LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES

Managing Editor: Professor J.W.S. Cassels, Department of Pure Mathematics

and Mathematical Statistics, 16 Mill Lane, Cambridge CB2 1SB, England
The books in the series listed below are available from booksellers, or, in
case of difficulty, from Cambridge University Press.
4
5
8
11
12
13
16
17
18
20
21
23
24
25
26
27
29
30
31
32
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51

Algebraic topology, J.F.ADAMS

Commutative algebra, J.T.KNIGHT
Integration and harmonic analysis on compact groups, R.E.EDWARDS
New developments in topology, G.SEGAL (ed)
Symposium on complex analysis, J.CLUNIE & W.K.HAYMAN (eds)
Combinatorics, T.P.McDONOUGH & Y.C.MAVRON (eds)
Topics in f i n i t e groups, T.M.GAGEN
D i f f e r e n t i a l germs and catastrophes, Th.BROCKER & L.LANDER
A geometric approach to homology theory, S.BUONCRISTIANO, C.P.ROURKE
& B.J.SANDERSON
Sheaf theory, B.R.TENNISON
Automatic continuity of l i n e a r operators, A.M.SINCLAIR
Parallelisms of complete designs, P.J.CAMERON
The topology of S t i e f e l manifolds, I.M.JAMES
Lie groups and compact groups, J.F.PRICE
Transformation groups, C.KOSNIOWSKI (ed)
Skew f i e l d constructions, P.M.COHN
Pontryagin d u a l i t y and the structure of LCA groups, S.A.MORRIS
I n t e r a c t i o n models, N.L.BIGGS
Continuous crossed products and type I I I von Neumann algebras,A.VAN DAELE
Uniform algebras and Jensen measures, T.W.GAMELIN
Representation theory of Lie groups, M.F. ATIYAH e t a l .
Trace ideals and t h e i r a p p l i c a t i o n s , B.SIMON
Homological group theory, C.T.C.WALL (ed)
P a r t i a l l y ordered rings and semi-algebraic geometry, G.W.BRUMFIEL
Surveys in combinatorics, B.BOLLOBAS (ed)
Affine sets and a f f i n e groups, D.G.NORTHCOTT
Introduction to Hp spaces, P.J.KOOSIS
Theory and applications of Hopf b i f u r c a t i o n , B.D.HASSARD,
N.D.KAZARINOFF & Y-H.WAN
Topics in the theory of group presentations, D.L.JOHNSON
Graphs, codes and designs, P.J.CAMERON & J.H.VAN LINT
Z/2-homotopy theory, M.C.CRABB
Recursion theory: i t s generalisations and applications, F.R.DRAKE
& S.S.WAINER (eds)
p-adic analysis: a short course on recent work, N.KOBLITZ
Coding the Universe, A.BELLER, R.JENSEN & P.WELCH
Low-dimensional topology, R.BROWN & T.L.THICKSTUN (eds)
F i n i t e geometries and designs,P.CAMERON, J.W.P.HIRSCHFELD & D.R.HUGHES (eds)
Commutator calculus and groups of homotopy classes, H.J.BAUES
Synthetic d i f f e r e n t i a l geometry, A.KOCK

52 Combinatorics, H.N.Y.TEMPERLEY (ed)

54
55
56
57
58
59
60
61
62
63

Markov process and related problems of analysis, E.B.DYNKIN

Ordered permutation groups, A.M.W.GLASS
Journees arithmetiques, J.Y.ARMITAGE (ed)
Techniques of geometric topology, R.A.FENN
S i n g u l a r i t i e s of smooth functions and maps, J.A.MARTINET
Applicable d i f f e r e n t i a l geometry, M.CRAMPIN & F.A.E.PIRANI
Integrable systems, S.P.NOVIKOV e t a!
The core model, A.DODD
Economics for mathematicians, J.W.S.CASSELS
Continuous semigroups in Banach algebras, A.M.SINCLAIR

64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110

Basic concepts of enriched category theory, G.M.KELLY

Several complex variables and complex manifolds I , M.J.FIELD
Several complex variables and complex manifolds I I , M.J.FIELD
C l a s s i f i c a t i o n problems in ergodic theory, W.PARRY & S.TUNCEL
Complex algebraic surfaces, A.BEAUVILLE
Representation theory, I.M.GELFAND e t a l .
Stochastic d i f f e r e n t i a l equations on manifolds, K.D.ELWORTHY
Groups - St Andrews 1981, C.M.CAMPBELL & E.F.ROBERTSON (eds)
Commutative algebra: Durham 1981, R.Y.SHARP (ed)
Riemann surfaces: a view towards several complex variables,A.T.HUCKLEBERRY
Symmetric designs: an algebraic approach, E.S.LANDER
New geometric s p l i t t i n g s of classical knots, L.SIEBENMANN & F.BONAHON
Linear d i f f e r e n t i a l operators, H.O.CORDES
Isolated singular points on complete i n t e r s e c t i o n s , E.J.N.LOOIJENGA
A primer on Riemann surfaces, A.F.BEARDON
P r o b a b i l i t y , s t a t i s t i c s and analysis, J.F.C.KINGMAN & G.E.H.REUTER (eds)
Introduction to the representation theory of compact and l o c a l l y
compact groups, A.ROBERT
Skew f i e l d s , P.K.DRAXL
Surveys in combinatorics, E.K.LLOYD (ed)
Homogeneous structures on Riemannian manifolds, F.TRICERRI & L.VANHECKE
F i n i t e group algebras and t h e i r modules, P.LANDROCK
S o l i t o n s , P.G.DRAZIN
Topological t o p i c s , I.M.JAMES (ed)
Surveys in set theory, A.R.D.MATHIAS (ed)
FPF r i n g theory, C.FAITH & S.PAGE
An F-space sampler, N.J.KALTON, N.T.PECK & J.W.ROBERTS
Polytopes and symmetry, S.A.ROBERTSON
Classgroups of group r i n g s , M.J.TAYLOR
Representation of rings over skew f i e l d s , A.H.SCHOFIELD
Aspects of topology, I.M.JAMES & E.H.KRONHEIMER (eds)
Representations of general l i n e a r groups, G.D.JAMES
Low-dimensional topology 1982, R.A.FENN (ed)
Diophantine equations over function f i e l d s , R.C.MASON
V a r i e t i e s of constructive mathematics, D.S.BRIDGES & F.RICHMAN
Localization in Noetherian r i n g s , A.V.JATEGAONKAR
Methods of d i f f e r e n t i a l geometry in algebraic topology,
M.KAROUBI & C.LERUSTE
Stopping time techniques for analysts and p r o b a b i l i s t s , L.EGGHE
Groups and geometry, ROGER C.LYNDON
Topology of the automorphism group of a free group, S.M.GERSTEN
Surveys in combinatorics 1985, I.ANDERSEN (ed)
E l l i p t i c a l structures on 3-manifolds, C.B.THOMAS
A local spectral theory for closed operators, I.ERDELYI & WANG SHENGWANG
Syzygies, E.G.EVANS & P.GRIFFITH
Compactification of Siegel moduli schemes, C-L.CHAI
Some topics in graph theory, H.P.YAP
Diophantine analysis, J.LOXTON & A.VAN DER POORTEN (eds)
An introduction to surreal numbers, H.GONSHOR

London Mathematical Society Lecture Note Series. 110

An Introduction to the Theory of

Surreal Numbers
HARRY GONSHOR
Rutgers University

The right of the

University of Cambridge

If; w

all manner of books

was granted by
Henry VIII

in

1534.

The University has printed

and published continuously

since 1584.

CAMBRIDGE UNIVERSITY PRESS

Cambridge
New York New Rochelle
Melbourne Sydney

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www. Cambridge. org
Information on this title: www.cambridge.org/9780521312059
Cambridge University Press 1986
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 1986
Reprinted 1987
Re-issued in this digitally printed version 2008
A catalogue record for this publication is available from the British Library
Library of Congress Cataloguing in Publication data
Gonshor, H.
An introduction to the theory of surreal numbers (London
Mathematical Society lecture note series; 110)
Bibliography: p.
1. Numbers, Theory of. I. Title. II. Series
QA241.G63 1986 512'.7 86-9668
ISBN 978-0-521-31205-9 paperback

CONTENTS
Page
Preface
Acknowledgements
Chapter 1

Introduction

Chapter 2
A.
B.
C.

Definition and Fundamental Existence Theorem

Definition
Fundamental Existence Theorem
Order Properties

Chapter 3
A.
B.
C.
D.

The Basic Operations

Addition
Multiplication
Division
Square Root

13
13
17
21
24

Chapter 4
A.
B.
C.
D.

Real Numbers and Ordinals

Integers
Dyadic Fractions
Real Numbers
Ordinals

27
27
28
32
41

Chapter 5
A.
B.
C.
D.
E.

Normal Form
Combinatorial Lemma on Semigroups
The a) Map
Normal Form
Application to Real Closure
Sign Sequence

52
52
54
58
73
76

Chapter 6

Lengths and Subsystems which are Sets

95

Chapter 7

Sums as Subshuffles, Unsolved Problems

3
3
4
8

104

Chapter 8
A.
B.

Number Theory
Basic Results
Partial Results and Unsolved Problems

page
111
111
114

Chapter 9
A.
B.
C.
D.
E.

Generalized Epsilon Numbers

Epsilon Numbers with Arbitrary Index
Higher Order Fixed Points
Sign Sequences for Fixed Points
Quasi e type Numbers
Sign Sequences in Quasi Case

121
121
124
129
135
138

Chapter 10 Exponentiation
A.
General Theory
B.
Specialization to Purely Infinite Numbers
C.
Reduction to the Function g
D.
Properties of g and Explicit Results

143
143
156
167
175

References

191

Index

192

PREFACE

The aim of this book is to give a systematic introduction to

the theory of surreal numbers based on foundations that are familiar to
most mathematicians. I feel that the surreal numbers form an exciting
system which deserves to be better known and that therefore an exposition
like this one is needed at present. The subject is in such a pioneering
state that it appears that there are many results just on the verge of
being discovered and even concepts that still are waiting to be defined.
One might claim that one should wait till the theory of
surreal numbers is more fully established before publishing a book on
this subject. Such a comment reminds me of the classic joke about the
person who is afraid of drowning and has vowed never to step into water
until he has learned how to swim. In fact, the time is ripe for such a
book and furthermore the book itself should contribute to developing the
subject with the help of creative readers.
The subject has suffered so far from isolation with pockets
of people in scattered parts of the world working on those facets of the
subject that interest them. I hope that this book will play a role in
eliminating this isolation and bringing together the mathematicians
interested in surreal numbers.
The book is thus a reflection of my own personal interest.
For example, Martin Kruskal has developed the theory of exponentiation
from a somewhat different point of view and carried it in different
directions from the presentation in this book. Also, I recently received
correspondence from Norman Ailing who has recently done work on a facet
of the theory of surreal numbers not discussed in the book. With greater
communication all this and more could play a role in a future edition.
The basic material is found in chapters 2 through 5. The
later chapters are more original and more specialized. Although room for

future improvement exists everywhere, chapters 7 and 8 are in an

especially pioneering position: this is where the greatest opportunity
seems to exist for knowledgeable readers to obtain new results.
ACKNOWLEDGEMENTS
I would like to thank the following people for their help in
connection with the manuscript.
First, there is Professor Larry Corwin who took time from his
very busy schedule to do a great deal of proofreading. He is responsible
for many improvements in the exposition throughout the manuscript. On
the other hand, I take full responsibility for any faults in the
exposition which still remain. Professor Joe D'Atri and Jim Maloney, a
graduate student, have also helped with some proofreading. Also, I
should mention Professor Barbara Osofsky who much earlier had read a
preliminary draft of chapters one and two and made many valuable
suggestions.
Finally, I mention the contribution of two secretaries of the
Rutgers Mathematics Department. Mary Anne Jablonski, the co-ordinator,
took care of numerous technical details, and Adelaide Boulle did an
excellent job of turning my handwritten draft into typescript.

INTRODUCTION

The surreal numbers were discovered by J.H. Conway. He was

mainly interested in games for which he built up a formalism for generalizing the classical theory of impartial games. Numbers were obtained
as special cases of games. Donald E. Knuth began a study of these
numbers in a little book [2] in the form of a novel in which the characters are trying to use their creative talents to discover proofs.
Conway goes into much more depth in his classic book On Numbers and Games
[1].
I was introduced to this subject in a talk by M.D. Kruskal
at the A.M.S. meeting in St. Louis in January 1977. Since then I have
developed the subject from a somewhat different foundation from Conway,
and carried it further in several directions. I define the surreal
numbers as objects which are rather concrete to most mathematicians, as
compared to Conway's, which are equivalence classes of inductively
defined objects.
The surreal numbers form a proper class which contains the
real numbers and the ordinals among other things. For example, in this
system ca-l, /w, etc. make sense and, in fact, arise naturally. I
believe that this system is of sufficient interest to be worthy of being
placed alongside the other systems that are being studied by mathematicians. First, as we shall see, we obtain a nice way of building up the
real number system. Instead of being compelled to create new entities at
each stage and make new definitions, we have unified definitions at the
beginning and obtain the reals as a subsystem of what we already have.
Secondly, and more important than obtaining a new way of building up a
familiar set such as the real numbers, is the enrichment of mathematics
by the inclusion of a new structure with interesting properties.

AN INTRODUCTION TO THE THEORY OF THE SURREAL NUMBERS

In fact, it is because the system seems to be so natural to

the author that the first sentence contains the word "discovered" rather
than "constructed" or "created." Thus the fact that the system was
discovered so recently is somewhat surprising. Be that as it may, the
pioneering nature of the subject gives any potential reader the
opportunity of getting in on the ground floor. That is, there are
practically no prerequisites for reading this book other than a little
mathematical maturity. Thus the reader has the opportunity which is all
too rare nowadays of getting to the surface and tackling interesting
original problems without having to learn a huge amount of material in
advance.
The only prerequisite worthy of mention is a minimal
intuitive knowledge of ordinals, for example familiarity with the
distinction between non-limit and limit ordinals. For a fuller
understanding it is useful to be familiar with the basic operations of
addition, multiplication, and exponentiation.
The results and some of the proofs in the earlier chapters
are essentially the same as those in [1] but the theory begins with a
different foundation. The later chapters tend to be more original. The
ideas in Chapters 6 and 7 are new as far as I know. [1] contains several
remarks related to chapter 9 where the ideas are studied in detail. Part
of the material in chapter 10 was done independently by Kruskal. At
present, his work is unpublished. I would like to give credit to Kruskal
for pointing out to me that exponentiation can be defined in a natural
way for the surreal numbers. Using his hints I developed the theory
independently. Although naturally there is an overlap at the beginning,
it appears from private conversations that Kruskal did not pursue the
topics in sections C and D.

DEFINITION AND FUNDAMENTAL EXISTENCE THEOREM

DEFINITION

Definition. A surreal number is a function from an initial segment of

the ordinals into the set {+,-}, i.e. informally, an ordinal sequence
consisting of pluses and minuses which terminate. The empty sequence
is included as a possibility.
Examples. One example is the function f defined as f(0) = +, f(l) = and f(2) = + which is informally written as (+-+). An example of
infinite length is the sequence of u> pluses followed by u> minuses.
The length Jt(a) of a surreal number is the least ordinal a
for which it is undefined. (Since an ordinal is the set of all its
predecessors this is the same as the domain of a, but I prefer to avoid
this point of view.) An initial segment of a is a surreal number b
such that i(b) (a) and b(a) = a(a) for all a where b(a) is
defined. The tail of b in a is the surreal number c of length
(a)-(b) satisfying c(a) = aU(b)+a]. Informally, this is the
sequence obtained from a by chopping off b from the beginning, a
may be regarded as the juxtaposition of b and c written be.
For stylistic reasons I shall occasionally say that a(a) = 0
if a is undefined at a. This should be regarded as an abuse of
notation since we do not want the domain of a to be the proper class of
all ordinals.
Definition.
follows:
a and

If a and

b are surreal numbers we define an order as

a < b if a(a) < b(a) where a is the first place where

b differ, with the convention that - < 0 < +, e.g.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

It is clear that this is a linear order.
essentially a lexicographical order.

In fact, this is

B FUNDAMENTAL EXISTENCE THEOREM

Theorem 2.1. Let F and G
a e F and b e G => a < b.
length such that a e F =*> a
is an initial segment of any
(Note that F or G may be

be two sets of surreal numbers such that

Then there exists a unique c of minimal
< c and b e G =>> c < b. Furthermore c
surreal number strictly between F and G.
empty.)

Note. Henceforth I use the natural convention that if F and G are

sets then "F < G" means "a e F and b e G =* a < b," MF < c" means
"a e F => a < c" and "c < G" means "b e G =* c < b." Thus we may
write the hypothesis as F < G.
Example. Let F consist of all finite sequences of pluses and G be
the unit set whose only member is the sequence of u> pluses. Then
F < G. It is trivial to verify directly that c consists of w pluses
followed by a minus, i.e., F < c < G and that any sequence d satisfying F < d < G begins with c.
This theorem makes an alternative approach to the one in [1]
possible. In [1] the author regards pairs (F,G) as abstract objects
where the elements in F and G have been previously defined by the
same method, as pairs of sets. (It is possible to start this induction
by letting F and G both be the null set.) Since different pairs can
give rise to the same number, the author needs an inductively defined
equivalence relation. Theorem 2.1 gives us a definite number corresponding to the pair (F,G) so that we dispense with abstract pairs.
Proof. Clearly, it suffices to prove the initial segment property.
Case 1.

If F and

Case 2. G
a e F

G are empty, then clearly the empty sequence works.

is empty but F is nonempty.

Let a be the least ordinal such that there does not exist
such that a($) = + for all 3 < a. Thus a cannot equal zero,

DEFINITION AND FUNDAMENTAL EXISTENCE THEOREM

since any

a vacuously satisfies the condition

a(3) = + for all

5
3 < 0,

Subcase 1. a is a limit ordinal. I claim that the desired c is the

sequence of a pluses, i.e., i(c) = a and c(3) = + if 3 < a.
Since, by choice of a, no element a of F exists such
that a(3) = + for all 3 < a, every element of F is less than c.
Now let d be any surreal number such that F < d.
Suppose y < a. Then y + 1 < a, since a is a limit
ordinal. Hence, by choice of a, there exists a e F such that
a(3) = + for all 3 < y+1, i.e. 3 y. Since a < d, d(3) = + for all
3 <_ T. In particular, d(y) = +. Thus c is an initial segment of d.
Subcase 2. a is a non-limit ordinal, y + 1. In this case there exists
an a e F such that a(3) = + for all 3 < y but there is no a e F
such that a(3) = + for all 3 <_ Y- Hence any a e F satisfying:
(3 < Y => a(3) = +) must satisfy: (a(y) = - or 0 ) . If all such a
satisfy a(y) = - then it is easy to see that the sequence of Y pluses
works for c. If there exist such an a e F such that a(Y) = 0, i.e.
the sequence of Y pluses belongs to F, then the sequence of ( Y + D
pluses works for c.
Case 3. F is empty but G is nonempty. This case is similar to Case 2.
Case 4. Both F and G are nonempty.
Let a be the least ordinal such that there do not exist
a e F, b e G such that a(3) = b(3) for all 3 < a. Again a * 0.
Subcase 1. a is a limit ordinal. Suppose Y < <*; then Y + 1 < <*. Hence
there exist a e F, b e G such that a(3) = b(3) for all 3 _< Y .
The value a(Y) is well-defined in the following sense. If (a ,b ) is
another pair satisfying the above properties then a(3) = a (3) for all
3 _< Y . Otherwise, suppose 6 <_y is the least ordinal for which
a(3) * a x (3). Without loss of generality assume a(6) < a ^ a ) . Then by
the lexicographical order b < a l f which is a contradiction since b e G
and a x e F. Thus there exists a sequence of length a, such that for
all y < a there exist a e F and b e G such that
a(3) = d(3) = b(3) for 3 < y.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

By hypothesis on a, d cannot be an initial segment of an

element in F as well as an element in G. Furthermore, an element of
F which does not have d as an initial segment must be less than d.
(Otherwise we obtain the same contradiction, as before.) Similarly an
element of G which does not have d as an initial segment must be
greater than d.
It follows that if d is neither an initial segment of an
element of F nor an initial segment of an element of G then d works.
Now suppose F has elements with initial segment d. Then
G does not have such elements. Let F1 be the set of tails with
respect to d of all such elements in F. Apply case 2 to F1 and 4
to obtain d 1 . Then the juxtaposition dd1 works.
First, as before the required inequality is satisfied with
respect to all elements in F or G which do not begin with d. Since
F1 < d1 it follows from the lexicographical order that dd' is larger
than all elements in F beginning with d.
On the other hand, let e be any element satisfying
F < e < G. Recall that for all y < a there exist a e F and b e G
such that a(3) = d(3) = b(3) for 3 <_ y. This implies by the
lexicographical order that e(3) = d(3) for 3 < a. Thus d is an
initial segment of e. Again using the lexicographical order the tail
e must satisfy F1 < e 1 . Hence d1 is an initial segment of e 1 .
Therefore dd1 is an initial segment of e.
A similar argument applies if G has elements with initial
segment d.
Subcase 2. a is a non-limit ordinal y+1. Then there exist a e F,
b e G such that a(3) = b(3) for all 3 < Y but there do not exist
a e F, b e G which agree for all 3 y. As before, the values a(3)
are well-defined, and we obtain a sequence d of length y. Again, as
before, any element in F which does not have d as an initial segment
must be less than d and an element in G which does not have d as
an initial segment must be greater than d.
Let F1 be the set of tails with respect to d of elements
in F which begin with d and similarly for G 1 . Then as stated
previously, there do not exist a e F 1 , b e G1 such that a(0) = b(0).
[Note that in contrast to subcase 1, F1 and G1 are non-empty although

DEFINITION AND FUNDAMENTAL EXISTENCE THEOREM

one of these sets might contain the empty sequence as its only element.]
Since F1 < G 1 , it follows that a(0) < b(0) for all a e F 1 , b e G 1 .
Now suppose d e F and d e G. This means that neither F1
1
nor G contains the empty sequence, i.e. a(0) and b(0) are never
undefined. Since a(0) < b(0), we obtain: a(0) = - and b(0) = +. It
is then clear that d works.
Since F and G are disjoint, d belongs to at most one of
F and G. Suppose that d e G. A similar argument will apply if
d e F. Then every a in F1 satisfies a(0) = -. Let F" be the set
of tails of F1 with respect to this -. (Such an iterated tail is,
clearly, the tail with respect to the sequence (d-.) Apply case 2 to
F" and <j> to obtain d 1 . Then the juxtaposition c = d-d1 works. We
already know that c satisfies the required inequality with respect to
those elements that do not begin with d. Since no b e G1 has
b(0) = -, this takes care of all of G. The choice of d1 takes care
of all elements in F beginning with d (the next term of which is
necessarily - ) . On the other hand, any element e satisfying F < e < G
must begin with d. Since d e G, the next term must be -. By choice
of d1, it must be an initial segment of the tail of e with respect to
d-, i.e. e must begin with d-d1.
This completes the proof.
Definition.
F < c < G.

F|G

is the unique

c of minimal length such that

Remark. A slightly easier but less constructive proof is possible.

First one extracts what is needed from the above proof to obtain an
element c such that F < c < G. Although this is all that is required
for the conclusion, the proof does not simplify tremendously. Nevertheless, it simplifies slightly since there is no concern about the initial
segment property. Once a c is obtained, the well-ordering principle
gives us a c of minimal length. At this stage it is useful to have a
definition.
Definition. The common initial segment of a and b where a * b is
the element c whose length is the least a such that a(a) * b(a) and
such that c(3) = a(3) = b(3) for 3 < a. If a = b then c = a = b.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

If one of a or b is an initial segment of the other, then

c is the shorter element. If neither is an initial segment of the
other, then either a(y) = + and b(y) = - or a(y) = - and b(y) = +.
In either case c is strictly between a and b.
Now let c satisfy F < c < G and be of minimal length.
Suppose F < d < G. Let e be the common initial segment of c and d.
Then F < e < G. Since c has minimal length and e is an initial
segment of c, e = c. Hence c = e is an initial segment of d.
C

ORDER PROPERTIES

Theorem 2.2.

If G = <j> then

F|G

consists solely of pluses.

Proof. This follows immediately from the construction in the proof of

theorem 2.1. It can also be seen trivially as follows. Suppose c has
minuses. Let d be the initial segment of c of length Y where Y
is the least ordinal at which c has the value plus. Then clearly
F < d and d has shorter length than c. This contradicts the
minimality of the length of c.
Theorem 2.2a.

If F = <f> then

F|G

consists solely of minuses.

Proof. Similar to the above.

Note that the empty sequence consists solely of pluses and
solely of minuses!
Theorem 2.3.

(F|G) _< the least a such that Va[a e F u G => (a) < a ] .
This is trivial because of the lexicographical order, since
otherwise the initial segment b of F|G of length a would also
satisfy F < b < G contradicting the minimality of F|G.
Note that < cannot be replaced by _< . For example, if
F = {(+)} and G = {(++)}, then F|G = (++-). The result also follows
from the construction in the proof of theorem 2.1. In fact, the
construction gives the more detailed information that every proper
initial segment of F|G is an initial segment of an element of F U G .
(An initial segment b of a is proper if b * a ) .
Theorem 2.3 has a kind of converse.

DEFINITION AND FUNDAMENTAL EXISTENCE THEOREM

Theorem 2.4. Any a of length a can be expressed in the form
where all elements of F1JG have length less than a.

9
F|G

Proof. Let F = {b: Jt(b) < a and b < a} and G = {b: i(b) < a and
b > a}. Then F < a < G and every element of length less than a is,
by definition, in F or G so that a satisfies the minimum length
condition. Note that the argument is valid even if a is the empty
sequence.
The last result is a step in the way of showing the
connection between what is done here and the spirit of [1], since the
result says that every element can be expressed in terms of elements of
smaller length, thus every element can be obtained inductively by the
methods of [1]. The next theorem shows that the ordering in [1] is
equivalent to the one used here.
Theorem 2.5. Suppose
and F < d.

F|G = c and

F'|G' = d. Then

c _< d iff

c < G1

Proof. We know that F < c < G and F1 < d < G 1 . Suppose c _< d; then
c _< d < G' and F < c _< d. For the converse, assume c < G1 and F < d.
We show that d < c leads to a contradiction. This assumption yields
F < d < c < G. Hence c is an initial segment of d. Also
F1 < d < c < G1 so d is an initial segment of c. Hence c = d which
contradicts d < c.
This last result is of minor interest for our purpose. Its
main interest is that together with theorems 2.1 and 2.4 it shows that we
are dealing with essentially the same objects as in [1] although here
they are concretely defined. Since the present work is self-contained
this is not of urgent importance, although it is worthy of noting.
Of fundamental importance here will be what I call the
"cofinality theorems." They are analogous to classical results such as:
In the e,6 definition of a limit, it suffices to consider rational e;
and in the definition of a direct limit of objects with respect to a
directed set, a cofinal subset gives rise to an isomorphic object.
Definition. (F',G') is cofinal in
(VaeF)(3beF')(b>a) A (VaeG)(3beG'

(F,G) if

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

It is clear that (F,G) is cofinal in (F,G), and
(F",GM) cofinal in (F',G') and (F',G') cofinal in (F,G)
(F\G") cofinal in (F,G). Also if F C F 1 and G c G 1 then
is cofinal in (F,G).
The following theorems are important although they
trivial to prove.

10
that
implies
(F'.G1)
are

Theorem 2.6 (the cofinality theorem). Suppose F|G = a, F1 < a < G1, and
(F'.G1 ) is cofinal in (F,G); then F'|G' = a.
Proof. Suppose (b) < (a) and F1 < b < G 1. It follows immediately
from cofinality that F < b < G, contradicting the minimality of (a).
Hence a = F'|G\
Theorem 2.7 (cofinality theorem b). Suppose (F,G) and (F',G')
are mutually cofinal in each other. Then F|G = F'|G'.
Note that it is enough to assume that F|G has meaning since
F < G => F1 < G 1 .
Proof. {x:F<x<G} = {x:F'<x<G'}. Hence the element of minimal length is
the same.
Although the two above theorems are closely related they are
not quite the same. Theorem 2.6 will be especially useful in the sequel.
I emphasize that in spite of the simplicity of the proof it is more
convenient to quote the term "cofinality1' than to repeat the trivial
argument every time it is used. Also it is convenient often with abuse
of notation to say that H1 is cofinal in H. However, this is
unambiguous only if it is clearly understood whether H and H1 appear
on the left or right, i.e. we must consider whether we are comparing
(H,G) with (H'.G1 ) or (F,H) with (F'.H 1 ). This is usually clear
from the context.
Cofinality will be used to sharpen theorem 2.4 to obtain the
canonical representation of a as F|G. Of course, the representation
in theorem 2.4 itself may be regarded as the "canonical" representation.
The choice is simply a matter of taste.
Theorem 2.8. Let a be a surreal number.

Suppose that F1 = {b: b < a

DEFINITION AND FUNDAMENTAL EXISTENCE THEOREM

11

and b is an initial segment of a} and G = {b: b > a and b is an

initial segment of a}. Then a = F'|G'. (In the sequel F'|G' will be
called the canonical representation of a.)
Proof. We first use the representation in theorem 2.4. Then F'cF and
G'cG. Since it is clear that F1 < a < G1, it suffices by theorem 2.6
to show that (F',G') is cofinal in (F,G). Let b e F. Then
JZ,(b) < (a). Suppose c is the common initial segment of a and b.
Then b c < a. Hence c e F 1 . A similar argument shows that G1 is
cofinal in G.
The above representation is especially succinct. It is easy
to see that F1 is the set of all initial segments of a of length 3
for those 3 such that a(3) = + and similarly G1 is the set of all
initial segments of a of length 3 for those 3 such that a(3) = -.
Thus the elements of F1 and G1 are naturally parametrized by
ordinals. Furthermore, the elements of F form an increasing function
of 3 and the elements of G form a decreasing function of 3. Thus by
a further use of the cofinality theorem we may restrict F1 or G1 to
initial segments of length y where the set of y is cofinal in the set
of 3. For example, let a = (++-++). Then
F1 = {( ),(+),(++-),(++-+)} and G' = {(++),(++-+),(++-+-)}. To avoid
confusion it is important to recall that the ordinals begin with 0. So,
e.g., a(3) = +. Hence the initial segment of length 3 =
a(0),a(l),a(2)] = ++-. In other words, this terminates just before
a(3) = + so that it really belongs to F 1 . Note the way F' and G1
get closer and closer to a in a manner analogous to that of partial
sums of an alternating series approximating its sum. However, the
analogy is limited by the possibility of having many alike signs in a
row; e.g., in the extreme case of all pluses, there are no approximations
from above. As an application of the last remark on cofinal sets of
ordinals we also have a = {(+),(++-+)}|{(++),(++-+-)} or even more
simply as {(++-+--)}|{(++-+-)}, since any subset containing the largest
ordinal is cofinal in a finite set of ordinals.
In view of the above it seems natural to use F1 and G1 for
the canonical representation of a. F and G on the other hand appear
to contain lots of extra "garbage."

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

12

Finally, we need a result which may be regarded as a partial

converse to the cofinality theorem. First, it is unreasonable to expect
a true converse; in fact, it is surprising at first that any kind of
converse is possible. If a = F|G choose b so that F < b < a. Such
b exists by theorem 2.1. By the cofinality theorem FU{b}|G = a.
However, F is not cofinal in F U { b } by choice of b.
Theorem 2.9 (the inverse cofinality theorem). Let a = F|G be the
canonical representation of a. Also let a = F'|G' be an arbitrary
representation. Then (F',G*) is cofinal in (F,G).
Proof. Suppose b e F. Then b < a < G 1 . Since a has minimal length
among elements satisfying F1 < x < G1 and b has smaller length than
a, F' < b is impossible, i.e. (HceF 1 )(c>b). This is precisely what we
need. A similar argument applies to G and G 1 .
The same proof works if the representation in theorem 2.4 is
used. At any rate, we now have what we need to build up the algebraic
structure on the surreal numbers. It is hard to believe at this stage,
but the relatively simple-minded system we have supports a rich algebraic
structure.

13
3

THE BASIC OPERATIONS

A ADDITION
We define addition by induction on the natural sum of the
lengths of the addends.

Recall that the natural sum is obtained by

expressing the ordinals in normal form in terms of sums of powers of u>

and then using ordinary polynomial addition, in contrast to ordinary
ordinal addition which has absorption.

Thus the natural sum is a

s t r i c t l y increasing function of each addend.

The following notation w i l l be convenient.
the canonical representation of
and a"

a, then

is a typical element of

G.

If

a = F|G

is a typical element of

Hence a 1 < a < a".

is
F

We are now

ready to give the d e f i n i t i o n .

a + b = {a'+b, a+b 1 }|{a"+b, a+b"}.

Definition,

Several remarks are appropriate here.

F i r s t , since the

induction is on the natural sum of the lengths, we are permitted to use

sums such as

a'+b

in the d e f i n i t i o n .

is needed for the beginning.

Secondly, no further d e f i n i t i o n

Since at the beginning we have only the

empty set, we can use the t r i t e remark that

of

f.

{f(x):xs(j)} = <>j regardless

For example, <|>|<>

j + <|)|<j> = <t>|<}>. Thirdly, there is the a p r i o r i

p o s s i b i l i t y that the sets

not satisfy

F < G.

F|G = u

G used in the d e f i n i t i o n of

if

F|G = u for some special symbol

u e FUG.

if

F < G

In the sequel when a d e f i n i t i o n of an

operation is given in the above form, we w i l l show that

less than

a+b do

To make the d e f i n i t i o n formally precise, one can

use the convention that

and that

F and

F is always

G so that the operation is really defined, i . e .

u is never

obtained as a value.
[1] is followed somewhat closely in building up the algebraic

AN INTRODUCTION TO THE THEORY OF THE SURREAL NUMBERS

operations.

14

However, some differences are inevitable because of the

d i f f e r e n t foundations.

We have a specific system with a specific order.

[ 1 ] deals with abstract elements and an order which is inductively

defined by a method which corresponds to our theorem 2.5.
Note that since we use a specific representation of elements
in the form

F|G, the operations are automatically well defined.

Nevertheless, in order to advance i t is necessary to have the fact that

the r e s u l t is independent of the representation.
Let us i l l u s t r a t e the d e f i n i t i o n with several simple
examples.
by

1.

Denote the empty sequence

Now

(+) = {0}|(|>.

the empty set and

sequence.

( )

by

0 and the sequence

( I t is easy to get confused.

Note that

(+)
G is

is the u n i t set whose only element is the empty

They are thus not the same.)

{0+<|>|<|>}|<|> = 0> 14> = 1.

Similarly

Then

0+1 = 1.

{0+1,1+0}|<|> = (l}|cj) = {(+)}|<j> = (++)

1+0 = {0} |<j> + <|>|<>

j =
Also

1+1 = {0}|<j> + {0}|<|> =

which is natural to call

"2".

I t does look rather cumbersome to work d i r e c t l y with the

d e f i n i t i o n , but so would ordinary arithmetic i f we were forced to use
{<!>}, I<1>,{$}}> instead of

1 , 2, etc.

and go back to inductive

definitions.
Theorem 3 . 1 .

a+b

is always defined ( i . e . never

b>c=>b>a+c

and

u) and furthermore

b>c=>b+a>c+a.

Remark 1. Although the first part is what is most urgent, we need the
second part to carry through the induction.
Remark 2. As a matter of style, one can prove commutativity first (which
is trivial) and then simplify the statement of the above theorem and its
proof.

However, it seems preferable to prove that

a+b exists as a

surreal number before proving any of its properties.

Proof.

We use induction on the natural sum of the lengths.

words, suppose theorem 3.1 is true for all pairs

numbers such that

(a) + (b)

is less than

In other

(a,b) of surreal

a. We show that the

statements remain valid if we include pairs whose natural sum is a.

Now
that

a + b = {a'+b, a+b 1 } |{a"+b, a+b11}.

F < G. Since

First, we must show

< a", it follows from the inductive hypothesis

THE BASIC OPERATIONS

a 1 + b < a"+b.

that

< a + b"

and

a + b

< a" + b

a + b' < a + b < a + b".

b

and

a + b1 < a + b".

Similarly
1

By d e f i n i t i o n

of

15

< a" + b.

b > d > c.

Hence

a+b

+ b < a + b < a" + b

is defined.

and

segment of the other.

Now suppose that neither

be the common i n i t i a l

Hence

This proves the required inequality when e i t h e r

is an i n i t i a l

the other and such that

a 1 + b < a 1 + b11

Also

U a ) + lib)

segment of

a + b > a

+ d>a

nor

is an i n i t i a l

_< a

and (a) + i{c)

and

c.

+ c

I t follows immediately that

Now assume

and b + a > d

a > b

and

segment of

<_ a.

Let

b > c.

Hence

+ a > c + a .

c > d =>

a + c > b + d.

Theorem 3 . 2 .
{g+b, a+k}

Suppose
where

d e f i n i t i o n of
of

Remark.

and

a = F|G

and

b = H|K; then

f e F, g e G, h e H, k e K.

a+b

I.e.

a + b = {f+b,a+h}|
although the

is given in terms of the canonical

representation

b, a l l representations give the same answer.

We shall c a l l this "the uniformity theorem for a d d i t i o n , " and

say that the uniformity property holds for a d d i t i o n .

Proof.

Let

a = A'|A\

a = F|G, b = H|K.

Suppose the canonical representations are

b = B'|B".
By the inverse c o f i n a l i t y theorem (theorem 2 . 9 ) , F

cofinal in

and s i m i l a r l y for the other sets involved.

{f+b, a+h}|{g+b, a+k}.

a + b

f + b < a + b.

is cofinal in

Also suppose

The betweenness

follows immediately from theorem 3 . 1 , e . g .

a'+b

in the canonical representation of

1

Consider

I t is now easy to check that the hypotheses of

the c o f i n a l i t y theorem (theorem 2.6) are s a t i s f i e d .

property of

is

( 3 f e F ) ( f _> a ) .

is one of the typical lower elements

a+b

as in the d e f i n i t i o n .

By theorem 3 . 1 ,

f + b^ a

similar argument applies to the other typical elements.

Since
1

+ b.

F
A

Hence the

c o f i n a l i t y condition is s a t i s f i e d so by theorem 2.6 we do get

a+b.

This technique w i l l be used often to get uniformity theorems

for other operations.
operations.

Such results f a c i l i t a t e our work with these

In p a r t i c u l a r ,

they permit us to use the methods of [ 1 ] in

dealing with composite operations, as we shall see, for example, in the

AN INTRODUCTION TO THE THEORY OF THE SURREAL NUMBERS

16

proof of the associative law for addition.

Theorem 3.3. The surreal numbers form an Abelian group with respect to
addition. The empty sequence is the identity, and the inverse is
obtained by reversing all signs. (Note that one should be aware of
potential set-theoretic problems since the system of surreal numbers is a
proper class.)
Proof: 1) commutative law. This is trivial because of the symmetric
nature of the definition.
2) associative law. We use induction on the natural sum of the lengths
of the addends
(a+b) + c = {(a+b)'+c, (a+b)+c'}|{(a+b)"+c, (a+b)+cM}.
By theorem 3.2 we may use
for

11

(a+b) .

a+b1

and b+a' instead of (a+b)'

and similarly

i . e . i t is convenient to use the representation in the

d e f i n i t i o n of addition rather than the canonical representation.

We thus

obtain
(a+b) + c = {(a'+b)+c, (a+b')+c, (a+b)+c'}|{(a"+b)+c, (a+b")+c, (a+b)+c"}.
A similar result is obtained for

a + (b+c).

Associativity follows by

induction.
3) The i d e n t i t y :

Denote the empty sequence by

again use induction:

terms

Q',0",

Then

0 = <|>|<j>. We

a + 0 = {a'+0, a+0 1 }|{a"+0, a+0"}.

There are no

so this simplifies to

by the inductive hypothesis.

4) The inverse:
Again l e t

and

a"

{a'+0}|{a"+0}

We thus get

We use induction.

reversing a l l signs, and l e t

F|G

0.

Let

which is

-a

be obtained from

be typical elements of

F and

b is an i n i t i a l segment of

an i n i t i a l segment of

-c

and

representation of

may be expressed as

a + (-a) = {a'+(-a),

a by

be the canonical representation of

Note that, in general, i f

-a

{a'}|{a"}

a + 0 = a.

b < c => -b > -c.

G respectively.
c

then

-b

is

Hence the canonical

-a"|-a'.

a+(-a"}|{a"+(-a), a+(-a')}.

a.

Therefore

Since

a1 < a < a", i t

is clear from the lexicographical order that -a" < -a < -a 1 . Using
induction and the fact that addition preserves order, we obtain
a1 + (-a) < a1 + (-a1) = 0. a + (-a") < a" + (-a11) = 0.
a" + (-a) > a" + (-a11) = 0, a + (-a1) > a1 + (-a1) = 0. Hence in the
representation of a + (-a), as H|K, H < 0 < K. Since 0 vacuously
satisfies any minimality property, a + (-a) = H|K = 0.

