Repeated selection sort of Progranl 6.32 executed on the right
list node v~ctor r of a unifornl rooted h-way tree

6.4

228

Sorting

P3

(T'W 1/s)1

Z +-

I-origin indexing

5
6

>

PI

+-

+-

cD I
I

k(Tl)

(1)

+-

z . -e

Tree as in Program 6.32.

-e < k < e for all legiti

mate keys k.

Scan index.

Index of minimunl elig

ible node in
e2 jTwl/ s .

Minimum key.

9
10

(Tw1/m)1

11

S +-

12

v(s) : v(T)

14

z+-e

16

>

(T m)l

m ffi (0)

13

17

+-

<J)I

+-

+- S

+ WI

>

Minimum eligible key.

Input file terminated by
dummy item with final
partition A2

cD 21

Output file.

The program vector (Pi

is statenlent i of this
program).

Legend

w : k((Ts)l)

18

z : k((Ts)l)

19

20

21

z : e

22

w : -e

23

+-

k((Ts)l)

+-

+-

-e

Program 6.34

Repeated selection sort with replacelnent

6.4

Internal sorting methods

229

the original group. If not, the item must be allowed to retain its position
but must be excluded from consideration in the sorting process. In general,
sorting with replacement allows the production of longer output strings
than would be otherwise possible with the same internal storage capacity.
The expected increase is discussed in Sec. 6.5. Since the process terminates
with the original positions completely occupied by ineligible items, the
production of a further string can be begun simply by declaring all itenls
eligible.
.

Repeated selection sort with replacement. In the repeated selection sort it is

advantageous to apply the eligibility criterion at each level, i.e., to limit
selection to keys which equal or exceed the key of the item being replaced.
The item being replaced is, of course, the last on'e transferred (either to the
output file or to the preceding level). The top level items are replaced from
the input file or, when the file becomes exhausted, by dummy items.
However, the use of the "infinite" dummy key value e as in Program 6.32
raises serious difficulties, which are avoided by the use of the value
-(e + I). This is done in Program 6.34.
Program 6.34. This program is very similar to Program 6.32, and only the
essential differences will be remarked. The main scan loop (15-20) differs only in
the added comparison with w to prevent the selection of ineligi ble items. The
variable w is normally specified (step 13) as the key of the item just transferred
out of the position being filled. However, if all items are ineligible, then z
remains unchanged from its initial value established by step 14, and the branch
from step 21 to 22 occurs. The variable w is then set to -e to make eligible all
items except the dummy fills [with key value (-(e + 1))]. which enter on exhaus
tion of the input file. If only these dummies remain in the level scanned, the
process returns again to step 22. This time, however, the branch to step 3 occurs.
Step 3 is the file read (which replaces step 1 of Program 6.32). When the file
becomes exhausted, the branch to step 2 replaces program statement P3 by Pl' so
that step 3 thereafter provides the requisite dunlmy keys.

Bubble sort with replacement. A straightforward bubble sort with re

placement produces the same length strings as a first-order selection sort
with replacement, and, indeed, differs from it mainly in the additional
performance of item interchanges. Bubble sorting with replacement is
therefore of interest only if the order induced in the remaining items by the
interchanges can be used to reduce the number of items scanned. This can
be achieved by accumulating the ineligible items in a growing suffix of the
vector of items and restricting successive scans to the prefix of eligible
items. This method is shown in Program 6.35.
Program 6.35. In the loop 11-13, w denotes the current itenl with the smallest
key, and it is interchanged with Xi if k(w) exceeds k(x i ). At step 14, w is therefore

230

6.4

Sorting

k(w)

+--

P5

PI

+--

vex)

+--

-e

z : k(w)

7
8
9

i=

+--

+--

e
I-origin indexing

+I

<1>11

+--

vex)

+1

w~xJ

10

11

i+--i-l

12

k(w) : k(x i )

13

W~Xi

14

k(w) : e

15

<1>2

16

+--

+-

+--

Vector to be ordered.

j
i

Index of first ineligible itenl.

Current item.

Scan index.

j+--j-I
j

<D 1

Key of last recorded item.

1

Input file.

cD 21

Output file.

Vector of progranl statenlents.

-e < k < e for all legitimate

keys k.
Legend

k(w)

Program 6.35

Bubble sort with replacelnent

the item with the smallest key among those scanned and (unless k( w) = e) is then
recorded in the output file and used to define z, the key of the last recorded item.
The main loop is then repeated from step 5, where w is respecified by reading
an item from the input file. If the new item is eligible, the branch from step 6 to
step lOis followed; if not, the parameter j (denoting the index of the first in
eligible item xi) is decremented and w is interchanged with the last eligible item xi
before continuing to step 10. Since step 10 initializes i to the value j, the sub
sequent scan is limited to eligible items.
Ifj becomes zero on step 7, the entire set of items including w is ineligible.
Step 8 then restores j so as to make all items eligible and hence to start a new
string.
The branch from step 5 and the subsequent redefinition of step 5 on step 2
serve to introduce "infinite" keys after the input file is exhausted. The immediate
redefinition of j and z (steps 3 and 4) to make all items eligible may appear
redundant. It may be verified that this procedure (1) does not increase the number
of strings produced, and (2) avoids the potential error of omitting the last string.

Eraluation of internal sorting methods

6.5

231

Address table sort. A sorting process can, in principle, be accomplished in

two distinct phases, a determination of the permutation required and the
execution of the permutation. In serial sorting, however, the process
cannot be effectively divided in this way. For, because of the limitation
to rank-preserving operations, the items could not be transferred directly
to their final positions even if they were known. In internal sorting this
limitation does not apply. A sorting process embodying these two separate
phases is called an address table sort. The method offers advantages when
the time required to transfer a complete item is large compared to the time
required to transfer its key.
The permutation p must be so determined that the permuted vector
y = xp is ordered on the key. Let KI be the vector of keys associated with
the given vector x, i.e., K/ = k(xJ, and let K2 be the identity permutation
LI For example, if the successive keys of the vector x are 17, 9, 6, 11,4, 8,
and 3, then
9 6 11 4 8

~).

If the columns of K are reordered (by any desired sorting process), to

produce the matrix P such that P/ is monotone increasing in i, then
p = p2 is the desired permutation vector. In terms of the foregoing
example,
4 6 8 9 11
5

p = (7,5,3,6,2,4, I).

and

6.5 EVALUATION OF INTERNAL SORTING

METHODS
In evaluating internal sorting methods, both the execution time and the
internal storage requirements must be considered. For the execution
times of the internal merge sort and the internal colurnn sort, the analysis
of the corresponding serial sorting process applies directly. Since their
storage requirements have also been discussed in Sec. 6.4, the present
section will be limited to the special internal sorting methods.
The measures of interest in the evaluation of these methods are four:
I.
2.
3.
4.

the
the
the
the

scan length
number of stages required
number of transpositions required
storage ratio.

232

6.5

Sorting

The number of con1parisons executed is frequently used as a measure of

execution time of a sorting process. However, since the selection of the
keys to be compared (involving instruction modification and possibly
extraction of key digits) usually accounts for the largest share of the time
required in a comparison, the scan length (number of items scanned) is a
more suitable measure. In a process which scans m items serially (as in
each individual stage of a bubble or a ranking sort), the number of com
parisons is (m - 1) and differs but little from m. However, for m small
(as in the successive selections in a repeated selection sort); the ratio
m -7- (m - 1) is significant, and the scan length is the better measure.
Moreover, in a process which does not use a serial scan, the number of
comparisons may differ markedly fron1 the scan length. For example,
each half-stage of the interchange sort requires approximately ml2
conlparisons for a scan of m items.
The storage ratio is defined as the ratio of the number of item storage
locations required to the number of items entering the sorting process
(without replacement). The four measures will be determined for each
method in turn and then compared. Table 6.37 summarizes the main
results for an assumed random initial distribution of keys. The entire
analysis is based on the assumption of distinct keys.
Expected number of transpositions

Let yen) be the total number of transpositions required to order all of

the n! possible arrangements of n items. Since the (n + 1)th item added
in the last position may rank first, second, ... , or last, requiring n,
(n - ]), ... , or 0 additional transpositions, respectively, for each of the
11! arrangements of the n items, then
n

y( n

+ 1) = (n + 1) y( n) + n! 2

k =0

= (n

l)y(n)

+ (~)(t1 + 1)!.

The function

y ()
n =

n!(n2- n)
4

satisfies this difference equation as well as the obvious boundary condition

= 0, and the expected number of transpositions is therefore given by

yeO)

yen)

n!

(n 2

11)

The maximum and minimum number of transpositions are (n 2

zero, respectively.

n)/2 and

6.5

El'aluation oj' internal sorting f11ethods

233

A pair of items whose relative order differs from the final arrangement
will be called a disordered pair. Thus in the set (2, 4, 3, 1), the disordered
pairs are (2, 1), (4,3), (4, 1), and (3, 1). To establish order it is clearly
necessary that each disordered pair be transposed. The sequential
transposition of disordered pairs of adjacent items is also sufficient to
produce ordering. Any transposition method which transposes only
disordered pairs therefore achieves order with a minimunl number of trans
positions and hence with an expected number of (n 2 - n)/4. The ranking
sort, bubble sort, and interchange sort all fall into this category. The
sequence in which the transpositions occur will, however, normally vary
for different methods. Thus the sequence of arrangements realized in
reordering the set (2, 4, 3, 1) is (2, 4, 1, 3), (2, 1, 4, 3), (I, 2, 4, 3), and
(I, 2, 3, 4) for the bubble sort, and (2, 3, 4, 1), (2, 3, I, 4), (2, 1, 3, 4), and
(1,2,3,4) for the ranking sort.

Bubble sort
In the bubble sort, both the number of stages and the scan length depend
on the initial order. The minimum and maximum number of stages
required are one and (n - 1), respectively. The minimum and maXimUl11
scan lengths are nand (n 2
n - 2)/2, respectively.
The expected number of stages required for a randonl initial order is
determined * as follows. If, at any stage, all itenls are ordered except for
one which occupies a position below (later in the seq uence than) its proper
position, then one further stage will complete the ordering. On the other
hand, each item which appears above its terminal position at a given stage
will be ITIoved down by exactly one position in the next stage. Conse
quently, the number of stages required will be determined by d, the
maximum upward displacement from its final position of any item in
the original array. More precisely, if x is the given vector of items, y the
corresponding vector of ordered items, p is a permutation such that
y = xp ' and d = II - p, then d = ((e["d)/d)I.
If the maximum upward displacement is d, then (assuming that all
keys are distinct and that the final order is consequently unique) the last
item in the final order (i.e., Y 71) can initially occupy anyone of the (d + I)
components of the suffix W d +1/X. Similarly, y ii-I may initially occupy any
one of the remaining (d + 1) positions of the suffix w d t-2/ X, and so on for
each of the (n - d) items of the suffix wn-d/y. The number of possible
initial arrangements of the items ofwn-d/y is therefore (d + 1)(n-d). Since
the d leading items (i.e., ad/y) can occupy anyone of the d remaining posi
tions without restriction, the number of possible initial arrangements with a

* The exact expression

for the number of stages is due to Friend (1956).

234

6.5

Sorting
c(n)

II

0.798
0.853
0.877
0.893
0.904
0.918
0.927
0.934

5
10
IS
20
30
40
50
Table 6.36

Coefficients for z(n) ~ c(n)V 7Tn/2

displacement not exceeding d is clearly d! (d + l)(n-d). The probability

that the maximum displacement r does not exceed d is therefore given by

pr(r

s: d

1)(n-d) d!
.

n!

Similarly,

pr(r

(d

s: d)

d(n-d+l\d - 1)!

- 1) =

n!

and hence

pr(r

d)

pr(r
(d

s: d)

- pr(r

+ l)(n-d) d!

d - 1)

- d(n-d+l)(d -

I)!

n!
The expected value of r is given by

n-l

er

n-l

~ d

x pr(r = d) =

d=O

d x pr(r

d)

d=1

l(ni
d(d + l)(n-(l) d! n!
l

nil d X d(n-d+H(d -

d=1

A change of dummy variable in the second summand (d

the two summands to similar form and yields the result

er
where

zen)

I)!).

(/=1

= t+

I) brings

n - zen),
1

= -

n!

sn-s s !.

s= 1

It is shown in the appendix to this chapter that zen) approaches the value

V 7rn/2

for large n, and Table 6.36 gives coefficients c(n) for the approxi

mation

zen)

= c(n)"/7rn/2.

6.5

Ecaluation oj' internal sorting Inethods

235

Since one extra stage is needed to determine that the ordering has been
completed, the expected number of stages is given by

e(J = n

1 - ,zen).

Since the scan length in the jth stage of the bubble sort is known to be
n - j + ], a similar analysis can be used to determine the expected total
scan length es' The result is

es =

n2

n - 2

+ (n +

or approximately,
es
.

Ir ')

11 -

--

l)z(n

I) -

nz(n),

37T11

AI
.~

Ranking sort

The number of stages in the ranking sort is clearly (n - I). The ex

pected scan length on the kth stage is determined as follows. The item may
rank in anyone of (k + I) positions; first, second, ... , last, with eq ual
probability. There are therefore (k + I) cases requiring scan lengths of
2,3, ... ,k, (k + I), (k + I). The last case requires a scan length of
(k + I) rather than (k + 2), since the process terminates on comparison
with the last item regardless of the outcome of the comparison. The total
scan length for the (k + I) cases is therefore

k+2.)
(I r
r=2

- 1=

k2

5k

+2

'

and the expected scan length is consequently 2 + kj2 - Ij(k + I).

Summing over the (n - 1) stages and denoting the expected total scan
length by e yields the relation
11/-]
n-1
1
e = 2( n - I) + - I k - I -
2 7;=1
7;=1 k + 1
=

/12

+ 7/1
4

4 _

7;=1

!.
k

But

n I l
= y + loge n + -

I 1.'=1

approximately, * where y

e=

* See,

n2

+
4

211

711

- --.)