THE BASIC OPERATIONS

17

Thus we now know that the surreal numbers form an ordered

Abelian group.

The i d e n t i t y and inverses are obtained in a way which is

h e u r i s t i c a l l y natural.
B MULTIPLICATION
The d e f i n i t i o n of m u l t i p l i c a t i o n is more complicated than
that of a d d i t i o n ; in f a c t , i t took some time before the standard
d e f i n i t i o n for m u l t i p l i c a t i o n was discovered.
Definition,

ab = {a'b + ab1 - a V , a"b +ab" - a"b n }|

{a'b + ab" - a'b 11 , aHb + ab1 - a " b ' } .

As p a r t i a l motivation note that i f
1

< a, b < b, then

a,b,a',b'

are ordinary

real numbers such that

a'b+ab1 - a ' b 1 <ab.

Similar computations apply to get the appropriate

inequalities i f either

a1

i s replaced by

( b - b ' ) U - a ' ) > 0, i . e .

a"

or

b'

replaced by

b".

Thus the inequalities are consistent with what is desired.

Theorem 3.4.

ab

is always defined.

Furthermore

a > b and

c > d = ac-bc > ad-bd.

Proof.

We use induction on the natural sum of the lengths of the

factors.

We shall refer to the inequalities

P(a,b,c,d)

and to the expression

proper i n i t i a l segments of
that

a0

i s of the form a 1

and
or

ac - be > ad - bd as

ab + ab - abo
b
a".

where

respectively as

a,b

f(a,bo).

In the former case we call

are
Note
a0 a

lower element and in the l a t t e r case an upper element.

I t follows immediately from the d e f i n i t i o n that the relation
P is t r a n s i t i v e on the l a s t two variables.

Since, at this point, we

can freely use the properties of addition in ordinary algebra

P(a,b,c,d)
that

i s equivalent to

ac - ad > be - bd.

This makes i t clear

P is t r a n s i t i v e on the f i r s t two variables.

Now l e t

consider

b o and

f(ao,b2<>) - f t a o . b ^ ) .

be i n i t i a l segments of

b and

This is

(aob+ab2o-aob2<>) - (aob+ab^-aot^o) = (ab20-a<>b20) - ( a b ^ - a o ^ o ) .

a > a

and

b2

> b ,

the inductive hypothesis says that the above

expression is p o s i t i v e , i . e .
increasing function of

If

P(a,a,b ,b o ) , so
if

a<> < a.

If

f(a,bo)

a<> > a and

is an

b ^ > b 2 ,

AN INTRODUCTION TO THE THEORY OF THE SURREAL NUMBERS

18

the above expression may be written ( a ^ o - a b ^ ) - (ab20-ab20) which,

again, is positive. Hence if a > a, f(a<>,bo) is a decreasing function
of b. Similarly, f(a,bo) is an increasing function of a 0 if
b < b and decreasing if b > b.
To show that ab is defined, we must check inequalities
such as f(a|,b') < f(a 2 ,b"). This follows easily from the above. If
a' = a'
this is immediate since b1 < b" and a' < a. More generally,
1
let a be max (a|,a'); then f(a'.b') f(a',b') < f(a\b n ) f(a 2 ,b").
We now consider f(a",b") and f(a 2 ,b') and let a" = min (a",a 2 ).
Then f(a",b") f(a n ,b n ) < f(a n ,b') <_ f(a 2 ,b'). Since the remaining
cases are similar to the ones we already checked, this shows that ab
is defined.
To prove the second statement of the theorem, assume first
that in each of the pairs {a,b},{c,d} one element is an initial segment
of the other. By definition f(a',b') < ab, f(a u ,b n ) < ab, f(a\b") > ab
and f(a",b') > ab. Now ab - f(a',b') = ab - (a'b+ab'-aV)
= (ab-a'b) - (ab'-a'b 1 ). Thus the first inequality says P(a,a',b,b').
Similarly, the other inequalities give P(a",a,b",b), P(a,a',b",b),
P(a",a,b,b'). This proves the statement in this special case.
Next, remove the above restriction on {a,b}, but still
assume that one of c,d is an initial segment of the other. Let e be
the common initial segment of a and b. Then a > e > b. By the above
P(a,e,c,d) and P(e,b,c,d). Hence by transitivity we obtain
P(a,b,c,d).
Finally, suppose neither c nor d is an initial segment of
the other and let f be their common initial segment. By the above, we
have P(a,b,c,f) and P(a,b,f,d), so we finally obtain P(a,b,c,d) by
transitivity. This takes care of all cases.
Theorem 3.5 (The uniformity theorem for multiplication). The uniformity
property holds for multiplication.
Proof. This is similar to the proof of theorem 3.2 and, in fact, all
theorems of this type have a similar proof once we have basic
inequalities of a suitable kind.
Now that we have theorem 3.4 and, in particular, the fact
that the inequality stated there is valid in general, the same

THE BASIC OPERATIONS

19

computation as in the beginning of the proof of the theorem gives us the

behaviour of f(c,d) in general for c * a and d * b. (We no
longer require that c and d be initial segments of a and b
respectively.)
Suppose a = F|G, b = F'|G', c<> e F U G , d<> e F'UG 1 . Then
f(co,d) is an increasing function of d if c < a and a decreasing
function of d for c > a and similarly for fixed d<>.
We now check the hypotheses of the cofinality theorem. The
betweenness property of ab follows from the same computation as in the
latter part of the proof of theorem 3.4. For example, since
ab - f(c',d') = (ab-c'b) - (ad'-c'd 1 ), P(a,c\b,d') says that
ab > f(c',d'). The other parts of the betweenness property follow the
same way. Note that we are now going in the reverse direction to the one
we went earlier, i.e. we have P and we obtain the betweenness property.
To check cofinality let e.g. f(a',b') be a lower element
using the canonical representation of ab. By the inverse cofinality
theorem (3ceF)(c2.a') and (BdeF1 )(d>b'). Then f (c,d) >_ f(c,b') >_
f(a',b'). A similar argument applies to the other cases. Actually all
the cases may be elegantly unified by noting that f(c,x) maintains the
side of x if co < a and reverses it if c > a. Thus f(co,x)
maintains the side of x if and only if it is an increasing function of
x. Hence in all cases f(co,x) is closer to ab if x is closer to
b. Similarly for f(x,do). This is just what is needed to obtain
cofinality.
Theorem 3.6. The surreal numbers form an ordered commutative ring with
identity with respect to the above definitions of addition and multiplication. The multiplicative identity 1 is the sequence (+) = {0}|<j>.
Proof. Commutative law. Because of symmetry this is just as trivial
here as it was for addition.
Distributive law. We use induction on (a) + lib) + i{c).
(a+b)c = {(a+b)'c + (a+b)c' - (a+b)'c',...}|... . By theorem 3.5 we may
use a+b1 and a'+b instead of (a+b) 1 . For unification purposes we
consider a typical term (a+b)c + (a+b)c - (a+b)co i n the representation of a(b+c) where an element is lower if and only if an even

AN INTRODUCTION TO THE THEORY OF THE SURREAL NUMBERS

number of noughts correspond to double primes.

{a+b,a+b}

by the above.

(ao+b)c + (a+b)co - (ao+b)c<>

used instead of

and

(ac)o + be

we use

(a+b<>)c + (a+b)c<> - (a+bo)c<>.

By

ac + be + ac + be 0 - ac - be 0

A similar r e s u l t is obtained i f

a + b

is

+ b.

A typical term in the representation of

which is

by theorem 3 . 2 .

(a+b)o

Thus, typical terms become

induction the f i r s t term becomes

= aoc + ac - aco + be.

For

20

(aOc+acO-aOcO) + be.

ac+bc

is

Note that this is j u s t i f i e d

Since a similar r e s u l t applies i f we take

ac + (bc)o

and since the p a r i t y rule as to which element is upper or lower is the

same as before, we see that

(a+b)c = ac+bc.

We are now permitted to w r i t e

ab - (ab+abo-aObo)

as

(a-ao)(b-t>o).

Associative law.

We use induction on

term in the representation of

(ab)c

(a) + &(b) + l(c).

is

A typical

(ab)Oc + (ab)c - (ab)OcO

which by theorem 3.5 may be w r i t t e n

(ab+abo-aObo)c + (ab)co - (aob+abo-aobo)co.

Again, an element is lower

i f and only i f an even number of noughts correspond to double primes.

By

the d i s t r i b u t i v e law the above expression is

(aOb)c + (abo)c - (aObo)c + (ab)co - (a<>b)co - (abo) c o + (ao D o) c o.

Of

crucial importance is the following kind of symmetry in the expression:

the terms are a l l obtained from

(ab)c

by putting a superscript on a t

l e a s t one of the f a c t o r s , and the term has a minus i f and only i f

there

i s an even number of superscripts.

Exactly the same thing happens with the expansion of
except for the bracketing.

I.e.

a(bc)

the parity rules as to which term is

upper or lower, and which addends in a term have a minus is the same as
before.

In f a c t , we obtain for a typical term

a(bc) = a(bc) + a(b0c+bc-b0c<>) - aO(bOc+bcO-b<>c)

= ao(bc) + a(bo c ) + a(bco) - a(bo c o) -

ao(boc)

- ao(bco) + a o ( b o c o ) .

The r e s u l t now follows by induction.

The i d e n t i t y .

F i r s t note that

d i s t r i b u t i v e law.
Since

a*0 = 0.

This follows from the

I t also follows immediately from the d e f i n i t i o n .

0 = <j>|<j>, and a l l terms used in representing a product must contain

lower or upper representatives of each f a c t o r , a0 = <j>|<j> = 0.

t h i s was not the case for a d d i t i o n . )

(Note that

THE BASIC OPERATIONS

21

We again use induction to compute a*l. Since 1 = {0}|<j>

the expression for a-1 reduces to a-1 = {a'-l+a-O-a'-O} |{a"-l+a-0-a"-0}
which is {a'-l}|{a"-l}|. By induction this is {a'|a"} which is a.
Compatibility of ordering. Suppose a > 0 and b > 0. Then by theorem
3.4 we have P(a,0,b,0), i.e. a-b - 0-b > a-0 - 0-0. Hence ab > 0.
Thus we now know that the surreal numbers form an integral
domain. We saw that multiplication behaves somewhat more subtly than
addition. We shall see in the next section that division is handled much
more subtly. At any rate, it is remarkable that all this is possible.
It is possible to get a nice form for the representation of
an n-fold sum and product by inductive use of the uniformity theorems.
It is trivial that ax + a 2 + + a n may be represented as
{a 1 l +a 2 -.-+a n ,a 1 +a 2 l ---+a n ,--.a 1 +a 2 ---+a n l }|
{a 1 "+a 2 ...+a n ,a 1 +a 2 ll ---+a n ,a 1 +a 2 ---+a n 11 }.
We claim that a a a n may be represented by terms
a 1 a 2 ...a n - (a 1 -a 1 )(a 2 -a )---(a n -a n ) where an element is lower if
and only if an even number of noughts correspond to double primes.
The identity ab - (a o b+ab-ab 0 ) = (a-a)(b-b) may be
written in the succinct form ab - (ab) = (a-a)(b-b). Theorem 3.5
allows us to use this representation for ab if we multiply by other
factors. Thus it is clear by induction that
(a 1 a 2 ...a n ) - (a 1 a 2 ...a n ) = ( a ^ a ^ M a ^ 0 ) . . . (a n -a n ). The parity
rule is easily checked. In fact, it is essentially the same as the one
in ordinary algebra for multiplying pluses and minuses.
It is possible to use the above computation to prove the
associative law. Of course, one would have to be more cautious with the
bracketing before one has that law.
C

DIVISION
We w i l l define a reciprocal for a l l

a > 0.

As usual,

induction w i l l be used, but the d e f i n i t i o n is more involved than the

ones for the e a r l i e r operations.

Let

a = A'|A"

be the canonical

representation.

One naive attempt would be as follows.

where

= { 0 aT * } H Ta I }

since

a > 0.)

a" e G and

a* e F - { 0 } .

Unfortunately this does not work.

We try

(Note that
Although

0 e A1
has some

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

22

of the properties expected of a candidate for , xa * 1 in general. It

a
turns out that more elements are needed to get a representation for the
reciprocal. Roughly speaking, the idea is to insert as many elements
into the representation of x as is needed to force the crucial
inequalities, i.e. in the standard representation of ax as a product we
want the lower elements to be less than one and the upper elements to be
greater than one.
What is needed is more complicated. First we define
objects <a ,a f--,a > for every finite sequence where
a-j e A1 A" -{0}. For arbitrary b we define bai as the unique
solution of (a-a-j)b + a-|x = 1. This exists by the inductive hypothesis
which guarantees that aj as an initial segment of a has an inverse.
Uniqueness is automatic. Now let < > = 0 and
<a
a
a
> = <a
i ' a 2 ' * " ' a n > O a n + i For e x a m P l e ^ I *= a i = a i
1 2"" n+i
We now claim that a = F|G where F = (<a 1 ,-,a n >: the number of
a. e A1 is even) and G = (<a ,--fa >: the number of a. e A1 is
odd).
Theorem 3.7. The surreal numbers form a field.
Proof. We first show that x e F ==> ax < 1 and x e G = ax > 1. This
will show that F < G. Since < > e F, < > = 0, and a-0 = 0 < 1, the
result is valid for < >. We now use induction on the length of the
finite sequence. In other words it is enough to show that if b has
this property so does x = bai. Now by definition (a-a^b + a x x = 1.
Clearly (a-a^b + a ^ = ab. Since a } > 0 it follows that x > b iff
1 > ab. Also it follows from the above identity that
ax = 1 + (a-a-j)(x-b).
Now for fixed a l 9 the map b - ba 1 preserves being in F
or G iff a is upper. For example, b e F and a e A* + ba e G.
In that case ab < 1 by the inductive hypothesis, thus x > b. Since
a > &l it follows that ax = 1 + (a-a )(x-b) > 1. Hence x e G and
ax > 1 as desired. The other three cases can be unified by noting that
any change in b or ax from lower to upper leads to a change in ba 1
and hence reverses the desired inequality. At the same time, any change
in a x or b reverses the sign of a-ax and x-b respectively so it

THE BASIC OPERATIONS

23

also reverses the inequality we actually have.

desired, so that now we know that
meaning.

This proves what we

F < G and therefore

F|G has

Let F|G = c.
Finally we compute

ac. A typical element used in the

definition of the product has the form

ai e A ' U A " and c 1 e F[JG. First,

axc + acx- a c
1

0 e A

where

and 0 e F. Thus we get a

lower element in the representation of ac by choosing

Hence

0c + a0 - 0*0 = 0
Suppose

is a lower element.

a : = 0. Then the elements in the definition of the

product reduce to ac
c

a x = c = 0.

and ac

is an upper element for ac iff

e G.
However, we know that

c e F > ac
acx

< 1. Hence i f

is a lower element
Now suppose

in

FtjG,

ac

c x e G - a c x > 1 and
is an upper element

ax * 0.

c1oa1

Then

and satisfies the equation

a^ + ac^ a c

> 1 and i f

and

is defined, is contained

(a-a 1 )c + a ^ = 1. Now

is a lower element for

the same side of

ac

acx < 1.

ac

iff

respectively i f f

c a

and

are on

e G iff

c a

> c.

(This follows from the e a r l i e r statement regarding the map b - ba 1 *)

Since

l^l

s a t i s f i e s the equation
o

the inequality

c1 a1 > c

(a-a 1 )c 1 + a x x = 1 and ax > 0

is equivalent to

( a - a ^ c ^ axc < 1. The

left-hand side of this inequality is nothing but

Hence the lower elements for
elements are greater than
that

cla1 * c

ac

1.

are less than

(Note that since

so that the negation of

">"

axc + a c ^

a^.

1 and the upper

c a

e F(JG, i t follows

may be taken to be

"<"

in the proof.)
We have shown that in the expression for
the betweenness property but
1 = (+)

ac, 1

Hence

is the number of minimal length satisfying the betweenness

property.

Therefore

ac = 1.

Thus f i n a l l y we know that we have a f i e l d .

Although we don't

need the information i t is of passing interest to note how

as a function of
for

satisfies

being in the lower part does not.

ba

varies

a . F i r s t , by solving the defining equation

1-U-a^b
i-at,
ba, we get ba, =
= h + ^ * The first expression
i

implies that

b and

ba 1 is an increasing function of b iff a < a x and

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

24

the second expression that ba1 is an increasing function of a iff

ab > 1. Hence bax is an increasing function of one of the variables
iff the other variable is upper. At the same time the function preserves
sides iff the fixed variable is upper. All this can be unified by
saying that ba1 gets closer to c if b gets closer to c and if
a } gets closer to a. As we already saw in dealing with addition and
multiplication, this is essentially what is needed to prove a uniformity
theorem. Since the details are routine and since we don't need it, this
will not be pursued.
Finally it is recommended to any reader who is confused by
the unified arguments to think first in terms of individual cases, e.g.
in the above assume b and a are upper and regard ba as a
function of b for fixed a .

D. SQUARE ROOT
Frankly, i t is not of extreme importance to obtain the
existence of square roots at this time, since that can be obtained from
the theory of i n f i n i t e series which w i l l be developed l a t e r .

However, in

view of the elegance of the theory, i t is worth seeing how square roots
can be obtained d i r e c t l y without further machinery.

In the bottom of

page 22 in [1] credit for this is given to Clive Bach.

We assume

a > 0 and use induction.

a l l i n i t i a l segments of
square roots.
elements in

Let
A'UA"

a = A'|A"

be the canonical representation.

have square roots.

with product denoted by

Let

0, then

b e A'UA11

then

Then a l l

H be the free groupoid,

, generated by the elements of

shall define inductively a partial map from

If

I.e. we assume that

(they are necessarily non-negative) have

A'UA".

We

H into the surreal numbers.

f(b) = / b .

If b, c e H, f(b) and f(c) are defined and are not both

a+f(b)f(c)
f(bc) =
f(b)+f(c)

(In analogy with this case it was possible to use the concept of free
semigroup to deal with division, but we preferred to be more concrete.
Here we are stuck with this formalism since we are dealing with nonassociative juxtaposition.) By induction (Yx)[f(x)>0]. Furthermore
f(x) > 0 unless x = 0.
F and G are now defined inductively as follows.

THE BASIC OPERATIONS

b e A1

If
f(bc) e G

then

25

f(b) e F. If b e A" then

if f(b) and f(c) are both in F or both in G. If one of

f(b) and f(c) is in F and the other in G, then

we are not assuming that
that

F and G

Theorem 3.8.

f(bc) e F. Since

f is one-one, a priori, it may seem possible

have elements in common.

However, we shall prove that

F < G which in particular guarantees that

Proof.

f(b) e G.

F and G are disjoint.

Every positive element has a square root.

We will show that

x e F =* x 2 < a

F|G = /a". First we show that

and

x e G =* x > a. This is clearly true for x e A ' U A " .

In order to

carry through an induction it is necessary to study the behaviour of

xy = a * xy

as a function of x and y. First

(y2-a)(xrx2)
.
x ^ y - x 2 y = (x + y ) ( x + y ) Hence, for fixed
iff y 2 > a iff y e G

function of x

.
.
.
y, xy is an increasing

using induction.

Also

y eG

iff the map x - xy preserves presence in F or G according to the

above definition.

Thus we have a similar desired situation to one we

have previously, i.e. preserving sides is equivalent to being an increasing function if one variable is fixed.
x 1 o x 2 e G. First, if * 1 x 2 e F, let

Now assume
x = max(x 1 > x 2 ).

Then, by the above

x ,x e G we take
V OV

a+x
2X

x 1 o x 2 _> xx.

x = min(x ,x ) and obtain

x e G, then

x > a. In either case,

a 2 + 2ax 2 + x 4 > 4ax 2 .

(x1x2)

fortiori

>

is symmetric in
< a

and

Therefore

x e F,

x2 < a

then

(a-x 2 )

x * a. Hence
(^i)
^ 2x J

and i f

> 0.

a2+2ax

*+xl* > a.
4x2

xy e F.
x

and

Without loss of generality

y) suppose

x e F

and

(because

y e G.

Then

> a.

Assume f i r s t that

xy = a.

Then

xv =

xy

and

( x oy) 2 =

Now assume
xy

x x >_ xx. Now

By the inductive hypothesis, i f

Thus

Similarly, if

4xva
. Clearly x 4= y since
(x+y) 2
Therefore (xy) 2 < a as desired.

x+y
x+y
x 2 < a < y 2 ; hence

4xy

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

If xy * a, then either
apply the above to ^ and y
since membership in F or G
want is that ()2 < a. (Even
can get by even if all we have

x <

or x > .

26

If x < , we

to obtain f^-y)2 < a. The above applies

is not required in the argument. All we
the latter is not really needed since we
is (y)2 a.) Since y 2 > a, xy is

an increasing function of x and since

x < ~,

we have

U ' y ) 2 < ( ^ y ) 2 < a.

If x > ,

then y >

(x is necessarily different from 0 ) .

a

Then we apply the earlier argument to x and and obtain

Since x 2 < a, xy is a decreasing function of y. Hence
2
( x o y ) 2 < ( X o} < a.

a 2

(x-) < a.

This finally shows that x e F => x 2 < a and

x e G => x > a. Since x ^ . 0 this shows that F < G. Hence F|G has
meaning. Let F|G = c.
We now compute c 2 . Then a typical term in the representation of c 2 is c c + c c - c c . This is lower iff c and c are
on the same side of c iff c 1 o c 2 is an upper element, i.e.
2

a+CiCo

c < c1oc2 = c

+Q

iff ctCj+Cg) < a +c x c 2

iff c x c + c 2 c - c ^ < a.

The argument breaks down if c = c = 0 since

undefined. But this case leads to a lower element which is
less than a. So lower elements are less than a and upper
greater than a, i.e. a satisfies the betweenness property
Now 0,/T" e F. Hence one of the lower elements
representation of c 2 is c /IT + c(0) - (/aTMO) = c /a1" _>
Now /a71" e G. Hence one of the upper elements is
c /aT + c(0) - (/a^MO) = c /a71" <_ /F1" /F11" = a".
By the cofinality theorem c 2 = a.

c c is
0 which is
elements
for c 2 .
in the
/a"1" /a"1" = a 1 .

Note that as in the case of division, what we did was to

insert just enough terms into F and G in order to force the betweenness condition. Again, just as in the case of division, we have what is
needed to prove a uniformity theorem.

27

REAL NUMBERS AND ORDINALS

INTEGERS
The main task of this chapter is to show that the surreal

numbers contain both the real numbers and the o r d i n a l s .

(The d i s t i n c t i o n

as to whether the surreal numbers contain the real numbers or a set

isomorphic to the real numbers is very much l i k e the d i s t i n c t i o n as to
whether the I l i a d was w r i t t e n by Homer or by someone else of the same
name.)

Along the way we shall see the e x p l i c i t representation of o r d i -

nary numbers as sequences of pluses and minuses.

additive i d e n t i t y
identity

is the sequence

ordered f i e l d , the expression

the positive integer

So far we know that the

is the empty sequence and that the m u l t i p l i c a t i v e

n.

(+).

Since the surreal numbers form an

n times
(1+1+1 )

may be i d e n t i f i e d with the

We now have the following r e s u l t which is

consistent with one's h e u r i s t i c expectations.

Theorem 4 . 1 .

The positive integer

is

n times
(+++
).

Proof.

We use complete induction, i . e . suppose the theorem is true for

n t1#mes
n+1 times
all integers m _< n. Then 1+1+1
= (1+1+1 ) + 1
n times
= (+++ . . .
) + (+) by the inductive hypothesis. By applying
n times
cofinality to the canonical representation we know that (+++
)
may be expressed as
hypothesis.

(+)

(+++

is c l e a r l y

)|<j> = {n-l}|<j>
{0}|<|>.

by the inductive

Hence by d e f i n i t i o n

n+1 = { ( n - l ) + l , n+0}|<|> = {n}|(j> = (+ e S )|<i>

which is

(++++...imeS)

again by applying c o f i n a l i t y to the canonical representation of

n+1 times
)

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Note that the symbol

28

has been used in two different

senses, once for addition, and once as one of the symbols used in the
ordinal sequences which we consider.
algebra where, for example, in
meanings.

This happens also in ordinary

+(a+b)

the two pluses

have different

However, the meanings are related in such a way that i t is

convenient in practice to use the same symbol for each.

use of the symbols

+ and

In our case the

is consistent with the intuitive feeling

of order, i . e . plus is above zero is above minus.

In any case, the

meaning should be clear from the context.

Corollary.
Proof.

The negative integer

-n

is

n times

).

This is an immediate consequence of the theorem and the formula

for the additive inverse obtained previously.

B DYADIC FRACTIONS
Since the class of surreal numbers contain the rational
numbers, i t seems natural to consider them next and even to conjecture
that the rational numbers correspond to f i n i t e sequences of pluses and
minuses.

Since

0 = ( ) < (+-) < (+) = 1 , i t is natural to conjecture

that

(+-) = o- A heuristic guess for (+) would be a toss-up

1
1
1
1
between -j and j Actually (+-) =-^ and (+) = ^
I t turns out that the f i n i t e sequences correspond to the
dyadic fractions, i . e . rationals of the form

Although they form a

proper subset of the rationals, they are dense in the reals. Thus they
can be used just as well as the rationals as building blocks later in
developing the reals.
Lemma 4.2.

Proof.

Let

addition.
a < c < b.
property.

If

{a}|{b} = c.

Then

then {a}|{b}

Hence
Now

a+b

a + c < a + b < b + c

{2a}|{2b} = a+b.

2c = c+c = {a+c}|{b+c}

We show that the l a t t e r is

2a < a + c and
2c = a+b;

{2a}|{2b} = a+b

by definition of

by cofinality.

First,

which is the betweenness

Also i t follows from a < c < b

b + c < 2b. Thus we have the cofinality property.

therefore

{a}|{b} = c =

that
So

a+b

The above is the key lemma for dealing with dyadic fractions.

REAL NUMBERS AND ORDINALS

29

For example 1 = {0}|<j> = {O}|{2} by cofinality. Hence the hypothesis of

lemma 4.2 is satisfied if a = 0 and b = 1. Hence {0}|{l} = \ So
| = {( )}|{(+)} = (+-). This says that the hypothesis of the lemma
is valid for a = 0 and b = j Hence {0}|{-i} = j So
X = {( )| {(+-)} = { ( + ) } . This sets up an induction, but only numbers
of the form will be reached. However, we also have, for example,
3 = {0,1,2} |<|> = {2} | {4} by cofinality. Again, by the lemma we obtain
|= (
We now show that this process enables us to determine the
rational number which corresponds to an a r b i t r a r y surreal number of
f i n i t e length.

Theorem 4 . 2 .

Surreal numbers of f i n i t e length correspond to dyadic

fractions.

Specifically, let d be a surreal number of length

m+n which satisfies

d(m) * d ( 0 ) .
Define

b(i)

as follows.

b(i) = 1
b(i) = -1
b(D =

if

i < m and

if

i < m and
1f

2i-m+i

d(i) = + .
d(i) = - .

i i l m and

d(i)= + .

and

m+n-1
Then

d =

\
1=0

b(i).

Remark. The above says informally that a plus is counted as a 1 and a

minus as -1 until a change in sign occurs at which point the sequence
of pluses and minuses is treated like a binary decimal (with 1 and -1
rather than with 1 and 0.) For example,
+. is 3 _| +

j.|.2|.

Proof. Let d(0) = +. A similar argument applies if d(0) = -. (As an

alternative one can take the result for d(0) = + and take the negative
of both sides.)

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

The case

n = 0,

30

which is the case where there is no change

in sign, is e s s e n t i a l l y the statement of theorem 4 . 1 .

We do the case
special.

Here we have

followed by a minus.
with

n = 1

i n d i v i d u a l l y since this case is

d(m) = - .

The sequence consists of

and the length is the least ordinal for which

defined.

m pluses

To avoid confusion r e c a l l that the ordinals begin

d

is not

(This may seem unnatural in the f i n i t e case, but is required i f

one wants a general d e f i n i t i o n . )

In any case, the two "unnatural" con-

ventions cancel so that the length is r e a l l y the number of terms in the

sequence!
Of course

m> 1 [d(m)*d(0)L

Now

2m-l = { 0 , 1 , 2 , . . .2m-2} |<|>.

This is the canonical representation by theorem 4 . 1 .

obtain

2m-l = {2m-2}|{2m}.

obtain

{m}|{m-l} = m - j

We can now apply lemma 4.2 with

m pluses.

m-1

consists of

Any surreal number between

lexicographical order begin with

have

as an i n i t i a l

a = m-1

and

b = m to

I t is easy to see d i r e c t l y from the

d e f i n i t i o n t h a t {m}|{m-l} = d.
of

By c o f i n a l i t y we

segment.

(m-1)

m-1

and

pluses and m

m must by the

m pluses followed by a minus,

Hence

d = m- j ,

i.e.

which is exactly what

the theorem says.

We now use induction on
for a l l

n <_ r

and l e t

n.

Assume that the theorem is true

n = r+1.

We f i r s t note that an immediate application of induction to

lemma 4.2 shows that {2a}|{2b} = a+b - + f \ } | { ~ - }
integers

s.

Let
1

d -

and
d1

c+1.

Hence

{^-} | f | ^ }

be the i n i t i a l

segment of

since

has length

s i m i l a r argument applies i f

= - + d

of length

m+n = m+r+1.

d = d'-.)

done separately, we can assume that

m pluses

positive

We already noted that the hypothesis is v a l i d for consecu-

t i v e integers

d = d'+ or

= " ^ 7 for a l l

Since the case

r _> 1 , i . e .

m+r.

Assume

that

n = 1
d'

Then

d = d'+.

(A

has been

begins with

followed by a minus.
Now l e t

d = FIG

be the canonical representation.

are f i n i t e , so by c o f i n a l i t y
l a r g e s t element in
d = d'+, clearly

and

x = d1.

a
y

has the form

{x}|{y}

the smallest element in

We cannot be as e x p l i c i t with

where
G.

F
x

and G
is the

Since
y , since in

REAL NUMBERS AND ORDINALS

31

this general s i t u a t i o n we have very l i t t l e information about the minuses

in

G.

We know that there is a minus a f t e r

m pluses; hence

Also, we can apply the inductive hypothesis to

d

= x =

for some integer

c.

and

y.

I f we can show that

can apply the above formula to obtain

y m.

Hence

y =

d = { x } | {y} = +

then we
which is

exactly what we need to prove the theorem.

Our apparent lack of control over
principle.

Let

greater than

m+r

c a r d i n a l i t y of

w i l l be overcome by the box

be the set of a l l surreal numbers of length not

t h a t begin with

every element of

is
H

m pluses followed by a minus.

1 + 2 + 2 ...2r-!.
is of the form

By the inductive hypothesis

for some integer

lexicographical order is s t r i c t l y between

precisely
form

2^-1

e H

unless

= m.

- > = d1.

Since

that

Now

r
2

Since

fact,

r = d
2

$y-

m-1

and

GCHUW.
1

Since there are

< d

and

H.

i[

In p a r t i c u l a r ,

1 < m+r.

Now

Hence every element of

d < G,

r
2

segment of d

and

which is a contradiction.

has the form

is a lower bound to

segment of

d.

G.

In

Otherwise, by consider-

we obtain an element of

Since

e G and hence the l e a s t element of

earlier,

m.

and by the

i t follows from the lexicographical order

is a c t u a l l y an i n i t i a l

-
2T

and

m is in

In e i t h e r case

d = d' +

ing the common i n i t i a l

below

m-1

such numbers, by the box principle every number of the

s t r i c t l y between

- > d.

The

G,

-
> d, i t follows that
2r
i.e.

y = -~-

As we said

this is what we need to complete the proof.

During the proof we showed that a l l dyadic fractions are

obtained this way.

Also i t is easy to see how to express a dyadic

fraction constructively as a sequence.

H e u r i s t i c a l l y speaking, we always

go in the r i g h t d i r e c t i o n to close in on the f r a c t i o n .

consider
Since
+.

2g-.

2g- < 2y

Since

2 < 2g < 3,

we want a minus next.

F i n a l l y , +++--+

with

Now

+++--

+++-.
is

h i t s what we desire on the nose.

can set up an elementary induction.

form

we begin with

For example,
This is
2j

2j

so we need a

More formally, one

We assume that a l l fractions of the

odd correspond to sequences of length

m+n.

Consider

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

,7 Precisely one of b and
2n+i

b+1

32

is odd.

Thus depending on which is odd we take

+ r
2n
2n+1
2n
2n+1
the corresponding sequence of length m+n and juxtapose a plus or minus,
e.g. if b is odd we juxtapose a plus to the sequence of b. Note the
lack of choice. Since b+1 is even in that case,
is a sequence of
length less than m+n. Tacking on a minus will thus not give

This is, of course, what we expect from the beginning since the numbers
are sequences and not equivalence classes of sequences. Anyway,
whatever appearance there may be of choice, it is clearly deceptive.
The whole idea of a shift from ordinary counting to a binary
decimal computation at the first change in sign may seem unnatural at
first. However, such phenomena seem inevitable in a sufficiently rich
system.
It is an amusing exercise in arithmetic to add numbers in
this form. Carrying exists as usual but since we deal with pluses and
minuses and do not have zeros, an adjustment is necessary if we would
otherwise obtain 0 in a place. Specifically, 0+ must be replaced by
+-. Also, one must be aware of the dividing line where the shift from
ordinary counting to the binary decimal computation occurs.
Finally, a suitable succinct way of expressing the result of
arithmetical operations on dyadic fractions given in the above form may
be useful in studying certain problems in the theory of surreal numbers.
Although the dyadic fractions look like a drop in the ocean of surreal
numbers, they form an important building block. As we shall see, for
example in chapter seven, it is possible to ask questions which are
non-trivial even for a set such as the dyadic fractions.
C

REAL NUMBERS

At this stage we develop a theory which is somewhat analogous

to that of Dedekind cuts. However, there is at least one important
difference. All the objects and operations are already present as a
subsystem of the surreal numbers. The analogy arises because of the
theory in chapter two. First, roughly speaking, the fundamental
existence theorem gives a well-defined element for every cut. Secondly,

REAL NUMBERS AND ORDINALS

33

the canonical representation gives a natural cut associated with any

element.

This correspondence, of course, occurs at every stage,

although the resemblance to Dedekind cuts is closer in some stages than

others.
Definition.

A real number i s a surreal number

f i n i t e length or is of length

which is either of

GO and s a t i s f i e s

(VnoJGniMHnzHCni^noDACna^noMCaCni) = + ] A [ a ( n 2 ) = - ] } .
In other words, the d e f i n i t i o n simply requires that the terms
of the sequence

a(n)

of pluses and minuses do not eventually have

constant signs.

This is analogous to the s i t u a t i o n for ordinary decimals

where one might rule out an eventual sequence of nines to ensure that
each number has a unique representation. In our case a sequence consisting eventually of pluses w i l l be a surreal number outside the set of
reals.
To show that the set of real numbers forms a f i e l d ,
suffices to check the closure properties.

it

However, i t is convenient to

have several lemmas in order to carry t h i s through.

Note f i r s t that the d i s t i n c t i o n between surreal numbers of
length

w which are real and those which are not can be expressed in

terms of the canonical representation.

representation, then
empty, F

a=F|G

i s real precisely when

has no maximum and

plus, then

If

is the canonical

F and

G are non-

G has no minimum; e.g. i f there is a l a s t

has a maximal element.

( I t is clear from what we already

know that t h i s element is i n f i n i t e s i m a l l y close to

but this issue

w i l l not be pursued now.)

The elements of

FUG

have f i n i t e length so they are

dyadic f r a c t i o n s .
Lemma 4.3.
that

Let

F and

F < G, F

G be non-empty sets of dyadic fractions such

has no maximum, and

G has no minimum.

Then

F|G

is a

real number.
Proof.

F|G

exists by the fundamental existence theorem.

By theorem 2.3
(a) = u) and
sign is

+.

&(a) _< w.
a

Let F|G = a.

I t suffices to rule out the p o s s i b i l i t y that

has eventually constant sign. Suppose that the constant

(A similar argument applies i f i t is

-.)

The p o s s i b i l i t y

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

34

that a consists exclusively of pluses is ruled out since a < G and G

is non-empty.
Now suppose a(n 0 ) = - but n > n 0 + a(n) = +. Let b be
the initial segment of a of length n 0 . Then b > a. By theorem 2.9,
(3ceG)(c<b) since b is an upper element in the canonical representation of b. Since G has no minimum (3dsG)(d<c). Since b,c, and
d are all dyadic fractions we can choose m so that d <_ b -
Let c be an initial segment of a of length n for some
n > n 0 . Then c is a lower element. Also, c-b has the form
+
=
for some r and s. (Actually
r ... +
2r
2r+1
2r+s
2r+s
s = n-n o -l.) In any case, for n sufficiently high, c > b r
Hence

a > c > b

Also

2m

a < d < b -

2m

2m
Thus we have t h e

desired contradiction.
Lemma 4.4. Let a = F|G. Suppose that (VxeF)(3a positive dyadic
fraction r)(3y)(y>x+rAyeF) and (VxeG)(3a positive dyadic fraction r)
(3y)(y_<x-rAyeG).
Also let F1 < a < G'. Suppose that (V positive dyadic r)
Then

a = F'|G'.

Proof. It is enough to check cofinality. We do this for

for G1 is similar. Suppose x e F. Choose y and r
is positive dyadic, y j> x+r and y e F. For that same
w e F1 and z e G1 such that z - w _< r. Since F < a
a > y 2. x+r. Since G1 > a we have w > z-r > a-r _> x.
this proves cofinality.
Note that the hypothesis does not require any
consist only of dyadic fractions.

F 1 . The case
such that r
r choose
we have
Since w e F1
of the sets to

Lemma 4.5. There are an infinite number of dyadic fractions between any
two distinct real numbers a and b.
Proof. It clearly suffices to obtain one dyadic fraction. If neither a
nor b is dyadic then the common segment works. More generally, if

REAL NUMBERS AND ORDINALS

35

neither a nor b is an initial segment of the other, we can use the

common segment. Suppose a is a proper initial segment of b. (a is
necessarily dyadic.) If b is dyadic the result is trivial. Otherwise
consider the canonical representation F|G of b. a e F O G . Since F
has no maximum and G no minimum we obtain a dyadic fraction between a
and b whether a e F or a e G.
Remark. Note that if a has a last plus and b is the initial segment
of a which stops just before the last plus, then there is no dyadic
fraction between a and b. Thus the requirement in the lemma that the
numbers be real is essential.
Lemma 4.6. Let a = F|G be the canonical representation of a real
number a which is not a dyadic fraction. Then for all positive dyadic
r there exist b e F, c e G such that c-b _< r.
Proof. Since there is no last + and no last - in a, then for all
n there are elements b e F, c e G which agree in the first n terms.
Thus c-b is bounded above by an expression of the form
+ -^- + +1 _< --= Since s can be made arbitrarily large by
a suitable choice of n this proves the lemma.
Note that it is easy to see that the requirement that a be
real can be relaxed but this is of no special concern.
We are now ready to check the closure properties.
Addition. Let a and b be real numbers. If both are dyadic fractions
then so is the sum. Suppose a is a dyadic fraction and b is not.
Let a = F|G and b = F'|G' be the canonical representations. Then
a+b = {a'+b, a+b'}|{a"+b, a+b"}. We claim that numbers of the form a+b1
are cofinal on the left. Consider a number of the form a'+b. Since
a,a1 are dyadic fractions and a > a 1 , a-a1 is a positive dyadic
fraction. By lemma 4.6, (3b'eF1 )(3b"eG' )(b"-b'<a-a'). Then
a+b1 2l a'+b" > a'+b. Similarly, we can see that numbers of the form
a+b" are cofinal on the right. By the cofinality theorem
a+b = {a+b'}|{a+b n }. Note that a+b' and a+b" are dyadic fractions.
Also since F' = {b 1 } has no maximum, neither does {a+b 1 }. Similarly

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

36

{a+b11}

has no minimum. By lemma 4.3, a+b is a real number.

Now suppose neither a nor b is a dyadic fraction. Again
let a+b = {a'+b, a+b1}|{a"+b, a+b"}. As before it is clear that the
left elements have no maximum and the right elements no minimum. More
specifically, since the numbers a', a", b 1 , b11 are all dyadic the above
representation of a+b satisfies the first condition of lemma 4.4. Now
let F1 = {a'+b1} and G1 = {a"+b"}. Then F1 < a+b < G 1 . Also lemma
4.6 together with the usual " j argument" shows that F1 and G1
satisfy the other conditions of lemma 4.4. Hence a = F'|G'. Finally,
F and G1 satisfy the hypothesis of lemma 4.3 so a is real.
0 and 1 are clearly real. The additive inverse of a real
is real since the definition is symmetric in pluses and minuses.
Multiplication. Let a and b be real numbers. If both are dyadic
fractions then so is the product. Now let a be a real number which is
not a dyadic fraction. There is no restriction on b. A typical element
in the representation of ab is c = ab - (a-a)(b-bo). [Recall that,
e.g., a 0 is either of the form a1 or a".] Since a is real we can
choose a which is on the same side of a as a and also closer to
a (we are using a unified argument). If we use the same b we obtain
the element cx = ab - ( a - a ^ M b - b O ) . Then c ^ c = (a^-aOMb-bo). Now by
applying lemma 4.5 to an arbitrary positive real r and 0 we obtain a
positive dyadic fraction d satisfying d < r. For a product of
positive reals r and r we obtain in this manner the positive
dyadic fraction d ^ 2 satisfying d d < r r . Thus we have what we
need to show that the first condition of lemma 4.4 is satisfied for the
representation of ab. The choice of F1 and G1 would depend on the
signs of a and b; however it suffices to assume that a,b > 0.
Let F1 = { a b : an is dyadic A 0 < a < a b is dyadic A
ii

0 bx< b}

and

~~

G = { a ^ : ax is dyadic Aa < a 1 Ab 1 is dyadic Ab < b^.