1211

0.5772 ... is Euler's constant. Hence

1
1.577 - log/, 11 - 211

+ -1-2 ,
1211

for example, Cramer (1951) p. 125, or Woods (1926) p. 171.

236

6.5

Sorting

correct to two decimal places for all n > 1. The maximum and minimum
scan lengths are (n 2 + n - 2)/2 and 2(n - 1), respectively.
The ranking sort takes advantage of initial order and the minimum scan
length of 2(n - 1) is achieved for an initially ordered set of items. Ranking
by insertion with binary search, as described by Program 6.29, requires
approximately 2 + [log2 jl comparisons on the jth stage. Hence if n = 21.:,
the number of comparisons required is given by

=
~

(2

(2

+ k)2

k)21.: - (2
k

21.:

+ 22 + 23 + ... +
= (k + 1)21.:.

In general, then,

e . n(log2 n

21.'-1)

+ 1).

For a random distribution, ranking by insertion with binary search

requires fewer comparisons than any other method, and, in the form
described by Program 6.29, also takes advantage of initial order. The
insertion operation requires, of course, a number of rotations of relatively
lengthy vectors.

Odd-even transposition sort

Estimates of the efficiency of the transposition sort may be obtained as
follows. Each half-stage requires the scanning of (approximately) n items
in nl2 comparisons, and results in reducing the displacement (either up or
down) of each item by at most one. The fact that the reduction in the
displacement may be zero for certain items can be illustrated with the
initial arrangement (5, 6, 1,2, 3,4). The number of half-stages must
therefore equal or exceed the maximum displacement, which, in turn.
equals or exceeds the maximum upward displacement d used in the analysis
of the bubble sort. Moreover, one final half-stage is required to determine
that order has been achieved, and the expected number of half-stages will
necessarily exceed the corresponding value obtained for the bubble sort,
namely, n + 1 - zen). Since the number of items scanned per stage in the
transposition sort exceeds the corresponding number in the bubble sort, it
follows that the transposition sort is much less efficient. Moreover, the
transposition sort does not allow sorting with replacement. I ts sole
advantage resides in the possibility of executing all operations in a given
half-stage in parallel.

Repeated selection sort

The number of items scanned per item selected in a pth-degree selection
sort is eq ual to the sum of the branching ratios of the associated tree. and,
as already demonstrated, a minimum scan length is provided by a tree

6.5

Evaluation of internal sorting 111ethods

237

with a common branching ratio m. Since m P :2:: n, the nunlber of items

scanned per output item is given approximately by

s = mp . m logm n.
The resulting expression for the total number of items scanned (nm logm 11)
is similar in form to the corresponding result for the m-way merge sort.
The indicated optimum choice of m is Napier's number e.
The optimum integral value of m is three, and its efficiency differs from
the theoretical optimum by less than 1 %. The value 111 = 2 simplifies the
required program and increases the expected amount of scan by only about
5 %. This case (m = 2) is referred to as a tournament sort. Larger values of
m may prove advantageous where the groups of items to be scanned are
contained in a serial store whose scan time is not significantly reduced by
reducing m.
Since p itelTI transfers are required per item selected, a more realistic cost
function for determining the optimum value of m may be given by the
function
c = (m + a)p = (m + a) logm n,
where a is the ratio of the time required for an item transfer to the time
required for the scan of a single key. As remarked in Sec. 6.2, the optimum
value of m is obtained as a solution of the equation
loge m

a
+ -.
In

The amount of itenl storage required for a pth-degree selection sort is

given by
111 1)+1 1
2
x

+ m + m + ... + 111J) =

In -

mn

m - l'
since n

m P

The storage ratio rem) is therefore given by rem) =

m/(m - 1), a function which decreases with m, rapidly approaching unity.

This ratio also represents the increase in execution time entailed in the
initialization of the lower level positions. If sorting with replacement is
used, initialization is required for the first string only.
The ratio rem) changes significantly for the first few values; thus

r(2)

2.00;

r(3)

1.50;

r(4)

1.33.

If the expected scan time sCm) is taken as a second criterion, then (since
5(2) = 5(4) > 5(3)), the value m = 2 will be eclipsed by both 3 and 4.

N
W

QO

Stages

Item Transfers or
Transpositions

Scan Length

Sorting Method
Expected

Max.

Min.

Expected

-- --- -Merge (m-way)

Column (q digits base

b)

Bubblet

logm n

+I-

zen)

In + 1 -

zen)

Interchanget

IOgm(i)

n -1

Ranking

Ranking by insertion (binary search)

n-l

n-l

n-]

With optimum allocation (2 fields).

Replacement possible.

])~

zen) = ~ Sll- ~s!

n! .s=l

== \

n - 2

--2--

In -]

I n~ In~

IOJm(~)

n2 -

n - 2
--2--

n2

+ (n + 1)2(n + 1) -

n.z(n)

+ (n + ])z(n + ]) -

lI:(n)

+ 7n

- - 4 - - 1.6 -logl'n-

n(log2 n

+ 2)

Min.

nq

nq

n2

Expected

IOglfl(~)
nq

+n- 2

.)

n- - n

--2-

.)
n- - n
n
2n
4

+n-

2 I 2n _ 2

2
n(log2 n

Min.

Max.

n'2 -

n log mn

nq

nq

n- -

-2
n2 -

n'2- n

+ 2) I

2n -

n- - n

1
nmp = nm logm n

I S~

-- ---

n logm n

nq
n2 -

p-th degree selection+

(branching ratio m)

11

Max.

Storage
Ratio

nmp

nmp

np

n'n:n I

0
np

I~
m-l

t Parallel operation possible: estimates are lower bounds.

All values exclusive of initial fill.

TTfl/2

Table 6.37

Characteristics of internal sorting methods (for random distribution of distinct keys)

CO')
Q"\

U.

Appendix

239

Sorting with replacement

The use of replacement with any internal sorting method offers the
advantage of increasing the expected length of the strings produced.
Gassner (1958) has shown that for n item positions and random keys, the
expected length of the first string is (e - I) x n == 1.718n, and that for
later strings the expected length rapidly approaches 2n.
Comparison of internal sorting methods
The results of the preceding sections are summarized in Table 6.37. The
ranking, bubble, and transposition processes show the most favorable
storage ratio, exactly one half of the best attainable in the merge and the
column sort processes. In the case of repeated selection, the storage ratio
depends on m and ranges from two to a limit of one.
The execution time is approximately proportional to the function

fen) = s

ct,

where s is the scan length, t the number of item transfers or transpositions.

and c the ratio of the cost (in time) of one item transfer to the cost of a scan
of one key. For any given method with variable parameters (such as the
repeated selection sort), these parameters may be chosen so as to nlinimize
j'(n). The choice between various methods may then be nlade (subject to
storage considerations) so as to minimize j'(n).
Since the ranking, bubble, and odd-even transposition methods share
the same number of transpositions, the choice between them depends on
the scan length and auxiliary factors. The odd-even transposition sort is
inferior to the bubble sort in this regard, and the bubble sort is, in turn.
inferior (by a factor of two for large n) to the ranking sort for n > 8. The
bubble sort retains the advantage that sorting with replacement may be
used, and the transposition method allows parallel operation.

APPENDIX TO CHAPTER 6
The following derivation of the limit of the function
z( n)

= -1

n!

L. s! sri

_,
.'i

s=l

arising in the evaluation of the bubble sort was suggested by Robert

Church. Clearly

L(n)

in

g(s) ds <: zen) <: U(n)

g(s

1) ds,

240

Sorting

s!
n!

where g(s) = - sn-s. Moreover,

V(n) =

+ 1)n-Sg(s) ds,
J(no (s-s-

and since g(s) is monotone increasing with a large positive derivative for
large values of s, only the upper end of the integral is significant. It can
therefore be shown that
lim V(n) = lim L(n).
n-+oo

n-+oo

Consequently,
lim L(n) = lim zen).
n-+oo

n-+oo

Applying Stirling's formula,

, L()'
n ===
I1m
n-+oo

Setting t

n ~-2 s+~ -s n-s d

7T S
e S
S
0
'V/27Tn n+~-n
e

in

n
- (n + ~)
e
n s n + 1(,
/~ e

ds.

1 - s/n yields

!~r: L(n) =

enn-tn+!-i)

= n

Ll

= n

Ll

Ll

nn+!-i(l -

t)n+!-ie-nerll n

dt

(1- t)"+Herddt,
ent+tn+!-il]Og, n-tl

dt.

Expanding loge (1 - t) yields

lim L(n)

"

e-t/2-(n+~)(t~/2+t 13+ , .. )

dt.

n-+ 00

For n large, only small values of t will be significant, and all terms in the
exponent may be dropped except -en + 1/2)t 2 /2. Similarly, the upper
limit of integration may be extended to infinity. Hence,
lim L(n)

1
00

n-+oo

e-(n+~)t !2dt

n(7T/(2n

l))~i

. (7Tn/2)~.

REFERENCES
Ashenhurst, R. L., (1953), '"Sorting and Arranging," Theory of Switching, Report No.
BL-7, Harvard Computation Laboratory, Section I.
Betz, B, K., and W. C. Carter, (1959) '"New Methods of Merge Sorting," 14th A.C.M.
Conference, Cambridge, Mass.

Exercises

241

Carter, W. C., (1962), "Mathematical Analysis of Merge Sorting Techniques,"

Congress of International Federation of Information Processing Societies, Munich.
Cramer, H., (1951), Mathenlatical Methods of Statistics, Princeton University Press.
Friend, E. H., (1956), "Sorting on Electronic Computer Systems," l.AC.M., vol. 3,
pp. 134-168
Gassner, Betty Jane, (1958), "Proof of a Conjecture Concerning Sorting by Replace
ment Selecting," unpublished.
Gilstad, R. L., (1960), HPolyphase Merge Sorting-An Advanced Technique," Proc.
Eastern loint Cotnputer Conference.
Goldstine, H. H., and J. von Neumann, (1948), "Planning and Coding Problems for an
Electronic Computing Instrument," Part 2, vol. 2, Institute for Advanced Study,
Princeton University Press.
Hildebrandt, P., and H. Isbitz, (1959), HRadix Exchange-An Internal Sorting 1V1ethod
for Digital Computers," l.A.C.M., vol. 6, pp. 156-163.
Margenau, H., and G. M. Murphy, (1943), The Mathenlatics of Physics and Chelllistry,
Van Nostrand, New York.
Mauchly, J. W., (1948), "Sorting and Collating," Theory and Techniques for Des(f{n 0./
Electronic D(fjital COlllputers. Lectures given at the Moore School, July 8-August
31, 1946, vol. III, University of Pennsylvania.
McCracken, D. D., H. Weiss, and Tsai-Hwa Lee, (1959), Progranllllit{f{ Business
Cotnputers, Wiley, New York.
Nagler, H., (1959), "Amphisbaenic Sorting," l.A.C.M., vol. 6, pp. 459-468.
Woods, F. S., (1926), Advanced Calculus, Ginn & Company, Boston, Mass.

EXERCISES
6.1 A given file contains a set of 15 items Xl' .. , X I5 with associated decinlal
keys 68, 9,60,14,60,73,79,15,67,5,9,41,57,9,41. For each of the serial files
used, show the contents at the conclusion of each stage of the following processes:
(a) a two-phase classification and merge using
(i) four files.
(ii) three files.
(iii) six files.
(b) a string classification and merge using four files.
(c) a single-phase merge using
(i) four files.
(ii) six files.
(d) a single-phase merge without rewind.
(e) a base ten column sort using eleven files.
(f) a base ten column sort using twenty files.
(g) a base ten column sort using four files and the partial pass column sort.
(h) a column sort using four files and a base three representation of the keys.
6.2 Modify Program 6.4 so that it will work properly without dumnly terminal
items (i.e., each terminal partition A2 is to be associated with a legitimate item).

6.3

Write a program for the string-doubling merge sort.

6.4 (a) Write a program for a base b column sort which uses backward read to
eliminate rewinding.
(b) Program a variant of the two-phase colunln sort (b + 1 files) so as to

242

Sorting
eliminate the copying of the "zero items" in each merge phase. Deter
mine the relative efficiency of the method.
(c) Program an analogous variant of the nl-way two-phase merge sort so as
to eliminate the copying of part of the items in the classification phase.

6.5 Construct the matrix M (cf. Table 6.10) specifying an efficient partial pass
column sort for the following cases:
(a) base ten and three files.
(b) base ten and five files.
(c) base eight and three files.
(d) base ten and four files using no rewind, i.e., files are to be read backward.
(e) base ten and three files using no rewind.
6.6 Using a set of matrices 1M, 2M, ... , l JM, of the form of Table 6.10, write a
progran1 to generalize the partial pass sort to the case of a mixed radix key.
6.7 (a) Reprogram the amphisbaenic sort (Progran1 6.13) so as to reverse the
order of the final output. (This is the form used in Nagler, 1959.)
(b) Program a generalization of the amphisbaenic sort which makes use of
partial passes within columns of the key.
6.8 (a) Program a modification of the bubble sort process which on odd
numbered stages bubbles the smallest item to the leading end and on
even-numbered stages bubbles the largest item to the tail end.
(b) By examining all 4! cases show that for four items the eApected nunlber
of stages is slightly less for the method of part (a) than for the un
modified bubble sort.
(c) Program a bubble sort using a string indicator s as in Program 6.24 but
using backward scan only.
6.9 (a) Program a modification of Program 6.17 to specify So = 2 and to auto
matically double So and restart if necessary.
(b) Compare the efficiency of the program of part (a) with that of the
straightforward program in which the number s of strings is assumed
equal to the number of items.
6.10 (a) Derive the relation loge 171 = 1 + alnl which must be satisfied by an
optimal value of n1 in order to minimize the cost function c = (n1 + a)
logrn n (cf. Sec. 6.2).
(b) Determine the optimal integral value of 171 for each of the cases a = 0,
1, e 2

6.11 For the amphisbaenic sort on a set of bt items with distinct keys belonging
to the set lO(b t ), determine
(a) the number of item transfers required.
(b) the number of file reversals (change of direction of read and record)
required.
6.12 Write a program describing the odd-even transposition sort in terms of
vector operations so as to show clearly the parallel nature of the process. Treat
all items with odd indices (plus a dummy item) as a vector and all even itenls (plus
a dummy item) as a second vector.

Exercises

243

6.13 (a) For each of the following sorting methods, determine whether initial
order of the items is preserved, i.e., whether the relative initial order of
all item pairs wjth duplicate keys is preserved.
(i) sinlple classification and merge.
(ii) classification and simple merge.
(iii) ranking sort.
(iv) bubble sort.
(v) odd-even transposition.
(vi) radix exchange sort.
(b) Prescribe a simple modification of the key which will ensure the preser
vation of initial order.
6.14 For the sequence of keys given in Exercise 6. I, show the explicit execution
of the following internal sorting processes:
(a) bubble sort with replacement (six item positions).
(b) pth-degree selection sort with replacement, with p = 3 and 111 = 2.
(c) ranking by insertion.
6.15 If the key is represented in a base b nunlber system, with each digit repre
sented in turn in a ranked binary code, then ordering can be achieved by a base
two column sort on successive columns of the underlying binary representation.
(a) Show more generally that ordering can be achieved by a base 21.' column
sort on (the base two value of) successive groups of k binary digits.
(b) Program the process suggested in part (a), including the deternlination of
an optimum value of k for a given nunlber n of available files. Assume an
original key of q digits, each represented by r binary conlponents. Do not
neglect the problenl of terminal conditions.
(c) If b = 10, r = 4, and if the successive decinlal digits are coded alternately
in a (ranked) bi-quinary and qui-binary systenl, the binary digits can be
grouped in twos and threes so as to allow column sorting with a nlaxinlunl
of five output files. Progranl a suitable process.
6.16 Program a sequence of rotations of infixes of the vector x which will
reverse the order of its components. (See the Ranking by insertion progranl
(6.29) for the case when the key defines a complete reversal of order.)
6.17 Assuming that an item transfer takes c times as long as a comparison of
two keys, determine a criterion for the choice of 171 in an 171-way internal revision
merge for the following cases
(a) assuming 2111 comparisons per item (n1 comparisons for eligibility and 111
for minimization).
(b) assuming that a ranking sort is used on the 111 itenl positions.
6.18 (a) Let z be a vector key of dinlension three whose successive components
represent department nunlber, job nunlber, and nanle, respectively.
Two lists are to be produced, one ordered by name within department
within job and the other by departnlent within job within name. Deter
lnine a mapping vector p such that y = zp is the vector key of least
dimension which contains the two required orderings.
(b) Let yi = Zpi be a set of vector keys defining a set of related orderings.

244

Sorting

Determine a vector key of nlinimum dimension which contains all of

the related orderings for the case pI = (3, 4, 1, 5), p2 = (3, 1, 5, 6), and
p3 = (1, 4, 3, 5).
(c) Analyze the effect of related orderings on the expected execution time
of a merge sort. (Consider the effects of duplicate keys on expected
string lengths.)
6.19 Let D be a O-origin matrix of the form of Table 6.14 which describes the
cascade sort, i.e., row Dj describes the distribution of strings at the completion of
stage j. Using the special matrices of Sec. 1.13, write concise expressions for the
lnatrices F and B such that
(a) Dj-I = B ~ Dj
(b) Dj = F ~ Dj-I
(c) Show that F and B are inverse.
(d) Determine the dominant eigenvalue of B when v(B) = 3, and show its
relation to the power of the cascade sort for four files. (cf. Sec. 6.1 and
Carter (1962).)
6.20 Determine the relative efficiencies of serial column sorting and serial merge
sorting for the following conditions. Internal sorting, with a maximum of 100
item positions, is to be used as an aid to each of the processes, and the time for the
internal sorting is assumed fixed. There are 10,000 items with 4-digit decimal
keys, and each key value is associated with at most four items. The initial arrange
ment contains 3500 maximal (increasing) strings.
6.21

Program an address table sort.

6.22 (a) The determination of the permutation vector required in the address
table sort can be considered as a mapping from each item onto its rank
in the set. Show that for distinct keys this mapping can be performed
by counting for each item the number of items having a smaller key.
(b) Progranl the method of part (a). (This is known as a counting sort.)
6.23

(a) Program a two-phase internal merge sort.

(b) Program a two-phase internal column sort.

6.24 Program an extension of Program 6.33 (pth-degree selection executed on

the right list node vector e 2 /]T) to
(a) sorting with replacement.
(b) the case of a singular homogeneous tree with dispersion vector veT).
(C) cover both cases (a) and (b).
6.25 I f the transfer from a serial file can proceed concurrently with other opera
tions, it is frequently advantageous to associate two fields of internal storage
(called buffers) with each file and to transfer the next group of items to one of the
fields while executing necessary operations on the items of the other. Buffers may
be used similarly for output files.
(a) Program an nz-way single-phase merge using two buffers for each of the 2m
serial files.
(b) Program a base b single-phase column sort using two buffers for each of the
2b serial files.

Exercises

245

6.26 If the nunlber of input buffers serving 111 files is reduced to I1Z -t- 1 sonle
advantage nlay still be gained by "predicting" the file whose (currently) associ
ated buffer will next be exhausted, and initiating a transfer from it to the idle
buffer. Repeat parts (a) and (b) of Exercise 6.25 for (111 + 1) and (b + 1) input
buffers, respectively. (See Friend, 1956.)
6.27 Repeat parts (a) and (b) of Exercise 6.25 with the nunlber of output buffers
also reduced as in Exercise 6.26.
6.28 Since a given initial arrangement may be easier to bring to descending than
to ascending order on the keys, and since a final reversal of order nlay be easy to
achieve (by backward read in the case of serial files or by address modification in
the case of internal storage), it nlay be advantageous to choose ascending or
descending order according to some estimate based on the initial arrangenlent.
Write a program which first counts the number of ascending strings and then
executes a ranking sort by insertion to produce either ascending or descending
order according to which appears to be the easier to achieve.
6.29 For the first few values of n, compute and compare the following alterna
tive evaluations of the expected nunlber of stages in a bubble sort
'It
sn-s s !
(a) n + 1 - zen), where zen) =
.'i
1
n.

2: -,- .

(b)n+l-v1Tn/2.

(c) c/n!, where c is the total count of all stages required for the n! possible
initial arrangements of n distinct keys.

chapter

THE LOGICAL CALCULUS

The present chapter develops two fundamental areas of symbolic logic:

canonical forms and the basic procedures of decomposition. a-origin
indexing is used throughout.

7.1

ELEMENTARY IDENTITIES

Certain elementary identities will first be summarized for reference.

The first of them (equation 7.1) merely defines a matrix of operators
employed in equation 7.4.

(V\=F ~)

o=

/\ Ix = V Ix
=FIx = =Ix

De Morgan's laws

(7.1)
(7.2)
(7.3)
(7.4)

=FIx

21 +Ix

(7.5)

=FIx = ((x ~. y) =F (x
-# / X

r y) -# (X ~ y))

(7.6)
(7.7)

(q A r)

(a /\ q) A (c /\ r)

(7.8)

= (\I Ix)

/\

(V Iy)

(7.9)

=FI-#I/(x 7 y) = (=FIx) /\ (=Fly)

(7.10)

= ((X

y))

/<=

V Iv

(a

f c) /\

II(x r, y)

Identities 7.2, 7.3, and 7.5 may be established by induction on the

dimension of x. Equation 7.4 summarizes the sixteen identities obtained
by extending equations 7.2 and 7.3 to arrays. For example, if 0 i h and
0 k j are the operators -# and /\, respectively, then equation 7.4 becomes
X ~ y

X ~ Y.

246

7.2

Canonical forms

247

The foregoing relation may be verified as follows:

(X

< Y)/ == =l=/(X i

== == I( Xi
== ==/(X i
==

/\

Yj)

1\ Y j )

(by equation 7.3)

V YJ

(by equation 7.2)

(X = Y)/.

Equation 7.7 is a direct extension of equation 7.6, which is itself derived

as follows. Since the operator =1= is associative and commutative, then

=1= Ix

==

((:/= ICyIX)) i= (=1= I(yIX)).

Moreover,

i=(ylx == 2 I( + Iylx) == 2 I( + I(x

1\

y) == =f/(x

1\

y),

the first and second and the third and fourth limbs being related by
equation 7.5. Consequently,

== (( =1= I(x /\ y)) =1= (~I (x /\ y))

== (( X ~\ y) i= (x <= y)).

=1= Ix

The following argument establishes equation 7.8. By definition,

Z/ ==

Consequently,

==
==
==

(a i 1\ c j ) 1\ (qi 1\

rJ

(a i 1\ qi) 1\ (c j 1\ r j )

(a 1\ q)i 1\ (c 1\ r)j.
(a 1\ q) c; (c 1\ r).

Equation 7.9 IS obtained by noting that if M == x X y, then M j

Yj.
Then
(v IIM)j == V IIMj == (v Ix) 1\ Yj'

==

1\

(vII M ) == (v/x)e 1\ y.
v/vll M == (v/ x) 1\ (vly).

and
Finally,

The derivation of equation 7.10 is similar.

7.2

CANONICAL FORMS

Intrinsic vector

Any function defined on a finite domain can be specified by listing each

possible value of the argument together with the corresponding function
value. For a logical function of n variables, the n arguments may be
considered as the components of a logical vector x of dimension n, and the

248

7.2

The logical calculus

domain of the function is then represented by the (rows of the) matrix

T(n) of dimension 2 n x n defined as follows:
~Tk =

For n

k.

3, for example, T has the form shown in Table 7.1.

P = T /\

i(f)

0 0
0 1
1 0
1 1
0 0
0 1
1 0
1 1 1

1
0
1
0

Xo 1\
Xo 1\
Xo 1\
Xo 1\

X o 1\ Xl

0
0
0

Xo

0
0
0
0
1
1
1

1\

Xl
Xl
Xl
Xl

Xl

X o 1\ Xl
Xo

1\

Xl

1\
1\
1\
1\
1\
1\
1\
1\

x2
x2

x2
x2

x2

x2

x2
x2

.((X) = i(() XP
=

(x o 1\

1\

Xl

Table 7.1

x2)

(x o 1\

Xl

1\

x2 )

V (x o 1\ Xl 1\

x 2)

The disjunctive canonical form

A logical function I can therefore be specified by its intrinsic vector i(/)

defined by:
Table 7.1 shows i(f) for the function

I(x) =
=

(x o 1\ Xl 1\ x2 ) V (x o 1\
(x o V Xl) 1\ x2

Xl 1\

x2 )

(x o 1\

Xl 1\

x2 )

A pplying the usual notation for operations on variables to operations

on functions as well (e.g., j denotes the function inverse to f, and f 1\ g
denotes the conjunction of the functions I and g) permits the expression
of certain easily derived identities concerning intrinsic vectors:

it!) =
=

i(1 V g)

j(f)
i(/) V i(g).

More generally, the intrinsic vector of any function of functions is the

same function of their intrinsic vectors.
The two trivial functions identically one and identically zero will be
called, respectively, the unit function and zero function, and will be denoted
by 1 and O. Thus i(l) = , and i(O) = e.

Canonical fornls

7.2

249

Characteristic vectors

A vector which represents a function j' is called a characteristic rector of

The intrinsic vector i(f) is but one of several useful characteristic
vectors.
The expression Y ~ x denotes the function obtained by first negating
each Xi for which Yi = 0 and then taking the conjunction (that is, and
function) over the resulting vector. Such a function is called a minterm in
x, and the components of the minterm rector

comprise all possible minterms in x with no repetitions.

The component Pi is a function of x which assumes the value one if and
only if x = Ti. Consequently, for any functionJ,

f(x)

= V /(i(f)jp)
= V /(i(f) /\ p)
= i(f) p.

This relation is illustrated by Table 7. I.