(This formulation is convenient since we can obtain the required conclusion i f

b is a dyadic fraction even though lemma 4.6 does not apply

in that case.)
such that

For arbitrary dyadic

b 2 -b x _< r

and a 2 ~ a i r #

r
^

we can choose
e

usua

a ,a ,b ,b

^ argument for proving

that the limit of a product is the product of the limits shows that the
other conditions of lemma 4.4 are satisfied.

Note that we use the fact

REAL NUMBERS AND ORDINALS

that for every real number

|a| _< n.
F|G

37

there exists an integer

such that

This is clear since in the canonical representation of

both

F and

G are non-empty.

number of length

as

( I n c i d e n t a l l y , for the surreal

u> which consists solely of pluses the argument would

break down.)
Hence by lemma 4.4
F1

addition,

G1

and

ab = F ' | G ' .

Again, as in the case of

s a t i s f y the hypothesis of lemma 4 . 3 , so that

ab

is r e a l .
Note that during the proofs of closure we obtained nice
representations of sums and products of reals with the help of lemma 4 . 4 .
Since such i n t u i t i v e - l o o k i n g representations f a i l in general, i t is
i n t e r e s t i n g that they work in the special case of real numbers.

Reciprocals.

I t is best to ignore the previous construction of

reciprocals.

I t suffices to consider the case where

larger than
G = (d: d

0.

Let

Hence

0 e F.

Clearly F < G, F

Also by lemma 4.5 we can find

fraction satisfying

da < 1 .

the closure p r o p e r t i e s , and

1 - da >

has no maximum.

Then

1 - da

1 - da > 0.

a <

Suppose

Now we already noted e a r l i e r that

| i< 1 " d a -

Thus

is a dyadic

is a real number because of

above by an i n t e g e r which we may j u s t as w e l l c a l l

iSPn

m so that

By lemma 4.5 one can f i n d m

2m

Hence

and
and G

2ma > 1 .

We now show that

so that

is a real number

is a dyadic f r a c t i o n A da < 1}

i s a dyadic f r a c t i o n A d a > 1 ) } .

are non-empty and

> 1_.
2m

F = {d: d

( d +

n
2 ,

iSTn)a<1-

is bounded

n
i.e. a < 2 .

Similarly

has no minimum.
By lemma 4.3 b = F|G is a real number. By closure of
multiplication ba is a real number. Now let n be arbitrary. We
would like to show that there exists an element c in G such that
ca < 1 + Choose m such that a < 2m.

- zx\

We now consider the set P = { r : ( 1 a > 1}. As a non1

^22 m + n element s. Then by
empty set of positive integers, P has a least
definition - 5 e G. Furthermore (-^r-)
V m + n a < 1. Hence
2m+n
2
( .!+ n ) a _< 1 +
c'a > 1

Hence

2n

S i m i l a r l y there exists an element

1 - < c'a < ba < ca < 1 +

2 n *~

~"

~~

2n

c1 e F

such that

This shows that

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

38

|ba1| < r for any positive dyadic fraction r. It follows from lemma
4.5 that ba = 1. This completes the proof.
Note that reciprocals of dyadic fractions are not necessarily
dyadic fractions. Thus a naive attempt consisting of expressing a as
F|G and using something like -j?\j to obtain the reciprocal would cause
dfficulties. Our proof circumvents that problem.
Now we know that the real numbers form a field containing
the dyadic fractions. They therefore contain all the rational numbers.
Finally, we would like to prove the l.u.b. property, i.e.
that every bounded non-empty set of reals has an l.u.b. within the set
of reals. It's important to note that in our development the real
numbers form a proper subset of the system we are studying so that care
is needed in stating the l.u.b. property. In fact, as a special case of
theorem 2.1 we see that any set which has no maximum has no_ l.u.b. in the
class of surreal numbers.
Incidentally, it is immediate from lemma 4.5 that every real
number is the l.u.b. of all dyadic fractions below it but this is not
what we really need. Since the "ordinary" real numbers can be
characterized as an ordered field with the l.u.b. property, the latter
is all that remains to be proved.
Suppose H is a non-empty set of real numbers bounded above.
Let G be the set of upper bounds which are dyadic fractions and let F
be the set of all other dyadic fractions. F and G are clearly nonempty. By lemma 4.5, F has no maximum. If G has a minimum b, then
b is already an l.u.b. to H and we are done. (Again, lemma 4.5 is
used since it guarantees that a least upper bound among the set of dyadic
fractions is automatically a least upper bound in the set of all real
numbers. So suppose G has no minimum. Then by lemma 4.3 r = F|G is
a real number. We now show that l.u.b. H = r. First, r is an upper
bound to H. Otherwise, (3aeH)(r<a). Let d be a dyadic fraction
satisfying r < d < a. (We are getting our money's worth out of lemma
4.5!) Since d < a e H, d e F. However r < d, contradicting the fact
that F < r. Finally, suppose s < r and s is an upper bound to H.
Let s < d < r for some dyadic d. Then d is also an upper bound to
H. Hence d e G. But d < r. This contradicts the fact that r < G.
We have finally shown the following.

REAL NUMBERS AND ORDINALS

39

Theorem 4.3. The real numbers form an ordered field with the l.u.b.
property (i.e they are essentially the same as the "reals" defined in
more traditional ways.)
We close this section with several remarks. Let a be a
real number which is not a dyadic fraction, thus a has length u>. If
a = F|G is the canonical representation, we know that F < a < G and
that the elements of F form an increasing sequence and that the
elements of G form a decreasing sequence. Since F has no maximum and
G no minimum, both sequences are infinite. It is clear from this and
from lemma 4.6 that if a n is the initial segment of A of length n
then lim a n = a. Of course, in the definition of a limit it makes no
n+
difference whether e is taken to be real, rational, or dyadic; however,
we certainly cannot use general surreal numbers.
Finally, we compare our definition with the one used in [1]
which makes no reference to sequences. It is convenient for motivation
to consider a third definition which is, roughly speaking, intermediate
in spirit between the two definitions.
Theorem 4.4. The following three conditions are equivalent.
(a) a is a real number (by our definition);
(b) For some integer n, -n < x < n and a has no initial segment
a a such that |a-aa| is infinitesimal;
(c) For some integer n, -n < x < n; and
a = {a-1, a - j, a - j - ' H ( a + l , a + j9 a + |} (the definition in [1]).
Remarks. An element a is infinitesimal if it is nonzero but for every
positive rational r, |a| < r. It is clear, for example, from what we
already know that the element a of length u which begins with + and
then after that consists only of -'s is an infinitesimal. The heuristic idea in definition (c) is that of the possibility of writing a in
the form F|G without forcing either F or G to be "too close" to a.
Of course, by the cofinality theorem, if a = F|G then one still gets a
if F and G are both enlarged to include elements closer to a. The
challenge lies in the opposite direction. How far away can F and G
be from a and still have a = F|G? As a rough rule of thumb, the
larger the length of a, the closer F and G must be to a.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

40

Proof,
this is
Not (a)
even if

(a) => (b). Since all initial segments are dyadic fractions,
clear.
= not (b). (This contra-positive notation seems natural here
it may be unusual.)
Suppose a is not real. Then (a) _> u>. If for all n
less than u>, a has a fixed sign then the condition -n < a < n fails,
as is clear from the ordering, so that case is clear. Now let a w be
the initial segment of a of length a>. Assume first that a^ consists
eventually of pluses only. (A similar argument will apply if o> consists
eventually of minuses only. ) Then we can use an argument which is essentially the same as the one used in the proof of lemma 4.3. We let a n
be the initial segment of a^ obtained by stopping just before the last
minus. (We already ruled out the case where a w contains no minuses.)
Then a n > a. For all positive rational r we can find m sufficiently
high such that the initial segment of a m of a of length m satisfies
a m < a and a n - a m < r. As in the proof referred to above, we use a
computation of the form - + -^- + r+j + ... Hence a n - a < r
for all r. Since n is fixed, we have |an-a| is infinitesimal.
Note that this part of the argument is independent of whether a is the
same as a^ or not.
Now assume that a^ does not eventually have constant sign.
Since a is not real, a * a w ., i.e. a w is a proper initial segment of
a. Let r be an arbitrary positive rational. Suppose a w = F|G. Then
by lemma 4.6 there exist b e F, c e G such that c-b _< r. Now
b < a^ < c. It is also clear that b < a < c by the lexicographical
ordering. (In fact, b and c occur in the canonical representation of
a as well as a^.) Hence la^-a) < r. Therefore la^-al is
infinitesimal.
(b) =* (c). Express a canonically as F|G. Then the conclusion is
immediate by the cofinality theorem. Given a1 e F, then a - a'
is not infinitesimal. Choose n so that a - a1 > - Then a -- > a 1 .
A similar argument applies to a1 e G.
(c) = (b). Let a = {a-1, a - |, a - |,... }|{a+l, a + |, a +j,...} and
let a = F|G canonically. Then by the inverse cofinality theorem the set
{a-1, a - j9 a --=-...} is cofinal in F. Let a1 e F. Then for some
n, a " 2 . a ' 1"-e- a"a> 2.* Therefore
A similar argument applies to G.

a -a1

is not infinitesimal.

REAL NUMBERS AND ORDINALS

41

In line with what was said earlier, note as a point of

caution that the theorem does not rule out the possibility that a is
real of the form F|G with elements in F or G infinitesimally close
to a. This is ruled out only for the canonical representation.
D

ORDINALS

Our next task is to show that the surreal numbers contain

the ordinals. Once this is done it will be legitimate to deal with
expressions of the form u>-l or h> First, of course, we have to make
precise what we mean by the statement that the ordinals are a subclass
of the surreal numbers. Recall that ordinary addition and multiplication
on the ordinals are not commutative. However, there do exist natural
commutative operations on the ordinals that have been considered in the
literature and they do correspond to the operations we defined on the
surreal numbers.
We identify the ordinal a with the sequence a a of length
a such that (Vn<a)[a(n) = +3.
First note that by theorem 4.1 this is consistent with the
situation for positive integers. Also it is immediate from the
lexicographical order that a < 3 => a a < a$. Furthermore, the canonical
representation of a a is {a$: 3<a}|<j>. Again, if H is a set of
ordinals then { a ^ } ^ = a a where a is the least ordinal such that
a > H. If H has a maximum 3 then a = 3+1. If H has no maximum
then a = l.u.b. H. a is called the sequent of H and denoted by
seq H in the literature. In summary, as far as the order properties are
concerned, the identification is quite reasonable. Thus for convenience
of notation we use a instead of a a . So the ordinals are the sequences
which consist only of pluses. (Incidentally, note that our definition of
cofinality [see p. 9] is consistent with the usual one for ordinals.)
We now show that addition and multiplication correspond to
what is commonly called natural addition and natural multiplication.
They are obtained by taking the usual expansion in powers of w and
operating as if they are ordinary polynomials (i.e. no absorption). In
order to state the next theorem precisely we tentatively use + for
ordinary addition, for natural addition and + for surreal
addition.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Theorem 4.5.

For any ordinals

42

a and 3, a+3 = c@3.

Proof. We use induction as usual. In view of our earlier remarks

a+3 = seq (a+Y, 6+3) = seq (ay, 603) by the inductive hypothesis. The
Y<3,<5<a
y<3,6<a
problem is now reduced to an elementary exercise in the ordinary theory
of the ordinals. Specifically, we may express a and 3 respectively
in the forms a'+w and 3'+w (r or s may be 0 ) . Then typical
lower elements are (a'+/) ( 3 ' + Y ) with Y < w S and (a'+6) (3'+u)S)
with 6 < u>r. Without loss of generality suppose r >_ s. Then the set
(a'+tj ) + ( 3 * + Y ) with Y < w is cofinal and we clearly obtain
r
s
(ct'+w ) (3'+w ) as the sequent which is what we want.
In order to state the next theorem precisely we tentatively
use for natural multiplication and x for surreal multiplication.
Theorem 4.6.

For any ordinals

Proof. Again we use induction.

I a) in. with
1 = 1

a and 3 ax3 = o3.

Now every ordinal has the form

r. > r. . for all

1

i and all

n-j integers. This may

k
r.
also be w r i t t e n in the form
I a> i with r. > r. . by breaking up a l l
1
r.
Ta) Tn. for which n. > 1 . The sum is also a natural sum and therefore
r
s.
by theorem 4.5 a surreal sum. Suppose a = GO 1 and 3 = o> i . By
r.. s.
the d i s t r i b u t i v e law for surreal numbers we obtain 01x3 = J w TXID J ,

I f at least one of

a and

3 is not a power

of

i,j
a>, then we can use the

inductive hypothesis to obtain

r. s.
ax3 = I a) l0o) J = a3
I.j
by the distributive law for natural multiplication over natural addition.
Now suppose that both a and b are powers of u>. Let
r
s
a = w and b = oj . Then by definition
ax3 = {((/x6) + (Y><W S ) - (6xY)}|(j) where 6 < OJS and y < </. Note that
unlike in the case of addition, we are stuck with the surreal operation
- which does not correspond to an operation on ordinals.) First,
since there are no upper elements, by theorem 2.2 ctx3 is an ordinal.

REAL NUMBERS AND ORDINALS

43

r*
(u> x6)
r
which by the inductive hypothesis and theorem 4.5 is (w Q6) +
s
r
v^s
Since 6 < w
and y < w . This is clearly less than u>
.
rS&
a)
satisfies the betweenness property in the definition of

Hence
o3.

Hence

The typical lower element is certainly not larger than

Ji(axg) a / ^

and since
1

Now if

< r

and

ax3 > i/ nxoaS = i/ rS

have

this is

</ ^ n .

ax$ is an ordinal,
n

ax$ w

is an arbitrary positive integer we

by the inductive hypothesis.

Similarly if

*
s
+ (yxw )
s
(>Qw ).

< s

we have

ax$ > J^

Furthermore
n.

We have

already noted during our proof of theorem 4.5 that seq(r'@s, rs') = rs.
By the basic properties of the expansion of ordinals in powers of
seq((/ S n, a)r+S|n) = < / + s .

follows that
ax3 = J5

Hence

axg :> (/+s.

o>, it

So finally

= o@e.
In view of theorems 4.5 and 4.6 we no longer need symbols

such as
and

and

x.

For convenience we will even drop the symbols

since it will be clear from the context whether natural or

ordinary operations are intended.

In fact, as a rule of thumb, in dis-

cussing elements we use natural operations which, as we have just shown,

are the same as the surreal operations.

On the other hand, when discus-

sing lengths and juxtaposition of sequences, the ordinary operations are

appropriate.

This issue involving choice of notation will occur in other

places as well.

In general, readability and reliance on context will

take precedence over a picayune attitude which complicates notation

unnecessarily.
We now know the exciting fact that the surreals form a field
containing both the reals and ordinals.
as

w-1

and

ju

have meaning.

So, for example, elements such

Incidentally, our

do with a meaning used in the literature, namely

l+o) = a) where

stands for ordinal addition.

sign sequence of several such strange elements.

cases of the general theory in the next chapter.

w-1

has nothing to

w-1 = u>, since

We shall compute the
These are all special
However, it is worth-

while to see some elementary concrete examples before getting involved

with a general representation theorem.

In fact, there is some

pedagogical value for the reader to experiment with other elementary

examples before coping with a more general and abstract situation.
We begin with
"exotic" element.

This is

w-1

which is about the simplest looking

a)+(-l).

The canonical representation of a>

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

is

{n}|<j>

where

- 1 = 4>| -CO>.

44

stands for an a r b i t r a r y non-negative integer and

Hence using the d e f i n i t i o n of addition

w-1 = {n-1} |{u)+0}.

By c o f i n a l i t y theorem b (theorem 2 . 7 ) , the left-hand side may be replaced

by

{n};

hence we have

{n}|{oj}

which we can see immediately from the

d e f i n i t i o n is the sequence of length

w+1

which consists of

pluses

followed by a minus.
We make several remarks before continuing with other
examples.

I t is often convenient to use theorem 2.7 to simplify expres-

sions of the form

F|G.

Even though there is an option of using theorem

2 . 6 , theorem 2.7 has the convenience of avoiding specific reference to

a = F|G

but r e f e r r i n g only to

and

themselves.

This helps to

streamline a computation; in f a c t , in future we shall use theorem 2.7

f r e e l y without quoting i t i f i t s use is obvious ( j u s t as in elementary
algebra one does not bother to quote the d i s t r i b u t i v e law every time i t
i t is usedi).

For example, an expression such as

where

m and
)

{n + y}|{oj

-}

run through a l l positive integers can be replaced by

Note that the l a s t part of the computation can be done in two

ways.
ing

One can d i r e c t l y use the d e f i n i t i o n and see that any

n < x < w

length

w+1

for a l l

satisfy-

must necessarily begin with the sequence of

r e f e r r e d to above.

answer and note that

tion.

{n}|{oj}

One can also take a good guess at the

happens to be the canonical representa-

This is worth emphasizing, because in more complicated situations

we often have a representation which is cofinal in a canonical representation.

In these cases i t is much easier to use the second method.

Recall that by the uniformity theorems, the computations may

be performed using any representations; thus we may choose ones which

appear to be most t r a c t a b l e .
use the representation
f r a c t i o n s below

and

For example, for any real number

a = F|G
G

where

is the set of a l l

If

we may

is the set of a l l dyadic fractions above

This is a u n i f i e d formula which applies whether or not

fraction.

dyadic

is not a dyadic f r a c t i o n , this is mutually cofinal with

the standard representation, but i t d e f i n i t e l y is not mutually cofinal

a

a.

is a dyadic

i s a dyadic f r a c t i o n .

I.e.

if

in the l a t t e r case the representation

cannot be j u s t i f i e d by theorem 2 . 7 , although theorem 2.6 c l e a r l y applies.

REAL NUMBERS AND ORDINALS

45

As obvious as these above remarks may be, they are worthwhile to state and get out of the way once and for a l l , since i t would be
clumsy to make them in the middle of a proof.

With this in mind, we can

now present computations and proofs more e f f i c i e n t l y ,

i.e.

using con-

venient representations and s i m p l i f i c a t i o n s without e x p l i c i t l y

stating

the obvious j u s t i f i c a t i o n .
Now consider

w-2.

This is

u)+(-2) = {n}|<j)+cf)|{-l} = { n - 2 } | { o ) - l } = {n}|{ur-l}

length

w+2

integer

n.

consisting of

Using induction we can compute

u)-(m+l)

for a positive

In f a c t , o)-(m+l) = {n}|<j>+<j>| {-m} = {n-m-1} |{a>-m}.

confusion note that

{n}|{o)-m}

which is the sequence of

u> pluses followed by two minuses.

m is fixed although

varies.)

(To avoid

This is

which, using the obvious inductive hypothesis, gives the

sequence of length

oo+m+l

consisting of

pluses followed by

m+1

minuses.
I t is natural now to ask what
minuses represents.

infinite, i.e.

pluses followed by u>

The pattern suggests naively that i t may be

but t h i s is obviously impossible.

cal order.

That element i s , of course,

co-w

still

above every positive integer because of the lexicographi-

In f a c t , no sequence of minuses no matter how large can undo

the e f f e c t of the f i r s t

co pluses.

We shall return to the above example

shortly.
The computation for
as

co.

For example

w-m

(co +6u)) - 3

followed by 3 minuses.

works for any l i m i t ordinal as well

is the sequence of

w2 + 6u>

pluses

One can also apply the computation to non-limit

ordinals to check the consistency of what we have already done.

example

(oo+6) - 1

you the sequence of

correctly.
a

is simply
w+6

w+5.

pluses followed by a minus i f i t is done

The step to beware of is the following.

we can simplify a lower set such as

{3: 3<a}.

For

The computation w i l l thus not give

{3-1:3<a}

For a l i m i t ordinal
by replacing i t by

This i s , of course, not v a l i d for non-limit ordinals, since

the former set is not cofinal in the l a t t e r .

We now compute a> + -z* This is
1
{n}|<j> + { 0 } | { l } = {n+2 a>+O}|{a>+l} = {w}|{u>+l}.
w+1

pluses followed by a minus.

This is the sequence of

In l i n e with our e a r l i e r computations,

this example i l l u s t r a t e s the contrast between the cases where

is a

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

l i m i t ordinal and where
sequence with

46

is a non-limit ordinal in the value of a

pluses followed by a minus.

In the same manner induction can be used to evaluate

for any positive real number
canonical representation of

r.
r.

co+r = {n}|<j> + F|G

This is

the left-hand side may be replaced by

where

w+r

F|G

{Go+F,n+r}|{oj+G}.

is the

Since

oo+F and we obtain

0 e F,

{GO+F}|{GO+G}.

By the lexicographical order and the inductive hypothesis, this is the

sequence with

u> pluses followed by the sequence for

r.

Incidentally,

this reasoning with the lexicographical order and juxtaposition is

similar to what has already been used back in the proof of theorem 2.1
and represents an important s k i l l to f a c i l i t a t e computation.
This argument works just as well when
as

is not an integer, in which case

is negative as long

is empty.

The l a t t e r case

has in fact been done e a r l i e r and the general conclusion is s t i l l valid

although

GO+F

can no longer be used as a lower set.

As before, similar results apply i f

GO is replaced by other

l i m i t ordinals.
We now consider

|w.

This is

({0}|{l})x({n}|<j>).

By the

definition of multiplication this is

n+OnO^^n+lnl}

= { r O l I w --gn} = {n}|{u)-n}.

result for w-n, this is the sequence of length

pluses and followed by

GO minuses.

GO. 2

Using the e a r l i e r

beginning with GO

This answers an earlier question

which was l e f t open.

I t is instructive to see another proof.
sequence of
Hence

2a = a+a = {n+a} |{Go-n+a}.

cofinality.
Hence

GO pluses followed by

w minuses.

Since

be the

a = {n}|{Go-n}.

We now prove that this is

We know that for a l l positive integers

n+a < GO.

Let

Then

a > n, i t follows that

satisfies the betweenness condition.

Since

c o f i n a l i t y condition is s a t i s f i e d ; hence

n,

GO < Go-n+a.

GO = {n}|<|>

u> by

n < a < Go-n.

So OJ

and

n+a ^ n, the

2a = {n+a} |{u)-n+a} = w.

The f i r s t proof involves more computation but is more

routine.

The second method requires a good guess at the answer and being

able to use cofinality in spite of a limited knowledge of

2^-1

and more generally

Go+r.

for arbitrary real

be handled similarly to

u>-l

is the sequence for

followed by the sequence for

ya

and

-^o+r

a.
r

In f a c t , the sequence for

r.

can
yu+r

Such simple

REAL NUMBERS AND ORDINALS

47

r e s u l t s make the subject quite t r a c t a b l e .

However, the reader is warned

t h a t crude juxtaposition does not work in every case.

Several more remarks are worthwhile to mention a t this point.
F i r s t , we take for granted obvious i n e q u a l i t i e s involving i n f i n i t e or
i n f i n i t e s i m a l elements ( e . g . , for a l l positive real
wr-n

i s positive i n f i n i t e )

cofinality.
M

and integers

n,

and apply them to obtain information about

Secondly, r e c a l l the d e f i n i t i o n of m u l t i p l i c a t i o n and

a = A1 |A"

consider

ll

b = Bl|B".

and

a"b+ab -a b }|{a b+ab"-a b",

Then

ab

a'^+ab'-aV}.

is

{a'b+ab'-aV,

I t is often convenient to

think of the lower sums in the form

a'b + ( a - a ' ) b 1

and

and the f i r s t upper sum in the form

a'b + ( a - a ' ) b "

or

a"b - (a"-a)b"
ab" - a ' ( b " - b ) .

In spite of the t r i v i a l i t y of the algebra, a suitable form supplies a

considerable gain in i n t u i t i o n .
We now compute
dyadic f r a c t i o n s .

-rw.

-r = (+-+) = { ^ I d }

by our work on

Hence

| u = ( { | } | { l } ) x ( { n } | c j ) ) = {u + ( | - | ) n } | { l . a ) - ( l - f ) n } .
the above remarks.
be w r i t t e n as

Also note that a typical lower term may a l t e r n a t i v e l y

j n + -2(w-n),

magnitude more c l e a r l y .
ju = {-^o+n}|{u)-n}.

but the form we have exhibits the order of

In f a c t , i t is immediate by c o f i n a l i t y

I f we accept the juxtaposition results for

see t h a t t h i s is the sequence of length

followed by

Note the use of

w minuses and then

u).3

consisting of

that
-^to+n we

pluses

pluses.

By a similar process we can use a double induction to obtain

the sequence for

wr, one for

to be j u s t l i k e the sequence for

a)

times.

in
r

wrn

and one for

r.

I t turns out

except that each sign is repeated

The e a r l i e r remarks on m u l t i p l i c a t i o n which are petty in an

individual numerical example gain in value as we generalize.

Now we consider
the induction procedure for
that

u) -a)

consists of

u>2-a).
co2-n

I t is easy to see that in this case

extends even to

pluses followed by

u)2-u)

so we obtain

w minuses.

The key

point is the elementary f a c t about ordinals which says that

a < a)2 + a+a) < a) 2 .

The reader can experiment with obvious generaliza-

tions in t h i s d i r e c t i o n and see that there are no further dramatic

surprises.
Before studying the ultimate representation theory, i t is
worthwhile to see what happens in the other d i r e c t i o n , i . e .
expressions such as

0)

with

We prefer to ignore our previous construction of

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

reciprocals j u s t as we did in our study of the real numbers.

48
Our primary

i n t e r e s t in the l a t t e r construction was, of course, in the issue of

existence since i t is usually not convenient to use such an inductive
construction for computation even i f i t is possible.
Instead, we use the good guess approach.
sequence of length
minuses.

Let

e be the

w which consists of a single plus followed by

This is clearly a positive i n f i n i t e s i m a l .

In f a c t , i t is

immediate that i t is the unique positive i n f i n i t e s i m a l of length

hence i t may be regarded as the canonical i n f i n i t e s i m a l .

co;

Heuristically,

t h i s is a reasonable candidate for being the reciprocal of the canonical

i n f i n i t e l y large number

w.

canonical representation
ew

= (0|^}) (W|<J0

of

Now

0|{}.

1 = {0}|<j>.

We now check the conditions of the

A typical upper element -w - -m + em

Hence 1 s a t i s f i e s the betweenness condition.

c o f i n a l i t y part requires minimal work.

we do obtain

Hence

Since e i s i n f i n i t e s i m a l , a typical lower element

which is em i s less than 1.

clearly i n f i n i t e .

is

Note that the

= (0|{~})x({m}|<j0 = {em+0ur-0u)}|{em + ia> _ im} =

{em}|{em + p j -pjiri}.
c o f i n a l i t y theorem.

We now prove this f a c t .

Regardless of

is

The

m, em >^ 0.

Hence

1 as required.
Note that in spite of the existence of i n f i n i t e s i m a l s there

i s no connection with nonstandard analysis.

transfer p r i n c i p l e .

So far we have not had any

In f a c t , no model-theoretic ideas of any kind have

played a r o l e .
Let us now consider
First, let

r + e where

be a dyadic f r a c t i o n .

suitable u n i t sets where

and

Then

is a real number.

has the form { s } | { t }

are also dyadic.

for

Then

r + e = { s } | { t } + { 0 } | { ~ } = { r + 0 , s + e } | { r + J , t + e } = { r } | { r + ^ } = {r}|G
where

G is the set of a l l dyadic fractions above

r.

I t is immediate

from the d e f i n i t i o n that t h i s is the juxtaposition of the sequence for

for

with the sequence for

Note that t h i s argument applies to

a l l dyadics including integers since by c o f i n a l i t y we can always insert

an

or

t.

Note also that we can now f i l l a gap which remained since

our discussion of the real numbers, since we see from the above that a l l
sequences whose terms are eventually minus are obtained this way.
a l l sequences of length
dyadic

r, or of the form

u> are either r e a l , w, of the form

1
r - for dyadic r.
(A)

Thus

r +
w

for

REAL NUMBERS AND ORDINALS

Now l e t

49

be non-dyadic.

Let

r+e = R'|R" + { 0 } | { ~ } = { r + 0 , r ' + e } | { r + p

where

G i s the set of a l l reals above

set of dyadics instead.

r = R'|R".

r.

(G may be taken to be the

I t makes no d i f f e r e n c e . )

I t is immediate from

the d e f i n i t i o n that we now obtain the sequence for

single plus.

Thus we have a sequence of length

counter-example to naive j u x t a p o s i t i o n .
is worth only

e!

is dyadic or not.

w+1. We have here a

In f a c t , the f i n a l "poor" plus

r

This brings some subtlety to the subject.

2e.

This is

{ 0 } | { ^ } + { 0 } | { ^ } = {0+e}|{|+e} = {e}\{h-e}
Hence

2e

= {e}|{-J}

i s the sequence for

the contrast between t h i s case and that of

by mutual

followed by a plus.

2u>.

Note

Again the f i n a l plus has

e.
Consider

followed by a

but the sequence added on depends on whether

Next, we compute

value only

Note that in both cases we are juxtaposing a sequence

to the sequence for

cofinality.

Then

r"+e} = { r } | { r + | } = r|G

|e.
=

This i s ( { 0 } | { l } ) x ( { 0 } \ { } )

{ 0 }

{ e }

Th1s

1s

the

se( uence

which is
of

length

w+1

which is that of e followed by a minus. We can say that the value of

the last sign is only -ze. The inflation rate is growing rapidly! Again
6

note the contrast between t h i s case and that of -*& where u> minus
1
signs took care of the y .
Next we consider re f o r a t y p i c a l positive real number
which is not an integer.
M

R'|R x{0}|{J} = { e r \ e r

This has the form

- ^(r"-r)}|{er",er'

c o f i n a l i t y t h i s s i m p l i f i e s to
pattern by i n d u c t i o n .

er',er"}.

+-J(r-r')}.

By mutual

This enables us to obtain the

We obtain the sequence for

followed by the

t a i l of the sequence for r a f t e r the f i r s t +. For example

3
3
x = (+-+). Hence -je i s the sequence for e followed by (-+).

To

j u s t i f y the induction we need to v e r i f y the pattern for positive integers

as w e l l .
instead to

In that case
{er'}|{}.

R"

i s empty and the expression s i m p l i f i e s

However, the pattern works in any case although

here there are no i n f i n i t e s i m a l upper elements as before.

A natural expression to consider next i s

e 2 = ?

In

analogy with e = one natural candidate for the sequence is a plus

followed by w2 minuses. But
{ }

{!})

= {0> +

follows from the previous result for

. 1_ } ( { }

rw

{ O } ! ^ }^.

It

that the result is the sequence

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

50

consisting of a single plus followed by u>-2 minuses. In particular,

we do not have w 2 minuses as one may naively guess. Of course, the
contrast between the behaviour of ru and re should make an alert
reader suspicious of such a guess at the outset.
It is next of interest to investigate e+e 2 . Is it the
sequence for e followed by the sequence for e 2 ? Well
{0}|{^-} + {0}|{^} = {e,e2}|{e + ^ , e 2 + -} = {e}|{de} where d is the
set of all dyadic fractions larger than 1. Using the previous result
for re thus leads to the sequence for e followed by the sequence for
e! Thus, again, juxtaposition behaves subtly. The e 2 term is contributed by the annexation of the term for e.
The examples we have done illustrate the most basic tricky
phenomena that occur in attempting to find the sign sequence for various
algebraic expressions. We close by considering the expression /u>.
We ignore our previous construction of square roots and use a
good guess method. Let b be the sequence consisting of u> pluses
followed by w 2 minuses. We consider the canonical representation and
obtain a cofinal upper set by restricting ourselves to those sequences
for which the number of minuses is a multiple of OJ. So by earlier
computations we have b = {n}|{~-} = {n}}{~}. Note first that this is
a priori a plausible candidate for /a!" since
/w

n < /uT and

n
b 2 = {bn, + bn o - n,n o ,
1

mi

} I {nb + ~ } '
m2

^1^2

Now a) = {n}|<f>. So we must verify the conditions for the cofinality

theorem. As usual in a proof of this form all we know about b in
advance is that n < b < , i.e. that b and T- are both infinite.
There is a trap which is a temptation to use circular reasoning (i.e.,
that b 2 = a)). In this respect this proof is similar to the one dealing
with and the second proof for the sign sequence for -jwThe conditions are easily verified.
bn x + bn 2 - n ^ 2 <_ b(n 1 +n 2 ) < w since j- is infinite. For the same
reason + < 0 < u>. Also
m^ m 2
m1m2
b is infinite.

nb + S - ^ >
2m 2m

A
2m

> a> since

REAL NUMBERS AND ORDINALS

51

This verifies the betweenness condition. If we let

ni = n 2 = 1, we obtain 2b-l as a lower element which clinches the cofinality condition since b is infinite. In fact, the same element 2b-l
works for all n. This shows that b 2 = w.
The supply of surreal numbers is very rich. Continuing in
the above manner is like using a teaspoon to empty an ocean. It is time
now to get some sort of hold on more general elements.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

5

52

NORMAL FORM

A COMBINATORIAL LEMMA ON SEMIGROUPS

So far we have more or less accepted everything we needed
from outside the theory of surreal numbers since the material was very
elementary. However, in order to study the normal form we need a
combinatorial lemma which is not as well known, and since it is interesting in its own right we shall prove it here.
Lemma 5.1. Let T be a set of positive well-ordered elements in a
linearly-ordered semigroup. Then the set S of finite sums of elements
of T is also well-ordered and each element of S can be expressed as a
sum of elements of T in only a finite number of ways.
Proof. Both parts will follow if we show that any sequence {s n } of
elements of S such that s n _> sn+i for all n in which the representations are distinct must eventually terminate. We will assume an
infinite sequence and obtain a contradiction.
Case 1. Suppose that we have only binary sums. Let s n = a n +b n where
a n e T and b n e T. Since T is well ordered, there exists a subsequence ai of a n such that ai
> a-j for all n. We can obtain
n
n+i ~ n
such a subsequence as follows. Choose i such that a-f is the least
value of the a's. If i' < i* ... < i n have been chosen, then choose
in+l > in sucn that ai
is the least value of the a's with index
n+i
larger than i n . Similarly there exists a subsequence of bj of bf
such that
for all

biJ
n.

> bi
n+i - Jn
Since aiJ

for all

n+i

+ bi

Vi

n
n
Then ai
> ai
and bi
> bi
J
J
J
J
n+i n
n+i ~ n
< aiJ + bi , we obtain that

n.

NORMAL FORM

53

a-j
= ai
and bi
= bi , contradicting the hypothesis that the
J
J
J
n+i
n
n+i
V
representations are distinct.
Case 2. Suppose that we have only sums of k terms for fixed k. Then
the proof is essentially the same, but the notation is minutely more
complicated. If s n has the form a n i + an2 sink "then we simply
iterate the process of taking subsequences for all fixed j _< k.
Case 3. If only a finite number of k's occur, then we can reduce to
case 2 since at least one of these k's must appear infinitely often.
Case 4. Suppose an infinite number of k's occur, i.e. the value of kis unbounded. Then we can choose a subsequence {sn 1 } of {Sn> as
follows: S J 1 = aii + a-f2 ... + ain- where ni is a strictly increasing
function of i. In particular, nj >^ i. Also, we express the sums in nonincreasing order, i.e. a-ji >^ ai2 2. 1. ain- We now choose a subsequence of {s n '} such that a-ji is monotonic increasing, a subsequence
such that ai2 is monotonic increasing, etc. By the usual diagonal
method we obtain a subsequence s n " such that a-jj for fixed j is
monotonic increasing as a function of i for i >_ j. [Since ni >_ i,
a-f j is defined for i _> j.] First we show that necessarily ni > i
for all i. Suppose ni = i. Then consider
si" = aii + ai2 ... aii
and

Since ai+i^j >_ ai 5 j for all j _< i, and si+i" contains an extra
term a-j+i^-f+i which has no analogue in si", it follows that
s

i + l " > s i " . T n is contradicts the assumption that the sequence is

monotonic decreasing.
For arbitrary i we now compare:
Si" = aii+ ai2 . . . aii + a i , i + i . . . * i n .
7
with
s

n."

= a

n.i + a n .2 ... an.i + a n .,i+i ... a n .n.

Since a n .j >^ aij for j _< i, since s n " contains more terms than si"
and since s n " is monotonic decreasing, there must exist k such that

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

i < k n-j with

aj k > a ^ .

Hence

aji >_ aj|< > a ^ k I a ^ i y

54
By

induction we may now define a function b(i) as follows: b(l) = 1 and

b(i+l) = nb(j). Then {^b(i),b(i)^ 1S an infinite strictly decreasing
sequence of elements of T, which is our final contradiction. This
completes the proof.
B THE a) MAP
Up to now we have considered real numbers, ordinals, and
algebraic combinations of these. What we need now is a more tractable
way of looking at a general surreal number. We begin by studying orders
of magnitude, a concept which has meaning in any linearly ordered field
containing the real numbers.
We define an equivalence relation on the positive surreal
numbers.
Definition, a - b iff (3 integer n)(na _> b and nb >_ a). This is
trivially an equivalence relation. The equivalence classes are called
"orders of magnitude." Related to this is another definition.
Definition, a b iff (V integers n) (nb <_ a), a b iff b a.
We say in words that a has a higher order of magnitude than b.
Clearly we have a b, b a, or a ~ b.
We shall assume that the reader has no trouble in seeing the
most obvious consequences of these relations, so that they will be freely
used without a detailed explanation when needed.
a b = na b, a c and b c =< a+b c. One property of
special interest is a ~ b and a < c < b =*> a ~ c, i.e. the equivalence
classes are convex.
One basic fact which is special about the surreal numbers is
that each equivalence class has a canonical number.
Theorem 5.1. Let a be a positive surreal number. Then there exists a
unique x of minimal length such that x ~ a.
Proof. The argument uses only the convexity. Because of well-ordering
there certainly exists an x of minimal length such that x ~ a.

NORMAL FORM
Suppose

x ~ a ~ y.
property

and

Let

z - a.

55

are d i s t i n c t , both have minimal length and

be the common i n i t i a l segment.
Also

l(z)

< (x)

By the convexity

which contradicts the minimality of

Remark. The same argument shows that the element of minimal length is an
initial segment of every other element equivalent to a.
Similarly to the above one can define additive orders of
magnitude using addition rather than multiplication.
Notation,

x ~ a iff (3 integer n)(a+n j> b) and (b+n _> a ) .

Since this is less important for our purpose and, besides, is
similar to the case of multiplicative orders of magnitude including the
possession of the convexity property, we do not give the details. It
suffices to note that here also every equivalence class has a canonical
member which is defined in a similar way.
We now come to one of Conway's most remarkable discoveries
[1, page 31]. The canonical elements can themselves be parametrized by
the surreal numbers in a natural way.
For every surreal number b we shall define an element
written u> which may be thought of as the canonical element of the bth
order of magnitude. (Although there are philosophical objections to the
use of the exponential notation which have some validity, there is
enough in common with exponentiation to make the notation psychologically
convenient.) As usual, we use induction and assume that a)C has been
defined for all proper segments of b. Then.
h

Definition.

k11

k '

u> = {0,ro>

}|{sw

of a l l p o s i t i v e real numbers, and

where
b1

and

and
b"

run through the set

run as in our usual

notation through the lower and upper elements of the canonical

representation of

b.

By c o n f i n a l i t y we can, of course, l i m i t
s to dyadic f r a c t i o n s with numerator
Theorem 5 . 2 .

b < c > w << OJ .

to integers and

1.

i s always defined and greater than

0.

Furthermore,

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

56

Proof. We prove this by induction on the length of b. Since b1 < b11

~"~

, i

we have

,n

u>

<< o>

p o s i t i v e reals
defined.

by the inductive hypothesis.

Since

and
0

s, rw

< sw

Hence, for a l l

Also, 0 < sw

Hence

to

is

is a lower element in the d e f i n i t i o n , 0 < w .

To conclude the proof, we use a method which is similar to

the one we used for the arithmetical operations.
immediate.
d = b

or

Suppose
c

and

is the common i n i t i a l

a>

If

u)C.

The uniformity theorem holds for

a> ,

i.e.

an a r b i t r a r y representation the same formula holds,

cob = { 0 , r / } | { s o ) G } .
Proof.

segment.

then the conclusion is immediate from the d e f i n i t i o n .