The expression

f(x)

i(f)

P = iCf) > (T ~ x)

is called the disjunctive canonicalform of the function f since it displays.r as

a (unique) disjunctive (that is, or) function of minterms, each of which
involves all variables and conjunction and negation only. The disjunctive
is one of several canonical forms of interest, each of which assumes the
form
f(x) = Y(f)g~ sex),
where the characteristic vector y(f) is a function off only, and the specific
vector sex) is a function of the argument x only. Each of the four forms of
interest is characterized by the particular over-all operator 0 1 occurring
in it and is therefore called the disjunctice, conjunctice, exclusice disjunctice,
or equicalence canonical form * according to whether 0 1 is the operator

v,A,i=,or=.
The characteristic vector and the specific vector appropriate to each forn1
will be characterized by its over-all operator 01' Thus

* The functions x V !J, x A y, (x i-: Y), and (x = y) are, except for the trivial functions,
the only associative commutative functions of two variables.

til

~
~

I i(n

II y(j./\) I

T~X

II y(j. \/) I

~X

y(f/)

ITO

y(f.=) I

XX

;<7)

I Y(f,V)

X'o /\ Xl /\

Xo

xl

Xo A xl A x 2

Xo

xl \j

Xz

Xo A xl !'\

Xo

X2

Xo /\

xl

!\

X2

Xo

V Xl V

1 /\ xl 1\ X 2

Xo

/\

xl

I, XZ

Xo

xl

\I x 2

Xo

!\ 1 1\ 1

Xo

Xo

!\

xl

!,

Xz

Xo

Xl

Xz

Xo

!'\ 1 A X z

Xo V

Xo

xl

!\

X2

Xo

Xl

\I x 2

Xo

;\

xl

!\

X2

Xo

\I Xl V

(Xo II
_

j(x)

(x o V

_
Xl)

!\

X2

Y(f"AJ

y(J,ic )

y(f.=)

r:;'

-
~
~

o
o

aO'Q

Xl /\

i;)

xl

V (0) V

(1) !'I (X O V Xl V

X2 )

V Xz

(XO /\

1\ (1) ;\

xl

Xo

X2 )

!'I

(X O

Xl

V (0) V
V

X2)

(X O

Xl

!\

/\ (1) /\

!\

(X 2)

# (0) # (0) # (0) F (0) #- (x o A

Xl)

#- (x o A

[(1)

(x2)

Xl)

(1)

(1)

Table 7.2

(1)

(1)

= (Xo V

Xl

V 0

V 0 V 0

0 V

XOVXIVO

X2

Xo

X2 )

Xl

s:::

0
0

0
0

V (0) V (0) V (0) (Disjunctive)

Xl

Xl /\

(xo V Xl

X2

X'2

o V Xl V X2

(XO V

0 V 0 V

Xl ;\

[(1)

oV

!\ 1

X o /\ xl /\

X2
Xl

1 /\ 1 !\ x 2

1 /\

o I

oV0 V0

1/\1 A 1

x2 ) A (xo V Xl

x 2)

(Xo

Xl

x2 ) (Conjunctive)

x 2 )] (Exclusive disjunctive)

x2)] (Equivalence)

Intrinsic, characteristic, and specific vectors

fD)

-.I

Canonical forms

7.2

25]

The forms are defined formally by the following expressions. *

f(x) =

Y(f, V) X (T ~- x)

(Disjunctive)

(7.11a)

y(f, /\) 0 (T ~ x)
y(f, efe) :(c (f 0 x)

(Conjunctive)
(Exclusive disjunctive)

(7.11b)

Y(f, =) v (T X x)

(Equivalence)

(7.11d)

(7.11c)

Table 7.2 shows the intrinsic vector of the function f of Table 7.1
together with corresponding pairs of characteristic and specific vectors.
These may be substituted in equations 7.11 (a-d) to verify that they do
represent the function f
Since x = Tk = Tk for some k, equation 7.1la may be written as

ik(f) =f(Tk ) = y(f, V) X (T).

Tk ).

Consequently,

i(f)

y(f, V) X (T ~

T) =

y(f, V) X S(V)

(7.12a)

Y(f, A) 0 (T

T) =

y(j~ A) '0 S( /\ )

(7.12b)

y(f, #) ~~ (T~,

T) =

Y(f, #) ~ S( # )

(7.12c)

y(f, =) ~ (T /~ T) = Y(f, =) v S( = )

(7.12d)

Each of the matrices S( 0) appearing in the right limbs of equations

7.l2(a-d) is a fixed function of T and is called the o-specific (e.g., dis

junctive specific) matrix. Since i(f) is a function of y(f, 0) and S( 0), the

relation between the intrinsic vector and each characteristic vector is

determined by the corresponding specific matrix.

Since S( V) = (f ~ T) = (T ~ T), it is clear that S( V) = I. Conse
quently, Y((, V) X S( V) = y(f, V), and therefore,

y(f, v) = iC[)
Similarly, S( /\)

= I,

and y(f, A)

0 S( /\) =

(7.13)

y(f, A), and again

y(f, A) = i(f).

(7.14)

An explicit expression for S( #) may be obtained by induction on the

dimension of the corresponding argument, and, to facilitate this, the
notation T(n) and S( #, n) will be used for the matrices appropriate to an
argument x of dimension n. T(n + 1) may be written in partitioned form
as

T(n

+ 1) = [E<,:__ !~!!}]'
E:

T(n)

expression T ~ x used here for the specific vector in the disjunctive form is
equivalent to the expression T !;;. x used earlier. Its use increases the uniformity of the
expressions for the canonical forms.

* The

7.2

The logical calculus

252

where and

are both column vectors. Hence

S(i=, n

+ 1) =

[:-I--~~~~] 0 [-rt)- -T~~)]

1\ S( i=, n) . E 1\ S(i=, n)]

= [E-:.------------------------[--------------------------,

f 1\ S(#, n)! f 1\ S(#, n)

since

E \~

EvE

f, and EvE

S(#, n

+ 1) =

f. Finally,

[S(i=,
n) S(i=, n)]
-------~----------------------.

Since S( i=, 1) =

[~ :] it is clear that

1 1 1 1
0
S( =1=,2)

(7.15)

S(#, n)

1 0

0 1

0 0

, S( =1=,3)

1 0

1 0

1 0 0

1 1

1 0

1 1

1 0

1 0

1 1

0 0 0 0 0 0 0 1
and so forth. *
The following useful properties of S(#) are easily verified for the fore
going examples and may be established generally by formal induction.
The matrix is self-inverse with respect to the operations (~), that is,

S( =1=) ~ S( =1=) = I.

(7.16)

Moreover, since every row of the transpose S( #) save the zeroth has an
even number of ones,
(2) I+IS(#) = EO.
Hence by equation 7.5,

#IS(#)

EO.

(7.17)

Since (by equation 7.4) S( =) = S( #) and since S( #) is of even dimension,

the same result holds for S( =).

* This result was first obtained by Muller (1954), who employed the matrix C of
binomial coefficients and showed that S( i~) = (2f)l oC. Also see Calingaert (1960).

Canonical forms

7.2

253

Equation 7.12c gives i(f) as a function of y(f, i=) and S("*). This relation
is more commonly written in transposed form and with i(f) replaced by the
equivalent y(f, V) as follows:

= S( =F)

y(f, V)

/\ y(f, =F).

(7.18)

Since S( =F) is self-inverse (equation 7.16), premultiplication of equation

7.18 by S( =F) ~ yields an identical expression for y(f, =F) as a function of
y(f, V), namely
y(f, =F) = S("*) :il Y(f, V).
(7.19)
The characteristic vectors of the identity function 1 and of the zero
function 0 may now be derived. Clearly i(l) = E and i(O) =. Hence by
equations 7.13 and 7.14,
y(l, V) = y(l, !\) = E,
and
yeO, V) = yeO, !\) = e.
Moreover,
y(I, =F) = S(#) ~ E
=

"*/S("*).

Hence, by equation 7.17,

Similarly,

y(l, #) =

EO.

yeO, "*) =

e.

(7.20)

The relations between the characteristic vectors of a function f and of its

inverse f may now be obtained. Since
i(l)

= (/),

(7.21)

then, by equations 7. 13 and 7.14,

y(f, V)

= ref,

V),

and

y(j~ !\)

= r(f,

!\).

Moreover,

y(f, #) =

S( =F) ~: i(j) = S( =F) ~~

l(f),

by equations 7.13,7.19, and 7.21. Hence

(7.22)
by equations 7.7, 7.17, 7.19, and 7.13. Characteristic vectors of a function
and of its inverse are displayed in Table 7.2.
The relation between y(f, =F) and y(f, =) may now be obtained by
applying equation 7.22 to equation 7.12c to yield

t(f)

= iC/)

(y(f, =F) #

EO)

~ S(=F).

i([)

S( #-) ~ Y(f,

i([)

i([)

Y(f, #-)

S( #-) ~

Y(f, =)

S( =)

i(/)

i(/)

v i([)

1(/)

i(/)

Y(f, =)

Y(f, #-)
#-)

S( =)

vY(f, =)

S( =) v r(f,

1(/)

Y(f, #-)

(Y(f, =) = eO)

S( #-) ~

(Y(f, #-) = eO)

Y(f, =)

S( =) v Y(f,

S( #-) ~ y(f,

=)

~
~

S( =) v 1(/)

r(f, #-)

(y(f, =) #- eO)

i(/)

S( #-)

=)

~.
~

-...
~

#-)

=)

(Y(f, #-) #- eO)

y(f, =)

S( #-) ~

y(f, =)

S( =) v 1(1)

y( f, #-)

(Y(f, =) #- eO)

S( =) v i(/)

i(/)

~ y(f, #-)

S( =) v y(f,

=)

y(f, #-)

(y(f, =) = eO)

(y(f, =) #- eO)

y(f, =)

~
~

Y(f, V) = Y(f, /\) = i(/)

S( #-, n)

= (T(n)

0 T(n

('~(F' n
=

S( =)

1)

S( #-)
Table 7.3

I :

Y(f, V) I E

i([)

~-~~--~-=--~-;- -~-~~-~-=--~-;

- c ~)

(3) S(#-, 1) =

(4)

S( #-) ~ r(f,
y(f,

S( #-) ~ 1(/)

(2)

#-)

(y(f, #-) #- eO)

.(/)

y(f, #-)

(1)

(.It
~

y(f, =)

y(f, #-)

Y(f, /\)

Y(f, #-)

I :0

Y(f, =)

EO

Zero and identity functions

Relations among characteristic vectors
CD)

-...J

7.3

255

Decomposition

Application of De Morgan's law and of the fact that S( =) = S(=F) yields

i(f) = (Y(f, =F) =F eO)

7 S( =).

Comparison with equation 7.12d shows that

y(f, =) = (y(f, =F) =F eO)

= (y(f,

=F) = eO).

The relations among the various characteristic vectors are summarized

in Table 7.3.

7.3

DECOMPOSITION

A logical function f(x) is said to be decomposable on u if it can be written

in the form
f(x) = g(h(ulx), ulx),
where g and h are logical functions. Since.f, g, and h are functions of
v(u), (1 + +Iu), and +Iu variables, respectively, then if +Iu > 1, both
g and h are functions of fewer variables than f Decomposition on u such
that + lu > 1 thus permits f to be expressed in terms of simpler functions
g and h and provides an important basis for simplification techniques.
Every function is decomposable on u = e. Moreover, if + lu = 1, then
u = i for some i, and

f(x)

[Xi 1\ f(x 1\ i)] V

[Xi 1\

f(x vel)].

Since bothf(x 1\ e) andf(x V e ) are expressible as functions ofeilx, then

the foregoing expression is of the required form with h(ulx) = Xi. Con
sequently, all functions are trivially decomposable for + Iii =
or
+Iii = 1.
i

Disjunctive canonical form

Ashenhurst (1957) determines nontrivial
arraying the intrinsic vector i(f) in a 2 +/u
follows:
F/ = ik(f),
where

and

(2e)

i = (2e)

deco~positions

of f(x) by
2 +/u matrix F defined as

(ulk)

(2e) ~ (ulk).

If, for example, u = (1, 0, 1, 1, 0), then the index k of each component ik(f)
appearing in F is given by the matrix* C of Table 7.4. The table also shows

* Ashenhurst (1957) calls the matrix C a decomposition chart and represents F by

circling each element of C which corresponds to a nonzero component of ;([).

N
Ut

o
1

3
4
5
6

7
8
9
10
11

12
13
14
15
16

17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

0 0
0 0
0 0
0 0
0 0
0 0
0 0
0 0
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
1 0
1 0
1 0
1 0
1 0
1 0
1 0
1 0
1 1
1 1
1 1
1 1
1 1
1 1
1 1
1 1

0
0
0
0
1
1
1

0
0
1
1
0
0
1
1 1
0 0
0 0
0 1
0 1
1 0
1 0
1 1
1 1
0 0
0 0
0 1
0 1
1 0
1 0
1 1
1 1
0 0
0 0
0 1
0 1
1 0
1 0
1 1
1 1

0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1

;(f>

1
0
0
0
1
0
0
1
0
1
0
0
0
1
1
0

1
0
1
0
1
0
1
0
0
1
0
1
0
1
0
1
1
0
1
0
1
0
1
0
0
1
0
1
0
1
0
1

1
0
0
0
1
1
0
1
1
0
0
1
0
0
1

'f
/\

0
0
0
0
0
0
1
1
0
0
0
0
0
0
1
1
0
0
0
0
1
1
0
0
0
0
0
0
1
1
0
0

Table 7.4

Xr
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
0
0
0
0
0
0
1
1
0
0
0
0

0
0
1
1

dVr
/\
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
1
1

0
0
0
0
0
0

Xr
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
0
0
1
1
1
1
0
0
0
0
1
1

nVr
/\
0
0
0
0
0
0
1
1
0
0
0
0
0
0
1
1
1

u=(l

0
'0

c=

0)

4
5
12
13

6
7
14
15

2
3
10
11

16
17
24
25

18
19
26
27

20
21
28
29

22)
23

30
31

(~

0
0
1

1
0

b = (0

c=(l

d = (0

1
1
1

0
0
0
0

0
1
1
0

0)

1)

0)

m =(0

0)

n=(l

1)

F=

~.

~
-..
~

a
0
0
0
0

D (D

~
~

S2"
C")

0
0
1
1
0
0
I

1
0
0
1
1
0
0

Decomposition of the function I(x) on u

CD')

--I

7.3

Decomposition

257

corresponding values of i(/) and F.

-

Let

= T

1\
c:t-'

x,

q = T .~ (ii/x),
and

r =

T~

(u/x)

be the minterm vectors of x, ii/x, and u/x, respectively. * Then the matrix

P = q

~ r

contains the components of p arrayed in the same order as the components

of i(f) in the matrix F. Consequently, t

f(x)

i(f) XP

= v/v//(F;\P).

Decomposability depends on the structure of the matrix F. If each

column of F is either zero or equal to the vector a, there exists a vector c
such that
(7.23)
F = a ~ c.
Hence,

f(x)

= v/v// M,

where

(a ~ c) ;\ (q

A r).

(7.24)

But by equation 7.8,

(a ;\ q)

? (c

;\ r),

(7.25)

(v/(a;\ q);\ (\//(c;\ r

(7.26)

and hence by equation 7.9,

f(x)

= V/V//M =

= (a X q) ;\ (c X r).