Otherwise we have
Corollary.

b < c

Here the computation is

= {u)X: xeF}

(/

if

for

i.e.

and s i m i l a r l y for

As usual, this follows from the inequality

b = F|G

G.)

b < c - w < w

using the inverse c o f i n a l i t y and the c o f i n a l i t y theorems.

Theorem 5 . 3 .

An element has the form

i f and only i f i t is the

element of minimal length in an equivalence class under

~.

Proof.

F i r s t consider an element of the form w . We have

b\,r
b\
,
b
u> = {0,ro) }|{su)
} . TI f x ~ a)
then x also s a t i s f i e s
b'
b"
ro>
w

< x < so)

since

i s an i n i t i a l

and

segment of

are a r b i t r a r y positive r e a l s .

x, so

i(u

) <_

Hence

i{x).

For the converse, we show that every positive element is

equivalent to an element of the form
b < c -- w
induction.

Let

a = A'|A"

every element in
Let

F = {y:

x < a < z

A'UA"

F < G, F|G

0 e A .

u> + a 1
has meaning.

By the inductive hypothesis

and G = {y: (3xeA")(x~a) y )}.

Then

i t follows that

x > y.

We use

is equivalent to an element of the form

y e FUG.

Now suppose
x e F, y e G and
x
v
x > y + w

In view of the inequality

Then

(3xeA') U~J)}

FOG t ft and l e t
Since

u> .

u)C, such an element is unique i f i t e x i s t s .

a".

~ x e A1

and

u>y ~ z e A".

a ~ </.

FOG = <j>.
Then

We claim that

wX ~ a 1 e A'

and

F < G.

For suppose

wy ~ a" e A".

This is impossible since a 1 < a".

Let

w .

Suppose

z = F|G.

Then

Hence
Since

is a complete set of

representatives for the equivalence classes containing the elements of

A1 - {0} and similarly for a) with respect to A". We now consider

NORMAL FORM

57

three cases.
ra) _>. a

Case 1 .

fr

1
~~~~~""
""'~~~'
1

a1

e A'
a e
A' s a t i s f y
hence a

some

positive real

~ oo .

Then

ru> _< a

for some positive real

satisfy

a" ~ < / .

but

Let

a' ~ u ~ ru> ;

and some

ru>x <_ a a"

Then

x e F.

XX

a _< a <_ rw

a" e A"

but

x e G.

Let

rwX ~ </ ~ a";

a ~ a> .

Case 3.
rJ

and some
X

Case 2.

hence

Neither case 1 nor case 2 is s a t i s f i e d .

G

< a

< sa) .

particular,

e A

for some real

(3xeG)(a"~a) X ).
a = A'|A M

Now l e t

r, ru

- {0}.
1

>_ a .

Then

This says that

(3xeF)(a'~u) X >.

S i m i l a r l y for

a" e A"

s, sooX <_ a".

Hence for some positive real

In
we have

Since

t h i s shows that the c o f i n a l i t y condition is s a t i s f i e d for

{0,ru) }|{SUJ } .

Hence

a = {0,rw }|{su) } = w .

The theorem now follows immediately, for i f

length in i t s equivalence class, then
already shown, o>

a ~ u> -> a = u>

has minimal

since, as we have

has the minimal length property.

Our next r e s u l t gives some j u s t i f i c a t i o n for the exponential

notation.
Theorem 5 . 4 .

(a)

w = 1 , (b) w w = w

is the same as the ordinal

Proof.

By d e f i n i t i o n

usual.

Let

and

our u>

in the usual sense.

oo0 = {0}|<|> = 1 .

a = A'lA"

, (c) for ordinals

b = B'|B".

We prove (b) by induction as

Then by the formula for

addition and the uniformity theorem, using the facts that

u)a = {0,ra) a l }|{su) a M }

and a)b = { O . r ^ ' } ! { s ^ " } ,

a) a+b = { 0 , r a ) a < + b , r 1 a ) a + b I } | { s a ) a l l + b , s 1 a ) a + b " } .

we obtain that

S i m i l a r l y for m u l t i p l i c a t i o n

we obtain that
ab

a'b
b'a
a'b
b'a
a'b1
a"b
b" a
a) a) = {O,roj (o 9r^
a) ,rw co + r.oo a> - rr.u
a) ,sa) w + s.co a)
a" b1' T a" b
b" a
a1 b
b" a
a 1 b11
a" b
-SS.u)

0)

J-|iSo)

+ rui

a)

so)

a) ,S.o)
a)

0) ,rco

co

+ S.U)

rs.o)

a)

, So)

a)

}.

By the inductive hypothesis this may be written

s / '+a
s/'

+ a

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

We now show that the representations for

are mutually c o f i n a l .
u>a

u>

58

and

u> w

One direction is immediate since the terms for

are among the terms for

wau) .

Since

u)C

b > c o>

the other

d i r e c t i o n follows e a s i l y by elementary reasoning with orders of magni-

tude.

First

Next,
Next, soua

+ s^

- ss 1 o a

II , r II

dominates the other terms.

element

rwa

+ s^

- rs^

Now we consider a typical upper

.

Since the term containing

dominates the other terms, i t follows that for any

i s above

s2u>

s2 < s

Since the same argument applies i f

u>

the element
and

are

interchanged, this v e r i f i e s the c o f i n a l i t y .

(c) follows easily by induction.
purpose of this proof l e t us temporarily use
our order of magnitude, and use
ordinals.

Then

u>

Let
F(c)

a = {a'}|<|>.
instead of

For the
a>

for

in i t s usual sense in the theory of

F(a) = { 0 , F ( a * ) } |<j> = {0,n</ }|<j> = o)S.

is a basic f a c t concerning the ordering of the ordinals.

The l a s t equality
This completes

the proof.
Corollary.

u> o>

=1.

Theorem 5.4 gives some j u s t i f i c a t i o n for the exponential

notation.

However, the main j u s t i f i c a t i o n comes from the nice way the

operations behave on the normal forms of the surreal numbers which we

w i l l see l a t e r .
Remark.

Note that i f F contains no maximum, then rw

may be replaced
F
x1
simply by w
because of c o f i n a l i t y . 1 For suppose1 rco 1 e F.
If
v'
x
v
x
y e F and y > x
then ur >> w ; hence or > rw . A similar
P

remark applies to

sw .

NORMAL FORM

We now obtain something which is analogous to the normal form

for ordinals. Here we need transfinite sums. We shall define
expressions of the form

I W ">r-f where
i

(aj) is a strictly decreasing

NORMAL FORM

59

t r a n s f i n i t e sequence of order type

from

for a l l

i.

a and r-j

Case 1 .

is a non-limit ordinal.

Case 2.

is a l i m i t ordinal.

form

i s a real number d i s t i n c t

This i s done i n d u c t i v e l y on a.
Let a = 3 + 1 .

We obtain

Then

I OJ ir^
i<a

in the

F|G.
A typical element of

such that

s-j = r-j

for

has the form

i < 3, and

I u is-j, where 3 < a

i<3
where e i s positive

S3 = r$ - e

real.
S i m i l a r l y a typical element of
where

3 < a, such that

e is positive r e a l .
I w ir-j

s-j = r-j

as an a l t e r n a t i v e for
0

i < 3, and for

(We use the natural

I
i

i
If

for

w
l

1#

oa 1# Sj,
i<3
S3 = r3+e where

G has the form

notation
Sj.)

is regarded as a l i m i t ordinal the d e f i n i t i o n leads to

the empty sum being

<|>|<j> = 0.

ordinary f i n i t e sum.

For

For

f i n i t e , the expression is j u s t the

i n f i n i t e , a proof that

order to show that the d e f i n i t i o n makes sense.

F < G is needed in

In f a c t , we shall show

that the ordering on surreal numbers is consistent with the lexicographic

ordering with respect to the

a's

and

r's

in the normal form.

we define the lexicographical order on expressions

Let

x =

Let

I a) "ir-i
i<a

and

y =

y < min(a,3)

then

If

Y = 3, then

x > y

x > y
iff

iff
rY > 0.

a Y > bY
If

I a> i r - j .
i<a

][ u> "isj.
i<3

be the l e a s t ordinal such that

If

First

or

Y = a

(aY,rY) * (bY,sY).

a Y = bY
then

and

x > y

rY > s Y .
iff

sy < 0.
Note that this is consistent with the situation for the
normal form for o r d i n a l s .
Theorem 5.5.

The expression

decreasing sequences

(a-j)

03 ir-j is defined for a l l s t r i c t l y

i<a
and a l l nonzero real r-j.
The ordering is

given by the lexicographical order.

Furthermore for a l l

3 < a,

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

60

I I ^ r j - I u>afrj| o)aJ if j < 3. (We call the latter inequality

the "tail property".)
Proof. We use induction on a.
Case 1. a is a non-limit ordinal. Let a = 3+1. Then
a.
a.
a
1<3

a.
b.
To begin with, let x = J u irj and y = I w 1 s^ and
i <a
i <a
suppose x > y in the lexicographical order. We must show that x > y
as surreal numbers. If (Vi<B)C(a-f,r-f) = (bj,si)] then either a3 > bg
or a3 = b3 and r3 > S3. In either case u> ^ 3 > u> &S3 by elementary
reasoning with orders of magnitude. Since addition preserves order, it
follows that x > y.
Next assume (a Y ,r Y ) * (b Y ,s Y ) for some y < 3 but
(a6>r,s) = (b6,ss) for 6 < y. Then either a Y > b Y or a Y = b Y and
r Y > s Y . In either case u> Y r Y - u Y s Y _> Y t for some positive real t.
Hence

I u 's-f = ( ), w "r-f + w r Y ) - [ I o> 1s-f + u> sYJ

I u> 'r-,- -

3y

by

= o) T r Y - a)

Ts

I I a) *rj
i

I
i l

ay

si

y ^ a) T t .

I <*> ^ j and
i<3
property by the inductive hypothesis. Hence
1#

rj| a) Y

and

less than

t.)

1s

nave

tne

for a positive real t 1 .

Again, since
1#

bo
b
a
a> p m Y co Y,

.
(We can

we obtain

I a) V j - I a) Si = J u rj - I 10 s-i + w &>3 - u ^S3 _> u Y t"

i<a
i<a
i<3
i<3
for a positive real t". In particular x > y.

x-y =

1#

Xw
i<3

J u> 1#Si| o>

i l

\ I u isj
i

J OJ 1#r| - I u " s-f >^ w Y t '

i<3
i<3

Therefore
use any

a)

However

1#

The same argument applies i f only one of

representation of length

a.

i f one is picayune since

ae

case the superiority which

x and y

have a

A slight difference in notation is needed

and be

are not both present, but in any

x gains in the

Yth

stage is necessarily

maintained since whichever one of the above is present is s t i l l of lower

aY
order of magnitude than u> . (There is even a possible case where aY is

NORMAL FORM

not present.

61

Then we use

bY
u> '

We now check the t a i l

Consider

| I co i r j - I to ""r-f | .
i <a
i <y

| w 3 r 3 + ( I to 1r-f -

i<3

I a) ">r-f) | .

instead.)
property.

Let

y < a

and

j < y.

This i s

If

y = 3

t h i s reduces t o

|to &ra|

i<y
a ;

co J .

which is certainly of lower magnitude than

a

certainly s t i l l of lower magnitude than

| a) i r j - ^ a) " r-f |
i<8
i<y

Otherwise

|to ^r^|

is

to , and so is

by the inductive hypothesis.

Hence the absolute

value of the sum is also, and we are done.

Case 2.

a is a l i m i t ordinal.
By the inductive hypothesis the ordering for elements in

FUG

is given by the lexicographical ordering; hence

We now verify the t a i l property.

We must consider

| I co i r j - co " r-f | .
i<a
i<3

F in the representation of

J u Vj
i<a

F < G.

3 < a and

j < 3.

Now among the typical elements

a.

of

Let

a.

is

a-;

I u> " r-f - co J e.

i<j

This is

I co ir-f + co J r-j = I co i r j . (One

i<j
1<j
must be cautious in reasoning with these i n f i n i t e "sums". Other results
immediate from the definition, since

which may appear to be just as obvious might require a technical proof

because of our specialized definition of i n f i n i t e sums.) Similarly,
a.
aj
among the typical elements of G is
[ w T J + u e, Hence
i<j
a.
a u;
a.
a . + w aej B t n e
w 1r
w l r
I co ' r j - co e < 2,
j < I
i
y
lexicographical
i<j
i<a
i<j
order, and the inductive hypothesis, we have
I co i r j - a) Je < I co i r j < I co i r j + co J e. Therefore
i<j
i<3
i<j
a.

a.

aj

| I co V j - I to " m-f | < co J (2e). Since e is an arbitrary positive

i<a
i<8
a.
a.
a j as
real this shows that
X ^ 1 r i ~ I w 1 p i l <K w
desired. The proof
i<a
i<8
that the lexicographical order is the correct order is similar to the
proof in the non-limit ordinal case. As before, let x = I co i r j
i

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

and

y =

J u isj

and suppose

x > y

i n the l e x i c o g r a p h i c a l

62

order.

i<a
(a$,r$) = (b,s,ss)

Suppose

for a l l 6 < y for some y < a but

a.
b.
aY
( a Y , r T ) * ( b Y , s Y ) . Then
J u ^ j - I u> ">sn- _> u> t for some positive
iy
iy
real t . By the t a i l property,
a.
|

a.

I a) "'r-j

i<a
Therefore

ay

I a) "> m-f |

a)

and

I I to ""s-j -

J ^ S J I

by
u

av
<u

i<y
i<a
i<y
~wa l. r
b.
atY f o r s o m e ~~
w l s
w
).
i " I
i 2.
p o s i t i v e real t .
i <a
i <a
I f only one of x and y has a representation of length

a,

then the e a r l i e r remark i n the n o n - l i m i t ordinal case remains v a l i d .

This completes the proof.
The next theorem gives us the importance of the t r a n s f i n i t e
sums we have been d i s c u s s i n g .
Theorem 5 . 6 . Every surreal number can be expressed uniquely i n the form

Proof.

Uniqueness is immediate from the fact that the ordering is given

by the lexicographical ordering.

Now l e t

by Theorem 5.3 that

real numbers

|x| ~ cu

such that

i s bounded above.

Let

for a l l positive real

follows that
uniquely.
A(x) = war

be an arbitrary non-zero surreal number.

r t 0.

r = .u.b. S.
e;

\ a) r 1 *.
i<a
ordinals.

hence

Let S be the set of a l l

Since
Then

|x| ~ </, S is non-empty and

(r+e)w

|x-w a r| wa.

> x and

Since

(r-e)co

|x| ~ ( /

<x

it

I t is clear that the above property determines

For convenience in the proof we shall use the notation

if

x * 0.

Now assume that

1

for some a.

su>a <_ x.

We know

x cannot be expressed in the form

We define a sequence
Suppose

A [ x - I a) 1 ^ 1
i<a

(a-j,r-j)

(a-j,r-j)

where

i s defined f o r a l l

runs through a l l

i < a.

the

Then

= a) <*ra.

Intuitively speaking, we are getting better and better

approximations to x as a is increasing.
are decreasing, so that the sums make sense.

We first show that the a's

NORMAL FORM

63

First let a = 3+1. Then

a
a
A[x- I u> ir-j] = u Prp.
i<3

a
a.
a.
a
a
o> a r a = Afx- X a) ""r-jl = Af (x- I u> 1 r 1 # ) - a) ^ 3 ] a> & by the inductive
i<a
i<3
d e f i n i t i o n of (agjrg). Hence a a < a$.
Now l e t a be a l i m i t ordinal and l e t 3 < a. We already
know that

|x- I u> i r - j | ~ Afx- I w " r-f ] u> &. By the t a i l

i<3
i<3
a.
a.
a
a.
a
property | I w ir-j - w Tr-j | u> &. Hence |x- J u 1 r - j | w P.
i<a
i<3
i<a
a
a.
ao
Therefore u a r a = A(x- w l r i ) <<: w S o f i n a l l y a a < a$.
i<a
Since by hypothesis x cannot be expressed as a sum,
I a) ir-j
i<a

has meaning for a l l

a.

We next show that

a[ J w ir-j) > a for any general sum.

i<a
Although this inequality is crude i t suffices for our immediate purpose.
Let
representation of

a < 3.

Then the elements of

J o> ""r-f

i<a

F and

G used in the

are also used in the representation of

a.
I a) ir-j. Hence the former is an initial segment of the latter and thus
i<3
it has smaller length. The uniqueness of representations as sums guarantees that the length is strictly smaller. This is enough to verify the
inequality since every strictly increasing function f from ordinals to
ordinals necessarily satisfies f(x) >_ x for all x.
If a is a limit ordinal we already know that
a.
a
|x - I a) ir-f I a> & for any 3 < a. This shows that x satisfies
i<a
F < x < G for the F and G used in the representation of J u ir-j.
i<a
a.
Hence a( I u> ">r-,-) < (x). This is true for all a. In view of the
i<a
earlier inequality this implies that i[x) is above every ordinal, which
is absurd. This contradiction completes the proof.
We have now established the normal form for surreal numbers.
The usual representation of ordinals in terms of powers of u> is a
special case of this, since finite sums correspond to ordinary addition

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

64

and this agrees with ordinal addition i f terms are arranged so that there
is no absorption.
Next we shall show the fundamental fact that the basic
operations can be performed on elements in the normal form analogously to
usual operations on polynomials.

This is the main justification for the

summation as well as the exponential notation.

If

x = J in i r j , we shall call a the normal length of

i<a
n(x). This is quite different from i{x)
which was

abbreviated

defined at the beginning.

We shall study

n(x)

in more detail

x,.

later.

We now need some lemmas which will help us deal with the
normal form.
Lemma 5.2.

u> r = {w (r-e)}|{a) (r+e)}

where

is an arbitrary positive

real number.
We know that
a>a = {0,su>a }|{ta) a }
positive r e a l .

Hence

r = {r-e}|{r+e}

by definition where
u*r

for any real

s and

r.

Also,

are arbitrary

has the form

e}.

Let
OJ

a)

a) ( r - e ^

a)

el
,

be positive, r e a l , and less than

e.

Since

i t is immediate that the lower terms are below

and the upper terms above

w (r+e^.

Hence the result follows

by the cofinality theorem.

The above proof is a good typical example of reasoning with
orders of magnitude and c o f i n a l i t y .
normal form i t s
Lemma 5.3.

This technique helps to give the

tractability.

a.
a.
a
a.
a
I w ir-j = { I u Vi+w a ( r a - e ) } | { I w Vi+ca a ( r a + e ) } .
i
i
i

Proof. Assume first that a is a limit ordinal. Then since

a.
a.
a
a.
1 w ' n = I a) 'r-| + a) a r a we can obtain a representation of J u 1r-\
i<a
i<a
ix
by using the definition for the f i r s t addend and lemma 5.2 for the second
a

addend.
and
i

2. u> 1r1- + u> a ( r a - e ) .

J u i ^ - w e + <D a r a where 3 < a

i<e
The l a t t e r terms are clearly cofinal by the

Typical lower terms are

NORMAL FORM

65

lexicographical order. Since the argument is similar for the upper terms
terms, the result follows in this case.
In general, every ordinal a has the form a'+n where a'
is a limit ordinal. We can now use induction on n. For convenience of
notation we can assume the result for
I

I u> ir-j and then prove it for

i<a
irj. The proof is now similar to the above. The only difference

is that the typical lower terms which are discarded because of cofinality
a.
a
now have the form
J u V j - w a e by the inductive hypothesis; the
i<a
upper terms are similar.
It is convenient to extend the definition of I u ir-f to
i<a
the case where r-j may take the value 0. In fact, we use exactly the
same definition, but we of course no longer have unique representation.
Lemma 5.4. Let r-j be a sequence of length a, and let {nj} for i < 3
be the subsequence of i's such that r-\ * 0. Furthermore, suppose
t>i = a n 1.
Proof.

and

si = r n ..
!

Then

a.
b.
i u ] H = I ul s 1 i<a
i<3

Although this appears to be completely t r i v i a l , a proof is needed

because of the special definition of infinite summation.

must show that including terms with

Essentially we

r-j = 0 does not affect sums.

We do

this inductively on the length of the partial sums of

J u> ir-j. For a

i<a
non-limit ordinal this is clear since we are dealing with ordinary
addition.
sion

For limit ordinals caution is required since in the expres-

I a) "ir-,*, the

B^ term plays a role even i f

r$ = 0, since i t

i<a
leads to elements in

FUG

such as

J m V j + ew P. However, by the
i<3
cofinality theorem we s t i l l obtain the same element. The trickiest case
which occurs is the following:
integers

n.

a = 3+w, r$ * 0 and rg+n = 0 for all

We want to show that

i
side is defined in terms of an
an ordinary sum.

F and

a> 1#rj = J to i r j .

The left-hand

i
G whereas the right-hand side is

A typical lower term in the definition of

u> ir-j

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

66

a.
Y
2. w l r i " w e W 1 ' t n y < &+a)> a n d there is a similar expression for a
i<T
typical upper term. By lemma 5.3 the right-hand side is

is

{ J w
i<3

1#

rj - a) &e}|{ I w 1#r-j + u &e}.

i<3

Hence by the c o f i n a l i t y theorem the

right-hand side equals the left-hand side.

The case j u s t considered represents a t r a n s i t i o n where we
invoked lemma 5 . 3 .

In a l l other cases c o f i n a l i t y is a l l that is needed.

I f we are given two surreal numbers the above lemma permits

a.
a.
us to w r i t e them in the form
i u> "*r-f and
I co 1 s-f, using the same a
i<a
i<a
and a-j's, by i n s e r t i n g zeros where needed.
I u> ir-j = ][ w * r i
i<a+3
i<a

Lemma 5.5 (the associative law).

Proof.

We use induction on

3.

If

I o>
j<3

ra+j.

i s a non-limit ordinal this is

j u s t an instance of the ordinary associative law.

If

is a l i m i t

ordinal we compute the right-hand side using the representations in the

d e f i n i t i o n for the r i g h t addend, whereas for the l e f t addend we use the
representation in the d e f i n i t i o n or in Lemma 5.3 depending on whether
i s a l i m i t or n o n - l i m i t o r d i n a l .

If

a+J

I u>
r a + j = F|G
j<3
c o f i n a l i t y the right-hand side may be expressed as

then by

{ I a) * r j + F } | { I to fr-f+G}.
(We invoke the lexicographical order and
i <a
i <a
reason as in the proof of lemma 5 . 3 . Thus a typical lower term has the

form

I u) ir-j + ( I u
i<
j<y

a+Jr

a+j

- ew a + Y )

for

y < 3 which by the

I w ir-j - ewa Y. A typical upper term can

inductive hypothesis is
i

be written similarly. But this is cofinal in the representation for

a.
I u) Tr-f so the proof is completed.
i<a+3
The above lemmas show that in spite of the apparently
artificial definition of infinite sums they behave in many ways in a
manner expected of sums. We now come to the important fact that formal
polynomial addition works.

NORMAL FORM

67

a.

Proof.

a.

a.

J u ] r j + I u> "" s-f = J u Mrj+s-j).

i <a
i <a
i <a

Theorem 5.7.

F i r s t note that lemma 5.4 allows us to express the fact that

formal polynomial addition works in this convenient form.

use induction on a.

If

a = 3+1 this is immediate.

As usual we

In fact,

X a) ">r-f + I a) 1#Sj = I u ^r-i + to &rg + w *isj + a> &S3. By the

i<a
i<a
i<3
i<3
inductive hypothesis this is

a.
a
a
a.
I a) Mr-j+sj) + a) &r$ + ca &S3 = I w "(r-f+s-j) using the ordinary
i<3
i<a
distributive law and the definition of summation.
Now suppose a is a limit ordinal. One typical lower
element of the sum is ( J u ir-j - u &e) + ( I w isj)
i<3
i<a
= ( J u i'rj - to ^ e ) + ( J u ">s-f +
I a) is-,*)
i<3
i<3
3<i<a

a.
(

I w
3<i<a

1s

by lemma 5 . 5 .

a
i

has the natural meaning

I
u> 3+1+1S3+1+-J.)
i<a-(3+l)

By the

I u> i(r-j+s-|) - (w e e ) +
I a> ""SJ. By
i<3
3<i<a
the lexicographical order this is mutually cofinal with

inductive hypothesis this is

I a) Mr-j+s-j) - a) Be#
i<3

By symmetry we obtain the same expression i f we

begin with a lower sum of the form

i

I to "r-j + ( J u
i

1#

s-f - w &e)9

and

we obtain a s i m i l a r expression for a typical upper sum. But this gives

us exact
us exactly
for some

a.
I u> Mr-i+s-j)

by d e f i n i t i o n , or by lemma 5 . 4 , i f

r-j+si = 0

i
We now turn to multiplication, and prove the remarkable fact

that formal polynomial multiplication works.

This i s , of course, the

main j u s t i f i c a t i o n for the exponential notation and the normal form.

F i r s t we prove a special case which can be thought of as the
i n f i n i t e distributive law.
Lemma 5.6.

b
a.
b+a.
o> [ I 00 " r- "| = J u
">r-j.
i <a
i <a

Proof. We use induction on a. I f a = 3+1 then we have

b
a.
b
a.
a
b
3.
b a
a) [ I a) i r - j ] = a) [ I a) " r-f + w & r $ ] = u> [ I OJ T r - f l + o> a> e r $ by t h e
i
i

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

68

ordinary distributive law. By the inductive hypothesis and theorem 5.4

b+a.
b+a
b+a.
this is I u Tr-j + a) prp = J u 'r-j. Of course we need the fact
i <a
i <a
that addition preserves order so that (b+a-j) is also a strictly
decreasing sequence for the above to make sense.
Now suppose a is a limit ordinal. We then compute the
product using the standard representations u> = {0,oo s}|{oo t} and
v a.

a.

a.

), u) 1r1* = ( J a) Tr-j-a) 3e} I { i a) ir-j+a) 3e}.

In order to simplify the

n o t a t i o n , i f the l a t t e r is w r i t t e n in the form d = { d ' } | { d M }

then both
a
a
d"-d and d-d 1 have the form oo 3e+c where | c | oo 3. Hence for
a
a
e < e < e they are between oo 3 e
and oo 3e
We can write the
{u> b d\u> b d'+u> bl s(d-d' h o o V - o o ^ ' t C d " ^ ) } |

product in the form

j oo d , oo d +oo

t (dd ) ,oo d oo s(d d) [

L

Since

II

oo >> oo

we obtain by elementary reasoning

with orders of magnitude that

b"
b1
b" a
b a
oo t ( d " - d ) + a) s ( d - d ' ) _> oo tu> 3e x > oo [oo H2e2)]
a) b d I I -a) b "t(d l l -d) oo b d l +oo b 's(d-d 1 ).

b
_> u [ d " - d ' ] .

Hence

Similarly

h"
h'
h
oo t ( d - d ' ) + oo s(d"-d) _> oo [ d " - d ' ] ; hence
oobd"-o)b s(d"-d) oo b d' + o)b t ( d - d ' ) .

Therefore by c o f i n a l i t y

wbd

can

d/

be

be expressed in the form

{ o o V , a) b d l +a) b l s(d-d I )H{a) b d l l ,o J b d M -a) b l s(d u -d)}.

Let

dx'

and

lower and upper elements respectively corresponding to ex for the same

b
b a
b1 a
b1 y
1
3. Then OJ (d x ' - d ) = oo [oo P U - e ^ ] > oo soo B(e 2 ) >_ oo s ( d - d ' ) . Hence
K

K'

oo d x ' > oo d'+oo

s(d-d').

Similarly

K'

w ( d ' - d ^ 1 ) > oo s ( d " - d ) .

Hence

w ^ j " u> b d 1 -u) b 's(d 11 -d).

Thus again we can simplify by cofinality and obtain

b

o) d = {u) d'}|{a) b d"}.

b
oo d

F i n a l l y we can use t h e i n d u c t i v e h y p o t h e s i s and o b t a i n t h a t

b
a.
a
b
a.
a
b+a.
b+a
G
{oo [ T oo ^r-j-oo Be]}j{co [ Y oo ^r-f+oo P e l } = { / OJ ^r-j-oo
j

b+a.
b+a
{ I oo ir-j+oo
B e } , which by d e f i n i t i o n i s

b+a.
I oo i r - j .

NORMAL FORM

69

We are now ready to consider formal polynomial multiplication.

If x = I w I'r-j and y = J u ^Sj then we define the formal product
i <a
i <3
a.+bi
to be J u 1
r-jsj.
i<a

xy

By lemma 5.1 each exponent

finitely many times and the set of all

the expression has meaning.

aj+bj

aj+bj

occurs only

is well-ordered.

Hence

(To be technical, when we consider an

J ^ l r i we are applying the lemma to the positive

i<a
elements ao-a-f. This is adequate because we are dealing with binary
products only. In more general situations we want a 0 to be 0 to
expression such as

avoid trouble.)
I t is well-known and easy to verify that with respect to
formal polynomial multiplication and addition one gets a ring.

In fact,

i t is an ordered ring with respect to the ordering we have.

a.
b.
a.+b .
( I u i r j ) ( I u> 1s1-) = J u 1 J r j s j
i <a
i <3
i <a
j<3
agrees with the formal product.

Theorem 5.8

Proof.

Again we use induction.

i . e . the product

Also we tentatively use the symbol *

for formal multiplication.

First suppose that either
Assume

a = Y+1.

a or

e is a non-limit ordinal.

The same argument applies i f

b

8 = Y+1.
b

Then
a

( i w^nH I u> iSi) = ( I <o iri+u/\Y)( I u> iSi) = ( I 1ri)( I , X i )

ix

+ u

KB

a-y

KY

i<8

1<Y

i<6

y ( I w 1 s i ) ^ t l i e ordinary distributive law. We now apply the

i<3
inductive hypothesis to the l e f t addend and lemma 5.6 to the right

addend to obtain
a +b
i
n aY+b.
f r
lr s
[ i ID i r-jsi + ]. w
Y i ) T ^ e argument is now completed by
i<Y
i<3
j<6
theorem 5.7 which tells us that formal addition works.
Now suppose that a and 3 are both limit ordinals. We
y = J i c V j , z = \ u> 1#SJ and
i<a
i<6
are lower or upper elements in the representation of y and z

simplify the notation as follows:

y ,z

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

respectively in the basic d e f i n i t i o n .

element of

yz

has the form

yz

70

Then a typical lower or upper

+ y z - y z .

(As a t times in the past

we use a u n i f i e d notation since the four kinds of terms involved are

d e a l t with s i m i l a r l y . )
yxz

+ y xz - y xz

By the inductive hypothesis this may be w r i t t e n

= (yxz) - (y-y ) x ( z - z )

by ordinary algebra.

(Recall that we already know that surreal addition agrees with formal
addition.)
if

Recall now that we get a lower element for

and

yz

are on the same side.

Now

y-y

u> Y e x + cx

has the form

for

some

av

p o s i t i v e real
6

a) e

+ c

|c |

e1

and some

for some

a)

|c 1 1

6 < 3, some

(y-y)x(z-z0)

Therefore

i f and only

co

Similarly

positive real e

has the form

y < a, some

6e

e
1

z-z

has the form

and some |c |
+

S*

wnere

By the sign rule for m u l t i p l i c a t i o n for lower elements

we have a plus in f r o n t of

u>

and for upper elements a minus.

aY+a

mutual c o f i n a l i t y we can now write

yz = (yxz-w

< r

are two real numbers then

u> r.

+ c

series for

yxz.

v a l i d even i f
form

aY+a<5

yxz

ay

< to r

$e).

| c 1 1 , | c 2 | u>
+ c .)
yxz.

Now

is a typical exponent in the

This again follows by c o f i n a l i t y .

has a l a s t term.

e)|(yxz+w

We must now show that the right-hand side is

a
a
yxz = (yxz-o) ye)|(yxz+w y e ) , where

By

aY+a

(Mutual c o f i n a l i t y follows from the observation that i f

and

u> 5 .

By lemma 5.3 this is

Now every exponent in

yxz

has the

(though the converse is not necessarily v a l i d because of the

possibi 1 i t y of c a n c e l l a t i o n ) .

Hence by c o f i n a l i t y

r i g h t - h a n d side in the representation of

yz, i.e.

yxz

does equal the

yxz = yz.

This

completes the

proof.

Remark.

that in view of the above remark about the converse we do

Note

not have mutual c o f i n a l i t y .

is t r i v i a l l y

Fortunately, since the required inequality

s a t i s f i e d , we don't need i t .

At any r a t e , although very

often i t makes no d i f f e r e n c e , in general we must be careful as to which

c o f i n a l i t y theorem is being used.

For example, in the f i r s t part of the

proof, we need mutual c o f i n a l i t y since otherwise we would require an

i n e q u a l i t y which is not a t a l l obvious.

NORMAL FORM

71

Theorems 5.7 and 5.8 give us a powerful tool for dealing with
surreal numbers. In fact, for many purposes we can simply work with
these generalized power series and ignore what surreal numbers are in the
first place. This is an example of the whole spirit of abstraction in
mathematics. However, there are limits to what can be accomplished by
general power series methods since the surreal numbers are somewhat
special. Note, for example, that the class of exponents is precisely the
class of all surreal numbers, which in itself is unusual.
Let us see what the power series methods accomplish. First,
we have an alternative way of dealing with inverses and square roots
which is much easier than the direct method used in chapter three. Let
us consider, for example, the inverse. The essential idea is as follows
follows.

Let

x = cj r 0 ( J u I'SJ) where bj = a-j-a0 and s-j = -7-.

r
i <a
o
a
-a
Since the inverse of w is u> 0 and since r 0 , of course, has an
inverse, it suffices to find the inverse of expressions of the form
I a) I'SJ where b 0 = 0 and s 0 = 1, i.e. of series which begin with 1.
i<a
In fact, if 1 + I co ir-j is a series beginning with 1, we get the
i<a
inverse by formally substituting
I u> ir-j for x in l-x+x 2 -x 3 .
i<a
First, by lemma 5.1 this leads to a series which has meaning so that we
obtain a surreal number. Then, theorems 5.7 and 5.8 guarantee that this
is the inverse of 1+x since (l-x+x 2 -x 2 ...)(l+x) = 1 for ordinary
formal series.
There is another method of using generalized power series to
obtain existence results which does not depend on familiarity with identities for ordinary formal series. We shall apply this method to show
that every positive surreal number has an n* n root for any integer n.
The same method can also be used to prove the existence of inverses. It
is a generalization of the well-known procedure for ordinary formal power
00

I a-jx1 where the coefficients of the various powers of x are

1=0
obtained recursively. I like this method because of its elementary selfcontained algebraic nature. We avoid any use of analysis and in
particular the binomial theorem for fractional exponents.
series

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Every positive surreal number has an n t h root for every

Theorem 5 . 9 .

positive integer
Proof.

72

n.

J. u

Let

form

GO r o [ l +

GO / n

a.l r

&e

surreal number.

I u> i s j " | .
0<i<a

Now GO r 0

This can be expressed in the

n t h r o o t , namely

has an

/ r ^ , from theorem 5.4 and the f a c t that

r0

is positive.

Hence

i t suffices to consider series which begin with 1 .

a.
Consider a series 1 + I GO i r j . We shai 1 express i t in the
(1+ I GO i y j ) n

form

by determining

ordinary power series

the i n t e g e r s .
example, 3

a = 3 = GO and the

terms

aj's

inductively.
and x j ' s

(For

are simply

In our case the s i t u a t i o n is s l i g h t l y t r i c k i e r .

might be d i f f e r e n t from

GO

where

For

a.)

(1+ I GO i y j ) n

Suppose that
all

XJ and y j

z _> XJ f o r some

agrees with

1 +

I GO i r j

for

i < Y , but

(1+ I GO i y j )
* 1 + I GO i r j . (Recall that in our generalized power
i <Y
i <<*
series the exponents are decreasing.) Then we claim that there exists
x

and y such that x < XJ for a l l i < y

x.
x n
a.
(1+ 2, w 1yj+to y )
agrees with 1 + I GO i r j
i <Y
i <a

z >_ x.
aj

Furthermore, i f a l l

f o r a l l terms

GO

where

XJ are f i n i t e l i n e a r combinations of the

with i n t e g r a l c o e f f i c i e n t s , then so i s

not simply the same as

and that

aj

x.

(The f a c t that

xj

is

makes the process t r i c k i e r than the one for

ordinary power s e r i e s . )
In f a c t , l e t
X

be the f i r s t exponent for which the coefA

(1+ I GO i y j ) n and 1+ I GO i r j
d i f f e r . Then x < XJ
i <Y
1<a
i , and the respective coefficients, s and t , satisfy s * t .

ficients of
for a l l

in

Note that s or t may be 0. Now consider an expression of the form

x.
x n
a.
2
(1+ ). w !yj+oj y) . This agrees with 1+ J. GO i r j
for a l l terms G
O
where

z >^ Xj

for some

sible disagreement is
s * t

G
O

i.

The earliest term for which there is posand, in fact, its coefficient is

there exists a non-zero

does not concern us.)

satisfying

s+ny = t .

With the above values for

and

s+ny.

Since

(Uniqueness
y

the claim is

NORMAL FORM

73

clearly satisfied. Since x is either of the form a-j or a sum of

x-j's, the second condition is clearly satisfied.
We now assume that 1 + J w 1#rj does not have an n*n root
i<a
and obtain a contradiction. We use the claim above to define a sequence
( x iyi) inductively where i runs through the class of ordinals by
letting (xyyy) be the pair (x,y) obtained above. Since later terms
have no effect on the coefficients of earlier exponents the induction
works. However, since each xi is a finite sum of a-j's, the
collection of possible X J ' S is a set so that eventually the sequence
x-| must terminate. This contradiction proves the theorem.
Remark. It is interesting to compare this with a classical situation in
which one is interested in power series which permit fractional exponents
but only series of length u>. In that case one has the burden of showing
that the sequence of exponents approaches . Fortunately, we do not
have this problem. For example, consider
1 + GO l + uT^ + uT^" 1 ... +a)~(1)""n ... . This is a series of length GO.
If we compute the square root using the proof of theorem 5.9 we begin
1 -1
1 -1
1 "2
with 1 + -pU
and then obtain 1 + -sw - -gw , etc. It is clear that
we would need a sequence of length greater than GO. (For example, it
it would take us "forever and a day" to reach the GOW term!) However,
the proof shows that we must eventually terminate at some ordinal.
D APPLICATION TO REAL CLOSURE
We shall use the same technique as in the previous section to
show that the class of surreal numbers forms a real closed f i e l d .
Specifically, we adapt the classical Hensel's lemma argument to our
transfinite series.
n
f(x) = xn + I hix11"1 be
i=l
n in the surreal numbers where hj has the form

Lemma 5.7 (Variation on Hensel's lemma).

a polynomial of degree
r-j+dj

with

r-f

real and

d-j

Let

infinitesimal.

(Thus all terms in the

series expansion of h-j have non-positive exponents.) Suppose, furthern

more, that xn + 1 r-jx11 1 factors into two relative prime polynomials

i=l
n

Po

and Q o . Then

x + I hjx "
il

factors into two polynomials, P and

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

74

Q, where P and Q have the same degrees as P o and Q o respectively

and the first terms of the series expansions of the coefficients of P
and Q are the same as the coefficients of P o and Q o respectively.
Proof.

First, by regrouping we regard the polynomial f(x) as a series

a.
of the form
i w TSJ where s-f is a polynomial over the reals of
i<a
degree at most n-1 for i > 0. Since a finite union of well-ordered
sets is well-ordered, the a-j's are well ordered. By hypothesis ao = 0
n
and s 0 = x n + I n x " " 1 .
i=l
Let the degrees of P o and Q o be r and s respectively
so r+s = n. We now extend P o and Q o in an inductive manner similar
( u 1#P-j)( I w "iQ-j)
i<3
i<3
agrees with f(x) for all exponents y such that y _> bj for some
i < B where the Pi's and Qi's are polynomials of degrees at most
to that in our construction of nth roots. Suppose

r-1

and s-1 respectively for i > 0, but ( J u " P-f)( I u " Q-f) * f(x).
i<3
i<3
We shall find b$ such that a$ < a-f for all i < 3 and
polynomials P3 and Q3 of degrees at most r-1 and s-1 respectively
b
b
b.
b
( I a) iPj+w 3p B )( I a) TQi+o) % )
i<3
i<3
exponents y such that y _> b^.
so that

Let
of
all

b e be the first exponent

agrees with

f(x) for all

x for which the coefficients

u>x in ( I a) iPj)( I a) iQ-f) and f(x) differ. Then

i<3
i<3
i.

bg < bf for

Now consider the series

I w ip-j+o) ^G and
I w iQ-f+w ^H
1 <3
i <3
where G and H are polynomials to be determined l a t e r . Then
b
b
b.
b
( I a) Pi+w G G ) ( I w TQ^+U e H) agrees with f ( x ) for a l l terms up to
i
i

a) ^. The condition for agreement for the coefficients of a> ^ is an

equation of the form HP 0 + GQ 0 = S for some polynomial S of degree at
most n-1 because of the bounds on the degrees of P-j and Q-,*. Since
P o and Qo are relatively prime there exists G of degree at most rand H of degree at most s-1 satisfying the above equation. Let

NORMAL FORM

75

P3 = G and Q$ = H. Also, as in the case of n t n roots, if the b-j's

are finite sums of the a-j then so is b$. The rest of the argument is
identical to the argument for n^n roots.
We regroup at the end so that we end up with monic polynomials of degree r and s over the surreals. (The justification for
regrouping is the same as for ordinary polynomials in two variables. Of
course the existence of bounds for the degrees of the polynomials is
crucial for the regrouping to make sense.)
We are now ready for the main result of this section.
Theorem 5.10. Every polynomial equation of odd degree with surreal
coefficients has a root. Furthermore the exponents which occur in the
series expansion of the roots are rational linear combinations of the
exponents which occur in the series expansions of the coefficients of the
polynomial.
Proof. Let P(x) = b Q x n + ^x11""1 + b 2 x n " 2 ... b n be a polynomial of odd
degree. We may assume that b = 1 and b L = 0 by making the substin
b
n
tution x = y - The polynomial now has the form x + I a-jx11"1.
n
i=2
Now suppose that the normal form of aj begins with u>Clrj. Assume that
the polynomial is not simply

n c
n
i
x . Let c = max .
1=2

stitution

x = yoo0.

f i c i e n t of
c

(yw C ) n

The equation becomes

which can be w r i t t e n in the form

begins with

(u>

yn +
r-j)o)

We now make the sub-

n
+ Y a-jfyu)0)11""1 = 0,
i=2

I ajaT^y 11 "" 1 = 0.
i=2
.

By choice of

The coefc

we have

i
<_ c with equality for at least one i, i.e. ci - ic 0. Thus all
coefficients begin with terms with non-positive exponents and at least
one term begins with exponent 0.
If an odd degree polynomial is factored into irreducible
factors at least one of its factors must have odd degree. Hence to prove
the theorem it is enough to show that an irreducible polynomial of odd
degree must have degree one. If we apply the above construction to an
irreducible polynomial the polynomial remains irreducible. Hence by the

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

76

contrapositive of lemma 5.7 the real part of the polynomial does not have
two relatively prime factors, i.e. has the form (x-a) or (x +bx+c) .
Since the degree is odd the latter possibility is ruled out. Hence the
real part of the polynomial has the form (x-a) n . Since the coefficient
of x11"1 is 0, it follows that a = 0. Therefore the real part of the
polynomial has the form x . This contradicts the fact that at least one
term besides x11 begins with exponent 0.
Since the construction leads to a contradiction, the polynomial itself must be x11. (We are not talking about the real part.)
Since the polynomial is irreducible n must be 1.
The last part of the theorem follows from the same proof.
For this purpose we restrict ourselves to surreal numbers whose exponents
are of the form referred to in the statement of the theorem.
E SIGN SEQUENCE
Our aim in this section is to obtain a formula which
expresses the sign sequence for I u> ">r-f in terms of the sign sequences
for a-j and r-j.
It is natural to look first for the sign sequence for w .
However, in order to carry through an induction we need to know the sign
sequence for certain special finite sums along the way. Thus caution is
required with the induction in order to avoid circular reasoning.
Specifically, we deal with finite sums of the form
w 1#r-f,
i<n
where aj+i is an initial segment of aj for all i and rj is either
an integer or a dyadic fraction with numerator 1. (It is understood
that the aj's are strictly decreasing, since we are working with normal
forms.)
We first need some lemmas which are roughly variations on
lemmas 5.2 and 5.3. In proving these lemmas we used the fact that
r = {r-e}|{r+e}, which in the case where r is dyadic involves throwing
out information. (This representation is cofinal but not mutually
cofinal with the standard representation of r.) We now see what happens
if we do not throw out information. In order to cut down on duplication
later we shall deal with a general dyadic although for our immediate
purpose we need only integers and dyadic fractions with numerator one.

NORMAL FORM

77

Recall t h a t by applying c o f i n a l i t y to the canonical represent a t i o n of a dyadic f r a c t i o n

r

we obtained

i s not an integer and i n the form

integer.

{r-l}|<j>

For the special case where

i n the form { s } | { t }
if

r = with

if

i s a positive
n > 0 we have

Lemma 5 . 8 . ( a ) I f

r = {r'}|{r"}

and the set of lower elements of a

is non-empty, then

o> r = {w r'+w n}|{o> r"-u> n} where

a 1 as usual

is a t y p i c a l lower element i n the canonical representation of a and

n

i s an a r b i t r a r y positive i n t e g e r .

I f the s e t i s empty then

A = {a)V}|{u)V}.
b) If n is a positive integer greater than one then
a

a1

a"

a) n = {u> ( n - 1 ) + a ) m}|{a)
empty, where

e} i f the set of lower elements of a i s nona 1 i s as before, a"

m i s an a r b i t r a r y positive i n t e g e r ,

is as usual a t y p i c a l upper element in the canonical representation of

a, and e i s an a r b i t r a r y positive dyadic f r a c t i o n with numerator 1 .
I f the s e t i s empty then
Proof, ( a ) We compute

o>an
o> r

as in the proof of lemma 5 . 2 .

a1

Again

{o> a (n-l)} | {o>a e } .

is

u = {0,sw

a"

} | { t w } . Hence

u> r i s

{ A 1 , A ' + U c / ' M r - r 1 ) , AMtu)al')(rll-r)}|{cA11, A ' M s c /

a) r 1 + (to)

Mr-r1)}.

)(rll-r)9

By c o f i n a l i t y we can eliminate the t h i r d terms

among the upper and lower elements and replace terms such as
a
a'
a
a'
w r 1 + (soo )(r-ri)

by

w r 1 + w m.

Also i f the set of

empty we can eliminate the f i r s t terms by c o f i n a l i t y .

a1

i s non-

Thus we get the

desired form,
(b)

In t h i s case there i s no r".

The computation is the same except

that now we need the t h i r d term of the upper elements since the other
a"
a
terms are not present. Since u> >> o> , by c o f i n a l i t y this term may be
a"
replaced by u> e.
Note that negative integers can be handled by sign r e v e r s a l .
Also, r

= 0

f o r dyadic fractions with numerator one so that the

formulas s i m p l i f y .
element is

u> n.

I f the set of
Otherwise

a's

i s non-empty then a typical lower

i s the only lower element.

We are now ready to consider f i n i t e sums.

A convenient

representation for n-fold sums has already been mentioned in Chapter 3.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

n
a-,* can be expressed as
il
{a1+a2...+a .'...+an>|{ax+a2...+ak"...an>

78

In fact,

1 k n.

where

We now apply this to sums of the

to earlier and use cofinality to simplify.

for all

1 j <_ n and

1#
I w 1#
rj referred
i<n
I t is understood that r-\ * 0

form

i.

a.
2. w 1r1- be a surreal number where for all i , a . , , is
n l
i<n
an i n i t i a l segment of aj less than a-j and r-,- is a dyadic fraction.
a.
Then
I w ir-j can be expressed in the form F|G where a typical element
i<n
x in G is obtained as follows:
Lemma 5.9.

(a)

If

Let

is not a positive integer let

upper element in the canonical representation of

"

be the minimum

,.

Then

I to ir-j + a) n - 1 r
" - a) T"1 m where a
' is a typical
n i
n i
1<n-l
"
"
lower element in the canonical representation of a ,
and m is an

x =

arbitrary positive integer.

( I f there is no a

'

the last term is

omitted.)
(b)

If

is a positive integer but

for at least one

i , then let

rj

is not a positive integer

be the largest index for which

a.

not a positive integer.

Then

a ;

x = I oo r1- + u r j - u>

is a typical lower element in the canonical representation of

and

m is an arbitrary positive integer.

(c)

If

is

m where

is the minimum upper element in the canonical representation of

aj

rj

a ;'

rj

rj,
aj

is a positive integer for all i and a 0 is not an ordinal

a "
x = w 6 where a 0
is a typical upper element in the canonical

then

r-j

representation of

a0

and

is a positive dyadic fraction with

numerator one.
(d)

If

then

r-j

is a positive integer for all

G is empty.

(In fact, such an

A typical element of
Proof.

and a 0

is an ordinal

is clearly an ordinal.)

is obtained similarly.

This follows easily from lemma 5.8 and the representation of

n-fold sums discussed earlier using cofinality.

Specifically, in the

NORMAL FORM

79

I w ir-j we replace u k r by x for some k where x i s

i<n
a
an element in the representation of u> Kr k given by lemma 5 . 8 . We
expression

consider

G.

F can be handled s i m i l a r l y .
In p a r t (a) the s i t u a t i o n i s analogous to that of lemma 5 . 3 ,

i . e . by the lexicographical order, the terms obtained by replacing

a
a) n - i r
by x are c o f i n a l . This gives us the r e s u l t . For part (b)
a
l e t us analyze more e x p l i c i t l y what happens when u> k r k i s replaced
by

x.

If

rfc i s not a positive integer

I ou i r i
i<n

i s replaced by

I a) ir-j + a) krfc" - w k m + J u i r j , In p a r t i c u l a r , the f i r s t term

i<k
k<i<m
a,
a l t e r e d i s the u> k term. I f r^ i s a positive integer the sum i s

a "

a"

I <a ir-j + oo k 6 + I u> ir-j. The higher term u k

is
i<k
k<i<n
Thus in either case when co kr^ is replaced by x the

replaced by
introduced.

term that is altered is at least the

u k term.
a

Now i f

k = j,

one that is altered.

For

i t follows that the

k < j

term is the f i r s t

the term that is altered is necessarily

higher (either u> k or s t i l l higher). Thus the terms obtained by altera-;

ing a) J r j are cofinal with respect to the terms obtained by altering
a
w ^r^ for k < j . (So far this may be regarded as a more detailed proof
of part (a) i f
a^"

is replaced by

is an i n i t i a l segment of

i n i t i a l segment of
a

< a

we a

aj

and a^" > a^.

ak".

k > j.

Since

a^

a^11 > a j .

Furthermore

Recall that
is an
Since

Nevertheless the inequality follows from the nature

of i n i t i a l segments.
au(a) = - .

aj(a) = - .

Now suppose

PP e a r "to have an inequality going in the wrong direction to

apply transitivity.
have of

a^

so is

n.)

Hence

Suppose
Since

a^"

a^

has length

a.

Since

is an i n i t i a l segment of

a^" > a^, we

a-i, then

a-j
Thus the terms obtained by altering u> r j
a,
the terms obtained by altering u> krk for

a^" > a j .

are cofinal with respect to

k >j.

We have shown that the upper terms may be simplified to

v

"

2, a) 'r-j +o) u r j - a)
i<j
we need a l t e r is the

j'

m + I w "'r-j,
j<i<n
j ^ 1 term.

i . e . by c o f i n a l i t y the only term

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

80

b = m a x U j 1 , ai+ ) then for m1 s u f f i c i e n t l y high

b
aj
a.
a) m1 - a) m > | I a) 1 r j | . Hence i t is clear by mutual cofinality
j<i<n
that the above expression can be simplified to
If

a.

a -i

a -j'

I co ir-j + u) J r j - co m. This completes the proof of part ( b ) .

i<k
In part (c) a l l sums obtained have the form
,, a.
a. "
a.
2, o) i r j + a) k 6 + I
a) ir-f. I t is standard by now that this is
i<k
k<i<n
a
a
mutually cofinal with
I co ir-j + OJ k 6. During our proof of part (b)
i<k
we already saw that a^" i s an i n i t i a l segment of a 0 and that

a
k " > ao Hence the above expression i s mutually cofinal with co k 6.
a "
When k = 0 t h i s is simply co 6. F i n a l l y , since every element of the
ii

form

a|<

i s also of the form

a 0 , we need only terms for which

k = 0,

thus completing the proof of part ( c ) .

Part (d) follows immediately since there are no a11. (Recall
that

a0

being an ordinal is equivalent to

a 0 consisting only of pluses

which is in turn equivalent to the non-existence of any

a0".)

I t is i n t e r e s t i n g to contrast the situations in parts (a) and

(c) with regard to c o f i n a l i t y .

In part (a) the l a s t term contributes the

cofinal part whereas the reverse is true in part ( c ) .

We are now ready to determine the formula for the sign
sequence for
ment of

co .

Let

of length

aa

be the number of pluses in the i n i t i a l seg-

a,

and l e t

a+

be the t o t a l number of pluses in

a.
Theorem 5.11.
with a plus.

(a) The sign sequence of

Then for each

(b) The sign sequence of

u>a i s as follows.

a we have a s t r i n g of

co n, where

a * i

co

We begin
pluses i f

n i s a positive integer greater

than

1 i s obtained by beginning with the sequence for u> and followa+

a i
ing i t by u> (n-1) pluses. The sign sequence of u> where n i s a
positive integer i s obtained by beginning with the sequence for
following i t by by
reversal.
unaltered.)

co n minuses.

co

and

For negative coefficients we use sign

(Note that we s t i l l count the pluses in

since

is

NORMAL FORM

81
a.
I w 1 r 1 # , where

(c) The sign sequence of

r-j

i s e i t h e r an integer or a

dyadic f r a c t i o n with numerator one, and, for a l l

i , a-j+i

i s an i n i t i a l

segment of

a j , i s obtained by juxtaposing a l l the modified sign

a.
sequences for each i , where the modified sign sequence of o> ">r-f i s
obtained as f o l l o w s :

For i = 0 we use the rule in part ( b ) . For

b.
i > 0 we apply r u l e (b) to the element u i r j , where b-j i s obtained

from

a-j

by ignoring a l l minuses.
Before proving the theorem we i l l u s t r a t e with several

examples.

First let

f i r s t term in

a = (+-).

a, which is

Then

begins with a plus.

+, gives rise to

u>

- , gives rise to

= u>
2

gether we have

1+w = u> pluses followed by w

minuses.

minuses.

ing sequences ordinal addition is what is relevant.)

since

a =j

this is

Then the

= u pluses and the

1+1

second term, which is

So a l t o -

(In juxtapos-

Incidentally,

/w" (by the law of exponents), so that this is

consistent with an example which was done in Chapter 4.

compute

Now l e t a 0 = (-+-++), ax = (-+-+) and a 2 = ( - + - ) . We

a
a i
a
a) 5 + w ^j - a) 2 3 . By rule (a) we begin with a plus, w

minuses, u> pluses, w2

(The group of

w2

minuses, oo2 pluses, and f i n a l l y

w3

pluses gets absorbed by the UJ3 pluses.)

pluses.
By rule (b)

this is followed by a>34 pluses contributed by the "5". The contribua i

tion from u ^ follows. By rule (c) this is the sequence obtained from
b i

u> ^r where bx = (++). This consists of UJ2 pluses followed by u 2

minuses. (Note that the contribution of the j
is the same for
a i

b 1

u) ^j and a) ^ j . )

Finally we have

since we would have had w-3

w-3 minuses because of sign reversal

a2

pluses i f the term

was +00 3.

The example suggests that the formula can be simplified i f we

consider blocks of pluses and minuses in a surreal number
individual signs.

In f a c t , this can and w i l l be done l a t e r .

rather than
However,

the present form is appropriate for the inductive proof.

Proof (of Theorem 5.11).
sign sequence

g(x)

We do this by induction on the length of the

obtained from

statement of the theorem.

x = I <*> * r-f

We want to prove that

by the formula in the

g(x) = x.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

First we show that

82

a.
x = I GO ""r-j,

g is one-one. Suppose

1<m
b.
1
y
I " S j , and g(x) = g(y). Assume first that a 0 = b 0 . Then g(x)