Since the first and last terms on the right of equation 7.26 are, respec
tively, functions of ii/x and of u/x only, the function f(x) is decomposable.
The required functions are simply

h(ii/x)

f(x)

r =

T ~ (u/x)

a Xq

g(h, u/x)

= a /~ [T )
h ;\ [c X rJ

(u/x)]
(7.27)

Since equation 7.27 does not represent the most general possible function
of hand u/x, it appears that the characteristic matrix F = a A c does not

* Although denoted by the same symbol, the matrices T are of differing dimensions
as required by compatibility.
'j- Since i(/) is equal to
V ), it may be substituted for it in the disjunctive canonical
form.

yet,

The logical calculus

258

7.3

represent the most general function decomposable on u. Ashenhurst

(1957) has shown that the most general type of function decomposable on
u is represented by a characteristic matrix of the form *
F = (a

A b) v

(a

7 c)

V (e ~, e),

(E ? d)

(7.28)

where b, c, d, and e are mutually disjoint and collectively exhaustive, that

IS,

e=

E.

The fourth term of equation 7.28 is identically zero and is included only
for formal completeness; hence
F

(ii I~ b) V (a

1\ d).

(7.23')

A d) 1\ (q A r),

(7.24' )

c) V (E

Equation 7.24 now becomes

M

li A b)

V (a

A c)

V (E

and since conjunction is distributive over disjunction, equation 7.8 may

again be applied to yield
M = li 1\ q) l, (b 1\ r V a 1\ q) .~ (c 1\ r V E 1\ q) A (d 1\ r.

(7.25')
Moreover, since V I V II(X \/ Y)
may again be applied to yield
f(x)
=

I IIX)

(V V

I IIY), equation 7.9

V (V V

= V I V /IM

a X q)

1\ (b :~

a /. q) /\ (c X r

(E X q) 1\ (d X r.

( 7.26' )

Since q is a specific vector of the disjunctive canonical form (i.e., a

minterm vector), it is some column of the specific matrix S( V). Since
S(v) = I, q therefore contains exactly one nonzero component, and
consequently (li X q) = (a X q), and E X q
be rewritten in decomposed form as

= h(iilx) = a q = a /\ (1' .~ (iilx

f(X) = g(h, ulx) = (ii 1\ (b Xr V (h
r = T
(ulx)

1. Equation 7.26 can thus

/\ (c Xr V (d Xr)

(7.27')

It is interesting to note that no use has been made of the fact that b, c,
and d are disjoint. Relaxation of this restriction does not, however,
increase the generality of the matrix F, since a V E = E, a V = a, and
a V
= E. It does suggest, however, that the matrix F nlay be expressed
more compactly as
F = (a A m) V (a n n),

* In Ashenhurst's terminology, the matrix F must be of the following form: each

column is either all zeros, all ones, the vector u, or the vector ii.

7.3

Decornposi t ion

259

where m = b v d and n = c V d. The second line of equation 7.27'

then becomes
(7.28)
j'( x) = (h /\ (m /, r)) \j (h /\ (n /\ r)),
a reflection of the obvious fact that the third term (d r) of equation 7.27'
can be incorporated in the preceding terms.
Table 7.4 shows a complete example of the decomposition process for
u = (1,0, I, 1,0). The characteristic matrix F is obtained by applying
the matrix C to i(j'). Clearly
F

Consequently,
f(x)

where
Since

= (li
= (a

= (Fz /\

=a

q,

c) V (e

b) V (a
m) V (a

~)

!\

(b / r)) V (h /\ (c

= T

(Xl /\

X4 ,

~\ eii/x),
Xl /\

x4 '

!\

c)

n).

r)) V (d

and r
Xl /\

r),

= f '; (ujx).

X4 ,

Xl /\

x4 ),

a = (I, 0, 0, I),

and
then

Similarly,

b
c

and

'y

(Xl /\ X4 ) V (Xl /\ x 4)

(Xl

= (xo /\ x2 /\ x~J V (X o /\ x2 /\
= x2 /\ (x o =1= x3 ),
= (x o /\ x3 ) V (xo /\ x2 /\ x3 ),
d / r = X o /\ x2 /\ xa.

x 4 )
x:~)

Alternatively, the use of the vectors m and n yields the solution

j{x)

= {h /\

[(in /\ x2
V

/\

x3 )

(x o

/\

V (xo

/\

(x 2

x3 ))])

3 )]}

{h /\ [(xo /\

3)

The entire decomposition process is described by Program 7.5. Steps

1-7 determine the characteristic matrix F appropriate to the decomposition
u. The loop 2-7 is repeated for each value of k from zero to 2 (u) - I.
Step 2 determines k as the vector (of dimension v(u)) whose base two
value is k. Steps 3 and 4 then specify the indices i and j appropriate to k,
and step 5 specifies element F/.
Step II determines d as the vector which specifies all full column vectors
of F, that is, d j = I if and only if F j = e. Step 12 determines e as the
corresponding vector specifying the zero columns.
If d and e together exhaust the columns (that is, d V e = e), then band
c (and a arbitrarily) must be set to zero. Since this is done by steps 8-10,
the exit branch on equality at step 13 terminates the process correctly. If
J1

(d V e) -#- , then any column of the matrix (d V e) / F can be used to

specify a; step 14 uses the first column. Step 15 determines b as the vector

O-origin indexing

1~

(2e(v(u)))_!_k

,--------+

-<-

0
-<-

-<-

(2e) ~(ujk)

-<-

(2e)~(ujk)

Fji

k-<-k+l

7.3

The logical calculus

260

<

'

-<-

Logical function of v(u) variables.

Logical vector specifying the de

composition

.f(x)

il/.f)

k : 2 J1 (u)

= g(h(iijx),

ujx).

Characteristic matrix for the de

composition u:

l i (F) = 2 +/ii; v(F) = 2--1 lu.

Row index for F.

e( +ju)

Column index for F.

-<-

e( +ju)

Scan index for i([).

10

e +- e( +ju)

Base 2 representation of k.

11

d+-e~F

12

e+-e~F

13

e : (d V e)

14

a -(- ((d V e)jF)o

15

b~a~F

16

e ~a ~ F

17

Nontrivial column of F.

(bjF)j

(ejF)j

(djF)j

(ejF)j

= a.

= a.

= e.

= E.

Legend

e : (b V e V d V e)

Program 7.5

Decolnposition of.f on u

specifying the columns of F which are equal to 0, and step 16 determines c

correspondingly for the vector a.
The functionfis decomposable on u if and only if b, c, d, and e together
exhaust all columns of F. The left-pointing exit on step 17 therefore indi
cates nondecomposability on u.
The algorithm can be extended to test all possible values of u successively
and so determine all possible decompositions. Use can be made of the
obvious fact that the matrix F appropriate to decomposition on u is the
transpose of the matrix F appropriate to u.
Other canonical forms
Ashenhurst (1957) remarks that decomposability is an intrinsic property
of a logical function j' and is independent of the form of its representation.
It can also be shown that the particular algorithm of Program 7.5 is intrinsic

7.3

261

Decomposition

in that it applies (in a slightly generalized form) to the characteristic

vectors of all four canonical forms.
Since i(f) = YCf, A) = y(j~ V), it is clear that the decomposition
algorithm applies directly to the disjunctive and conjunctive characteristic
vectors. The case for the remaining forms will be developed for the
exclusive disjunctive form only.
Let F be defined as in Program 7.5 but with i(}') replaced py y(f, #).
Moreover, let
p = TO x

q
r

and

= T 0 (li/x)
= T 0 (u/x),

P = q A r.

Then, clearlY.r(x) = */=I=//(F A P).

As in the case of the intrinsic vector, the structure of F must be expressed
in terms of the characteristic vectors of a given function h, of its inverse 17,
and of the identity and zero functions. I n the exclusive disjunctive form,
y(I, *) = eO, yeO, *) = E, and if y(h, =1=) = a, then y(h, *) = (a =1= EO).
The term in E nlay again be disregarded and the form required of F for
decomposability may (following equation 7.23') be written as
F

[(a

EO)

A bJ

c) V (eO ~ d),

V (a

(7.23")

where b, c, and d are mutually disjoint.

The matrix M such thatf(x) = */*//M may now (as in the analogous
case of equation 7.25') be obtained by applying equation 7.8:

(((a

* eO) A q) A(b A

r))

(a

A q)

!I (c

V (eO A q)

r))

A (d

r)), (7.25")

Equation 7.25" will also be written as

M = U

Vv W,

where U, V, and W denote the successive matrices of the right limb.

Since b, c, and d are disjoint, so also are the matrices U, V, and W. For
any pair of disjoint matrices X and Y, it is easily shown that

#/*//(X

Y)

#/#//X)

Hence

f(x)

= #/*//M = {[(*/=I=//U)

=1=

(#/#//Y)).

* (*/*//V)] # (=I=/=I=/W)}.

The application of equation 7.10 to each of the nlatrices U, V, and W no\v

yields

f(x)

= {[(a

EO) ~ q) A (b ~ r))

a 7~ q) A (c /\ r))J

* eO ~ q) A

(d'~ r))}.

(7.26")

262

The logical calculus

7.3

Since a, (a
eO), and eO are the characteristic vectors of the functions
h, h, and 1, respectively, equation 7.26" may be written in the decomposed
form
h(lijx) = a .~ (T 0 (li/x

f(x) = (((h 1\ (b ~
-/\
r = T v (ujx)

(h 1\ (c ;

r)

(d

(7.27")

For the example of Table 7.4,

y(f, eTc) = (l, 1, 1, 1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 0).

Table 7.6 show its decomposition in the exclusive disjunctive form.

(i

b= (0

c =

(I

d= (0

Table 7.6

(a i~ eO)

1 0
1 0
1 0
0 0

0 0 0 1
1 1 1 1
1 1 1 1
0 0 0 0

I) (i) (I)

1)

1 0

0)

0)

Decomposition in exclusive disjunctive form

The only change required in Program 7.5 is in the specification of the

vectors d and b on steps 11 and 15. These may be replaced by the following
steps:
11 d +- 0 /\ ]?
15

b+-(a

eO) ~ F.

The program may be made completely general (i.e., for y(f,

by simply
replacing EO with y( 1, 0) in the foregoing steps, and replacing E with
yeO, 0) in steps 8, 9, 10, and 12.

REFERENCES
Ashenhurst, R. L., (1957), "The Decomposition of Switching Functions," Proc. 0.( all
International Symposiunl on the Theory of Switchin/{, Harvard University, pp. 74
116.

Exercises

263

Calingaert, P., (1960), "'Switching Function Canonical Forms Based on Commutative

and Associative Binary Operations," presented at the October meeting of the
American Institute of Electrical Engineers.
Muller, D. E., (1954), "'Application of Boolean Algebra to Switching Circuit Design
and to Error Detection," Trans. Institute of Radio Engineers, New York, vol.
EC-3, pp. 6-12.

EXERCISES
7.1

Use De Morgan's laws (equations 7.2 and 7.3) to establish

(a) the identity U

0V

[j

XV.

(b) the identity U ~ V = U V.

(c) equation 7.4.
7.2 Show that the expressions x = (y #- z) and y = (x #- z) are equivalent.
7.3 For the function !(X o, Xl' X 2 ) = (X o /\ (Xl #- X 2 V (x o /\ x2 ), derive
(a) the intrinsic vector iC,().
(b) the four characteristic vectors of f
(c) the intrinsic vector of f
(d) the four characteristic vectors of f
7.4 Use the matrix of equation 7.1 to summarize the canonical fonn expres
sions (equations 7.11 (a-d in a single equation.
7.5 Show that y(1, /\) = Y(f, /\ ).
7.6 Use De Morgan's laws and the result of Exercise 7.5 to derive equation
7.11b from 7.11a.
7.7 Show that if y /\ z = 0, then

~ (y V z) =

X ~ y)

#- (X ~ z.

7.8 Let ,( and g be two disjoint functions (i.e., ,(x) /\ g(x) = for all x), and
let h = ,( V g. Derive expressions for the four characteristic vectors of h in terms
of the four characteristic vectors of! and g.
7.9 Each of the sixteen logical functions of two variables may be characterized
by its intrinsic vector iC,() and be denoted by

,(x, y) = flex, i(f'), y).

For example, (x #- y) = flex, (0, 1, 1,0), y). (The function flex, iC,(), y) is some
times called the kth Boolean,(unction and denoted by fl k (x, Y), where k = ~i(l).)
This notation can be extended to vectors X and y so as to permit different func
tions to be specified for each component. Thus

+-

flex, F, y) <=>

Zi

fl(x i , F i , Yi),

where v(F) = vex) and p(F) = 4. Show that

(a) if X = (0, 0, 0, 0, 0, 0, 1, 1), Y = (0, 1, 0, 1, 0, 1, 0, 1),
and II F = (0,0,0,0,6,6,6,6), then (-l(x, F, y)
= (0,0,0,0,0,1,1,0).

264

The logical calculus

(b) ()(x, F, y) = ()(x, F, y)

(c) (-J(x, F V G, y) = f3(x, F, y) V (J(x, G, y)
(d) (J(x, F 1\ G, y) = (J(x, F, y) 1\ (3(x, G, y).

7.10 The functions (3(x, F, y) defined in Exercise 7.9 can alternatively be ex

pressed as (J(x,f, y), where f =Jl F. Develop relations on f corresponding to
those of Exercises 7.9 (b-d).

SUMMARY OF NOTATION

S.1 CONVENTIONS
Basic conventions
(a) I-origin indexing assumed in this summary.
(b) Controlling variables appear to the left, e.g., ujx, b

r x.

y, k

t x, and

(c) Dirnension n may be elided (if determined by compatibility) fronl

e(n), ek(n), ak(n), wk(n), and lj(n).
(d) The parameter j may be elided from operators Ij, OJ' .L, and i j , and
from the vector li if j is the index origin in use.
(e) The parameter k may be elided from k t x if k = 1.