i<n
and g(y) have the same tail after discarding the initial segments
a
b
corresponding to <*> o = w o. Now a + ,a +, ..., a + is a decreasing
=

sequence of ordinals since

of

a.

< a-j and a. + 1 is an initial segment

a-j. Hence the length of the tail has the form

J ID i n-j.
i<m
Thus the length of the tail determines a-j+ and n-,*
uniquely. Furthermore, a-j+ determines a-j uniquely. This is because
a-j is obtained from ao by stopping at a plus. Finally, the signs of
the various strings determine whether r-j is an integer or a dyadic with
numerator 1 and the value of n-j determines r-j. Thus x = y.
Now we rule out the possibility that a 0 t b 0 . If neither
a 0 nor b 0 is an initial segment of the other, then clearly a discrepancy between g(x) and g(y) arises at the point where a 0 and b 0
differ. Suppose without loss of generality that a 0 is an initial
segment of b 0 , and consider the tail following the sequence for GO .
a. +
The length of the tail of g(x) has the form
J u 1 n-j which is less
than

a ++i
u)
.

i<m
The t a i l of

identical signs.

g(x)

So certainly

I t is clear that
form

begins with a string of

a + +i
to

g(x) * g(y).
g(x)

has f i n i t e length only i f

has the

o)r = r, in which case the formula is consistent with what we

already know about dyadic fractions.

Now l e t
g(x) = F|G.

in the sequence for

plus in

x = GO .

An element of
g(x).

We look at the canonical representation

F
If

is obtained by stopping just before a plus

a

consists only of minuses then the only

is at the beginning so

rises from a plus in

F = {0}.

a at some place

the i n i t i a l segment of

a r i s i n g from that plus in

of length
a.

Otherwise a plus in

a where
a.

a(a) = +.

There are

GO

pluses where

c < to

selves to values of

.
c

Let

b be

pluses in

Then the typical element of

obtained by juxtaposition of the sequence arising from

is

b with

By c o f i n a l i t y we may j u s t as well l i m i t ourof the form

GO n

for positive integers

by case (b) and the inductive hypothesis, such an element is

n.

GO (n+1).

But

NORMAL FORM

83

Similarly an element of G is obtained by stopping just before a minus

in the sequence for g(x). Let b be an initial segment of a of
length

a where

a(a) = -. Then (again using cofinality) we obtain the

typical element of G by juxtaposition of the sequence arising from b

w b n minuses.
g(x)

one.

By the inductive hypothesis this is co (jn). Therefore

jO,u) nf Nco

IJf - co = x.
2"
Next l e t x = co n, where n is a positive integer larger than

Since

g(x)

is obtained from

have the same upper elements.

g(co )

For the lower elements we may l i m i t our-

selves to the contribution of the term

on

by adding pluses only, both

by c o f i n a l i t y .

pluses to the sequence for

f i n a l i t y we assume that
a+

ordinal less than

c
and

co , where c < co (n-1). Again by coa+

b
has the form co (n-2) + co m where b is an

m is a positive integer.

through a l l the i n i t i a l segments of

a l l ordinals less than
this is exactly
hypothesis.

a+,

so

less than

( I f there are no terms

Now l e t

runs

runs through
w

But

m#

a'

this reduces to
g(x)

In

Since this is similar to the previous

g(x)

has the same lower

The upper elements are obtained by adding on

coa (n-1) + coa m minuses.

Since

co (n-1).)

are just what we need by

g(x) = co n = x.

x = u>a(-).

g(u ) = co .

co () - co m.

a'

case i t suffices to outline the argument.

l+

g(co (n-l) + w m) = co (n-l) + u m by the inductive

lemma 5.8(b) to deduce that

a,

Now as

coa (n-2) +

has the form

any case the upper and lower elements for

elements as

Thus we add

By the inductive hypothesis this is

= {0}|()

we have just what we need by

lemma 5.8(a) to deduce that g(x) = co IJ = x.

We now l e t x = I co i r - j . The argument is a straightforward
application of lemma 5.9.
tation of

g(x).

Let

g(x) = F|G

Suppose f i r s t that
i > j .

be the canonical represen-

is not a positive integer but

r-j

is a positive integer for

In order to get a set

in

G, i t suffices to consider a set of minuses in

g(x)

G1

cofinal

which is

a r b i t r a r i l y far out. We obtain this from the contribution of the term

a-j
w rj.
Thus a typical element of G is obtained by juxtaposing the
sequence obtained from the truncated sum up to the

t n

typical upper element in the canonical representation of

term with a
g(co r j ) , where

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

84

bj is obtained from aj as in the statement of part (c) of the theorem.

In other words, the typical element of G1 has the form
g( I a ir-j + a) J rj" - w J 'm) where

rj", aj 1 , and m are as in the

statement of lemma 5.9(b). This follows by the earlier part of the proof
dealing with monomials. By the inductive hypothesis this is
a.
ai
a i'
j. u) irj + a) J r j " - w J m.
If r-j is a positive integer for all i, then g(x) is
a
obtained from g(o> o) by adding on pluses only. Hence g(x) and
a
a "
g(w o) have the same upper elements, namely u o e,
A similar argument applies to the lower elements. In all
cases the upper and lower elements obtained for g(x) are just what we
need by lemma 5.9 to deduce that g(x) = x. (For convenience we unified
various cases. For example, in lemma 5.9 part (a) may be regarded as a
special case of part (b) and part (d) of part (c). There is a pedagogical advantage in separating cases at the beginning for the sake of
concreteness, but at a later stage it is repetitious and tedious.)
We are now ready to determine the sign sequence of a general
sum
I u) i*rj. First, we define what we mean by a reduced sequence a$'
of

a$, where

3 < a.

The reduced sequence

a$ of

a3

discarding the following minuses occurring in

I

if

II

if

a$(6) = -

and there exists

is obtained from

a$ by

ae:

y < 3 such that

(Vx <^ 6)[a Y (x) = a g ( x ) ] , then the 6th minus is discarded

3 is a non-limit ordinal,

an i n i t i a l segment, and r

has

followed by a minus as

is not dyadic, then the last minus is

discarded.

For example, if a 0 = (+-++) and a x = (+-+-) then

i.e. the first minus is discarded but not the second one.
If a 0 = (+++) and a = (+++-) then a = (+++) if r 0
is not dyadic but a ' = a if r 0 is a dyadic fraction.

a^ 0 = (++-),

Roughly speaking, we ignore minuses which occur earlier;

however, the second part gives a special situation where even a new
minus is ignored.

NORMAL FORM
Theorem 5 . 1 2 .

(a)

85

The sign sequence f o r

o b t a i n e d by j u x t a p o s i n g the sequence f o r
from

o>ar
u

f o r p o s i t i v e real

is

w i t h the sequence obtained

by omitting the f i r s t plus and repeating each sign in

r,

times.
(b)

If

is negative the sign sequence in (a) is reversed.

(c)

The sign sequence for

I w irj
i<a

is obtained by juxtaposition of
0

the sign sequence for the successive

reduced sequence of a-j.
Remarks.

a.
o
u> " rT- where a-j

is the

Theorem 5.11 is, of course, a special case of theorem 5.12. In

fact, recall that in theorem 5.11 we were interested primarily in the

sign sequence for w

and therefore used only those sums which were

needed for the induction.

We illustrate theorem 5.12(a) with a simple example.
2

Consider

u {^-).

Here

a = (+-) and r = (+-+).

We already saw

earlier that GO2 gives rise to w pluses followed by w 2 minuses.

Since a + = 1 the contribution from r is w minuses followed by u>
pluses.
As a simple example of theorem 5.12(c) consider
= (+).
8

Since -x = ( +-)
2

we ignore the f i r s t minus in determining

k

the contribution of the term m8.

This therefore becomes

a> pluses

a>22 minuses.

followed by
Proof.

L
I
u>2 + w8.

We f i r s t consider the case where the surreal number has the form

I u) "ir-i with r-j arbitrary dyadic. This is a s l i g h t generalization of

i<n
theorem 5.11. The proof that g is one-one extends immediately to this
case.

For example, the signs of the strings s t i l l determine

uniquely.

r-\

Recall also that lemmas 5.8 and 5.9 deal with general dyadic

coefficients.

This gives us a head s t a r t in imitating the proof of

theorem 5.11.
The subcase where x = war with r a positive dyadic
f r a c t i o n but neither an integer nor a dyadic fraction with numerator one
i s similar to the cases
proof of theorem 5.11.

x = wan

and

In f a c t , l e t

x = ^("fj")

dealt with in the

n

r = {r'}|{r }.

Note that

r'

is

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

the i n i t i a l segment of
and

r"

86

obtained by stopping just before the last plus

by stopping j u s t before the last minus.

By hypothesis

begins with a plus following which there is at least one plus and one
minus.

Hence in the canonical representation of

g(x)

we obtain co-

f i n a l i t y on both sides by l i m i t i n g ourselves to the contribution of the

term

"r".

By further use of c o f i n a l i t y a typical upper element is

obtained by adding on

m minuses to

g(w r")

and similarly for

typical lower elements. By the inductive hypothesis such a typical upper

a
a'
u> r'-co m. This and a similar result for lower elements

element is

gives us what we need by lemma 5.8(a) to deduce that

g(x) = w r = x.

For f i n i t e sums the proof is identical to that of theorem

5.11(c) since no use is made there of the assumption that the dyadics

r-j

have numerator one.

Next, we consider the case
The l a s t part of the proof that
length, no longer works.
length

x = co r
g(w2)

For example,

by the formula.

where

is not dyadic.

is one-one, which depends only on

and

g(oj/7)

have the same

(Of course, i f one is interested, the proof

can be extended to the present case by noting that even i f the length of
a + +i
the t a i l of g(x) is u)
the signs are necessarily not a l l alike
so

g(x) * g(y).

On the other hand, this is not too important now since

i t is no longer urgent to know in advance that

Let
a

5.2

oj r

r = R1|R"

is one-one.)

be the canonical representation.

may be expressed as

aj R'|to R"

since

By lemma

is not dyadic.

(Note

the crucial s i m p l i f i c a t i o n for non-dyadic coefficients where we get by

with lemma 5.2 rather than lemma 5.8(a)).
the canonical representation of

g(x).

Consider the lower elements in

Since

does not have a last

plus we obtain a cofinal subset by taking only those i n i t i a l segments

which stop just before a string of pluses which correspond to a plus in
r.

But this is precisely of the form

r e s u l t for dyadic coefficients.

the form

Uar").

Hence

Similarly a typical upper element has

g(x) = codr'|coar" = coar = x.

The fact that i t is possible for

of

so f a r .

and s t i l l have

= w r 1 since we have the

g(w r ' )

n(x) > n(y)

to be an i n i t i a l segment

has helped to complicate the proof

We were unable to use induction on

l i k e the natural type of induction to use.

n(x)

which a p r i o r i seems

Fortunately, for the rest of

the proof we can use induction on a quantity closely related to

n(x)

NORMAL FORM

87

instead of the length of g(x).

We now define rni{x) (the reduced normal length of g(x).
Let

(a)
(b)

I oj ir-j be the normal form of x. First we define h(i)

i<a
i < a.
If r-j is not dyadic then h(i) = i+1 providing i+1 < a.
If r-j is dyadic then h(i) is the least j exceeding i such
that either rj is not dyadic or 3k(i k < j) and aj is not
an initial segment of a^) providing j < a ] , h(i) may be
undefined for some i.
We now obtain a subsequence dj of a as follows:
d 0 = 0.
d.+1=h(d.).

(c)

If 6 is a limit ordinal then

for
(a)
(b)

d$ = lim dy providing lim dy < a.

Y<3
Y<3
Finally, rn(x) is the ordinal type of the sequence {dj} (i.e the
least i:dj is undefined).
Note that the definition says that if h(i) = j then for
i <_ k < k+1 < j, a.
is an initial segment of a. . Hence necessarily
j = i+n for a finite n. Furthermore
I oi kr^ is a sum of the kind
we considered earlier since again by the definition all r^ are necessarily dyadic if the sum does not reduce to a monomial. Thus, informally, we obtain the reduced normal length of x by regarding all such
finite blocks in the expansion of x as a single term.
We now need a lemma which bears the same relation to lemma
5.9 that lemma 5.3 bears to lemma 5.2. Let x have the form
where for all

i _> a, a-j+i

dyadic.

I.e.

I u> ir-j
i <a+n
is an initial segment of a-,- and r-j is
a.

has the form

where

is a f i n i t e sum

a.
rn(x) _< rn( I a> ">r-j) + 1.
i<a
a.
Let x have the normal form
I co ir-j and express x as
i<a+n

of the kind considered earlier.

Lemma 5.10.

I u> "r-j + y
i<a

Note that

a.
I a) Tr-j + y. Suppose that y is a surreal number which satisfies the
i<a
hypothesis of lemma 5.9. Then x can be expressed in the form F|G

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

where a typical element
(1)

If y

x" of G is obtained as follows:

satisfies case (a) or (b) of lemma 5.9 then

x" = I w "rj + y" where

(2)

88

y" is as in the lemma.

If y satisfies case (c) and (3aa")(Yi(i<a + a a " < a i )

..
v ai
"
= J. w l r i + w 6

then

x
(3)

If y

satisfies case (c) and no such

a a " exists or if y

satisfies case (d) (in which case there certainly does not exist an
a a ")

then
a

x" is a typical upper element in the representation of

I a) ir-j

as given by the d e f i n i t i o n i f a

by lemma 5.3 i f

A typical element of
Proof of lemma.
where

z"

is a l i m i t ordinal and

is a non-limit ordinal.
F is obtained s i m i l a r l y .

Typical elements

x"

are

is a typical upper element of

I u> ir-j + y"

a.
I w ">r-f.

and

z"+y

In cases (1) and (2)

terms of the former type are clearly cofinal by the lexicographical order
thus proving the lemma in these cases.

In case (3) consider a typical

a .

element of the former type.

This has the form

a "

I OJ 1 r-f + w a e.

Then by d e f i n i t i o n o f case ( 3 ) ( 3 j ) ( j < a A a j a a " ) , Now

a.
a.
a.
a"
I a) nr-j + OJ Je 1 is clearly less than
I a> ">r-f + u> a e for e1 < e by
i<j
i<a
the lexicographical order. This guarantees that terms of the type
z" + y

are cofinal in this case.

As usual, by cofinality such terms

may be replaced by z". This completes the proof of the lemma.

The distinction between case (2) and case (3) can be
expressed in terms of the sign sequences of the a-j's.
(3) is characterized by the condition (Va a n )Paj)(j<a
the

Recall that case

aja a ").

Since

a-j's are decreasing and since the a a " are initial segments

corresponding to minuses in a a , an aj corresponds to a a " by the above

condition precisely if it either equals
minus as an initial segment.
satisfying

a a " or has a a M

followed by a

Furthermore, the existence of an aj

aj < a a " for a given

a a " is precisely condition I in the

definition of reduced sequence for discarding minuses for the minus a a

corresponding to this a a ".

NORMAL FORM

89

Let us m a i n t a i n the c o n d i t i o n f o r case (3) b u t assume t h a t

for some

aa"

sponding

aj

t h e r e i s no a j
satisfies

j < aa", j

satisfying

aj = aa".

aj < aa".

Then the c o r r e -

Since there is no

must be an immediate predecessor of

aj

a,

satisfying

i.e

a = j+1.

This

is condition ( I I ) in the d e f i n i t i o n of reduced sequence except for the

lack of reference to the nature of
sponding to

aa"

ra-i.

Furthermore the minus corre-

must be the last minus in

aa.

We now have what we need for the main induction on

First suppose that

rn(x)

is a non-limit ordinal.

where f(Vi^ot)(a-f+i

is an i n i t i a l segment of

aj

Let

and

rn(x).

I OJ n r-j
i <a+n
is dyadic) or

x =

r-j

x = I u> 1#r-j + y where y is a

i<a
f i n i t e sum of the kind considered e a r l i e r . By the inductive hypothesis
(n=l

and

r-j

is non-dyadic)l.

Then

a.
a.
00
we may assume the result for J u V j . We now use induction on Jig(y ))
i
a o
where y=
I
toa i r-j. The argument is similar to the one used in
y= a<i<a+n
I
i
the proof of theorem 5.11; however, there is a complication because of
the need to consider reduced sequences.

In this connection note the

obvious fact concerning juxtaposition of sequences that i f

where

F and

G are both non-empty then

SA = SF|SG.

As before, we desire to show that

regard

g as a function of

I a i"rj.
of

aa

A = F|G

g is one-one, but now we

using the reduced sequence for fixed

This can d i f f e r from the e a r l i e r case only by the contribution

to the sign sequence since a l l minuses occuring in

i > a are automatically ignored.

aa + a a

is one-one.

aj

for

Thus i t suffices to show that

The immediate reaction may be that i t is

unreasonable to expect this but recall that

(Yi<a)(aj>a a )

so that the

apparently obvious way of getting counter-examples f a i l s .

S p e c i f i c a l l y , suppose
b

*aU) * a(J)

but

aa(i) = ba(i)

a a * ba
for

i < j .

generality the "dangerous" case occurs i f

is ignored in

aa.

Hence

lexicographical order, since

but

a a = b a .

Without loss of

a a ( j ) = minus and the minus

(33<a)(Vk<j)[a B (k) = a o ( k ) ] .
ba < bg = a$,

Assume

ba(j)

By the

is necessarily minus

regardless of whether condition I or I I holds in the d e f i n i t i o n of

reduced sequence, thus leading to a contradiction.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

a.
g( I o> 1 r-j)
i <a+n

Now the sign sequence for

90

is the juxtaposi-

[g( I u> V j J j C g t y 0 ) ] . (Note that we already know by theorem 5.11

i<a
that g(y) = y and similarly for y 1 , y , etc. but i t is convenient to
maintain the notation g(y) for consistency of notation in dealing with
tion

juxtaposition.
Suppose
is negative.
by

aa.

is positive.

Suppose f i r s t that

This corresponds to

A similar argument w i l l apply i f

g(y )

as in lemma 5.10(1).

subsets cofinal in the canonical representation of

pluses and minuses in the segment within
then

( I

y = { y ' 0 } | { y " 0 }

O H g ]

contains a minus not contributed

g(y).

We can obtain

g(x)
Now i f

by considering
y = {y'}|{y"}

and furthermore

{ [ ( I ^ ^ i ) 1 [g(y') 1}I{[g( I ^ i r , - ) 1 [g(y"0)]}

i<ct
i<a
i<a
by t r i v i a l reasoning with juxtaposition.
By the inductive hypothesis this is
a.
a.
a.
{ 1 ^ i + y'}|{ I " l r i + y"} which is I u ^ + y as desired by
i <a
i <a
i <a
lemma 5.10(1). It is worth remarking that juxtaposition works trivially
in the argument here since the minuses ignored in y depend only on the
nature of a-,- for i < a so they are the same for all y 1 and y".
Now suppose all minuses in g(y) are contributed by a a ,
i.e. by minuses in a a which are not ignored. Such a minus corresponds
to an a a " . Assume first that case 2 of lemma 5.10 holds. Then by
cofinality we may limit ourselves to such a a " and the juxtaposition
argument is identical to that of the earlier case.
The most subtle case occurs when case 2 is not satisfied. By
the remarks following the proof of lemma 5.10 this can happen only when
a has the form j+1 and the minus corresponding to a a " = aj is the
last minus in a a . Since we are dealing with a minus in a which is not
ignored, r a -i is dyadic. (This case is the most delicate with regard to
the issue of ignoring of minuses.) We obtain a subset cofinal in the
upper elements of the canonical representation of g(x) by taking
+
a
sequences of the form g( I <*> i>j) followed by </a pluses and
a) a n minuses for some integer n. This follows because first all
minuses preceding the one corresponding to a a " are ignored in y .

NORMAL FORM

91

Next i t follows from the i d e n t i t y I a)1+1 = a/* that

+
i<a
with a)
pluses.

g(y)

begins

F i n a l l y , there are no minuses following the succeeding

a ++i

a) a
minuses contributed by the minus corresponding to a a . Now
.,

a.

a.

a ;

a.

1 a) ir^ = I a) i r j + w J r j . Also the t a i l of g( J u ^ j )

contributed
i<a
i<j
i<a
a.
aj+
aj
by a) J r j followed by the OJ
pluses and OJ n minuses i s exactly
a-;
the c o n t r i b u t i o n of a term of the form OJ (rj+e)
f o r some positive
g[ I OJ ir-j + u> J ( r j + e ) ] , which i s
I OJ i r j + OJ J e by
i<j
i<a
the main inductive hypothesis (the one on r n ( x ) ) . Hence g(x) has
a.
a.
a-j
the form { I OJ V - j + y ' } | { a > i r j + GO J e } which i s x by lemma 5 . 1 0 ( 3 ) .
i<a
Now suppose g ( y ) does not c o n t a i n a minus. Then any

dyadic

to

[g(x)]n

upper element

i n the canonical r e p r e s e n t a t i o n o f g ( x )
a.
g( OJ 1 r 1 - ) , i . e i s an upper element in the
i<a

corresponds to a minus in

canonical representation of the l a t t e r .

=

{ I w
i <y

l r

i "

w Ye

lll

I u
i <y

i r

+ w Ye

"}

Now g( J u V j ) = I OJ ir-j
i <a
i <a

where

Y<a

and e ' , e " are

positive dyadic with numerator 1 .

By the inverse c o f i n a l i t y theorem for a r b i t r a r y
there e x i s t s

and e

such that

I u i r f + w e = g[ I OJ V J + OJ Y e ]
i<Y
i<Y
hypothesis.

[g(x)]" >

(g(x)) M

(Note that since

by the main inductive

i s dyadic i t i s guaranteed that

rn[ I OJ i V j + o)Yel _< rni[ I OJ i r ^ ] . )

i<Y
i<a
We must now show that
[g( >] a i r i ) ] [ g ( y 0 ) ]
i<a
Of course
=

a.

9( 1
i

w 1 r

i).

a.
Y
gf I OJ V-j + OJ e]

i s greater than

= g(x).

g| ), u i r j + OJ e] = ]i u i r j + OJ e > I OJ Tr-j
i^Y
i^Y
i<a
Thus by the lexicographical order the only d i f f i c u l t y

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

92

II w V j + wYYee. In
i<Y
a.
aY
order for this to occur rY is necessarily dyadic.
I u> " r-f + 03 'e is
i
the sequence
I u> 1#rj followed by u>Y
pluses and w Y n minuses
i<y
a.
for some integer n. Therefore the contribution of
I u V-j to
a.
a) V j consists only of pluses. Necessarily y < i + r-j is a
i<a
positive integer and y _< i - ai+i is an initial segment of a-,*. The
latter follows since a-j+i < a-j as, otherwise, a-j+i would have a minus
not occurring in aj and, a fortiori, not in any aj for j < i, and
a.
would thus contribute a minus to J u V j . It follows that a = y+n
i<a
a.
_ a. +
for some integer n. So I w "*r^ contributes
I w " r-j pluses to
i
y<i<a
I to V j where y <_ i - a-|+ > a . j + i + . Furthermore g(y) contains
i<a
only pluses. Since r a - i i s dyadic a l l minuses in a a are contained in

arises when

I a> iir j
i<a

is
i an iinii tii all segment off

aa-i;

hence the number of pluses in g(y ) i s bounded above by a number

a +
of the form o> m for some integer m. Since a a + < a a _ i + this
g( J u ir-j)g(y)
i<a
aY+
followed by less than a>
pluses.

f i n a l l y shows that

consists of

the sequence

I u> V j
i<Y

This proves the inequality.

We now have what we need by the c o f i n a l i t y theorem to deduce

that

g(x)

may be expressed in the form

{ I w * H + y ' } | { I OJ ""rj + 03 Y e } . Case (3) of lemma 5.10 applies, so

i<a
i<y
the above i s a representation of x, thus f i n a l l y g(x) = x.
This completes the induction for the case where
a non-limit o r d i n a l .
rn(x) = n(x).
rn(x)
of

(In general, i f

has the form

rn&(x)

Now suppose

rnl(x)

rn(x)

is a l i m i t ordinal.

n(x) has the form

is

Then

u>a + m, then

coa + n since the blocks used in the d e f i n i t i o n

are a l l f i n i t e . )

Thus we may assume that

has the form

I o) "Jr.: for a l i m i t ordinal a and that the result is known to be

i<a
a.
v a l i d for
jj u V j for 3 < a. We shall show that the representations
i

NORMAL FORM

93

of x given by the definition and the canonical representation of g(x)

are mutually cofinal which is enough to guarantee that g(x) = x. It
suffices to consider upper sums since the argument is similar for lower
sums.
Let y be an arbitrary upper element of g(x). Then y is
an initial segment of x determined by a minus sign contributed by a
term u r$ for 3 < a.
Consider

Since

a is a limit ordinal, 3+1 < a.

a> ir-j + w & +1 e

z=

which is an upper element of x. By

the inductive hypothesis this is

an i n i t i a l
initial

g(z)

segment which in turn contains

segment.

Hence

Now l e t
has the form

z < y

g( I o> 1 # rj) as
i<3
followed by a minus as an

which contains
y

by the lexicographical

order.

be an a r b i t r a r y upper element of

I u> "ir-j +
l<3

3e

x.

where we may assume that

Then

is a

dyadic with numerator one.

We claim f i r s t that the set of
tributes a minus to

g(x)

is cofinal in

a.

exists a Y such that the contribution of

consists only of pluses.

i.

However, since

a-j+i

con-

Otherwise suppose there

J u ir-j
Y<i<ct

to

g( J u ir-j)
i<a

As in the proof of the case where

l i m i t ordinal we obtain that

Y

for which u Y r Y

is an i n i t i a l

segment of

is a nona-,*

for

is a l i m i t ordinal this already is a

contradiction.
I t follows that
since otherwise a l l

g(x)

cannot be an i n i t i a l

contributions of

u> ir-j

for

segment of

3 < i

to

g(x)

y
would

consist only of pluses which we j u s t noted is impossible.

Hence
g(x)
aY
a) r Y .

and

g(x)

differ.

Suppose

i s defined a t the l e a s t ordinal

The sign

(g(x))(j)

= +.

[g(x)](j)

for which

is contributed from a term

By the lexicographical order this would

imply t h a t
false.
6 > y

y < g( I u ir-j) = ( J w V j )
where 6 = max(3+1,Y) which is
i<6
i<<5
Hence ( g ( x ) ) ( j ) = - .
By what was said e a r l i e r there exists

such that

a) 6 r $

also contributes a minus to

determines an upper element

of

g(x).

g(z)

g(x).

contains

This
g( J u i'rj)

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

94

as an initial segment since 6 > y. Hence by the lexicographical order

z = g(z) < y which is precisely what we need.
This finally completes the proof that g(x) = x in all
cases.
Now that we have the fundamental relation between the sign
sequence of a surreal number and its normal form, we can study the
surreal numbers in more detail. For this purpose it will be convenient
to express theorem 5.11(a) in a form which considers contributions to the
sign sequence of u> by strings of pluses and minuses. Specifically, if
a string begins at the i t n place in a then the next string begins at
the j t n place where j is the least ordinal larger than i such that

Corollary 5.1. The sign sequence of a> consists of the following

juxtaposition. We begin with a plus. For each string of pluses in a
we have a string of u> pluses where b is the total number of pluses
in a up to and including the string. For each string of minuses in a
we have a string of w c minuses where b is the total number of
pluses in a up to the string and c is the number of minuses in the
string, a is regarded as beginning with a string of pluses.
Proof. This follows immediately from theorem 5.11(a). For pluses we use
the identity
I w1 l = w a for 3 < a. For minuses a a remains fixed
i
during a string. (No minus contributes a plus!) Finally if a begins
with a minus then a may be regarded as beginning with 0 pluses giving
rise to w = 1 plus in a> . Thus the last statement is a convenient
way of unifying the cases where a begins with a plus and where a
begins with a minus. In the former case the first plus is superfluous by
absorption, so the statement gives the plus precisely when it should.
To illustrate this we refer back to the second example
following the statement of theorem 5.11. We had a 0 = (-+-++). The
string of two pluses at the end gives rise to w 3 pluses since there is
a total number of three pluses up to and including that string.
Note finally how strings of pluses and strings of minuses are
treated in entirely different ways.

95
6

LENGTHS AND SUBSYSTEMS WHICH ARE SETS

Up to now we have considered basic material which in some

form is contained in [1], At this point we begin to consider problems
that are new and more specialized. In this chapter we are interested in
information regarding upper bounds of lengths of surreal numbers obtained
by various operations. This will allow us to obtain subclasses of the
proper class of surreal numbers which are actually subsets and closed
under desirable operations. Our first result is an easy one.
Theorem 6.1.

&(a+b) (a) lib).

Proof. We use induction as usual. A typical upper or lower element in

the canonical representation of a+b has the form a+b or a+b.
Without loss of generality consider a+b. By the inductive hypothesis
(a+b) <_ (a) S U ( b ) < (a) ( b ) . The result follows by theorem 2.3.
This result is sharp. In fact, we already know that in the
special case where a and b are ordinals we obtain equality. If a
and b are not ordinals then we usually have a proper inequality. For
example, SL[1) = 1 but i[j) = 2, hence i(l) < iij) + (~).
Incidentally, the sign sequence formula of the preceding chapter makes it
comparatively routine to study the question of when equality is obtained.
We do not pursue this here since we are at present more interested in
bounds.
Theorem 6.2.

a(ab)

(a)
3

(b)#

[We

are

using

ordinary ordinal

exponentiation.]
Proof. Using induction as in the proof of theorem 6.1 we must consider
elements of the form ab + ab - ab.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

96

Since (-a) = (a) and since theorem 6.1 extends trivially

to arbitrary finite sums we have:
(ab + ab-abo) <_ (a o b) (ab) Jt(abO). Let c = maxU(a) (b),
( b ) ] . Then by the inductive hypothesis (ab) _< 3 C and
_< 3 C . Clearly (a) i(b) < c. Hence (ab<>) < 3C. Therefore Ji(ab+ab-aObO) < 3c + 3c + 3c = 3C + 1 . Furthermore
c < (a) (b), hence c+1 _< (a) @ i(b). We have shown that an
arbitrary element in the canonical representation of ab has length less
then 3*(a) ^^K
thus the result follows by theorem 2.3. [Note that
C
3
is necessarily a monomial in the expansion of ordinals in terms of
powers of o so that and + agree with addends of the form 3 C ,
hence the above computation is justified.]
It is strongly conjectured that (ab) (a) * i(b) in
analogy with the case for addition. In fact, it appears as if a routine
but extremely messy proof is possible using the sign sequence formula
(theorem 5.12). However, I feel that if the conjecture is true there
should exist an elegant proof. [In fact, the referee of the original
manuscript claims to have a proof. Since, as is pointed out later,
theorem 6.2 is adequate for the future, this is not being pursued here.]
It is possible to prove theorem 6.1 using the sign sequence formula. In
view of the simplicity of our proof this would be silly. On the other
hand, such a proof would give us detailed information on the comparison
of (a+b) and (a) (+) i(b).
Fortunately theorem 6.2 suffices for important future results
in spite of the crudeness of the bound used. For example, to begin with
we have the following corollary.
Corollary 6.1. The set of surreal numbers with length less than a fixed
e number is a subring of the class of surreal numbers.
The situation for reciprocals is more complicated. The fact
that 1(3) = 3 and i(j) = JO makes it clear that a different type of
result is needed. Specifically we shall deal with the cardinality of the
lengths.
Theorem 6.3.
a

U(~)| XoU(a)|

[i.e. unless

(a) is a finite ordinal

LENGTHS AND SUBSYSTEMS WHICH ARE SETS

97

Proof. We use the construction in chapter 3C. Let d = max( |n(a)|,X 0 ).

The collection of all elements of the form <a 1 ,a 2 ,*,a n > used in the
above construction has cardinality bounded above by d since all aj are
initial segments of a.
Our main induction will be on (a) and we shall show that
|<a x ,a 2 , a n >| d using a subsidiary induction on n. Consider
<a 1 ,a 2 ,---,a n + 1 > and let b = <a 1 ,a 2 ,-- ,a n >. Then < a 1 a 2 ' " ' a n + i >
is the unique solution of the equation (a-aj+^b + a.j+1x = 1.
|(a-j +1)| _< d. |&(b)| <_ d by the subsidiary inductive hypothesis and
U(-

)| < d by the main inductive hypothesis. Hence by theorems

i+i
"
6.1, 6.2 and elementary facts about ordinals \l(x)\ _< d. (Note that for
infinite ordinals a, |3 a | = |a|.)
Hence the upper and lower elements in the representation of
a consist of at most d numbers each of which has length of at most
cardinality d. Hence there is a single ordinal a of cardinality d
which is an upper bound to the lengths of all the upper and lower
elements. (Such an ordinal is obtained as an ordinal sum of all the
lengths since d = d.)
a

Finally l() < a+1 by theorem 2.3 and certainly

a
|o+l| = |a| d.
"
Theorem 6.3 can also be proved by using the normal form in a
manner similar to the proof we shall use for theorem 6.4. I feel that it
is worthwhile to have proofs of both kinds, i.e. those that use the
normal form and those that don't. Even if there is a known proof of one
type for a certain theorem a search for a proof of the other type can
lead to certain new insights. It is tempting to make an analogy with
synthetic and analytic proofs in geometry, with analytic proofs corresponding to proofs which use the normal form. This makes some sense
although analogies always have their limitations.
Theorem 6.4. If the cardinality of the lengths of all the coefficients
in a polynomial of odd degree is bounded above by an infinite d, then
the polynomial has a root b such that \i(b)\ <_ d.
The result will follow from several lemmas dealing with
lengths of elements in normal form. All these lemmas follow from theorem
5.12.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Lemma 6 . 1 .
Proof.

If

|&Ua)| = |i(a)|.

i s not dyadic then

Since every term in

(in fact at least

gives r i s e to at least one term in

a> terms), i t i s clear t h a t

(a) (u )

Now given any surreal number

by replacing a l l minuses in
b = it(a), i t

a, l e t

by pluses.

be obtained from a

[Although formally we can say

is preferable to keep surreal ordinals and ordinals

regarded as lengths a p a r t , especially because the algebraic

are

co

hence

| i t ( a ) | <_ |*(a> a )|.

certainly

that

98

operations

different.]
S i n c e t h e c o n s t r u c t i o n o f t h e s i g n sequence f o r

c o u n t i n g t h e number o f p l u s e s a t v a r i o u s s t a g e s , c l e a r l y
|*(a)b)| = |it(b)| = | U a ) | .

However,

the f a c t t h a t

Lemma 6.2.

Proof.

I u)

\i{^

Since

a+1

r)\

Hence

\i{>/)\

i{u

involves

) <_ (co ) .

<_ | * ( a ) | .

(Recall

= <A)

= \i[u> )|

f o r any non-zero real

i s an i n i t i a l

segment of

u> r

i f a * 0.

i t follows

that

it( w ) ((o r).

Now

u> r

i s obtained by f o l l o w i n g

each of which is obtained from

a

(w r) U ( o ) ) ] [ u ) ] .
a

has i n f i n i t e

Lemma 6.3.

If

Therefore

length.

w r

by ignoring minuses.
a

| i ( w r ) | X0\i{^)\.

Hence
If

a * 0

that t h i s is not as obvious as i t looks.

In general the d i f f i c u l t y

segment of

then

then

I f OJ r

so the r e s u l t is

is caused by the f a c t that because

of the ignoring of minuses the c o n t r i b u t i o n of

b

a> s t r i n g s

This completes the proof.

i s the f i r s t term then i t i s an i n i t i a l

of

by at most

i s one of the terms in the normal form of

Note f i r s t

immediate.

ID

to the sign sequence

i s not necessarily the sign sequence of u .

Consider the example

followed by

w+1

pluses and

followed by

u>.2

minuses.

-uT^uT2

-w" 1 + uT 2 .

u> minuses,
Thus

uT2

w"

This consists of a minus

2

consists of a plus

cannot be a subsequence of

but they both have the same length.

Nevertheless the idea of the proof is not too hard.

If a

LENGTHS AND SUBSYSTEMS WHICH ARE SETS

99

minus is ignored in the term u> r there is a corresponding minus which

makes a contribution earlier. (There is an exception which is handled
separately, namely that of the extra minus ignored.)
Proof. We show that UJ r is a subsequence of b providing the
distinction between pluses and minuses in the sign sequence is ignored.
Consider a minus occurring in a which is ignored in the contribution of
w a r to b. Then the initial segment of a consisting of the sequence up
to and including this minus occurs as an initial segment of an a a for a
a
a
a
aa
term w a r a preceding the term w r,
r or
or the
the term
term ww r has a
a
predecessor which has the form w a r a with a a the initial segment up
to but not including the minus and with r a non-dyadic. For each minus
such that the first condition holds we choose the first term satisfying
the condition. Thus corresponding to the sequence of minuses in w r
which are ignored, except possibly for the last one, we obtain a sequence
t 1 ,t 2 # **t3. If y < 6 then ty _< t<$ since the condition for agreement
is more stringent for larger y. It is not necessarily strictly
increasing. For each t in the sequence let Dt = {a:t a =t}. Then Dt
is a partition of all the minuses ignored except possibly the last one.
If s < t, y e D s and 6 e Dt then y < 6. We now obtain the following
subsequence of b by juxtaposition. From each term a> V t we extract
a
the contribution to w t of those signs such that the first minus not
preceding it corresponds to a term in Dt. By brute force this juxtaposition gives the part of the sign sequence of w r up to and including
the contribution of all minuses in a which are ignored except possibly
for the qualification mentioned above.
We still must consider the complication caused by minus signs
which are ignored although they do not occur previously. This may happen
in

a
wrrt

as well as

w r.

In this case we use the rest of the c o n t r i -

bution of the preceding term.

The lack of any contribution from the

ignored minus sign is precisely made up by the contribution from the

preceding non-dyadic c o e f f i c i e n t . '
if
to

d
a) .

is

has

signs.

pluses and
ooa

signs

This is the same as

contributed by a non-dyadic
contributing

Recall that i f

followed by a minus then the minus contributes

ooco which is the number of signs

r

to

wr

since

has

u> signs, each

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Lemma 6.4.

a
Let a = I u> <*ra. Then

100

|e| _< |*L.u.b. U(a a )u>]|.

Remark. It turns out that this lemma is not needed for the proof of the
theorem. However, I feel that it is of interest in itself and fits in
naturally in our present discussion of lengths. A similar remark applies
to lemma 6.5. What is really needed is lemma 6.5a which strengthens both
the hypothesis and conclusion of lemma 6.5. Lemma 6.5 is of interest
because of the surprising fact that the bound is sharp in spite of the
lack of any hypothesis on the length of 3.
Proof. If (a a ) is finite for all a then there can be at most w
terms so we need only consider the case where at least one a a has
infinite length.
Let d = |i.u.b. (a a )|. Then all a a terminate at d or
earlier. By a simple cardinality argument there are at most 2^ such
sequences. This is not good enough for our purpose, and in fact we shall
show that this can be strengthened using the fact that the a's are
well-ordered.
In fact, assume that there exists a well-ordered decreasing
sequence of surreal numbers (a a ) containing d + > d members where d +
is the successor of d. We show that this leads to a contradiction.
First consider the sequence a a (0). Since this can take on
only the values plus, minus, and 0 by the lexicographical order there
exists an a a such that a a (0) is fixed for a > a 0 . [E.g. if a a (0)
is minus for any a it must be minus for all larger a.] If the fixed
value is 0 we have an immediate contradiction since this says in
particular that a a = a a +x = 0. (Recall that 0 stands for
"undefined.")
We now construct a monotonic increasing sequence (aj)
defined for all i _< fc.u.b. (a a ) satisfying a a (i) = aai.(i) for
all a > aj.
Suppose otj is defined for i _< j. Since aj
is monotonic increasing we know that a a (i) = a a .{i) for all a > aj and all
i _< j. Consider the subsequence a a (j+l) where a > aj. By the lexicographical order this is monotonic decreasing. Hence the same argument
as in the case a a (0) shows that there exists an aj+ x such that
a(j+l) = a a # (j+1) for all a 2 . a - + i Note "tnat the subsequence

LENGTHS AND SUBSYSTEMS WHICH ARE SETS

101

considered still has d + members.

Now suppose j is a limit ordinal and a-j is defined for all
i < j. Since | j| _< d the set of all a-,- for i < j is bounded above
by a certain 6. By the same argument as before considering the
subsequence of a a for a_> e we can find a suitable aj.
Finally consider the case i = a.u.b. (a a ). The construction
gives us an a-j such that for all a > a-j we have a a (j) = a a .(j) for
j _< i. However, a a (j) is undefined for j > i (even for j=i). Hence
&OL = a a . for a > a.. Since all a a are distinct this is a
contradiction.
Remark: Note that the theorem refers to the cardinality of the A.u.b. of
a set of ordinals. This is not the same as the .u.b of the cardinalities of the set of ordinals, so caution is required in the statement of
the theorem. To see the distinction it suffices to consider the case
where the set consists of all the countable ordinals. Following this
a
through if we let a = I u> a where a a consists of a plus followed by
a<ojj

minuses then we see t h a t the conclusion of lemma 6.4 cannot be

expressed as

Lemma 6.5.

| 3 | _< . u . b .

|(aa)|

a
If a = \ o> a r a
a<3

then

since the r i g h t - h a n d side i s

Xo.

|(a)| <_ |.u.b. it(aa), u>|.

Proof. By lemmas 6.1 and 6.2 we know that

a
U(o) a r a ) | _< |*(a a )| X o 1 d where d = |.u.b (a a ), u|. Hence the
a
contribution of <J a r a to the sign sequence of a has at most
cardinality d since signs may be ignored but no extra signs added. By
lemma 6.4 there are at most d terms in the expansion of a. Hence
there are at most d = d signs in the sign sequence for a.
The following corollary is immediate from lemmas 6.3 and 6.5.
a
Corollary 6.2 If a = ). u a r a

Lemma 6 . 5 a .
then

If a =

|t(a)| < d.

where not all

a a are dyadic then

a
I co r a a n d i . u . b . ( | a ( B ) | , U ( a a ) | , X ) < d
a<6

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

102

Proof. This is the same as the last part of the proof of lemma 6.5.
Note that the bound here is slightly sharper than the one in lemma 6.5.
This bound we actually have to "pay for" although the early bound came
free thanks to lemma 6.4.
Lemma 6.6. If a x ,a 2 ... a n are arbitrary surreal numbers and
i> r 2 r n are rational then U(Ir-|aj)| _< |max (aj)|y 0 .

Proof. We know that for any real r i(r) _< u so |(r-j) X o for all
i. The lemma then follows from theorems 6.1 and 6.2. (Of course, all
that is used is a weakened form of these theorems which refer only to
cardinalities of lengths.)
We now have all we need to prove theorem 6.4. This may seem
strange since none of the lemmas have anything to do with polynomials of
odd degree! In fact, the proof of the theorem will not make direct
reference to such polynomials. The aspect of theorem 5.10 which is
crucial is that the exponents are rational linear combinations of the
given exponents, so that the same proof works for reciprocals.
Proof of Theorem 6.4. This follows easily from the lemmas by a kind of
back and forth argument.
Let a be an ordinal of infinite cardinal d which is an
upper bound to the lengths of all the coefficients. By lemmas 6.3 and
6.1 (used in that order) a is also an upper bound to the lengths of the
exponents occurring in the normal forms of the coefficients. Since
n(a) <_ (a) for any surreal number, a is also an upper bound to the
normal lengths of all the coefficients. By lemma 6.6 d is an upper
bound to the cardinalities of the lengths of the exponents occurring in
the normal form of the constructed root, d is also an upper bound to
the cardinality of the normal length of the constructed root. (By
elementary cardinal arithmetic, the cardinality of the set of all finite
rational linear combinations of an infinite set S has the same
cardinality as S.) By lemma 6.5a the cardinality of the length of the
constructed root is bounded above by d.
Remark. A proof using lemma 6.5 is also possible although our approach
seems simpler. For such a proof we need a strengthened form of lemma 6.6

LENGTHS AND SUBSYSTEMS WHICH ARE SETS

103

which makes fuller use of theorems 6.1 and 6.2. This will enable us to
obtain an ordinal upper bound of cardinality d to the lengths of the
exponents occurring in the normal form of the constructed root. In fact
the least e number larger than a works.
Finally, as a culmination of the results of this chapter we
have shown that the subset of surreal numbers a such that |(a)| d
for any fixed infinite cardinal d is a real closed field. Since all
operations concerned depend on only finitely many elements the condition
(a) _< d may be replaced by (a) < d. [The latter formulation gives
more fields.] These are all "honest" fields since their carriers are
sets.
Not all subfields have the above form. In fact, the two most
well-known fields found in nature, the rationals and reals, both consist
of all surreals of finite length together with some but not all surreals
of length w.
The field of all surreals of countable length should be a
worthwhile object for further study.

104
7

SUMS AS SUBSHUFFLES, UNSOLVED PROBLEMS

This chapter i s s t i l l

a t a pioneering l e v e l .

I have a strong

f e e l i n g t h a t there i s a r i g h t way of looking a t the subject which when

discovered w i l l

g r e a t l y enrich the theory of surreal numbers.

A sequence
<a

o> i> "

>

and

<c 0 , c l 5 " c^>

< b o , b 1 " *b3>

<ii,i2>m ' #1a>

sequences

and

<

if

is a s h u f f l e of the sequences
there e x i s t s t r i c t l y

JiJ2*##J6>

sucn

increasing

t n a t

(Vk _< a)(a|< = CJ ) A ( V k _< 3)(b|< = CJ )

and such t h a t every ordinal

l a r g e r than

j's

i s one of the

i's

or

but not both.

not

A subshuffle

of two sequences i s a subsequence of a s h u f f l e of the two given

sequences.
This d e f i n i t i o n i s consistent w i t h the i n t u i t i v e meaning of a
shuffle.

For o r d i n a l s i t i s known t h a t

a + 3, where

r e f e r s to

surreal a d d i t i o n , i s the l a r g e s t o r d i n a l which can be obtained as a

s h u f f l e of
"trivial"

and

8.

shuffles

However, the s h u f f l e
the
3)

u> from
gives

w2+8

For example, i f

ct+3

and

3+a

o)+u)+5+3

a = OJ+5

give

w2+3

3 = w+3
u2+5

( i . e . we f i r s t take the

and then take the r e s t of

which i s

and
and

then the

respectively.

co from

then

followed by the r e s t of

a+3.

This chapter i s devoted to the proof of the f o l l o w i n g theorem

which strengthens theorem 6 . 1 .

Theorem 7 . 1 .
of

and

For any surreal numbers

and

b,

a+b

i s a subshuffle

b.

Proof,

(a)

finite,

using i n d u c t i o n .

We do t h i s f i r s t in the case where

&(a)

and

Jt(b)

are

(This seems easier than the use of a r i t h m e t i c of

dyadic f r a c t i o n s as discussed a t the end of chapter 4B, where the strange

nature of the c a r r y i n g causes c o m p l i c a t i o n s . )

For the purpose of the

SUMS AS SUBSHUFFLES, UNSOLVED PROBLEMS

105

proof it is convenient to think of a subshuffle of a and b informally

as a sequence obtained by moving along, selecting a sign from either a
or b with the restriction that the signs from a and b must be taken
in the order they occur although signs may be omitted.
Now a+b = {a'+b,a+b'} |{a"+b,a+b11}. Suppose a+b ends with
a plus (a+b is a finite sequence). A similar argument will apply if it
ends in a minus.
Let a+b = d+. By the inverse cofinality theorem there
exists an a 1 such that a'+b j> d or there exists a b1 such that
a+b' j> d. Since both cases are similar it suffices to assume that
a'+b _> d.
Since a+b > a'+b _> d, a'+b must have d as an initial
segment. (This includes the possibility d = a'+b.) By the inductive
hypothesis a'+b and hence a fortiori d is a subshuffle of a' and
b. Now a1 is by definition a subsequence of a obtained by stopping
at a plus, i.e. this is a plus in a occurring after all the signs in
a'. Thus d+ can be obtained as a subshuffle of a and b by using
the representation of d as a subshuffle of a1 and b followed by
that plus.
As a technical detail for the future we need the following:
If a > 0, b < 0, and a+b > 0 then the subshuffle for a+b may be
chosen so that the first plus comes from a. This is not a trivial
requirement. For example, let a = (+-), b = (-++) and c = (++--+).
Then c can be expressed as a subshuffle of a and b by beginning
with the segment (++) of b. However, if one tries to begin with the
plus in a one gets stuck at the last plus in c. Of course c * a+b
since it is immediate that c > 1 and that a+b < 1.
By hypothesis a begins with a plus, b with a minus, and
a+b with a plus. In the inductive hypothesis we consider terms such as
a, a+b, etc. (Recall that a is either of the form a1 or a".)
First suppose that a > 0, and a+b > 0. Then the triple (a , b,
a +b) satisfies the technical hypothesis. Since the subshuffle used for
a+b obtained in the proof begins with the subshuffle used for a + b it
follows that the first plus in a+b is taken from a.
To complete the proof of the technical detail we must examine
the cases where terms such as a and a +b do not satisfy the hypothesis. Recall in the proof that a +b has d as an initial segment where

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

106

a+b = d followed by a plus [or d followed by a minus if a has the

form a " ] . Hence the only cases where terms such as a 0 and a+b fail
to satisfy the technical hypothesis occur when terms such as
a , b , or d are 0.
If b = 0, it certainly contains no pluses so the first
plus in a+b
is necessarily taken from a (i.e. no inductive hypothesis is needed).
If d = 0 then a+b = (+) [a+b > 0 ] . The plus can certainly
be taken from a since a is positive and thus contains at least one
plus, for example the first sign.
If a 0 = 0 then a+b = b < 0. Since a+b > 0 and both
a+b and a +b begin with d it follows that d = 0; thus we are back
in the previous case.
Of course a similar result applies if all signs are reversed.
(b) It is now easy to see that the result is valid for all real a and
b. Since the finite case is the same as the dyadic case it suffices to
assume that a is not dyadic. This case is now trivial. Since a has
both pluses and minuses arbitrary far out and a+b has length ao we can
obtain a+b from a alone; in fact no matter what we do we can't
possibly get stuck because of the availability of signs from a. In
particular, the technical detail referred to earlier can be satisfied.
(c) We now prove this for numbers of the form u> r and w s. Recall
that the sign sequence for w d r is given by a contribution from w a
followed by the contribution from r which resembles r itself except
that the first sign is ignored and each other sign from r is repeated
u> times where a is the number of pluses in a. If the first sign
were not ignored in the rule the proof would follow trivially from (b).
The subshuffle for w (r+s) could be simply obtained by selecting the