Branching conventions
(a)

jf

x:y~

The statement to which the arrow leads is executed next if (x/!!y)

= I; otherwise the listed successor is executed next. An unlabeled
arrow is always followed.

(b)

x : y,

The statement numbered Si is executed next if (xr i Y) = 1. The null

synlbol 0 occurring as a component of r denotes the relation which
complements the disjunction of the remaining relations in r.
~

(c)

Program a, b

Program a branches to its statement b. The symbol a may be elided

if the statement occurs in Program a itself.

Operand conventions used in summary

Logical
Integral
Numerical
Arbitrary

Scalar
u, v, w
h, i,j, k
x,y, :::

a,b,c

Vector

u,v,w
h, i,j, k
x,y, z
a,b, c
265

Matrix
V, W
H,I,J,K
X, Y,Z
A,B,C
[T,

Tree
U,V,W
H,I,J, K
X,Y,Z
A,B,C

266

Summary of notation

S.2 STRUCTURAL PARAMETERS, NULL

v(a)

Dimension
Row dimension

v(A)

Column dimension

II(A)

Number of components in each

column vector A j

Height
Moment
Dispersion vector

v(A)
v(A)

Length of longest path in A

Number of nodes in A
v 1 (A) = number of roots of A;
viA) = maximum degree of nodes
on levelj - I; v(v(A)) = v(A)

Moment vector

~(A)

II(A)

Number of components in vector a

Number of components in each row
vector Ai

1.5

1.23

<5(A)

= number of nodes on levelj

of A; v(~(A)) = v(A)
Degree of node i of tree A
b(A) = max <5(i, A)

Leaf count

A(A)

A(A) is the number of leaves in A

Row dimension
of file
Column dimension
of file

v(<1

N umber of files in each row of a

1.22
file array
Number of files in each column of a
file array

Degree of node i
Degree

6(i, A)

~/A)

N ull character

11(<1> )

Null character of a set (e.g., space in

the alphabet) or null reduction
operator

I.3

267

SumnlalY of notation

S.3 RELATIONS
Equality

a=b

Membership

aEb
b;2a
as;b
b:J a
acb
b={1

Inclusion
Strict inclusion
Similarity

1.15

a and b are identical

a = b i for some i
aj E

b for aU j

b ;2 a and a ~ b
b ;2 a and a ;2 b

The relation which holds if and only

if ff/l does not. Examples of comple
mentary pairs~ E, ; ~, =P; >', >.
A list of relations between two vari
ables is construed as the or of the
relations. Thus xc:::> y is equivalent
to x :$ y. When equality occurs as
one of the ored relations, it is indi
cated by a single inferior line, e.g.,
~ and s;.

Complementary
relations
Combined (ored)
relations

S.4 ELEMENTARY OPERATIONS

Negation

*-

w=l<=>u=O

And

*-

u 1\ v

Or
Relational state
ment

*-

u V

*-

(a .jf b)

Sum
Difference
Product

Quotient
Absolute value

z*-x+y
z*-x-y
z*-xxy
z *- xy
c*-aXlI
c *-au
z *- x
y
z

*-/

xI

Ceiling

k
k

LxJ
*- rxl

j-Residue mod h

*-

Floor

L'

*-

Ii i

1.4

= 1<=> U = I and v = 1
= I <:=:> u = 1 or v = 1

= I <=> the relation

,jf

b holds

z is the algebraic sum of x and y

z is the algebraic difference of.r and y

z is the algebraic product of numbers
x and y, and c is the arbitrary character
a or zero according to whether the

logical variable u is one or zero.

z is the quotient of x and y
z

[(x> 0) - (x <:: 0)]

kS=:x<:::k+1
k
x
k - I
i = hq + k ~
integral.

.i ~ k <.i + h;

and q is

268

Sumn1ary of notation

S.5 VECTOR OPERATIONS

Component-by C
component exten
sion of basic
operation
Scalar multiple
z

a0

Ci

= ai

(a

~1.5

0 bi. Examples: x x y,

Ii

b), h

i,

1\

v, fxl.

+- x X y
+-xy
c+-aXu
C +-au

Reduction

+-

c = (0 0 0 a l 0 a 2 ) 0 a 3 ) 0 0 .) 0 (II')' 1.8
where 0 is a binary operator or rela
tion with a suitable don1ain. Examples:
+ Ix, x Ix, ~i: Iu. Reduction of the
null vector e(O) is defined as the identity
element of the operator o. Examples:

Ola

+ le(O)
/\ le(O)

= 0;

x /e(O)

= {,

V le(O)

0,

1.
1.16

Ranking
j-origin b-index

c = 0 if a b; otherwise c is thej
origin index of the first occurrence of
a in b.

ofa

j-origin b-index
ofa

a
a

= aj, wherej = v(a) 11 (i + k)

c i = aj, wherej = v(a) /1 (i - k)
z = +I(p x x),wherePl' = {,and
Pi-l = Pi X Yi

Left rotation
Right rotation
Base y value of x

C ~-k ~

COlnpression

+-

Expansion
Mask
Mesh
Catenation

c
c
c

+-

u\b

tilc

+-

la, u, b/

tilc = ~/a,ulc = ulb

+-

\a, u, b\ tilc=a,u/c=b

+-

ci

z+-Y~x

c is obtained from a by suppressing

each b i for which u i = 0

ulb

c +-a

0, ulc

c = (aI' a 2 ,
\a, WI'(b), b\

al'(a)' b I ,

...

bl'(b

1.6

1.14

1.9

SUlnnzary of notation

269

S.5 VECTOR OPERATIONS (continued)

Characteristic of

+-

Ey X

Wi

= (Yi

x) ~ v(w) = l'(Y)

1.15

xony
W

+-

Ej(h)

Wi

Full vector

+-

E(h)

Wi

=
=

Zero vector

W --

e(h)

W +-0

Wi

=0

Prefix of weightj

jth unit vector

1.7

(i =j)
1
V(lV) = h

aj(h)

First k of Wi are unity

where k = min (j, h).

Suffix of weightj w+-wj(h)

Last k of Wi are uni ty

where k = min (j, h).

+-

Maximum prefix

+-

'Y./u

Maximum suffix

+-

w/u

is the max length prefix in fl.

Example:

1.1

a/(I, 1,0,1,0, I) = (1,1,0,0,0,0).

is the max length suffix in u.

Example:

w/(I, 1,0,1,0,1) = (0,0,0,0,0, I).

Forward set
selector

W +-

a/a

Backward set
selector

+-

T/a

Maxima selector

+-

ufx

Wi

= I if a j

Wi

= ui

1\

/=

ai

for allj < i

(Xi

nl)

where

nl)

where

1.18

m = max (u/x)j
j

Minima selector

+-

ulx

Wi
nl

U i

Interval or
j-origin
identity permu
tation vector

j-origin permu
tation vector

j-origin mapping

c *- a b
c ~bLa

j-origin ordering

+-

lj(h)

1\

(Xi

= min (u/x)j

k = (j, j

+ 1, ... ,j + h

- I)

1.7

1.17

+-

()j/x

Ci

= if b i lj(v(a; otherwise
0

a bi in aj-origin system for a. In

the first form (that is, c +- a b ), the
origin cannot be specified directly.
Ci

= kLx is in ascending order and

original relative ordering is n1ain
tained among equal components, that
is, either Yi < Yi-[-l or Yi = )'i-1 1 and

k i < k i+ 1

270

Summary oj' notation

S.6a ROW GENERALIZATIONS OF VECTOR

OPERATIONS
Z~XOY

Z/ = X/ 0 Y/

Z ~

O/x

Zi =

CZ:/A

C = Al

+-

M~BlItA

l.16
l.6

~ktA

~k~A

Ci

~Ab

Ci

K ~ 0/t/X

+--

Ki

C ~u/B

Zi =

k i ~ Ai

yi _L Xi

l.14

C j = A bj

Sit A

l.8
1.15

= k i t Ai

A 2 (2'; ..
... ~) A"

Mi=Bil/lAi
Ci

z~Y-lX

1.5

O/Xi

l.17

= Bi S/tAi
= (),JXi

Ci =Ai

Bi

l.9

Ci = U/Bi

U/B

C =

Ul/Bl

... CD Ull/Bil
C~u\B

~U'\b

C ~'\A,u,B\
C --(- ',a, U, b'\

C ~
C

jA, u,

ja, U, bj

W~rJ./U
W~w/U

BI

/A, U, B/

a/U

W~T/U
W~UrX
W~U[X

= 0, u/C = B

ti/C
[J/C

ti/C

= A, ujC =

0, UjC

D/c = a, U/C = b
ti/C = ii/A, u/C = ujB
fJ/C = [J/A, U/C = U/B
C = jE.\a, U, E.\bj

Wi = rJ./Ui
Wi =w/Ui
Wi = a/Ui
Wi = T/Ui
Wi = Uir Xi
Wi = Uit Xi

1.10

l.18

Summary of notation

271

S.6b COLUMN GENERALIZATIONS OF VECTOR

OPERATIONS
Z+--XOY
Z

+--

o//X

+--

0//A

Zi
Zj
C

= X/ 0 y.iJ
= o/Xj
= AI OA2 (8

1.5
1.8
...

1.15

@AII
M

+--

tth

Mj

=
=
=

Bj

tl!

kj

t Aj

1.l6

Aj

1.6

C+--k,rA

Cj

C+--kVA
z<-YJlX

kj ~ Aj
Zj = Y j -L X j

1.14

+--

Ab

Ci =Ab,

1.l7

+--

+--

O,j/X

HI! A

C+--AGJffiB
C

<-

u//B

c+--V//B

Cj

Cj
Kj

=
=

B j Sil A j
(j,jX j

C j = u/B j

c = V1/B 1 8:) ...

... CD V..IBv

= 0, u//C = B
iJ//c = 0, V//C = b

+--

u\\B

<-

V\\b

+--

\\A, u, B\\

ii//C = A, u//C = B

+-- \

la, V, b\\
//A, u, B//

U//C

C~
C //A, V, B//
+--

+--

I/a, V, b//

ii//C

ii//C

= a, V//C = b
= iii/A, u//c =

W j = (X/V,

w
w

w//V

W+--a//V
W <- TI/V
W +-- VrrX
W <- VLlX

u//B

Ui/c = Vi/A, V//C = V//B

C = IE.I\a, V, E.\\b/

W+--(X//V
+--

1.9

C j = A j CD B j

j
j

a/Vj

Wj
Wj
Wj

T/Vj
V,rXj
V;lXj

=
=

I.IO

w/vj

1.18

Sunlnlary 0.[ notation

272

S.7 SPECIAL MATRICES

Full matrix

E(p, q)

W/=l

Zero matrix

W +- F.(p, q)
W+-O

W/ =0

+-

kl(p, q)

Superdiagonal

Identity

W +-I(p, q)

Upper left
(triangle)
Upper right
Lower left

+-

[](p, q)

+-

[)(P,q)

+- ~(p,q)

~-

+-

Wji
W

,u(W) = P
v(W) = q, for
p andq

= (i + k = j)
= IO(p, q)

:J

Wi

Lower right

1.13

W~+l-j

= (i + j ~

W~+l-iJm =

m),

min (p,

q))

integers.
Elision of
P and q if
dimensions
determined
by
compatibility.

Wf1+ 1 - i
v+l-i

[](p,q)

S.8 TRANSPOSITION

1.12

C.i
z =

C+-B

Diagonal

C+-B

Counter diagonal

/'
C+-B

cV(B)+l-~

Horizontal

C+-B

C!~(B)+l-i

f1(B)+l-z
J

Vertical

C+-B

Vector

Y
Y

+-

Bi
J

Cf'(B)+l-i

Yi = x v+ 1
i

+-

S.9 SET OPERATIONS

Intersection

c+-bna

Difference

c+-b~a

Union
Cartesian
product

c+-bua

+-

bi @
b n

c = EbOjb
c = EbOjb

c = b CD (a ~ b)

Cl +d l-(k-E) = (b k1 \ bk~ , ... bk~);

d j = v(b j ); 1 s:: k j s:: d j

Clearly, v( c) = n, and fi( c) = x jd

1.15

Summary oj' notation

273

S.10 GENERALIZED MATRIX PRODUCT

C/ = OI/(Ai O 2 B j ), where O 2 produces a vector 1.11
(Le., is not the operator l-), and 0 1 is a reduction
operator (and hence C/ is a scalar).
Ci = (Ai 0 b), where 0 is any operator which
produces a vector of dimension v(b).

C j = (a 0 B j ), where 0 is any operator which

produces a vector of dimension v(a).

C/ = (a i 0 bj).

S.ll FILES
File

<D/

A representation of a of the form 1.22

, av(a), PV(ll) + 1, 0,
Pv(a) -f 2, . , Pv(P), where Ph is the
partition at position It, PI = P"'(P) = A,
and (ttl /\ (1)/P S; A.
Set file to position It. Called rewind
if h = 1, and wind if h = v(p).
(Pv aI' P 2 , a 2 ,

Position file

7T(<D j i )

+-

Record (fronl position h)

Forward
o<D/ +-a, Ale a h +- a, Ph-t-l +- Ak ; stop at position

It + 1. Zero prescript nlay be elided

and Al may be elided.

Backward
a h- l
a; Ph-l +- Ak ; stop at
position It - 1. Al lllay be elided.
Read (from position h)
Forward
a, b +- o(D/ a +- aIL; b +- Ph-t--I; stop at position
h + 1. Associated branch is controlled
by Ph+I' and b may be elided. Zero
prescript may be elided.
Backward
a +- a h - I ; b +- Ph-I; stop at position

It - 1. Associated branch is con

trolled by Ph--I and b may be elided.