contribution of ooa from a suitable one of w a r or w a s and each block
of signs in the contribution of r+s from the corresponding place where
the corresponding sign in r+s is obtained. This is clearly a subshuffle if the contribution of u> is selected from a term whose coefcoefficient has the same sign as r+s. (At least one of r and s has
the same sign as r+s. Since the result is vacuously true when r+s = 0
we can ignore that case.)
We have to adjust the construction to take account of the
fact that the first signs of the coefficients are ignored in the sign

SUMS AS SUBSHUFFLES, UNSOLVED PROBLEMS

107

sequence rule. (Of course, we do not mean that it is totally ignored.

The first sign determines whether w a r is positive or negative and thus
affects the contribution of w .)
Let x,y, and z be sequences corresponding to w r, w s,
and u> (r+s) in which the first signs are not ignored so that by the
above z is a subshuffle of x and y. We show how this can be used to
obtain OJ (r
{r+s) as a subshuffle of to r and u> s. We break this up
into cases.
Suppose r and s are positive. Then so is r+s. The
contribution of u> consists of w pluses with possibly some minuses
interspersed among them. We need less but there is also less to choose
from. In expressing a sequence as a subshuffle of other sequences no
harm can ever be done by selecting at any time the earliest place where a
sign occurs in the sequence (since any continuation which was legal
before is a fortiori still legal). We may therefore assume that in the
representation of r+s as a subshuffle of r and s the first plus is
taken either as the first plus in r or the first plus in s. Suppose
without loss of generality that it is the first plus in r. The first
plus in s may also have been used in the representation of r+s. (If
not our work is easier.) If so, then s could not have been used for
earlier terms in the representation by the definition of a shuffle. We
now obtain the representation of a> (r+s) as a subshuffle as follows.
The contribution of u> is taken from the contribution of w in u r.
The section in z and x which is missing in OJ (r+s) and w r
respectively is ignored. The only possible difficulty remaining occurs
if the representation of z makes use of the segment of y which is
not contained in a s. i.e. the first plus in s is used in the
representation of r+s. We then still can obtain a subshuffle by using
the u a pluses contributed by w a in u>as since they have not yet been
used, and by the above, the ordering requirement is still maintained.
If r and s are negative the result follows by sign
reversal.
Now suppose r is positive and s is negative. Also suppose without loss of generality that r+s is positive. The construction
is similar to the previous one thanks to the technical condition referred
to in part (a). We may thus assume that the first plus in r+s is taken
as the first plus in r. The only possible difficulty now occurs if the
minus which s begins with is used in the representation of r+s. Just

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

108

as in the previous case we can now use the u minuses which a> contributes to a) s. (Since s is negative these w minuses are present
by sign reversal because of the u> pluses in to .)
(d) We now prove the result in general. If we allow the use of zero
a.
a.
coefficients, we may assume that x = I w "*r-j, y = ^ w 1s1- and
i <a
i <a
x+y = I co i(r-j+s-j). Now if no minuses were ignored in the sign
sequence formulas the result would be trivial. Juxtaposition of all the
subshuffles as we run through all i would lead to a subshuffle of x
and y. Furthermore, minuses "missing" in x+y cause no trouble; in
fact, this makes things easier. The difficulties occur only when minuses
in x or y are "missing." As in the proof of part (c) we must show
that other sources for these minuses exist so that x+y can still be
obtained as a subshuffle. (It is ironic that the main cases of difficulty occur when r-j+si = 0 for some i! Since the corresponding term
in x+y vanishes, the subshuffle requirement appears to be vacuous.
However, the issue of ignored minuses arises in later terms when the
above equality holds.)
We shall obtain x+y as a subshuffle of x and y by
a.
induction on a. We assume that I u> Mr-j+s-j) has been obtained
as a subshuffle of

a.
I w ir-j and

a.
I u> ">Sj and

I CJ i(rj+Sj) can be obtained as a subshuffle of

show that
I w ">r-f and

I GO

1#

SJ

as a continuation of the earlier subshuffle. (The last part is crucial

for reaching limit ordinals. Subshuffles for i < y for all y < 8 do
not lead in an obvious way to a subshuffle for i < 8 unless they
satisfy the continuation property.)
Suppose that there is a minus in 8 which is not ignored in
the contribution of w 3 to x+y but is ignored in the contribution of
OJ 3 to either x or y, say x. Otherwise, as noted before, the continuation of the subshuffle is trivial. We now use an argument which is
similar to the one used in the proof of lemma 6.3. Suppose that the
minus is ignored in the contribution to x because it occurs in an
exponent a Y for y < 3 in the normal form of x. The other possibility for ignoring minuses will be considered later. Then a Y does

SUMS AS SUBSHUFFLES, UNSOLVED PROBLEMS

109

not occur as an exponent in the normal form of x+y. Furthermore every

exponent between a Y and a$ also contains the minus so it cannot occur
in the normal form of x+y otherwise the minus would be ignored in the
a
contribution of u> 8. Thus y _< 6 < 3 => r<5+s<s = 0, i.e. r<$ = -s,$.
We can now use the same type of construction used in the
a.
proof of lemma 6.3 The contributions of w to x and y now give us
a supply of pluses and minuses which may be used to obtain the subshuffle
needed. As in the proof of lemma 6.3 we first consider the case where
all minuses considered do take part in the contribution of wa"Y for
for the least y in which they occur. Only one little detail is needed
to supplement the argument in lemma 6.3. The contribution of the concerned minuses in a3 to x+y depend upon the sign of r3+S3.
However, since a$ and b<$ have opposite signs, we have available
whichever sign we need.
We now consider a minus not ignored in the contribution of
o) to x+y but ignored in the contribution to x and not occurring in
an exponent y < 3 in the normal form of x. We know that x has an
immediately previous term with non-dyadic coefficient. The exponent of
that term may occur in x+y but if so it must also be an immediately
previous term with a dyadic coefficient (because of the rule for ignoring
minuses). Whether or not the exponent occurs in the normal form of
x+y it must be the exponent of the previous term of y and have a nondyadic coefficient. This is the situation we now have. The normal
a
a
forms of x and y consist of terms GO &~lro
and m &~lso
where
r

P -1

and

p-1

are non-dyadic
p-1

and

r.
p-1

+ s. .

P- I

is dyadic (in

p-1

particular it may be 0 ) . a o . may or may not occur as an exponent in

p-i

the normal form of x+y. In any case the inductively defined subshuffle
makes use of only a finite number of pluses and minuses in the
contributions of r 0 . and s o
to x and y respectively. (If one
p-i

p-i

is picayune, in order for the proof to be formally correct, this aspect

of the construction should be included in the inductive hypothesis.
However, an excess of formalism would complicate the exposition
unnecessarily.) Recall that a non-dyadic real has pluses and minuses
arbitrarily far out. Thus there is enough left over in the contributions
of r_
and s, . to the sign sequence of x and y respectively to
p-i

p-i

continue the subshuffle of x+y to include the contribution of the minus.

(The number of signs needed was calculated in the proof of lemma 6.3.)

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

110

There is still the problem of a similar situation occurring

for some y < 3. This can be handled as before by letting the (y-l)st
term play the role previously played by the ($-l)st term.
This finally completes the proof of the theorem.
By comparing the proofs of theorem 6.1 and 7.1 we see the
contrast between a proof using the normal form and one which does not. I
referred earlier to the analogy between this and the contrast between
synthetic and analytic proofs in geometry. Very often there is a choice
between a lengthy tedious but routine analytic proof and a quick
synthetic proof using an appropriate theorem. We see the contrast also
in our work on surreal numbers. The use of the normal form gives us a
proof which is somewhat routine but messy because of the quirks in the
sign sequence formula. However, even if there is a shorter proof which
does not use the normal form, the proof using the normal form still may
have value because of its constructive nature.
I conjecture that a result similar to theorem 7.1 is valid
for multiplication. This would involve orderings obtained from cartesian
products of sequences of signs. However, this is far beyond our present
pioneering level. Further insights on dealing with sign sequences are
needed before one can tackle such problems the right way.

Ill
8

NUMBER THEORY

A BASIC RESULTS
This chapter overlaps chapter five in [1] to some extent. We
study a subring of the class of surreal numbers for which there are
results analogous to those in number theory. By the results of chapter
six the theory is essentially unchanged if one restricts oneself to
certain suitable subsets of the class of surreal numbers so that we deal
with "honest" rings.
Definition. A surreal number a is an integer if the exponents in the
normal form of a are all non-negative, and if a zero exponent occurs
then the real coefficient is an ordinary integer.
For example,
this) whereas

M2-^-!

is an integer (the = does not prevent

u>2+^- is not. Also, w 2 + uT 3 7

is not an integer.

This definition, which is equivalent to the one used in [1],

may seem artificial at first. There certainly exists a wide choice of
other subrings. However, this definition leads to desirable theorems.
Moreover the existence of equivalent definitions, one of which can be
expressed simply in terms of the sign sequence and the other in terms of
the relation a = F|G, suggest on philosophical grounds that our
definition leads to a "natural" system.
Theorem 8.1. The following are equivalent
(1) a is an integer.
(2)

There does not e x i s t an o r d i n a l

a(a+l)

(3)

have opposite s i g n s .

a = {a-l}|{a+l}.

such t h a t

a(a)

and

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

112

Proof (1) =*> (2). This is immediate from the sign sequence formula. If
b > 0, to begins with an infinite number of pluses, and in <o r all
signs are repeated infinitely often. Fortunately, the ignoring of
minuses in sums does not cause any extra complications. For b = 0 we
know that an integer corresponds to a finite sequence of signs which are
all alike.
Not (1) = not (2). For b = 0, we know that if r is not an integer,
then r = (o r has length at most w and there is a change in sign
somewhere, i.e. we have not (2). If b < 0, then u> begins with a plus
followed by a minus. This almost proves not (2). What remains is to
consider the possibility that the minus may be ignored. Let b be the
first negative exponent occurring in the normal form of a. If the first
minus sign in b is ignored then there is a previous term with exponent
zero and non-dyadic coefficient so again we have not (2).
(1) =o (3). First, assume that all exponents are positive. Then the
result follows immediately by cofinality, where we use the definition if
there is no last term in the normal form and lemma 5.3 if there is a last
term. If a zero exponent does occur then it necessarily occurs in the
last term with an integer coefficient. The result then follows by
cofinality from lemma 5.10. (Actually, since 0 has no proper initial
segments we need only a simpler version of the lemma.)
Not (1) => not (3). Let the normal form consist of terms with negative
exponent and express a as b+c where c is the tail in the normal
form consisting of all terms with negative exponent. Then by the
lexicographical order a-1 < b < a+1. Clearly i[a) > lib) by the sign
sequence formula (in fact, b is a proper initial segment of a ) , so
a * {a-1}|{a+1}.
Now let us consider the other possibility, i.e a has the
,. b.
form
I a) Trj + r a where b-j > 0 and r a is not an integer. Then
i<a
if r a is replaced by either the largest integer less than r a or the
smallest integer greater than r a we obtain a number c satisfying
a-1 < c < a+1. One of these necessarily has smaller length than a, so
again a * {a-l}|{a+l}.
Each of the equivalent definitions has its own intrinsic
interest. (2) requires nothing more than the definition of surreal
numbers and thus can even be used in the beginning of chapter two. (3)

NUMBER THEORY

113

is used by Conway in [1], This is consistent with his style of making

the relation a = F|G fundamental. Our definition (1) is also mentioned
by Conway. The latter definition appears to be most useful for proofs.
In fact, many of the arguments depend only on the generalized group ring
structure given by the normal form and make no essential use of the fact
that the exponents are surreal numbers. Thus by taking theorems 5.7 and
5.8 as definitions for a set consisting of generalized power series we
can obtain a theory which is independent of the theory of surreal numbers
although the results will be valid for surreal numbers as one special
case. One convenient hypothesis for the exponents is that they form a
divisible abelian group.
The following result is immediate by theorems 5.7 and 5.8.
Theorem 8.2. The integers form a subring of the surreal numbers.
Remark.

This can also be proved easily using definition (3).

Theorem 8.3.
Proof.

a) c
-c
co a

Every surreal number is a quotient of two integers.

a.
Let a = i w V j .
i<a
Suppose c = <j>|{O,a-j}. Then c < 0 and (V i) (c < a-j). Then

is an integer.

Now co c a = I ID i""cr-j. Since

i<a

is also an integer.

Of course

a =

uTca

a-j-c > 0 for all i,

co"c
C.

Let

a = I w ir-j. For convenience, express this as y+r+z,

i<a
where y consists of all terms with positive exponent, z consists of
all terms with negative exponent, and r is a real number which may be
0.
Unless r is an integer and z is negative let p = y+n
where n is the largest integer in r.
If r is an integer and z is negative let p = y+r-1.
It is clear from the lexicographical order that p is the
largest integer not greater than a. It is also clear that p is the
unique integer satisfying p _< a and a-p < 1. Incidentally, this
remark on the spacing of the integers among the surreal numbers is
further justification for our definition of "integer."

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

114

Theorem 8.4. The division algorithm holds, i.e. if a and b are

positive integers there exist unique integers q and r such that
a = bq+r and 0 _< r < b.
Proof. Let q be the largest integer not greater than Then
~ >_ q and ^ - q < 1. If r = -^ - q it is immediate that 0 r < b.
Conversely, the conditions a = bq+r and 0 r < b imply
that 7- _> q and q - T- < 1. We have already noted that this characterizes q uniquely.
Fortunately, for finite, i.e. ordinary, integers divisibility
in the surreal sense is equivalent to ordinary divisibility. However,
strange things do happen soon so that our system differs in important
ways from ordinary number theory. For example, w is divisible by
every finite integer since is an integer for all finite n by our
n
definition. In fact, more generally every integer whose normal form does
not contain a term with zero exponent is divisible by every finite
integer.
B

PARTIAL RESULTS AND UNSOLVED PROBLEMS

We first note that the ring of integers is not Noetherian.

In fact, the principal ideals (a>), (w 2 ), (w 3 ) (GO11) form a strictly

ascending chain of ideals. Any reader who is dissatisfied with using
concepts such as Noetherian for rings which are proper classes should
reread the remarks in the first paragraph of this chapter.
Slightly less obvious is the following.
Theorem 8.5.
principal.

There exists an ideal with two generators which is not

Proof. In fact, let I = (w,aV2"). We show by contradiction that I is

not a principal ideal. Let U,oo/2) = (a). If d < 1 then u)d divides
both a) and co/2; hence it divides a. Hence no exponents occurring in
the normal form of d can be less than 1.
Now

a divides

GO, i.e.
a

is an integer.

NUMBER THEORY

115

Let a = I a) ir-j. If we write

ii a
clear that the normal form of

a in the form

1
begins with

w Or o [l+ w "is-j] it is
"b 1
a>

(as in dealings

a
r0
with ordinary power series). Therefore the normal form of 7- begins
a

with

a)

.
r0

that

b0 2. 1

contain any

Since

we

^ve

is an integer

b0 = 1.

b0 < 1.

Since we also know

Furthermore i t follows that

further terms so =

where

r0

can't

is the reciprocal of an

a VQ
a has the form ~ for some integer

integer, i.e.

n. But then a

does not divide w/2" so we have a contradiction.

We are interested in the classification of primes. By our
earlier discussion for ordinary integers, a number is a prime in our
system if and only if it is an ordinary prime. It is an open problem
whether other primes exist. However, we have partial results which tell
us that various surreal numbers are composite so that the search for
primes can be narrowed significantly.
b.
For example, let a = J u " r-j be a surreal integer. If
i<a
bi > 0 for all i then we can let c = {0}|{bi}. Clearly w C divides
a so that a is not a prime. Hence the normal form of all primes
necessarily end with a term whose exponent is 0. Furthermore, n
v

1 w -V j + n so that the last term in the normal form of a prime

i<a
is necessarily 1.
Consider next ordinary polynomials in w. Every polynomial
of degree larger than two factors since any factorization of polynomials
obviously leads to a factorization when the indeterminate x is replaced
by 00. Furthermore, any polynomial may be regarded as a polynomial of

divides

of degree larger than two in 00 ; hence any polynomial factors, e.g.

j_
2_ l_
U3+1 = (w3 + lJ((o7-(/+l).
The same idea can be extended to any finite sum where all the
exponents are rational multiples of a fixed surreal a.
The following result, which is a big help in narrowing down
our search for primes, is being honoured by being designated as a theorem.
Theorem 8.6.

. b.
Let a = 2.w l r i be an integer. Suppose that there exists
i

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

j

such that

Proof.

b 0 >> bj and

bj * 0. Then

116

is composite.

Let

k be the least j such that b 0 >> bj. Then we may write

b.
b.
a = x+y where x = \ w ir-j and where y begins with a> ^r^.
i<k
Now all exponents in y""1 are rational linear combinations
of the exponents bj. Since the b's are decreasing and non-negative,
every exponent bj satisfies b 0 >> bj. Hence all the exponents in y" 1
have lower order of magnitude than b 0 . Furthermore, by choice of k
every exponent in x has the same magnitude as b 0 . This shows that all
exponents in xy""1 are positive. (In fact, they have the same order of
magnitude as b 0 .) Hence xy" 1 is an integer.
By theorem 8.6 we can narrow our search for primes to those
integers all of whose exponents with the exception of the zero at the end
have the same order of magnitude, i.e. are finite non-infinitesimal
multiples of a fixed surreal a. Furthermore, since a factorization of
. ab.
a number of the form
I u> ir-j automatically leads to a factorization
i<a
, b.
of
l uj V j , we may just as well limit ourselves to finite noni<a
infinitesimal exponents.
We now mention a device which permits us to regard the
exponents as ordinary real numbers but with the price of being stuck with
a larger coefficient field.
The class of all surreal numbers in which the exponents
occurring in their normal form are all finite forms a subfield since the
set of all finite numbers is certainly closed with respect to finite
linear rational combinations. Similarly, the class of all surreal
numbers in which the exponents occurring in their normal form are all
infinitesimal forms a subfield.
Now consider a surreal number

a of the form

I u> irj

where

i<ct

bj is finite for all i. It is well known and easy to prove that every
finite number a can be expressed uniquely in the form r+e where r
is real and e is infinitesimal. (In the language of nonstandard
analysis r is called the standard part of a.) We now group together
all terms which have the same standard part. Specifically, corresponding
to every real number r we obtain an expression of the form

I w isj
i

NUMBER THEORY

117

where c-j) is the subsequence of (b-j) consisting of all b-j whose

real part is r. [For some r the set of Cj's may be empty; in fact,
because of the well ordering it is not too hard to show that the set is
empty except for countably many reals.] This can be expressed as
a) ( I w isj) where ej is a decreasing sequence of infinitesimals.
i<a
As r varies we may thus regard the whole sum as a generalized series with real exponents with coefficients themselves power
series with infinitesimal exponents. (Since the set of a's is wellordered, so is the set of r's.) It is clear from the nature of formal
multiplication of series that the latter point of view leads to a system
isomorphic to the original.
Thus in a search for primes it is natural to investigate
surreal numbers for which the exponents occurring in their normal forms
are real numbers.
Among the simplest looking numbers for which the problem
reamains open is u)/2+u>+l.
Open Problem.

Is wvT+w+l

a prime?

We suspect that no matter whether the answer turns out to be

yes or no, the proof will make no essential use of the number

/2 but in

fact will use only the fact that /2 is irrational. Moreover, the proof
would probably be extendable to more general polynomials. The following
unsolved problem is found in [1].
l

1 i
Open Problem.

Is

OJ+W

+w

u> +1 a prime?

(Do not let an earlier remark confuse you. We had stated that if all the
exponents in a polynomial are rational multiples of a fixed number then
the polynomial is composite. The discussion there concerned itself with
polynomials of finite length.)
It is of some interest to classify the factors of u>. Among
the obvious factors are finite integers and numbers of the form w
where r < 1. It is clear from the proof of theorem 8.3 that any number
such as

I w ir-j where
i

a0

is infinitesimal (e.g., uP + 1) is a

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

118

factor. With caution one can take certain products of these such as
1
j_
(o)"2"-7)(u)s+l). O f course one cannot take all possible products; e.g.,

l-wi)(OJ^+1)
I

(u)

> to, so it is certainly not a factor of w.

The question is whether there exists factors of a more subtle

a.
Specifically, can there exist factors such as J u V-j where
i<a
is not infinitesimal? I feel that this question is of interest in

kind.
aQ-a!

connection with the open problems concerning primes.

Although we have no

theorems to the effect that an answer to the latter question will imply
an answer to any of our questions on primes, I believe that the
techniques of the proof will supply insights to study the classification
of primes.
Furthermore, the above question leads to algebraic questions
of interest which are independent of the theory of surreal numbers. I
believe that generalized power series in which exponents may be
irrational and lengths may be larger than
their own sake.

to are worthy of study for

Generalized series do arise in certain subjects such as

valuation theory, but as a whole such objects have been neglected in the
1iterature.
We close this chapter by proving a partial result concerning
possible factorizations of GO.
Theorem 8.7.

to cannot be expressed in the form

b.
d = I to is-j
i<oo
numbers.

and

cd where

a.
c = I to ir-j
i<oo

are integers, and all the a's and b's are real

Remark.

No restriction on the field of coefficients is needed for the

proof.

Also we are tacitly assuming that the r's and s's are distinct

from

Proof.

0, i.e. the series are in normal form.

Suppose

a.
b.
w = ( I w 1r-,-)( I to 1 s 1 -).
i <to

By suitable normalization

i <to

t h i s may be e x p r e s s e d i n t h e f o r m

a.
b
1 = (1+ a i r - j ) ( i + u i S j )
i <w

o f course, not n e c e s s a r i l y

the same

a's, b's,

r's

and

with,

<OJ

s's.

Since we are c o n s i d e r i n g formal m u l t i p l i c a t i o n only we may

NUMBER THEORY

replace

co

by

x .

Also,

exponents are p o s i t i v e ; ,
bounded above by

1.

where

(bj)

(a-,-)

length

and

119

since

and

are

integers

hence i n the n o r m a l i z a t i o n

Thus we f i n a l l y

have

are s t r i c t l y

w) sequences of r e a l

all

the

original

the exponents

1 = (l+x i r - j X l + ^ x

is-,-)

i n c r e a s i n g bounded o r d i n a r y

numbers.

We would l i k e

(i.e.

to show t h a t

is impossible regardless of the f i e l d of coefficients.

are

( I t i s , of course,

t r i v i a l to satisfy the above identity i f we do not require that (a-j)

(b-j)

of

this

and

be bounded.)
Let

write r

sup(aj) = l
m

a +m

= (x -*+Ix i

and

sup(bj) = m.

-ViKl+IxbiSi).

sup(a-j+m-) = i+m-i = m = sup b j .

the f i r s t cancel.
two series

If

i < m then we can

Then

The equality says that a l l terms after

Thus we f i n a l l y have the following situation.

I u irj

and

I u> ">s-f

with

( C J ) and

bounded sequences of real numbers with the same

(dj)

sup I

r 0 = s 0 = 1.

6 = min[cj-Cj-1,(ci-dj:dj<c-|)].

product.

For convenience

We w i l l now obtain a contradiction.

We now choose an arbitrary but fixed

is f i n i t e . ]

both increasing
such that a l l

pairs cancel except for the product of the f i r s t terms.

we may choose

We have

[Since

CJ. Let

sup(cj) = sup(dj),

(dj:dj<cj)

6 > 0.

Now choose n 0 such that n _> n 0 - d n > -6.

c.
d
Now the term (x 1 r 1 -)(x n s n ) for n ^ n 0 must cancel in the

(We may j u s t as well assume that

n 0 _> 1, so that the excep-

tion consisting of the product of the f i r s t terms w i l l not arise.)

Hence there must exist a pair
Cj + d n = Cj + d m .
lation.)
k _< i-1
Cj

j .

m.

Hence

j > i.

is increasing.

l _> ( S l+ d m ,

i.e.

occur is f i n i t e .

Furthermore, i f

i.e.

CJ <^ A - 6 .

dm were less than

Hence

dm _> CJ .

6' = m i n [ c j + 1 - c j , ( d j - c j : d j > c j ) ] .

This exists since

Then CJ+ > CJ + d n = Cj + dm >_ CJ + 6'+d m .

dm <_ - 6 ' .

Thus i f we vary

Similarly, i f

n the set of

Hence
m's

which

dm > CJ we obtain

dm + A _> CJ + 6 '+ i _> CJ + 61 + d n = CJ + 6' + d m .

1

such that

cj + d n > cj + SL-& >_ dm + >_ CJ + dm

This contradicts the equality.

Now l e t

dj

(i,n)

Cj + d n > CJ + -6 _> CJ-X + I > C|< + dm for any

we would obtain similarly that

for a l l

d i s t i n c t from

(This is certainly a necessary condition for cancel-

Therefore
and any

(j,m)

Thus the set of

We have shown that for

j's

Hence

l _> 6' + c j ,

which occur is f i n i t e .

dm * CJ these are only f i n i t e l y many

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

120

pairs (j,m) such that Cj + d m = c-j + d n for any n _> n 0 . Hence there
exists an n x such that for n j> n x the only possible pair (j,m) which
satisfies Cj + d m = c-j + d n must have d m = Cj and hence Cj = d n .
Since i was arbitrary to begin with and also since the same argument
can be applied to dj this shows that CJ = d-} for all i. Furthermore, because of cancellation we must have r-,*sn = -r n sj for n_> n l t
If we begin with i = 0, since r 0 = s 0 = 1 this tells us
that for n 2. "i> r n = ~ s n Now let i = ni Tnen f r m 2. n 2 we nave
r s
m n = " r n % - Since r n = - s n * 0 this gives r m = s m .
Hence if n i ^ m a x ( n 1 , n 2 ) then r m = s m and r m = - s m ; i.e.
s m = 0 which is a contradiction.
Unfortunately, the proof breaks down for lengths other than
u). The limited interest of the theorem is due to the fact that it is
the only result we have which gives some restriction to factorizations,
thus giving us minimal hope for the existence of non-trivial primes.
All our other results give circumstantial evidence against the existence
of such primes.

121
9

GENERALIZED EPSILON NUMBERS

A EPSILON NUMBERS WITH ARBITRARY INDEX

On page 35 in [1] Conway makes some remarks on the possibility of extending the transfinite sequence of epsilon numbers to more
general indices, e.g. he gives a meaning to e- 1 # He also mentions
other interesting surreal numbers. He mentions that the equation
u) X = x has a unique solution and that there exist various pairs (x,y)
"~X

""V

satisfying oo = y and w = x. In this chapter we study this

systematically. Moreover, we also discuss higher order fixed points, e.g.
in the sequence of ordinary e numbers eo>ei there exist a such
that e a = a and such a can be parametrized by ordinals. It is
interesting that a general elegant theory exists for surreal numbers
which have such fixed point properties.
First, we summarize the situation for ordinals. It comes as
a surprise to the beginner that although w apparently increases much
faster than a, there exist ordinals such w = a. Such ordinals, known
as epsilon numbers, can be arranged in an increasing transfinite sequence
e
o s i '" 8cr Not only is this sequence defined for all a but there
exist a such that e a = a. Furthermore, this construction can be
extended indefinitely. Specifically, let us use the notation e o (a) = e a
and let e 1 (0) be the least a such that e o (a) = a. Then for every
ordinal 3, there is an increasing sequence e$(0) ,$[1), ,3(0))
where 3(a) runs through all ordinals which are fixed for every e^
with y < 3.
It is possible to do even more by using a kind of
"diagonalization.11 It turns out that the function e a (0) is continuous
as a function of a, so there exist a such that e a (0) = a. [If
is regarded as a double array, the sequence e a (0) looks more like a

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

122

first column than a diagonal but it plays the role of a diagonalization.

On the other hand, e a (a) is not a continuous function of a and is thus
not worth considering.] It is hence possible to form a new double array
and then continue as before.
We now turn to surreal numbers. First, we shall deal with
epsilon numbers. Although the results are special cases of theorems
about general fixed points which we shall prove later, we feel that it is
pedagogically reasonable to handle this relatively concrete case first.
Define w (a) inductively for all positive integers n and
a
OJ (a)
surreal numbers a as follows: ^ ( a ) = u> and w + 1 (a) = u n
. We
are now ready to define e, inductively for arbitrary surreal numbers
b. Let b = B'|B" be the canonical representation of b. Then
e. = (w (1), co [e. ,+1]}|{OJ [e.,,-1]} where n is an arbitrary positive
integer and as usual b1 and b" are general elements of B1 and B"
respectively.
Before stating the basic theorem we look at several examples.
The definition gives e = {m (1)}U = l.u.b. {a (1)} which is the
ordinary f i r s t

on1
number. We abbreviate

n
e as e.

Then

0)

e1

= {a) ( l ) } | { o ) n ( e - l ) >

Which

i s

( u ) , ^ , ^

,"'l||e-l,ai

-1

, u)W

,}.

Note that the sequence which comprises the upper set is decreasing
although this might run counter to intuition. This is so because
a)c
<< w e = e whereas e-1 ~ e. Moreover, once we know that
OJ
< e-1 the decreasing nature of the sequence follows by induction.
e.

e, is defined for a l l b and satisfies OJ D = e, . e, is

b
b
b
a s t r i c t l y increasing function of b and e > OJ (1) for a l l b and a l l
Theorem 9 . 1 .

positive integers
Proof.

n.

As usual we do this by induction.

Since

b1 < b", e, , < e. .

Hence e. , = </>' u> b" = e ,. Since e > w (1) = to i t is clear

that e. ,,-2 ~ e. . Therefore e.,,-2 > e. i , i . e . e.n-1 > e u i + l - Thus
b
b
b
b
b
b
for a l l positive integers n,

w (e. ,+1) < co (e.,,-1).

Also, w (1) < e .

Since both sides have the form w i t follows that

n ( 1 ) " V ~ e b " - 1 ' So u n ( 1 ) < e b"" 1 '

Now

e. , + 1
u) b

e. i
u) b =

~ e, , + 1 .
b

Hence

to

e
k
b

'+1

> e, , + 1 .
b

GENERALIZED EPSILON NUMBERS

123

Similarly uT b "
^b" = e ,, ~ e ,,-l, hence GO b"
implies that w U b ' + 1 ) is an increasing function of n
a) (e ,,-1) is a decreasing function of n. Hence if m
n b
arbitrary positive integers we have
u ( +1) < (
( +1)
1
1
Sit1Ce

< eb,,-l. This

and that
and n are

<o (1) 1s
n
certainly an increasing function of n we also have w (1) < w (e. ,,-1).
We now have exactly what we need to conclude that s is defined.
The fact that w (1) < e is immediate since w (1) occurs
among the lower elements. It is also immediate from the definition that
e < ^ ( e ,,-1) < e.,,-1 < e. . Similarly, e., < e. . The usual argument
using common initial segments shows that e. is strictly increasing.
Since the set {w n (1)}, {w n (e,b ,+1)} contains no maximum and
the set {OJ (e, ,,-1)} contains no minimum we know that
n b

mV

Vn V

^W^' * WV" *'

by cofinality.

This completes the proof.

Corollary 9.1.

The uniformity theorem is valid for

e, .

Remark. Recall that this means that any representation

give us the same result.

b = F|G will

Proof. This follows from the usual argument using the inverse cofinality
theorem and the cofinality theorem since e. is an increasing function.
We generalize the usual definition of epsilon numbers to
surreal numbers by defining an epsilon number to be any surreal number
such that a) = a. We thus have a class of epsilon numbers parametrized
by the surreal numbers.
Theorem 9.2. Any epsilon number between
initial segment.
Proof.

Suppose

e.i < e < e. H. Since

Hence e = w > w = 1 . Similarly

induction that e > w (1) for all

e. , and

e. has

e.

as an

u> = e it follows that e > 0.

e > w. We obtain immediately by

n. Assuming e > u> (1) we obtain

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

powers of
Similarly
But
But

0) (1 )

e = a) > a) n

that

= a) + 1 ( 1 ) .

Epsilon numbers are i n p a r t i c u l a r

u>. Hence e < e. => e e. . Therefore e < e b ""l.

e. , < e => e, , + 1 < e . Hence u> [ e , , + 1 ] < w ( e ) < w ( e , , , - l ) .
b
b
nb
n n b

a)
a) (e)
(e) =
= e
e

since

124

by
by induction.
induction.

Hence
Hence

e.
e.

ii s
an i n i t i a l
s ar

segment of

s a t i s f i e s the required i n e q u a l i t i e s .

Corollary 9 . 2 .

If

initial

segment of

Proof.

Let

is an i n i t i a l

segment of

then

is an

e .

b = B'|B"

lexicographical order

be the canonical representation.

1

< c < B".

Hence

e. , < e

By the

< e.,,

so that the

r e s u l t follows from theorem 9 . 2 .

Remark.

Note that the proof of theorem 9.2 makes no use of the f a c t that

the representation of
Theorem 9 . 3 .

Proof.

is canonical.

Every epsilon number

Let

e = F|G

i s of the form

be the canonical representation of

be the set of indices of epsilon numbers in

indices of epsilon numbers in
Let

a = C|D.

Now

e.

G.

e~ < e < e~

Since

and

for some

e.

Also l e t

and

the set of

F < G, i t follows that

by choice of

b.

D.

C < D.

Hence, by

theorem 9.2

e
is an i n i t i a l segment of e. On the other hand any
a
epsilon number of the form e
which is a proper i n i t i a l segment of e
must be contained in the set

erUen,

u
e

so i t cannot equal

e .

Hence

a = eThus the parametrization we have gives us the whole class of

epsilon numbers.
Note. Induction is not used in the above proof. We do not need the fact
that every epsilon number in F U G has the form e a for some a.
B HIGHER ORDER FIXED POINTS
We now prove a general fixed point theorem.
Theorem 9.4. Let f be a function from surreal numbers to surreal
numbers satisfying the following:

GENERALIZED EPSILON NUMBERS

125

(a) For all a, f(a) is a power of u>.

(b) a < b ^ f ( a ) < f(b).
(c) There exist fixed sets C and D such that if a = G|H
with G containing no maximum and H no minimum then
f(a) = [
Then there exists a function g which is onto the set of all
fixed points of f and which satisfies the above hypotheses with respect
to the sets f (C) and f (D), where n is an arbitrary positive
integer and f stands for the nth iterate of f.
Remark. Note that w satisfies the hypothesis if we let C = {0} and
D = <j>. Note also that the theorem permits us to obtain higher order
fixed points by induction, since the conclusion says that g satisfies
the hypothesis. Finally, (c) may be regarded as a generalization of
continuity for ordinal functions.
Proof. This is essentially a generalization of the proof of theorem 9.1.
We define g(b) inductively as
g(b) = {f n (C), f n [2g(B')]}Hf n (D), f n [|g(B n )]}. (Note that in contrast
to the proof of theorem 9.1 we are multiplying and dividing by two rather
than adding and subtracting one. This is better in general since multiplying or dividing by two is guaranteed to preserve order of magnitude.
Adding one does, not as the example w" 1 4 u)-1+l shows. Some functions
we consider do have such small values in the range. Anyway, any function
which preserves the order of magnitude and satisfies the required
inequalities can be used in the proof and it is easy to see by mutual
cofinality that the same result is obtained.)
We first show inductively that g(b) is defined for all b,
is an increasing function of b, and that f(g(b)) = g(b) for all b.
First, because of condition (a), it follows that
f(a) < f(b) + f(a) f(b). Now by the inductive hypothesis
g(b') < g(b"). Hence f[g(b')] = g(b') < g(b") = f[g(b")]. Therefore, by
the above remark f[g(b')] f[g(b M )], i.e. g(b') g(b"). Thus
2g(b') < 2 U ( b H ) . Also, since every element of C is below every element
in the range of f by condition (c), it follows that

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

c e C = ^ c < g ( b " ) ^ c g ( b n ) =$> c < ^ ( b " ) .

126

[Condition (c) i s a p p l i -

cable since any surreal a can be represented in the required form by

brute force in a t r i v i a l manner even i f the canonical representation does
not satisfy the hypothesis of the condition.]

Similarly

d D 2g(b ) < d.
Now i f
i t is a power of
a = f(a)
f(ja)

a
w.

f(2a).

a < 2a.

Therefore

f(a) = a.

C < range

is a fixed point of
Hence

Hence

then

f ( x ) < x.

an increasing function of
function of

if

function of

x,

then

is positive since

f(a) < f ( 2 a ) , thus

2a ~ a < f(2a).

fC-^a) < -^a.

f < D we see that i f

x e DUygCB")

Therefore

Similarly,

Combining this with the fact that

x e CU2g(B')

then

f ( x ) > x and i f

I t follows by induction that

if

x CU2g(B')

x D yg(B").

Since

fn(x)

is

and a decreasing

f (x)

is an increasing

by condition (b) and a t r i v i a l application of induction

we obtain

fnCCU2g(BM)] < f n+m [CU2g(B')] < V m [ D U i g ( B " ) ]

shows f i n a l l y that

f(b)

<

VDu|g(B")].

This

is defined.

I t is immediate that
using common i n i t i a l segments.

g is increasing by the usual argument

Moreover, in the d e f i n i t i o n of

g(b)

the lower terms have no maximum and the upper terms no minimum.

Hence

condition (c) applies and we obtain

f[g(b)] = {C,ffn(C),ffn[2g(B')]H{D,ffn(D),ffn[|g(B11)]}
by

which is

g(b)

cofinality.
g(a)

point of

is clearly a power of

is a f o r t i o r i in the range of

v a l i d by the usual argument.

g(b) = ( f n ( C ) ,

So i f

Now suppose
x

G has no maximum.
g(G).

is a fixed point of

G has no maximum there exists

g(y) > g(x).

hence

since a fixed

The uniformity theorem is

then

We claim that

then

f(2x) > 2x > x.

f [2g(x)]

y G such that

In f a c t , since they are both powers of

g(y) > 2g(x).

f g(y) > f [ 2 g ( x ) L
g(y) > f [2g(x)3.

Applying
Since

f [2g(G)]

is

One direction is clear since we already

other hand, consider any element of the form

Since

f.

b = G|H

fn[2g(G)]}|{fn(D),fn[|g(H)]}.

mutually cofinal with

know that i f

w for a l l

g(y)

u>,

On the

where

x G.

y > x.

Then

g(y)

g(x),

to both sides we obtain

is a fixed point of

This is exactly what we need.

this says that

A similar argument

GENERALIZED EPSILON NUMBERS

a p p l i e s to the upper elements.

obtain that
condition

g(b) = { f

127

Therefore by the c o f i n a l i t y

(C),g(G)}|{f

(D),g(H)}.

(c) w i t h respect to the sets

f (C)

Thus
and

theorem we

satisfies

f (D).

(Inciden-

n
tally, note that it is not necessary for n to run through the set of
all positive integers. By cofinality any subset containing arbitrarily
large integers will work as well.)
To complete the proof of the theorem we still must show that
every fixed point of f is in the image of g. We first show that the
analogue of theorem 9.2 is valid. In fact, let x be a fixed point
between g(B') and g(B"). Then c < x. Since g(B') < x it follows
that g(B') x, hence 2g(B') < x. Therefore f n [2g(B')] < f p (x) = x
and similarly f (C) < x. Since the same reasoning applies to the upper
elements, it follows from the definition that g(b) is an initial segment
of x. Also the analogue of corollary 9.2 as well as the remark
following the corollary remain valid. We now verify the analogue of
theorem 9.3 which is what we need. As in the earlier argument we let x
be a fixed point and let x = F|G be the canonical representation. Let
A be the set of all a such that g(a) e F and B the set of all b
such that g(b) e G. Then A < B. Let c = A|B. Then g(c) is an
initial segment of x. However, g(c) ^ g(A)Ug(B), hence g(c) f F(jG.
Thus g(c) is not a proper initial segment of x so g(c) = x.
Theorem 9.4 allows us to construct higher order fixed points
by induction. We begin with any f satisfying the hypothesis, e.g. w .
Suppose we call the fixed point function f . By induction there exist
fixed point functions {f } for all positive integers n where f +1
is obtained from f by the construction in the proof of the theorem.
We now indicate how the sequence can be extended to functions with
transfinite indices. For this purpose we extend theorem 9.4 to certain
ordinal sequences of functions.
Theorem 9.4a. Let f 0 be a function satisfying the hypotheses of
theorem 9.4. Then there exist functions f a for every ordinal a
satisfying the hypotheses and such that for a > 0, f a is onto the set
of all common fixed points of fg for 3 < a and satisfies condition
(c) with respect to the sets g(C) and g(D) where g runs through all
finite compositions of f$ for $ < a.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

128

Outline of Proof. We do this by induction. Specifically we assume that

we have functions f 3 for all $ < a satisfying the requirements and
show how to construct f a . Since the proof is similar to the proof of
theorem 9.4, to avoid tedious repetition we note only the minor
modifications required.
First let us note several consequences of the above given
properties of f 3 for 3 < a. First f 3 [f Y (x)] = f y (x) if 3 < y.
Furthermore, if 3 < Y every fixed point of fy is a fortiori in the
range of fy, hence is a fixed point of f 3 . This implies for example
that if a is a non-limit ordinal 6+1 then the common fixed points of
f 3 for 3 < a are simply the fixed points of f$.
We now look at functions g = f^lf^2"'f^n
more closely. By
the above we may assume that 3j >_ 3i+i for all i. Suppose that
f 3 (x) > x for all 3. It follows by induction and transitivity that
g(x) > x for all g. Let g have the above form with 3i > 3 2 and
let h be any composition with all indices less than 3 ^ Since f 3 (x)
is a fixed point for all fy with y < 3, it follows that
hf 3 l (x) = f 3 l ( x ) . Hence f 3 l (x) = hf 3 l (x) > h(x). Now let hx also be
any composition with all indices less than 3]_. Then
f$ h ^ x ) > f 3 ,(x) > h(x). By induction and transitivity
f

31m M * )

for all
sitions

> f

3i n h < x )

if

m > n

Similar results hold if

f 3 (x) < x

3. These give us the basic inequalities among all the compog for a fixed x.
Now if x e C or x has the form 2y where y is a common
fixed point of all f 3 then f 3 (x) > x for all 3 by our earlier
proof. Similarly if x e D or x has the form ^y where y is a
common fixed point then f 3 (x) < x for all 3. The above inequalities
now allow us to imitate the proof of theorem 9.4.
Incidentally, we must be careful in our reasoning with
inequalities. For example, assuming f 3 (x) > x for all 3, we easily
obtain f 3 f T (x) > f 3 (x). However, we do not obtain f 3 f T (x) > f Y (x). In
fact, if 3 < y we have equality.
We now define f a (b) inductively on b as follows.
fa(b) = {(gCCj^CZfafB'JlHlgfDl^E^-falB 11 )]}. In the case where
a = 6+1 the above inequalities show that we obtain cofinal subsets if we

GENERALIZED EPSILON NUMBERS

129

consider only g's of the form f n . Thus the formula for f a (b) is
consistent with the one obtained in theorem 9.4 starting with the
function f$. For general a the proof proceeds as before. The first
problem that arises is that we need a g bearing the same relation to an
arbitrary pair (g x ,g 2 ) which f m + n bears to the pair (f m fn) ln the
proof of theorem 9.4. If gl = f 0 m hx and g 2 = f$2n h 2 , then
9

f m "'/o

a \ will work because of the above inequalities. So f a

is

iiiaxvpiP2'

defined for all a.

The above inequalities also guarantee that the lower terms in
the definition have no maximum and the upper terms no minimum. This
allows us to write the following, as we did in the earlier theorem:
where g1 is the subset of g consisting of those products for which
all subscripts are less than 3. (We are, of course, using the inductive
hypothesis on the ordinals.) The above inequalities give us the required
cofinality to conclude that this is f a (b). The rest of the argument is
identical to the one used in proving theorem 9.4.
Although the most canonical example of the above is the class
of higher order epsilon numbers, we will later investigate another
interesting fixed point sequence. At any rate we have obtained a rich
supply of exotic surreal numbers.
In the next section we return to a more concrete situation
when we study sign sequences of epsilon numbers. We hope that the reader
is curious about the sign sequence of e_ l 5 for example.
C

SIGN SEQUENCES FOR FIXED POINTS

As an example let us compute the sign sequence of e-x

directly. Recall that e-x = {u>n(l)}|{a>n(e-l)}. a)n(l) consists of u n
pluses, e-1 consists of e pluses followed by one minus, w 6 " 1
consists of a>e = e pluses followed by w e + 1 minuses. (Note the
convenience of thinking in terms of blocks of like signs in computations
as referred to in chapter five.) We have here an explicit confirmation
of the fact stated earlier that u>
< e-1. to
= w u = eu. We can
now determine u> (e-1) by induction. The number of pluses remains
unchanged, i.e. is e for all n. As for the number of minuses at each

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

130

stage we premultiply by oo8 l = eta. (Recall that ordinal multiplication

is what is relevant here.) Hence w n (e-l) has (eoo) minuses following
the e pluses. This simplifies to e n w.
{e CJ} is mutually cofinal with {e }. It follows directly
from the definition that e-j consists of e pluses followed by e
minuses.
We now turn to general epsilon numbers.
For convenience of notation we regard an arbitrary surreal
number a as beginning with a 0 pluses followed by b 0 minuses, then
a x pluses, b x minus ... a a pluses, b a minuses, etc., i.e. we
partition the sequence a into strings of like sign. In this notation
a a , of course, may be 0 if a is a limit ordinal.
Theorem 9.5. a is an epsilon number if and only if a 0 * 0; all a a
different from 0 are ordinary epsilon numbers satisfying
a^u)
a a > l.u.b. a$; and furthermore b a is a multiple of u>
for a a * 0
3<a
and a multiple of u> a where c a = I a$ for a a = 0.
3<a
Proof. We compute w by Corollary 5.1 to the sign sequence formula.
The blocks in w a correspond to the blocks in a. The a t n block of
pluses in w d has order type w
and the a t n block of minuses
(c a +a a +i)
a
has order type m
b a . w = a if and only if the following
equations are satisfied:
c a +a a
(1)

03

(2)

a)

a
= aa

for a l l

such that

aa * 0

and

OJ = a 0 .

c a +a a +i

Note:

b a = ba

for a l l

a.

Since the sign sequence formula for a power of

plus, special consideration is needed for

a0.

Consider the f i r s t equation

Since

<D

= aa.

u> begins with a

w _> x

this can be expressed as a conjunction of two equations:

for ordinals

GENERALIZED EPSILON NUMBERS

131

The first says that a a is an epsilon number. Since epsilon

numbers are additive absorbing, the second condition can be expressed as
c a < a a . In particular, (a a ) is a strictly-increasing sequence.
Furthermore as epsilon numbers the a a are a fortiori powers of a>.
Thus we can use the above squeeze argument to deal with
I a$. In
3<a
general,
I u> _< (l.u.b. w )2. (We really need the "two" here. For
d<a
d<a
example, consider 1+oo+w +o) +u)W.) Hence
I a 3 1 c a 1 (l.u.b. 33)2. Since epsilon numbers are multiplicative
3<a
B<a
absorbing the condition c a < a a can be replaced by the condition
l.u.b. a$ < a a .
3<a
satisfied.

Now l e t us rephrase condition (2) given that condition (1) is

a a +i
Since ca + a a = a a this becomes w
ba = b a . We now use

elementary facts with regard to absorption of ordinals.

additively absorbs

An ordinal

i f and only i f x _> aw. Hence an ordinal y

a a +i
m u l t i p l i c a t i v e l y absorbs u>
i f and only i f a l l exponents in the
normal form are at least
reasoning with
Remark.

ca

(a a +l)w = a a .

replacing

(If

a a = 0 we apply the above

a a .) This completes the proof.

The above theorem with the aid of theorem 9.2 gives us an

immediate evaluation of

e-i

since i t is clear now that every epsilon

(A)

number less than e 0 must begin with e 0 pluses and w minuses.

One interesting question that can be asked is the following.
For which a is it possible to obtain an infinite sequence (a-j) where
a = a) l and for all i a-,- = u> 1 + 1 ? Epsilon numbers obviously have this
property. It may be surprising at first that this property characterizes
epsilon numbers since it looks substantially weaker than the equation
a = w . However, this follows easily from the sign sequence formula.
For ordinals we know that u>X >. x for all x hence the sequence
a.
terminates and it is immediate by induction that w " = a-,- implies
a) = a. In general the length of each block is monotonic decreasing and
hence is eventually constant. Since the n at which the length becomes
constant a priori may depend on a, a little caution is required.
However, we can use induction on a to show that the length has been

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