File array
(T>
Full
Array offiles <1>/, for i E ll(/Il),
j E ll(v(D)).
Row
Row of files <I>/, for} E ll(v(<l).
Column
Column of files <1>/, for i E l\/I(<I).
Compression
u/(f)
Row
Selection as in corresponding opera
u//(I>
Colunln
tions on nlatrices.
-0(-

274

Summar.y of notation

S.12 TREES
Path i

Node i
Subtree i

+-Ai

C 1 is the i 1 th root of A; c j is the ijth 1.23

node of the nodes on level j reach
able from node c j _ 1 .
C +- (Ai)ll(i) The final node of path Ai.

C +-Ai
C is the subtree of A rooted in node i.

Component-bycomponent
Path reduction

C +- O/A

Level reduction

C +- O//A

j-origin b-index

B +- b

j-origin mapping

C+-bLA

Full right list

n1atrix

C +-]A

Full left list

matrix

C +-[A

Right list matrix

Left list matrix

C +- a 2
C +- a 2 /[A

C+-AOB (Ci)l'(i) = (Ai)v(i) 0 (Bi)v(i).

lj

IlA}

Tree compression C

+-

U /A

Path compression C +- u/A

Level com
pression
Level mesh
Level mask

C +-u//A

C +- \ \A, u, B\ \
C +- //A, u, B//
Path catenation C +- A ''1 B

Reduction by operator or relation 0

on nodes in left list order.
Reduction by operator or relation 0
on nodes in right list order.
(Bi)u(i)

lj

A i }l'(i)

Rooted subtree C i is a single null

character node if b i Lj(~I(A)); other
wise C i = A bi , where A is treated in a
j-origin system.
The rows of the index matrix Ci2/C are
the right (left) justified index vectors
(with null fill to the common
dimension v(A)) arranged in increasing
order; C 1 and C 2 are the correspond
ing degree and node vectors of A.
The degree and node vector colun1ns
of the full right (left) lIst.
C is obtained from A by suppressing
node i if node i of U is zero and re
connecting so that for each remain
ing pair of nodes, the one lies in the
subtree rooted at the second if and
only if it did so in A.
C is obtained from A by suppressing
all nodes on levelj if u j = 0, and re
connecting as in the compression U /A.
C is obtained from A by suppressing
rooted subtree A j if u j = 0.

ii//C = A; u//C = B
ii//C = ii//A; u//C = u//B.
C is obtained by connecting roots of
B to leaves of A, allotting successive
groups of at most r~I(B) ~- A(A)]
roots of B to each successive leaf of A.

SUlnmary of notation

275

S.12 TREES (continued)

Full tree

Zero tree

Each node ofW is unity and the

structure of W is deternlined by
compatibility.

W +-- E(k)

Each node of W is unity; W is honlo

geneous (i.e., all nodes on any level
have a comnlon degree) and 'J(W) = k.

W +-- E
W +--0

Each node of W is zero and the

structure of W is deternlined by
compatibility.
Each node of W is zero; W is honlo
geneous and v(W) = k.

W
Path tree

+-

E(k)

W +-- UE
W +-- UE(k)

Level tree

~ 1.23

W +-- E

W +-- u E
W +--uE(k)

u/W = 0; u/W = E; structure ofW

determined by conlpatibility.
u/W =0; u/W = E; Wishonlo
geneous and 'J(W) = k.
u//W = 0, u//W = E; structure of W
determined by conlpatibility.
u//W = 0; u//W = E; W is honlo
geneous and 'J(W) = k.
W = U 1\ (A = 111E), where 111 is the
maximum (mininlum) over all nodes
of VIA.

Maximization
Minimization

W+--VrA
W+--VlA

Maximum path
prefix

Maximum path
suffix

W +-- w/V

Forward path
set selector

a/A

(Wi)l'(i) = 1 if (Ai),'(i) differs fronl all

preceding nodes of path Ai.

Backward path
set selector

W *--T/A

Maximum level
prefix

W ~- ex//V

(Wi)"(i) = 1 if (Ai)l'(i) differs fronl all

other nodes of its subtree.
j
j
e /W = a/e /V

Maximum level
suffix
Forward level
set selector
Backward level
set selector

W ~- w//V

ej/W

w/ej/U

W+--a//A

ej/W

a/ej/A

W ~- T//A

ej/W

T/ej/A

~-

~-

ex/V

W is obtained from V by zeroing all

nodes of every subtree rooted in a
zero node.
W is obtained fronl V by zeroing
every node which contains a zero
node in its subtree.

Index

O-origin indexing, 71, 186

O-residue, 14
I-origin indexing, 12, 16
I-residue, 14

Autonlatic programming, 71
Auxiliary variables, 38
Axis of transposi tion, 25

Backward chain, 113

Base address register, 74
Abelian group, 69
conversion, 190
Accumulator, 78
value, 27, 38, 146
index, 73, 79, 83
Active domain, 148, 153
Batching, 138
Betz, B. K., 198
Add, clear and, 76
Binary opera tion, 13
Additive indexing, 73
search, 141, 155, 221, 236
Address, 72
Birkhoff, G., 26, 28, 32
leading, 106
table sort, 176, 213, 231
Biunique transformation, 144, 146
Address-ordered pool, 117
Blaauw, G. A., 110
Addressing, indirect, 32, 74, 76, 84
Block chaining, 119
system, open, 153
sort, 191,213
Aids to serial sort, 204
transfer, 206, 213, 221
Algebra, rna trix, 24
Boehrn, E., 153
Algorithm, Euclidean, 67
Boldface letters, 10
Allocation, implicit, 108
Boolean function, 263
of memory, 105
Branch arrow, 2
Alternator, 184, 188
in files, 42, 181
Amdahl, G. 1\1., 153
in a graph, 46
Anlphisbaenic sort, 191, 195, 203
instruction, 73, 80
Analysis in p, 160
Inodifica tion, 6
Analyzer, 159
Branching, 75
Arbitrary variable, 10, 13
ratio, 47
Argunlent, 10
Brooks, F. P., 129
prograrll, 159
Bubble sort, 213, 217, 229, 233, 239
Ari thmetic instructions, 84
Bucket file, 154
opera tions, 11
Burks, A. \;\~., 53, 122, 163, 164
Arrow, 2
Busy indica tor, 94
Ashenhurst, !{. L., 188, 190, 191,255,258
Assembler, 159
Calingaert, P., 252
Associa tive memory, 104
Canonical form, 247, 249, 255, 261
277

278

Index

Cap, 29
Carr, J. \V., 171
Carter, \V. C., 198, 203, 244
Cartesian product, 30
Cascade sort, 191, 198, 203
Category, 189
Catenated scan, 135, 139
Ca tena tion, 20, 21, 29
Ceiling function, 12
Central register, 72, 78
Chain, 113, 117
filial, 128
list matrix, 121, 126
Chained representation, 110, 116, 135,
147, 221
stack, 113
Chaining block, 119
end-around, 112, 135
vector, 155
Change sign, 76
Channel-unit interlock, 95
Channel, cornn1and, 88
data, 88

input-output, 75, 95

instruction, 100

trap, 75, 97

Character, null, 31, 58

Characteristic matrix, 258
Characteristic vector, 29, 249, 253, 261
Check, parity, 98
Chien, R. T., 154
Church, Robert, 239
Circuit, 46
Class of operand, 8
Classification, 177
and simple rnerge, 178, 186,201,211
sirnple, 177, 20.1
string, 183
Clear and add, 76
Clustering of keys, 154
Coalescence, 144, 148, 152
Code, Gray, 63
Huffman, 51

operation, 72

point, 124

variable length, 124

Column-chained n1atrix, 121

Column, compression, 22
dimension, 14
expansion, 21

Column, list, 22, 108, 119

rnapping, 33
rnask, 21
rllesh, 20
of files, 43
operation, 8, 18
reduction, 17
sort, 178, 186, 191, 193, 204, 2 11
vector, 14
Comrnand, 71

channel, 88

register, 73

Cornparison, 2

of internal sorting, 239

with zero,S

Corl1patibility, fIrst-order, 171

in corllpression, 18

of trees,S 7

of vectors, 16

Con1piler, 159
Cornponents, 13, 22
Cornpound operation, 8
staternent, 160, 165
Corllpression, 18, 21, 110, 113
tree, 54, 60
Cornputer fdes, 44
instruction, 71
progran1, 71
Congruence, 12
Conjunctive canonical fonn, 249
Connection n1a trix, 46
Contained in an ordering, 190
Control register, 72
Controlled scan, 135, 141, 147
Controlling parameter, 8
Conventions, 7
Conversion, base, 190
Convert instructions, 87
Copy operation, 177, 189
Counter, instruction, 73
Crarner, I-I., 138, 235
Cup, 29
Cycles, disjoint, 32
Cyclic scan, 135
Da ta channels, 88
path, 101
De Morgan's law, 16, 24, 64, 103, 246
Decisions, leading, 6
Decon1position, 246, 255

Index
Decomposition, chart, 257

of a vector, 21

Decrement, 78, 83

Degree, 47

vector, 49, 53, 163

Derived key, 134, 144, 148

Design, hardware, 71

logical, 101

systern, 71

Dewey, G., 129

Dickson, L. E., 65

Difference, set, 29

Digital sort, 186

Din1ension, 5, 13, 38

-ordered pool, 11 7

colurnn, 14

elision of, 16

in 11', 106, 108

Directed graph, 46

scan, 135, 147

Director, 159

Disjoin t cycles, 32

vectors, 29

Disjunctive canonical form, 249, 255

Disordered pair, 233

Dispersion, suffix, 173

vector, 47, 59, 165

Distributor, 78

Domain, 10

active, 148, 153

Double chain, 113, 115

Dummy item, 179

variables, 40

Duplicate keys, probability of, 201

Dynamic relocation, 74

Elen1ent, null, 6, 10, 16, 32

of a rna trix, 14

Elen1en tary operations, 11, 13

staten1ent, 160

Eligible iten1, 179

Elision, 5, 8, 11, 16, 32

Enable, 75, 100

Encoding rnatrix, 105, 109, 134

End of ftIe, 90

End-around chaining, 112, 135

Entry arrow, 3

Equivalence canonical form, 249

class, 134, 147

function, 12

279

Equivalence, of staternents, 11

reduction, 17

Equivalent formulas, 166

Euclidean algorithm, 67

Euler's constant, 235

Evaluation, of internal sorting, 231

of serial sorti ng, 199

Exchange sort, 212

Exclusive disjunctive canonical fonn,

249, 261

Exclusive-or, 12

reduction, 17

Execu te instruction, 83

Execution, instruction, 72, 77

sequence, 2, 5

tin1e of a sort, 199

Exit arrow, 3

Expansion, 19, 21

Expected nurnber, of strings, 200

of transposi tions, 232

Falkoff, A., 104

Fanlilies of operations, 8

Fetch instruction, 72, 76

rHode indicator, 76, 95

Fields, 205

Fifo discipline, 117

File, 40, 11 5, 135, 171, 176

bucket, 154

compu ter, 44

coordinate, 177

end of, 90

rewind, 183

stack, 172

Filial chain, 128

vector, 128

Filial-heir chain list, 127

Filler cornponents, 22

First-order con1pa tibili ty, 171

Fixed scan, 137

Floor function, 12

Flowcharting, 1

Fonnat vector, 105, 134

Fonnula, 163

equivalent, 166

nlinimax fonn, 166

Newton-R.aphson, 64

red uction of, 164

singular, 164

280

Index

Fortuitous function, 133

Forward chain, 113, 117
Fraction, scan, 135, 140
Freiman, C. V., 154
Friend, E. H., 233, 245
Full list matrix, 50
rna trix, 22, 26

vector, 15

Function, Boolean, 263

ceiling, 12
equivalence, 12
floor, 12
fortuitous, 133
identity, 253
inverse, 37
key, 176
magnitude, 11
mapping, 34
program, 159, 169
sign, 11
trivial, 248
unit, 248
zero, 248, 253
Gap, record, 98
Garner, H. L., 67
Gassner, B. J., 239
Genera tor, 159
Gilstad, R. L., 198
Goldstine, H. H., 1, 178
Graeffe method, 65
Graph, directed, 46
index, 48
Gray code, 63, 104
Greek letters, 9
Grid matrix, 106, 118
Group, abelian, 69
rota tion, 26
Hanan, M., 154
Harary, F., 132
Hardware designer, 71
Height of a tree, 47
Homogeneous tree, 58, 121, 226
Huffman code, 51
Huffman, D. A., 123, 129, 132
Identities, logical, 246

Identity function, 253

matrix, 26

Identity, permutation vector, 33

Immediate branching, 181
Implicit allocation, 108
specification, 38
I nclusion, set, 28
Index, 31
accumulator, 73, 79, 83

graph, 48

list, 51, 60

matrix, 49, 53, 73, 125

origin, 16, 30, 32

register, 73, 108

systems, 14

vector, 48, 60

Indexability, 76
Indexing, a-origin, 71, 186
1-origin, 12, 16
additive, 73
j-origin, 14, 43
operation, 31, 36
Indicator, busy, 94
fetch mode, 95
limit, 90
Indirect addressing, 32, 74, 76, 84
Infix vector, 16
Information store, 72
Initial node, 46
Initialization, 6
Input file, 41
Input-output channel, 75, 95
instructions, 88
I nsertion, of an instruction, 75
ranking by, 213
sort, 212
Instruction, arithmetic, 84
branch, 73, 80
channel, 100
cOlnputer, 71
convert, 87
counter, 73
enable, 75
execution, 72, 77
fetch, 72, 76
input-output, 88
logical, 78, 83
shift, 85
Integral variable, 10, 13
Interchange, 7
Interchange sort, 232
Interlock, channel-uni t, 95

Index
Internal sorting, 176, 204, 205, 212, 231,
239
Interpreter, 160
Interruption, 75, 81
Intersection, set, 28
Interval vector, 15, 33
Intrinsic vector, 248, 261
Inverse characteristic vector, 253
functions, 37
mapping, 34
permutation, 33, 139
Italics, 10
Iterns, 176
dummy, 179
eligible, 179, 181, 213, 229
relatable, 177, 189
Iteration, 4
Iverson, K. E., 69, 129

Key, 133, 176

clustering, 154
derived, 134, 144, 148
multiple, 143, 155
transformation, 134, 144, 147
vector, 176, 189
Kronecker delta, 12
Kunz, K. S., 64

Left list, 122

rotation, 8
Length, of a path, 46
scan, 135, 148
Level compression, 54, 60
of a tree, 46
tree, 57
of structure, 39
LIFO discipline, 117
Limit indicator, 90
Linear representation, 108, 110, 118
List, chain, 121
column, 22
index, 51, 60
leaf, 122, 145
left, 122
matrix, 49, 108, 121, 161
processing, 110
right, 226
row, 22
vector, 122, 163
well formed, 52
Literals, 9
Load and store, 78
Logic, symbolic, 24
Logical design, 101
iden ti ties, 246
instructions, 78, 83
operations, 11
trees, 60
variable, 10, 13
vector, 18
Loop, 2
Lukasiewicz, J., 160, 163
Lukasiewicz notation, 122, 161, 169

oC-formula, 163, 173

minimax form, 166
oC-notation, 163, 169
oC-phrase, 163
oC-reduction, 164
Lagrange multipliers, 226
Lamb, S. M., 146
Language, programming,
structure, 7
Last in first out, 117
Leading address, 106
decisions, 6
Leaf, 46
list, 122, 145

Maclane, S., 26, 28, 32

Magnitude function, 11
Mapping, 32, 34', 58, 133
Margenau, H., 69
Marimont, R. B., 132
Marked pool, 11 7
Mask, 8, 19
Matrices, representation of, 119
vector of, 39
Matrix, 13, 14
algebra, 24
chain list, 126
characteristic, 258
column-chained, 121

j-origin indexing, 14, 43

Jacobsen, W. H., 146
Jacobson, N., 28, 65, 69
Johnson, L. R., 121, 148, 155
Jordan box, 68
elimina tion, 68
Justification, 23

281

282

Index

Matrix, conlponent, 14

connection, 46

dimension, 5

element, 14, 75

enable, 100, 105

encoding, 109, 134

full, 22, 26

grid, 106, 118

identity, 26

index, 49, 53, 73, 125

list, 49, 108, 121, 161

nlultiplication, 4

product, 23, 33

row-chained, 120

sparse, 22, 120

special, 26

specific, 251

superdiagonal, 26

trace of, 27

triangular, 26

Mauchly, J. \V., 178, 191

Maximal path, 46
Maximization, 36, 58
Maximum prefix, 9, 22, 58
suffix, 22, 58, 165
McCracken, D. D., 195
Member of a vector, 28
Mernory, 72
allocation, 105
Merge, 177
sinlple, 177, 191,205
single-phase, 179, 183, 206
sort, 178, 199, 204
string-doubling, 186
two-phase, 179
IVlesh, 19, 110
Metaprogranl, 71, 159
Microprograms, 71
Mid-square method, 154
Minimax form of -formula, 166
Minimization, 36, 58
Minterm vector, 249, 257, 258
Miscellaneous symbols, 10
Mixed radix, 27
1\10ment, of a tree, 47
vector, 47, 60, 125
Moore school, 178, 191
Muller, D. E., 252
Multiple keys, 143, 155
scalar, 13

M ultiplica tion, elision of, 11

generalization of, 11
nlatrix, 4
Multiplier-quotient register, 78
1\1 urphy, G. 1\1., 69
Nagler, H., 195,242
Newton-Raphson formula, 64
Node, 46
vector, 46, 49, 226
Nonlinear represen ta tion, 109
N ormaliza tion, 23
Notation, Lukasiewicz, 122, 161, 169
parenthesis, 161, 169
parenthesis-free, 122, 163
Polish, 122, 163
sunlnlary, 7
translation of, 169, 173
Null character, 31, 58
element, 6, 10, 16, 32
scan, 25, 28
vector, 16, 17
N umber base, choice of, 190
NurTIber systerTI, 12, 14, 27
Numerical variable, 10, 13
Odd-even transposition sort, 213, 232
Oettinger, A. G., 171
Open addressing system, 153
Operand, classes, 8
structured, 8, 13, 45, 106
Operation, arithmetic, 11
binary, 13
code, 72
column, 8, 18
compound, 8
copy, 177, 189
count, 212
elerTIentary, 11, 13
families, 8
indexing, 31, 36
logical, 11
mapping, 58
ordering, 36
rank-preserving, 177, 189, 231
ranking, 31, 34, 133
revision, 183
row, 8, 18
set, 28
shift, 78

Index
Operator, statement as, 160
symnletric, 166
Optimum m in m-way sort, 200
Optinlum base for a sort, 202
Ordered set, 23, 28
tree, 45, 47
Ordering, contained in, 190
of stat emen ts, 5
on x, 29
opera tion, 36
related, 189, 200, 202
vector, 36, 176
Output file, 41
Outer product, 25
Overflow indica tor, 85
posi tions, 78
search, 149
toggle, 83
CP-notation, 163, 169
Palernlo, F. P., 154
Parameter, controlling, 8
Parenthesis, elision of, 8
notation, 161, 169
Parenthesis-free notation, 122, 163
Parity, 12
check, 98
Partial pass sort, 191, 203
Parti tion, 115
file, 41, 181
symbol, 115
Partitioned representation, 115, 118
Partitioning of a tree, 123
Path compression, 56, 60
data, 101
nlaximal, 46
tree, 57
zero extension of, 61
Permutation, 8, 32, 133, 139
of a tree, 58
vector, 33, 148, 176, 213, 231
Peterson, \V. \V., 154
Phase, 179
Phister, M., 63, 104
Phrase, 163
Physical vector, 105
Polish notation, 122, 163
Polynomial, 27
Polyphase sort, 199
Pool, 113, 116

283

Posi tion file, 41

Power of a merge sort, 199
Pre-count column sort, 211
Prefix, code, 123
maximum, 9, 22, 58
vector, 15
Probability of duplicate keys, 201
Product, Cartesian, 30
Ina trix, 23, 33
outer, 25
Program, 2
argument, 159
computer, 71
function, 159, 169
in p, 160
interaction, 7
meta-, 71
skeleton, 160
statement, 2
trace, 83, 160
utility, 160
vector, 7
Programming, automatic, 71
language, 1
Pseudo-english,
Q register, 78
Queue disci pli nes, 11 7
Quotes denoting literals, 10
Radix exchange sort, 191, 212, 213
mixed, 27
Random-access sort, 176
storage, 205
Range, 10, 38, 159
I~ank-preserving operation, 177, 189, 231
Rank, coordinate, 177
of an array, 39
Ranking, by insertion, 213, 236
operation, 31, 34, 133
sort, 213, 221, 235, 239
Ratio, branching, 47
storage, 205, 212, 231, 237
Read file, 41
Record, file, 41
gap, 88, 98
Reduction, 8, 16
of a formula, 164
of a tree, 58
Register, central, 72

284

Index

Register, command, 73
index, 73, 108
multiplier-quotient, 78
sequence, 73
Relatable items, 177, 189
Related orderings, 189, 200, 202
Relational statement, 11, 16
Relocation, dynamic, 74
Reordering, 32
Repeated selection sort, 213, 223, 229,
236
Replacement sort, 213, 226, 239
Representation, of variables, 105
chained, 110, 116, 135, 147,221
linear, 108, 110, 118
nonlinear, 109
of matrices, 119
of trees, 121
partitioned, 115, 118
solid, 106
Residue, 12
Revision operation, 183
Rewind, 90
file, 41, 183
Richards, R. K'., 63
Right list, 226
Ring, 101
Riordan, J., 69
Roman letters, 10
Root, 46
vector, 52
Rooted scan, 135
tree, 47
Ross, I. C., 132
Rota tion, 33
group, 26
left, 8, 14
Row-chained matrix, 120
Row, compression, 18, 21
dimension, 14

expansion, 21

list, 22, 108, 119

mapping, 33

mask, 21

mesh, 20

of files, 43

operation, 8, 18

reduction, 17

vector, 14

Run signal, 81

Rutishauser, H., 69
San1uel, A. L., 153
Satellite vector, 155
Scalar n1ultiple, 13
Scan, 133
catenated, 135, 139

controlled, 135, 141, 147

cyclic, 135

directed, 135, 147

fixed, 137

fraction, 135, 140

length, 135, 148, 231

null, 25, 28

of equivalence class, 147

rooted, 135

shortest, 139

to nearest terminal, 139

Schay, G., 154

Search, 133
binary, 141, 221, 236
overflow, 149
single table, 149
Selection, 8, 17, 32, 38
set, 30
vector, 22
Selector set, 23, 58
Self-indexing, 40
Sequence, of execution, 2, 5
register, 73
Serial access, 41, 135
sort, 176, 199,204
Set operations, 28
selector, 23, 58
Shaw, J. C., 110
Shift instructions, 85
opera tion, 78
Sign function, 11
Similar vectors, 29
Simple classification, 177, 205
classification and merge, 178, 200, 206
merge, 177, 191,205
Single table search, 149
Single-phase merge, 179, 183, 206
vs two-phase, 200
Singular formula, 164
tree, 47, 60, 165
Skeleton program, 160
Skip, 75, 80
Solid representation, 106

Index
Sort, address table, 176, 213, 231
amphisbaenic, 191, 195, 203
block, 191, 213
bu~ble, 213, 217, 233, 239
cascade, 191, 198, 203
column, 178, 186, 204
digital, 186
exchange, 212
execution time, 199
insertion, 212
internal, 176, 204, 205, 239
merge, 178, 204
odd-even interchange, 232
optimum base, 202
partial pass, 191, 203
polyphase, 199
radix exchange, 191, 213
random-access, 176
ranking, 213, 221, 235, 239
repeated selection, 213, 223, 229, 236
serial-access, 176
string-doubling, 211
tournamen t, 237
transposition, 212, 219, 236, 239
two-phase column, 187, 193
two-phase vs single~phase, 200
with replacement, 213, 226, 239
Sparse matrix, 22, 120
Special logical trees, 60
matrices, 26
vectors, 15
Specific matrix, 251
vector, 249, 258
Specifica tion, arrow, 2
implicit, 38
program, 160
Spread, 144, 146, 154
Spruth, W. G., 154
Stack, 113, 117
vector, 166, 170
file, 172
Stage, 179, 231
Statement, compound, 160, 165
elementary, 160
equivalence, 11
number, 3
ordering,S
program, 2
relational, 11, 16
Stirling's formula, 240

Storage, allocation in sorting, 209

random-access, 205
ratio, 205, 212, 231, 237
Store, information, 72
String-doubling merge, 186, 211
Strings, 178
classification, 183
distribu tion of, 198
expected number, 200
Structure, levels of, 39
of the language, 7
Structured operands, 8, 13, 45, 106
Subpath, 46
Subroutines, 39
Subtree, 47, 48, 51, 58
Suffix, dispersion, 165, 173
maximum, 22, 58
vector, 15
Summary of notation, 7
Superdiagonal matrix, 26
Supervisor, 159
Symbolic logic, 24
Symbols, miscellaneous, 10
Symmetric operators, 166
System designer, 71
Tag vector, 79
Tape units, 88
Tensor, 39
Terminal node, 46
parti tion of a file, 41
Tie, 10
Toggle, enable, 75
Tou~nament sort, 237
Trace, of a matrix, 27
program, 83, 160
Transfer, 75, 80
block, 213, 221
Transformation, key, 134, 144
spread, 154
Translation of notation, 169, 173
Translator, 159
'[ransposition, 231
axis, 25
sort, 212, 219, 236, 239
Trap, 75, 81, 97
Tree compatibility, 57
compression, 54, 60
homogeneous, 58, 121, 223
level, 46, 57

285

286

Index

Tree, logical, 60

moment, 47

ordered, 45

parti tioning, 123

path, 57

permutation, 58

reduction, 58

representation, 121

rooted, 47

singular, 47, 60, 165

uniform, 58, 145

Triangular matrix, 26

Trivial functions, 248

Two-phase, column sort, 187, 193

merge, 179

vs single-phase sort, 200

Uniform tree, 58, 145

Union, 28

U nit function, 248

vector, 15, 30

storage ratio, 212

Univac, 106

Utility program, 160

Variable, 9

arbitrary, 10, 13

auxiliary, 38

integral, 10, 13

length code, 124

logical, 10, 13

numerical, 10, 13

representation, 105

Vector, 13

catenation, 20, 21, 29

chaining, 155

characteristic, 29, 249, 253, 261

coalescence, 148, 152

degree, 49, 53, 163

disjoint, 29

dispersion, 47, 59

filial, 128

format, 105, 134

Vector, full, 15

identity permutation, 33

index, 48, 60

infix, 16

interval, 1.1, 33

intrinsic, 248, 261

key, 176, 189

list, 122, 163

logical, 18

mapping, 134

minterm, 249, 257, 258

moment, 47, 60, 125

node, 46, 49, 226

null, 16

of rna trices, 39

ordering, 36, 176

path, 46

permutation, 33,148,176,213,231

physical, 105

prefix, 15

program, 7

representation, 78

root, 52

satellite, 155

selection, 22

similar, 29

special, 15

specific, 249, 258

stack, 166, 170

suffix, 15

tag, 79

unit, 15, 30

Vocabulary, 163, 169, 170

Von N eu man n, J., 1, 178

\Vell formation, 161

\Vell forn1ed, 52, 164

\Vind file, 41

\Voods, F. S., 235

\Vord of memory, 72

\Vright, H. N., 12

Zero, extension of a path, 61

function, 248, 253