132

constant from the beginning. For the sequence of a a 's we can use the
same argument we used for ordinals using the fact that the earlier a a 's
are constant. The same argument works for the b a 's. To see this note
that the above argument for ordinals works for any function f which
satisfies f(x) ^> x for all x (the argument makes no use of any
special properties of u> ) so that it applies in particular to
functions such as u
x.
We next obtain a formula for the sign sequence of e a . The
work is facilitated by the next lemma.
Lemma 9.1. Let f and g be strictly increasing maps from the surreal
numbers onto the same class S which preserve the initial segment
property. Then f = g.
Remark. A function f preserves the initial segment property if for all
a and b such that a is an initial segment of b, f(a) is an initial
segment of f(b). For example, corollary 9.2 says that the function e x
has this property.
Proof. As usual we use induction. Let b = B'|B" and suppose
f(x) = g(x) for all x e B'LJB11. Now g(b) e S. Therefore g(b) = f(c)
for some c. g(B') < g(b) < g(B"). Therefore f(B') < f(c) < f(B").
Since f is an increasing function B1 < c < B". Therefore b is an
initial segment of c thus f(b) is an initial segment of f(c) = g(b).
Similarly g(b) is an initial segment of f(b). Therefore f(b) = g(b).
Since e. is an increasing function which preserves the
initial segment property by corollary 9.2 it suffices because of theorem
9.5 to find a sign sequence formula using "good judgment."
Theorem 9.6.

Using the notation of theorem 9.5 let d a = ag. Then

the ath block of pluses in e a consists of

block of minuses of (sa. ) b a minuses.

e.

pluses and the

atn

Proof. Let f(a) be the function given by the above sign sequence rule.
It is immediate from theorem 9.5 that the above is an epsilon number for
all a. (Note that in general a>ea) = {if) = e w .) It is also clear

GENERALIZED EPSILON NUMBERS

that

is onto the class of all epsilon numbers.

number the inverse image with respect to

(1)

133

a a = d a - l.u.b. dg

where

In fact, given any

is obtained as follows:

is the length of the

ath

block

of pluses.
(2)

ba

is the unique solution of the equation

is the length of the

a"th

It is clear that
segment property.

(e^J^x = c where

block of minuses.
f

is increasing and has the initial

This is because roughly speaking

blocks into larger blocks.

Hence

f maps larger

f(a) = e a .

The above theorem gives us a "concrete" way of looking at the

epsilon numbers.

The term "concrete" is of course relative since this

depends on regarding the ordinary epsilon numbers for ordinals as

concrete objects.

To illustrate the relativeness of the terms "concrete"

and "abstract," we may think of the integer "5" as abstract compared to

the phrase "5 apples."

On the other hand, an abstract category is a

concrete example of an internal category!

We next show that higher order fixed points can be handled in
a similar manner, i.e. we desire to express the lengths of the blocks in
the sign sequences explicitly in terms of corresponding higher order
fixed points for ordinals.

It turns out that it is easy to generalize

theorems 9.5 and 9.6.

Theorem 9.7.
numbers.
that

Suppose that

be a function from surreal numbers to surreal

g

and

are functions on the ordinals such

is strictly increasing continuous with image in the class of

powers of
0.

Let

and

h arbitrary except that it never takes on the value

Finally assume that the sign sequence for the function

is given

as follows:
(1)

The

a^n

string of pluses of

(2)

The

atn

string of minuses of

f(a)
f(a)

has length
has length

we are using the notation of theorem 9.6.

for all

Proof.

a, g(a a ) = a a , a a > l.u.b. a$, and

3<a

Then
ba

This is similar to the proof of theorem 9.5.

g(d a ).
h(d a )b a
f(a) = a

where
iff

is a multiple of

In fact, it is

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

134

easier in the sense that we are given a "handicap," i.e. the analogue of
some of the results obtained on the way in the above proof are contained
here in the hypothesis.
f(a) = a if and only if the following equations are
satisfied:
(1) g(d a ) = a a .
(2) [h(d a )]b a = b a .
As before, equation (1) can be expressed as a conjunction of
two equations g(a a ) = a a and c a + a a = a a . The first condition says
that a a is a fixed point. This implies in particular that a a is a
power of a). As before, with the help of the first condition the second
condition can be replaced by a a > l.u.b. a$. (In the earlier proof we
3<a
did not need the full multiplicative absorbing property. It was enough
to absorb 2 multiplicatively.)
The equation [h(d a )]b a = b a is equivalent to the condition
[h(d a )] divides b a by the theory of ordinals. This completes the
proof.
Remarks. We do not require that the range of h consists only of powers
of OJ although this does occur in the main examples. Strictly speaking,
continuity of g is not used in the proof. Continuity guarantees fixed
points for g, hence fixed points for f.
Other variations of the theorem are possible and in fact one
will be considered later. At this time we are concerned specifically
with higher order epsilon numbers. We now state an analogue of theorem
9.6.
Theorem 9.8. Let f satisfy the hypotheses of theorems 9.4 and 9.7.
Then the formula for the sign sequence of the fixed point function f1
given by theorem 9.4 is as follows.
The a t n block of pluses in f'(a) has length g'(d a )
1
where g is the fixed point function of g in the theory of ordinals
and the a t n block of minuses has length [hg'(d a )] b a .
The proof is identical to that of the proof of theorem 9.6.
Note now that the conclusion says that f satisfies the hypothesis of
theorem 9.7 to make an induction possible. We let g' play the role of

GENERALIZED EPSILON NUMBERS

g

(hg')w

and

the role of

h.

135

By theorem 9.6 the function

satisfies

the hypothesis of theorem 9.8 so we can use the theorem to determine the
sign sequence of

en

for f i n i t e

n.

As a convenient reference for computation and recognizing a

pattern l e t us use the diagram

[g,h] > [g',(hg') ]

parameters f o r the given f u n c t i o n .

el

corresponds to

fixed points of

{ei,[(ee j"]^}.

Then

e0

which express the

corresponds to

Since the image of

by d e f i n i t i o n , this simplifies to

ex

[e,e W ] so
consists of

[ e ^ U j ) " * 1.

By our e a r l i e r remarks we now have a setup for induction

which allows us to express the higher order fixed point functions of
f i n i t e index for surreals in terms of the corresponding functions for
n+i
ordinals. Thus the pair corresponding to en 1S [ n( n)
1We now outline the continuation to t r a n s f i n i t e indices. The
a
pair corresponding to ea is [ e a , ( e a ) w ] . The term ea in the pair is
a
clear. The j u s t i f i c a t i o n of ( e ^ ^
follows from the remark that any
multiple of

u>

for a l l

necessarily a multiple of
8 < a

8 < a where

is a l i m i t ordinal is

01

w , hence a multiple of

is necessarily a multiple of

x* w

for a l l

I t is nice that such a comparatively e x p l i c i t description of

higher order fixed points exists.
D QUASI e-TYPE NUMBERS
In this section we introduce other interesting surreal
numbers.

F i r s t , we prove the existence of a solution to the equation

oTx = x.

In f a c t , we l e t

a = 0 and
o

a , = u>
n+i

Then heuristically

speaking, (a ) is an alternating sequence converging to the unique

solution of the above equation.
(x,y)

such that

eo

We shall also construct other pairs

"V

"X

=y

and w

= x. For convenience let

f(x) = w" and g(x) = ff(x).

Theorem 9.9. tox = x has a unique solution which is obtained as follows:
-a
Let a = 0 and define a
inductively by a
= w n. Then (a )
is a strictly increasing sequence, (a2
sequence, and a 2

< a2

) is a strictly decreasing

for all n. The unique solution is then

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

given by

136

{a2r)}|{a2n+1}.
~"X

Proof.

Since

the equation

is a decreasing function of

= x

strictly decreasing function of

increasing function.
g(a ) = a
sequence.
a

2n+3

"

a2

2n)

f(a

<

2n)

decreasing sequence.
g

is a s t r i c t l y

ao
< ao , .
2n
2n+i
a

2m -

is

x, hence

= a
2 n +

< a

2m+2n

f(x) = w

is a

g(x) = ff(x) is a strictly

u> hence

is a strictly increasing

t h e r e f o r e

i s

s t r 1 c t 1 y

2n+1

Again, &l

i s a power of

co so

< a .

Since

increasing function we obtain by induction t h a t

I t follows t h a t for a r b i t r a r y
J

it is immediate that

Now

certainly a power of

> 0 = a . So by induction
a
From ao 1O > ao
we obtain
2n+2
2n

f(a

has at most one solution.

2m+2n+i -

ao
2m

and

ao ,
2n+i

we have

2n+r

Since a l l even terms are less than a l l odd terms,

{a

2n}'{a2n+i}

-a = { - a 2

has

mean1n

+1>U"a2

}.

L e t

9-

{a

2n}'{a2n+i}*

Then

Since the lower terms have no maximum and the

upper terms no minimum, we have

a

= (0,f(a2n+1)}|{f(a2n)}

Instead of beginning with

that

g(a)

i s between

"convergence."
a

and

f(a)

one can begin with any

and s t i l l

such

obtain the same kind of

Our l a t e r work w i l l enable us to e a s i l y f i n d examples of

which s a t i s f y the above condition as well as examples which do not.

I t i s easy to see t h a t

theorem 9.4.
that

a = G|H

-a = -H|-G

with

s a t i s f i e s the hypothesis of

hence

G containing no maximum and

uT

-uT a = {-u>" G }|{0,-uT H }.

{O,g(G)}|{l,g(H)>.
and

In f a c t , conditions (a) and (b) are immediate.

H

= {0,a>~ } | { u T } .
Finally,

Now suppose

no minimum.

Then

Therefore

g(a) = o)" w " a = {0,arw~ G } | {u,o,aruT H } =

Therefore condition (c) i s s a t i s f i e d with

c = {0}

D = {1}.
Recall t h a t for the e a r l i e r f i x e d points

D was the empty

s e t , so t h a t i t would have been adequate to state condition (c) i n the

form

f(a) = [C,f(G)]|[f(H)]

f o r the immediate a p p l i c a t i o n s . The more

GENERALIZED EPSILON NUMBERS

137

general form was used in the statement of theorem 9.4 with the present
example in mind.
At any rate, we now know that we have a fixed point function
1
g for g. Note that since g(x) = ff(x), any pair of the form (x,y)
where x is a fixed point of g and y = f(x) satisfies u> x = y
y

and

00=

x.

If

t h a t the i n d i c e s of
Theorem 9 . 1 0 .

Proof.

g'(-b)

i s a f i x e d p o i n t of
x

and

f(x)

so i s

f(x).

We now show

are r e l a t e d i n a simple manner.

= f[g'(b)].

By the general formula

g'(0) = [g (C)]|[g (D)]. This is

n
n
precisely {a2 }|{a2
} in the notation of theorem 9.9. Hence g'(0)
is the unique solution of oT x = x. Hence f[g'(O)] = g'(0) = g'(-0) so
the theorem is true when b = 0. (Although we did not need a separate
proof for the case b = 0 it is of interest to see it explicitly,
especially since the proof is so simple.)
We now need several inequalities before setting up the
induction as usual.
We first claim that 2f(x) < f(~x) < ff[2f(x)3 for any
fixed point x of g. Since a fixed point x is necessarily larger
than 0, we have yx < x. Hence f(x) < ftax). Since the range of f
1
consists of powers of w and 2f(x) ~ f(x), we even have 2f(x) < f(-^-x).
Also 2f(x) > f(x), hence f(2f(x)] < ff(x) = x since x is a fixed
point of g(x) = ff(x). Again since we are dealing with powers of u>
we have f(2f(x)] x. Therefore f[2f(x)] < jx so finally
f(jx) < ff[2f(x)].
Similarly we obtain the inequality
First,

ff[jf(x)] < f(2x) < -^fU).

f(2x) f(x) so f(2x) < ~f(x). Also |f(x) < f(x); therefore

f[^f(x)l ff(x) = x. Hence

f[-|f(x)] > 2x. Finally we obtain

ff||f(x)l < f(2x).

We are now ready for the induction. By theorem 9.4
g'(b) = {g n (0),g n [2g l (b l )]H{g n (l),g n [|g'(b 11 )}. As we saw in the proof
of the latter theorem, the lower elements have no maximum and the upper
elements no minimum, hence

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

f[g'(b)] = {0,fg n (l),fg n f|g l (b")l}Ufg n (0),fg n [2g 1 (b')]}.

138

(Note the

reversal of sides since f(x) = u" x .) Since g(x) = ff(x) and f(0) = 1
the right-hand side is
{g n (O),fg n [|g'(b")]H{g n (l),fg n [2g'(b l )3}. Also -b = {-b"}|{-b-'} hence
g'(-b) = {g n (o),g n C2g 1 (-b")]}|{g n (i),g n [|g'(-b 1 )]}
= {gn(0),gn[2fg1(b")J}|{gn(l),gn[-|fg'(b1)]}
by the inductive hypothesis.
The inequalities we obtained earlier are just what are needed
to check mutual cofinality. First we have 2f(x) < f(jx) < ff[2f(x)]
for any fixed point, in particular for x of the form g'(b"). Since
g is an increasing function this gives
g n [2f(x)] < 9 n f(|x) < g n + 1 [2f(x)]. Similarly from the
inequality

ff[jf(x)] < f(2x) < ^f(x) we obtain

g n + 1 [|f(x)] < g n f(2x) < gn[-|f(x)]. Hence

g'(-b) = f[g'(b)].

The fixed points we obtained here are quite different from

the earlier kind. Generalized epsilon numbers are all above w (1) for
all n. The present ones are all infinitesimal. They may be regarded as
"large infinitesimals." Since they are all squeezed between a 2
and
a
in the notation of theorem 9.9, it is of interest to determine the
first few terms of (a ) to get an intuitive idea of the size of these
n

fixed points.

In fact,

a =0,

ax = 1, a 2 =

and

a 3 = oo . Thus

a 2 is the canonical infinitesimal, and a 3 being above u>

for all
positive integers n is a large infinitesimal. A study of the sign
sequences will enhance the intuition further.
E SIGN SEQUENCE IN QUASI CASE
First, we determine the sign sequence of the unique solution a
""X

of u) = x which we feel is a surreal number of special significance

(just like the first epsilon number for ordinals). There are essentially
two different approaches which can be used for the computation. One can
compute the sequence (a ) inductively and use theorem 9.9 or work
directly with the equation uT x = x. By direct computation the beginning

GENERALIZED EPSILON NUMBERS

terms of (a ) are as follows:

139

is the null sequence,

consists of a plus followed by u> minuses,

followed by co minuses and co
w^"1"1^ = oa^'2

by

minuses.

pluses,

consists of a plus

a^ consists of a 3

followed

(Note how the sign sequence so far suggests

the heuristic idea of "large infinitesimal.")

emerging which allows us to obtain
as

a x = (+), a 2

an

We now see a pattern

for all n by induction as well

a. For convenience regard a general surreal number as beginning

with

c pluses

d minuses, c1 pluses, dl pluses c pluses, d

minuses, etc.
Theorem 9.11.

The unique solution of w

= x has the following

d

sequence: c = 1; d = to; for each non-negative integer n, c

= to n
2
and d + 1 = (c + 1 ) ; and for a _> to both c
and d are zero.
Proof.

We f i r s t show that the above pattern is followed by a l l

the sense that

a
P

in

is the initial segment consisting of the first n

strings of like signs.

We already know this for all a

with

n _< 4

and use complete induction on n, i.e. assume that the pattern is

followed for all a where m _< n. We will show that the pattern
remains valid for a ,,.
n+i
Now a
is obtained from a
by adding a sequence of like
-a
-a
signs.

Hence

= u>

is obtained from

= GO n"1 by adding a

sequence of like signs by the sign sequence formula.

Because of the

minus in the exponent the signs are opposite to those in the last string
of a . We now separate the even and odd case to determine the length of
n
n-i

i=o di
the e x t r a s t r i n g .
d

= (c w)c .

a), i t

is clear that a l l

powers of
increasing.

fore

By the sign sequence formula

to.

cp = ^

Now, since the square of a power of

c's

and

d's

and
i s also a power of

w i t h the exception of

Also the sequence c , d , c , d ,

n-i
Therefore
I d . = d _ 1# Also c _> GO
2

to

are

is

strictly

if

n _> 1 .

c we simplifies to c . Thus the above becomes

n n
n

Thered

n x

c = to ~
n

and

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

d

140

= c 2.

There are now two ways to conclude, each of which is immediate. First {a 2 } and {a 2 + 1 > are cofinal sets in the canonical
representation of the proposed sign sequence. Thus the sequence is
{a 2 }|{a2
} = a. (It is possibly easier to go back to the definition
of F|G.) Second, one can ignore the individual a and note that the
above computation shows directly that the proposed sequence satisfies
-x
= x.
It may seem at first that this method can produce other
""X

solutions of OJ = x. This leads to the scary feeling that something is

wrong; however, fortunately the attempt fails and instead the computation
leads to information which is consistent with what we have so far. In
fact, uniqueness can be checked by sign sequence reasoning by a method
which is similar to what was used in the above proof (which, incidentally, is somewhat analogous to the technique of solving differential
equations by power series.) c and d are all uniquely determined.
If the sequence for x continues beyond this, i.e. if c or d t 0
w
-x
-x w
then a) must continue with an opposite sign so to cannot possibly
equal x. Thus the uniqueness follows directly from the sign sequence
formula.
-x
We next classify the fixed points of g(x) = oT w .
Theorem 9.12. x is a fixed point of g if and only if x begins with
a and then all c's and d's are epsilon numbers larger than e ,
each being larger than the least upper bound of the predecessors.
Remark. Note that in contrast to the earlier type of fixed points here
the c's and d's have similar restrictions.
Proof.

Note first that the length of a is e . This follows easily

o
from a Cantor-Bernstein type argument since i t is immediate that
<_ e and thatfc(a)_> e .
A fixed point of g must begin with a. Conversely, for any
sequence x beginning with a the first u> strings of like signs in
x and g(x) are alike. Thus it suffices to study the strings of like
signs beyond a.

GENERALIZED EPSILON NUMBERS

141

If c a and d a are lengths of strings in x let c a ' and

d a ' be the corresponding lengths in f(x) = <o
(the signs shift of
course) and c a " and d a " be the corresponding lengths in
*~x
g(x) = w~ w . (Note that by the sign sequence formula for a >_ u>, the
ath string of g(x) correspond to the ath string of x. This is in
contrast to the case a < w where the ath string in x gives rise to
the (a+l)st string in f(x).) We now compute c a M and da".
First

( 3<a
c a = (w

ca" = < / < a

Similarly

and

>
Jc
and. d.a
a

d a " = (u) 3<a

= ca

3<a

)da\

It would be messy to express c a " and d a " in terms of

c a and d a . Fortunately this is not needed in order to obtain the
decisive inequalities.
Suppose the sign sequence has the form described in the
statement of the theorem, then we can reason similarly as in the proof
of theorem 9.5. By absorption c a ' = c a and d a ' = d a . (Beware that
this does not say that w = x since all signs are reversed.)
Similarly c a " = c a ' and d a " = da'. Since the signs have been reversed
twice, this shows that g(x) = x.
Now suppose g(x) = x. Then c a " = c a and d a " = d a .
(Recall that we are dealing only with a > u>.) From the above formulas
c

a" 2. w

il w

Similarly
Furthermore

Hence

c a = c a " >_ w
da
d a = d a " _> d a ' >_ a>
hence

thus
da

ca

is an epsilon number.

is also an epsilon number.

d a = d a " >_ d a ' >_ I d$ and

3<a
c
a = c a" 1. I c e ' 1 I C 3- Tnis shows that the c's and d's have the
3<a
3<a
required properties when regarded as separate sequences. However, we
also have c a = c a " _> c a ' >_ I d
and d a = d a " >_ I c
' _> I c .
3<a P
3<a P
3<a p
Finally, if we take into account the terms in a we obtain that all
c a 's and d a 's are larger than e 0 . This is enough to complete the
proof of the theorem.
We can now continue in a manner which is similar to what we
did for ordinary epsilon numbers except that now things are easier

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

142

because of the similar treatment of pluses and minuses. We now no longer

discriminate against the poor minuses but use cumulative sums. Thus,
regardless of whether the ath string of x consists of pluses or
minuses, the a t n string of g'(x) beyond a is made up of the same
sign and has length e. , where b a is the total length of the sequence
of x up to and including the a*" string.
The fixed points we obtained seem interesting because of
their contrast with epsilon numbers. Higher order fixed points may be
constructed by theorem 9.4, but this does not seem to be interesting
enough to pursue in much detail. It suffices to note that we need a
variation of theorem 9.7. On the one hand we use cumulative sums rather
than d a and treat pluses and minuses alike so a function h is not
needed. On the other hand, we need a preliminary sequence such as our
element a. A fixed point of f necessarily begins with the juxtaposition [a,f (a),f (a),]. Finally, to guarantee a fixed point one must
make sure that the length of the latter sequence gets absorbed when added
to the fixed points of g to ensure that we obtain a fixed point of f.

143
10 EXPONENTIATION

A GENERAL THEORY
As was mentioned in the introductory chapter, Kruskal
discovered that a theory of exponentiation for the surreal numbers is
possible. Taking advantage of his hints I discovered that an elegant
natural theory does exist, i.e. exp x can be defined in a uniform way
for all surreal numbers x and it has the properties that are expected
of an exponential function. Note that the function w x is not suitable
as an exponential function even though the theory in chapter five makes
this notation convenient. For example, it is certainly not onto since no
two numbers in the range have the same order of magnitude. (The word
"exponent" used in the past is a convenient abuse of language.)
Although we begin with a unified definition of exp x the
subject breaks up naturally into three cases.
(a) x is real,
(b) x is infinitesimal,
(c) x is purely infinite (i.e. all "exponents" in the normal form
of x are positive).
The unified form is somewhat complicated to deal with, whereas
the theory simplifies in each of the above cases for different reasons.
(Note that any surreal number is uniquely a sum of three numbers each of
which satisfies one of these cases.) Case (c) is the only one which is
worthy of a substantial discussion. In case (a) it suffices to show that
the unified definition is consistent with the usual one and in case (b)
that the unified definition is consistent with the result by formal
expansion in the spirit of chapter five.
Although the notation 'u)x" will continue to be used with
its original meaning there should be no danger of confusion since the

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

144

exponential function will consistently be written as exp x. This is

because it is obvious how to define general exponentiation in terms of
exp x and its inverse function thus there is no need for us to consider
expressions such as a in particular
u>
in the sense of
exponentiation.
The motivation for the definition of exponentiation (as it
was for multiplication) is based on suitable inequalities satisfied by
ordinary real numbers. Usually the construction of a definition requires
some ingenuity; one must include enough inequalities to obtain the
desired result and avoid circular reasoning. We recall, for example,
that for ordinary real numbers we have the implication
(ax < a ) A ( b x < b) ab x + a x b - a 1 b 1 < ab.
Such implications make the definition of multiplication possible. Recall
also that in the study of reciprocals in chapter three we could not get
by with the obvious inequalities but we were forced to use some subtlety
instead.
In searching for a definition of exponentiation similar
difficulties occur. It is natural to attempt to use Taylor series
approximations. For example, if x is positive, 1+x + ^y "' "pT is a
lower approximation to exp x. However,

{0,1+1 + ^ 7 ... "ryll^ = 3 * exp 1

so this is not adequate. Also the sets

{l+u> + JJ } and

{1 + - ^ y j ... - ~ - j } are mutually cofinal so that caution is required

in framing a definition to guarantee that exp w t exp(w2). Actually
the idea of using Taylor series does work as long as one uses proper
refinements.
x
x^
We use the notation [x] for 1+x + ^ 7 --7 for an
arbitrary surreal number x. For negative x we will be interested only
in odd n for which [x] is positive. Thus when we use the notation
[xl 2n+i
,, when x is negative, it is understood that [ x2n+i
L ,, is
positive. This awkwardness when x is negative is a price that is paid
for unification. Note that in case (b) all [x] 2 +1 are positive and in
case (c) we have the other extreme where no [x] 2 +1 is positive. Thus
we have the ironic fact that case (a) which we already know all about
from elementary analysis is the one that causes the greatest awkwardness
when combined with the other cases! This is why we prefer as soon as

EXPONENTIATION

145

possible to split the subject into the three parts according to the
different cases. On the other hand, although the theory can be developed
by considering each case separately from the beginning, we prefer not to
do so since the subject then takes on an ad hoc form with definitions
that should really be theorems. In fact, the whole beauty of the subject
of surreal numbers from the very beginning lies in the possibility of
making uniform definitions which reduce to what is desired in various
special cases.
As in the past we use the canonical representation of an
element x to define exp x inductively as follows:

exp x = {0,(exp x' )[x-x']n,(exp x")[x-x"]2

X
n

LX

xj2n+1

The restriction of [x] 2

to positive values for the upper
elements is essential in order to have any hope for the above definition
to work since 0 is a lower element. On the other hand, by cofinality,
no harm is done if negative values are included among the lower elements.
(Of course, no good is done either!) Also, it turns out that expressions
involving [x] for negative x and even n are redundant. On the
other hand, examples such as ones we mentioned earlier suggest that we
need all the terms we have.
Before proving that exp x is defined and investigating its
properties, we mention several needed inequalities among elements of the
form [x] .
n
(1) For positive, x,y we have [x] n [y] n >_ [x+y] n and
[ x + y ] 2 p > [x] n [y] n .
The above inequalities depend only on ordered field properties and follow
immediately from the binomial theorem. Also we note the obvious fact
that [x] is an increasing function of n.
For negative x the situation is more complicated. Anyway,
for infinite x there are no [x]
to consider, so the problem
reduces to the case where x = r+e where r is real and e is
infinitesimal.
First, for real r we can apply plain ordinary analysis.
r2
-r
Since 1-r + yp ... e > 0 there certainly exists n such that
r2

' r2n+i

1-r + pT ... - >pn+i\i

if n
[r]

>

Necessarily

r < 2n+l. Hence it follows that

is the least n for which the latter expression is positive,

is defined precisely when n >_ n , [r]
is an increasing

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

_y

function of n and lim[r] 2

X"f"V

= e . Since

146

X V

= e e , it is a

trivial exercise in analysis to show that

Vm 3n o (n > n Q - C x ] 2 n + 1 C y ] 2 n + 1 > [x+y] 2|n+1 )
and
Vm 3n Q (n > n Q -* [ x + y ] 2 n + 1 >. [x] 2 m + 1 Cy] 2 r a + 1 ).
[ - x ] 2 n + 1 < e" x

Since
for

and

[ x ] 2 n + 1 < ex,

clearly

[-x]2n+1[x]2n+1 < 1

x > 0.

For infinitesimal e the work simplifies since we can reason

purely algebraically using orders of magnitude. Clearly [ e ] 2 n + 1 is
defined for all n and is increasing.
For negative infinitesimals x and y we also have the
inequalities [ x ] 2 n + 1 [ y ] 2 n + 1 > [x+y] 2 n + 1 and [ x + y ] , ^ > C x ] 2 n + 1 [ y ] 2 n + r
These follow from the binomial theorem by reasoning with orders of
magnitude.

In both inequalities

2n+2

is the lowest total degree of

terms occurring in the l e f t and not on the r i g h t , and a l l such terms have
positive coefficients.

Also, [-e]

+It

^ 2 +1

<

"^ ^ o r

e >

^'

~^1S

follows from the fact that on the left-hand side a l l terms involving
beyond

e = 1 for

c o e f f i c i e n t of

i 2n+l

cancel by the binomial theorem but the

is negative.

Furthermore
s1nce

Finally, we consider a finite element of the form r+e with

r real and e infinitesimal such that r > 0. Then r+e > 0. Hence
[-(r+e)]o , differs from [-rl , by an infinitesimal. If [-rV
2n+i

2n+i

is positive so is [-(r+e)]
for

n 2. n

so is [-(r+e)]

2n+i

. Similarly, if [-r]

is increasing

. A petty nuisance is caused by the

possibility that [-r]

equals zero for some n. However, we can use
the above argument for real r for the improper inequality
r2

fore

r2n+i

[-(r+e)]

I and stl'11 obtai'n an n0- ^ r ^2n+i and there~

will still be increasinq for n > n . Thus there may

2n+i

be an no for which

"

[-(r+e)L2n+i, is considered but not [-rl2n+i

al but

EXPONENTIATION

147

by the above t h i s can happen for at most one

The r e s u l t s r e l a t i n g
s t i l l remain v a l i d .
n

for a l l

and y

x,

Since

and no harm is done.

[ x ] 2 n + 1 , y] 2 n + 1 >

[x]

and

[x+yl2m+1

is a s t r i c t l y increasing function of

the improper inequalities in the above theorems for real

can be replaced by proper inequalities (by making

i f necessary).

The results for numbers of the form

r+e

larger

then follow

immediately from the results for the real parts r.

Again we have

[ - ( r + e ) ] 2 n + 1 [ r + e ] 2 n + 1 < 1.

This gives us the basic properties of [ x ] 2

for negative x.

+1

For the sake of symmetry between the cases for positive and negative x
and for the sake of uniformity, a l l of our results relating
and

[x+y]

[x] , [y]

can be expressed in the form given for negative real

numbers, i . e .

we have no need for the e x p l i c i t value of

several of the cases.

given in

What really matters for our purpose is what

happens for s u f f i c i e n t l y large

n.

Now that we are armed with the needed inequalities, we are

ready to prove the basic theorem.
Theorem 10.1.
then

Proof.
v r

is defined for a l l

exp x [ y - x ] n < exp y

tive integers

exp x
n.

Also

x.

and ex
exp y [ x - y ]

Furthermore, i f
< exp x

x <y

for a l l posi-

exp x > 0.

Recall that exp x = {0,exp x ' [ x - x ' ] ,(exp x " ) [ x - x " ]

n
\/

Ayr.

\/

{fv" vi > r ' i

} We use induction as usual. First we must show

LX - x j n LX - x J 2 n + 1
that the lower terms are below the upper terms, so that exp x
The inequalities involving

is defined.

0 are immediate in view of our

r e s t r i c t i o n that a l l the terms of the form [ x ' - x ]

are positive, and
the inductive hypothesis that exp x' and exp x" are positive.
We now compare terms of the form exp x ' [ x - x ' ] m , and . exp
.. r
F i r s t suppose

m = n.

Choose

so that

By the inductive hypothesis exp x ' [ x " - x ' ]

1

exp x ' [ x " - x ] [ x - x ]

mm

[x] n

< exp x",

i.e.

[x"-x']
< exp x".

exp x ' [ x - x ' ]

>_ [ x " - x ] [ x - x 1 ] .

Therefore
< T^Tr~-4-.
L X XJ

Since

m
is an increasing function of n, the result for general m and n

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

follows immediately.

148

In fact,

exp x'[x-x'] m < exp x ' C x - x 1 ] ^

, < Lx "-xJ P *
. 13^x7"
max(m,n)
n
Exactly the same argument shows that
exp x"[x-x"L m+1 < r ? x p n x ' (In fact, we deliberately ignored the
Lx ~xj
fact that

can be taken to equal

2m in the above proof in order to

make i t possible to use the same argument in both cases.

Note that we also have
positive integers
Hence

< 1 for arbitrary

2m+i
n
n by an argument similar to one used above.

m and

< r ?xpx

exp x ' [ x - x ' ]

n

Lx

[x-x1]

[x'-x]

However, this is not enough.

We need

"XJ2m+i

an inequality of the form

exp x ' [x-x '] <

Lx

where

x ' and

' J

x ' are two arbitrary lower elements not necessarily equal. It turns
out that we can use a device which is similar to what we used for
different subscripts m and n which will also be important later; in
fact, it is the same cofinality idea which we used, for example, in our
development of multiplication in chapter three.
Specifically, we desire to prove the following: Let
x

i'

< X

choose

2"

Then

Vm

n so that

^n!exP

X X

~ 1 '] exp x 2 '[x-x 2 '] }. In fact,

[x-x '] [x '-x '] > [x-x '] . Then

2

exp x 2 ' [ x - x 2 ' ]

1 'C

> exp x 1 l [ x 2 ' - x 1 ' ] [ x - x 2 ' ]

> exp x 1 ' [ x - x 1 ' ] , where we

used the inductive hypothesis in the f i r s t inequality.

Similarly, we

exp x 2 '
exp x x '
have Vm 3n [r , , < , , ,
1. Hence, if x ' = max(x \ x ' ) ,
Lx 2 -xj - L X X " x J 2 m + 1
3
1 2
then for a sufficiently high p we have
exp x 3 '
exp x 2 '
exp x '[x-x ' ] m <_ exp x '[x-x '] < r .
<_ r . -i

LX

3 "XJ2p+i

LX

2 "XJ2n+i

Similarly, if x 2 < x " we have the following inequalities:

Vm 3n[exp

1"t

and
Vm 3n r

exp x

""x1ll]2

+1

ex

EXPONENTIATION

so that we obtain

149

exp x " [ x - x " ]

<

exp x 2 "
H

Thus we f i n a l l y have what we need to conclude that

defined.
computing

x ' > x2'

we can dispense with

x 2 ' in

exp x.
Since

certainly

i s a lower element in the d e f i n i t i o n of

exp x > 0.

The inequality

exp x [ y - x ]

i s immediate from the d e f i n i t i o n i f either

segment of the other.

< exp y

or

x < z <y

where

for

exp y [ x - y ]

is that common i n i t i a l segment.

and

y.

Now

Hence

> exp x [ z - x ] Cy-z]

,, < exp x.

Corollary 1 0 . 1 .

x <y

Otherwise we can use the same technique as in

Vn 3 p [ z - x ] p [ y - z ] p > [ y - x ] R .
exp y > exp z [ y - z ]

i s an i n i t i a l

chapter three by considering the common i n i t i a l segment of

Proof.

is

In a d d i t i o n , the l a t t e r i n e q u a l i t i e s give us important

c o f i n a l i t y r e s u l t s ; e.g. i f

Let

exp x

> exp x [ y - x ] .

Similarly

This completes the proof.

The uniformity theorem is v a l i d for exponentiation.

This is immediate from the i n e q u a l i t i e s obtained in the proof of

the above theorem using as usual the inverse c o f i n a l i t y and c o f i n a l i t y

theorems.
We are now ready to study the main cases separately.
real

l e t us temporarily use the notation

For

for the ordinary

exponential f u n c t i o n .
Theorem 10.2.
Proof.

exp r = e

Note f i r s t that

for real

r.

exp 0 = {0}|<j> = 1 which is a good s t a r t

although we have a long way to go!

Since a l l proper i n i t i a l segments of reals are reals (in f a c t
dyadic) we may use induction.

Hence we have

exp r = { 0 , e r ' ( r - r ' ) n , e r " [ r - r " ] 2 n + 1 | f

, , -^r^
1n
2n+i

c o f i n a l i t y to show that t h i s expression gives

e .

any real number, can be expressed in the form

{e - e } | { e +e}

positive real.

We

use

Recall that

e , like

where

I t is immediate from our e a r l i e r i n e q u a l i t i e s that

s a t i s f i e s the betweenness property.

For example,

e
e

is

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

[r-r1]

< e

= er.

Cofinality is also easy.

0, which is one possible value of

1
[r]
and upper elements T T
n

If

r > 0

then

r1,

gives rise to lower elements

1
r
Since lim [r] = lim -i--^
= e ,

" rJ 2n+i

n>0

+~ L " rJ 2n+i

these elements are cofinal in the above expression for

r < 0, we can use

150

as a possible value of

r".

e .

Similarly, if

This completes the

proof.
Remark.

Note that we used the fact that lower terms in

upper terms in

exp r

and vice versa.

initial segment we needed was

0.

contribute to

Also, in this case the only

This is not true in general although

x1

cofinality results permit us to cut down to some extent on the set of

and

needed to compute
1

use of

= 0

exp x.

[For example, as we saw earlier the

only would not distinguish

For infinitesimal
this case the formal sequence
of chapter five.
spirit for real

exp u> and

exp(oo2).]

there is a different approach.

1+x + |y + ...

In

has meaning by the results

This is purely algebraic in contrast to the analytic

x.

Again we use the temporary notation

ex,

this time

using it for the formal sum.

Theorem 10.3.
has the form

exp x = e x
r+e

where

for infinitesimal
r

is real and

x.

More generally, if

is infinitesimal, then

exp(r+e) = e e .

Proof.

Recall that every finite number x

the form

r+e

where

is real and

can be expressed uniquely in

is infinitesimal.

already have theorem 10.2 we may assume that

has real part one and in general

[r+e]

e * 0.

Since we

Note also that

is finite and has real part

Since all initial segments of finite surreals are finite, we

may use induction.
sentation of
form

r+6

We show first that by cofinality that in the repre-

exp x

where

we can restrict ourselves to initial segments of the

is infinitesimal.

(Note that

p
upper
segments of the above form; e.g., 1+

r+e

might have no

has only numbers such as

as upper segments, so that this does not follow immediately from our
earlier cofinality results.)
segment of

r+e

The basic idea is that

and regardless of whether

is an initial

is positive or negative,

EXPONENTIATION

151

terms containing exp r contribute to both upper and lower terms of

exp(r+e).
Suppose y = s+5 is a lower initial segment of x = r+e
where s < r (necessarily 6 = 0 as is easily seen, but this fact is
irrelevant for our purpose). By the inductive hypothesis and our earlier
remarks the real part of exp x[x-y] is exp s[r-s] which is less
than e r . Similarly the contributions of y to the upper elements all
P

have real part larger than e . On the other hand, the contributions of
r
r
r such as e [e] all have real part e . Since the same reasoning
applies to upper terms this proves the desired cofinality.
Since all the upper and lower terms remaining in the reprer6
sentation of exp x now have essentially the form e e x[6 ] it remains
to study expressions such as e l[& ] n .

First, the identity e le 2 = e l 2 for infinitesimals

and x 2 follows from the identity for formal power series. Also, by

reasoning with formal power series we have

[x] < e and

Now if r+5 is a lower initial segment of r+e then we have

terms such as exp(r+6)[e-6] in the representation of exp(r+e). By
the inductive hypothesis this is (exp r)e [e-6] , which is less than
8
(exp r)e e e = (exp r)e .
A similar argument for the other terms in the representation of
exp(r+e) shows that (exp r)e
satisfies the betweenness condition for
these terms. Thus to complete the proof it suffices to show that the
terms are cofinal in a representation of (exp r)e e . We use the standard
representation dealt with in chapter five, i.e. if (exp r)e e has the
form

) a) ir-j then we have

i<a
(exp r)e e = { I a> frf-co ^6}|{ I GO ir-j+w $6} where 3 < a and 6 is
i<3
i<6
positive real.
Suppose x has the form, I w i*rj. Then every exponent b
i<a
n
e
in (exp r)e
is a finite sum of a's. Let b = I a-j . if

a^ = min(ai .)
J

then

b 2.

na

a.
Now y = I u> V-f
i<3

j=1

'
is an i n i t i a l segment

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

152

a.
I m Tr-j and hence participates in obtaining upper and lower terms
i<a
to exp x. Let z = x-y = I w "irj.
i>3
Suppose first that z > 0. Then exp y[z] n is a lower term
in the representation of exp x. Now

of

(exp y)e z - exp y[z] n = (exp y ) | " ^ + ^ , +"!. The first term in exp y
has exponent 0. Hence the first exponent in the latter expression is
(n+Da^ which is smaller than na$. (As a technicality, in case all
a-j . are zero, we can choose 3 to be 1 so that a^ < 0.) Hence
j

(exp y)e and exp y[z] n agree in all terms up to and including the
terms with exponent na$, a fortiori the terms with exponent b.
For z < 0 the argument is identical except for the fact
that we consider only odd n.
b.
Also note that in general if two expressions I w is-j agree
i<B
in all terms up to and including exponent c, then so do their reciprocals
since it is clear that by formally taking inverses terms with given
exponents cannot give rise to terms with earlier exponents. Therefore
the above argument applies equally well to expressions such as
~

and LZJ
fSU.
n
If y = r+6 then by the inductive hypothesis
exp y = (exp r)e . Hence
7

V~r
J

V~Y*~^'Z.
J

(exp y)e = (exp r)e e = (exp r)e

= (exp r)e . The same
argument applies to upper elements. Thus we have shown that for any
given exponent a g occurring in (exp r)e
there are terms in the
representation of exp(r+e) which agree with (exp r)e e for all terms
up to and including the terms with exponent ag. By the lexicographical
order such terms are cofinal in the above standard representation of
of (exp r)e . This completes the proof.
We now know that exp x agrees with what is expected for
finite x. Also exp x is clearly strictly increasing and satisfies
exp(x+y) = exp x exp y. This is clear from theorem 10.3 and the
corresponding identities for reals and infinitesimals separately.
2
For real x we have the function n(l+x)
which for sufficiently small x may be expressed as 1-x + y~ '' * Then

EXPONENTIATION

153

exp n(l+x) = 1+x. This identity is valid for formal series, hence it is
applicable to infinitesimal surreal numbers, i.e. it is possible to
define a log function in the natural way. For any infinitesimal
exp n(l+e) = 1+e. This shows in particular that the function x maps
the infinitesimals onto the class of all surreal numbers of the form 1+e
where e is infinitesimal. Combining this with the behaviour of exp
for ordinary real numbers, we see that exp maps the finite numbers into
the class of all positive finite non-infinitesimal numbers.
This essentially completes the theory for the finite case.
There were no surprises. On the contrary, the fact that the unified
definition agrees with what is expected gives some philosophical
justification for using it for the infinite case where we lack an a
priori alternative. This is the realm where exotic results occur and, as
mentioned earlier, the case of main interest.
From now on we would like to limit our study to the case
where x is purely infinite, i.e. where all exponents in the normal form
of x are positive. However, before doing so we need a "piecing
together" result.
Theorem 10.4. If x = y+z where y
finite, then exp x = exp y exp z.
Proof. We again use induction.

is purely infinite and

Note that if x has the form

z is

I u> irj,
i<a
then y necessarily has the form
I u ir-j and z the form I u> irj.
i<3
i>0
Hence y is an initial segment of x. We use an argument similar to the
one used in the proof of theorem 10.3 to show that in the representation
of exp x we can restrict ourselves to initial segments of the form
y+u where u is finite.
Regardless of whether z is positive or negative, terms containing exp y contribute to both upper and lower terms in exp(y+z).
Now suppose x-x
is positive infinite. Consider a term exp x [x-x ] .
Note that [x-x ] x [x-x ] by elementary reasoning with orders of
magnitude. Therefore exp x > exp x 1 [x-x 1 ] +1 exp x ^ x - x ^ . A
similar inequality applies if x-x
is negative infinite. On the other
hand, y contributes terms such as exp y[z] or exp y[z]
if z
is negative. Now [z] p _> 1 if z is positive. If z is negative we

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

154

can certainly choose n so that [z]

is not infinitesimal. In fact,
by our earlier discussion [z] 2
can be infinitesimal for at most one
value of n for each z. [z] 2
is certainly finite. Hence such a
term is of the same order of magnitude as exp y. It follows from the
above and the inequalities in the statement of theorem 10.1 that
exp y ~ exp x. Thus y contributes a lower term in the representation
of exp x of the same order of magnitude as exp x. This is thus larger
than exp x [x-x ] by our earlier inequality.
Since a similar argument applies to upper terms this proves
the cofinality.
Now the sign sequence of an element of the form y+u is the
sign sequence of y followed by the sign sequence of u. This follows
from the fact that all exponents in the normal form of y begin with a
plus whereas those in u do not, so that by the sign sequence formula y
does not contribute to minuses which are ignored in the sign sequence for
y+u. Hence the initial segments of y+z are of two types: proper
initial segments of y, and y followed by an initial segment u of z.
Since y is an initial segment of y+w for any finite w,
y+w is certainly not a proper initial segment of y, so the first type
cannot have that form. On the other hand, the second type consists
precisely of all numbers of the form y+v where v is an initial
segment of z. This says that our cofinality result restricts the
initial segments considered precisely to those of the second type, i.e.
those of the form y+u where u is an initial segment of z.
Thus the formula for exp x = exp(y+z) involves terms such
as exp(y+u)[z-u] which by the inductive hypothesis is
exp y exp u[z~u] . exp u[z-u] is, of course, a typical term used in
the computation of exp z. Similar remarks apply to the other terms so
that we finally obtain that exp(u+z) may be represented as
{exp y F}|{exp y G} where F|G is the representation used in the
definition of exp z. (exp y F stands for the set of all products of
exp y with elements of F and similarly for exp y G.) We must prove
that exp y exp z = {exp y F } | {exp y G} which heuristically is a kind
of distributivity. In general there certainly is no such "distributive
law" for multiplication. In the present case we make use of the special
known properties of y, z, F, and G.

EXPONENTIATION
F i r s t , since
and

155
0

i s a lower element of

exp y

then

exp y F

exp y G appear among the terms in the formula for the product

exp y exp z.

I t therefore suffices to show that these p a r t i c u l a r terms

are cofinal with respect to a l l other terms in the formula for t h i s

product.
For t h i s purpose we need some information regarding orders of
magnitude.

For convenience, l e t

used in the d e f i n i t i o n .
segment

of

Since

satisfies

t y p i c a l term of the form

exp z [ y - z ]

H|K
y

be the representation of

is purely i n f i n i t e , any proper i n i t i a l

|y-z|

is i n f i n i t e .

exp y [ y - z ] .

exp z [ y - z ]

exp y

< exp y.

If

z <y

we have a

As we already noted,
Similarly i f

z >y

we have

>> r X j % Z > exp y. Also r e c a l l that for negative i n f i n i t e x

L
yj
n
n+i
there are no terms of the form [ x ]
. Thus we have shown that in the
2n+i

yj

representation exp y = H|K, h e H =5> h exp y

We now look at the representation

and k e K ^ exp y k.

exp z = F|G. Suppose

u < z. Then u contributes terms such as exp u[z-u]

to F and
r,.e*i
to G. If z-u is not infinitesimal, then neither
LU ZJ
~ 2n+i
[z-u]

- [z-u]

nor -r-T
rT
is infinitesimal. Also,
L
J
2n+i
2n+3
exp(z-u) - [z-u] n _> [ z - u ] n + 1 - [z-u] n hence exp[z-u] - [z-u] n is
not infinitesimal.

Similarly

r
L

-i
2n+i

exp[z-u]

is not infinitesimal.

Since all the expressions considered are f i n i t e , we obtain by multiplication with exp u that
exp u ~ exp u[z-u] +1 - exp u[z-u] ~ exp u[exp(z-u)] - exp u[z-u] ~
exp u
]

[u z]

" 2n + i

expu
[

exp u exp(z-u) =

expu

course

, _ v = exp z.

What the above essentially says is

that the various orders of magnitude of differences of elements of

and

G with each other and with exp z

z-u ~ oj ,
z-u.

Then

If

z-u

i.e.

[z-u]

are the same.

is infinitesimal we need more caution.

Suppose

is the f i r s t exponent occurring in the normal form of

- [z-u]

~ u>

By formal multiplicaion

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

[z-u] - U - u ] ~ u
^r

" T^

Hence exp u[z-u]

156

. Again using formal reciprocals we see that

* (2n+2)a and

^ -

- (exp(z-u) ~ *< 2 " + 2 > a .

-exp u[z-u] n ~ exp z -exp u[z-u]

exp u

exp z - exp u[z-u] 2 n + 1 ~ [ u * j ^

exp z.

We have similar results if u > z.

Now we are ready to study the terms in the product
exp y exp z = (H|K)(F|G). Recall that this has the form
{h exp z + f exp y - hf, k exp z + g exp y - kg}|{h exp z + g exp y - hg,
k exp z + f exp y - kf} where h e H, k e H, f e F, and g e G. Recall
also that we desire to prove the cofinality of terms of the form
f exp y and g exp y. First, consider h exp z + f exp y - hf =
f exp y + h(exp z-f). f has a form such as exp u[z-u] n . Suppose
f = exp u[z-u] . Then f' - f ~ exp z - f . Also h exp y.
Therefore f exp y + h(exp z-f) < f exp y + exp y(f'-f) = f1 exp y
which gives us precisely what we need. Now consider
k exp z + g exp y - kg = g exp y - k(g-exp z ) . g has a form such as
F _XS
If we let f = exp u[z-u]
then we have
g - exp z ~ exp z - f . g - f ~ g-exp z. Also exp y k. Therefore
g exp y - k(g-exp z) < g exp y - exp y(g-f) = f exp y.
The upper terms in the representation of the product can be
handled the same way. For this purpose express h exp z + g exp y - hg
in the form g exp y - h(g-exp z) and k exp z + f exp y - kf in the
form f exp y + k(exp z-f).
This completes the proof.
We are now ready to study the purely infinite case.
B SPECIALIZATION TO PURELY INFINITE NUMBERS
We begin by inviting the reader to share with us the pleasure
of the simplification obtained in the case of purely infinite numbers.
Recall that a purely infinite number is a number in which all
the exponents in its normal form are positive. The number itself may be
positive or negative. Also 0 is a purely infinite number! This may

EXPONENTIATION

157

sound strange at first but it is consistent with the usual meaning in

logic of universal quantification over the empty set and is natural for
the theoretical development. Such numbers can be characterized also in a
way which does not mention the normal form. Just as surreal numbers of
the form o>a are the elements of smallest length among numbers of the
same multiplicative order of magnitude, purely infinite numbers are
elements of smallest length among numbers of the same additive order of
magnitude. Although this point of view does not play a significant role
in our work it does add a certain beauty to some later theorems such as
theorem 10.7.
If x is a purely infinite number then we already know that
every initial segment y satisfies |x-y| is infinite. Hence no terms
of the form [z]
where z is negative arise in the definition of
exp x. Thus lower initial segments of x do not contribute to upper
terms in the definition of exp x and vice versa. Therefore we have
exp = {0, exp x'Ex-x'imfffl K }. Also for positive infinite z
n
z n + 1 > [z] and [z] > z . Hence by mutual cofinality we may
simplify the representation of exp x to {0, exp x'(x-x')n }\{ex [

}.

\X ~X/

Thus we no longer have to bother with sums of the form [x] .

We now go one step further by showing that we need consider
purely infinite initial segments only. In fact, suppose x1 = y+z where
x1 is a lower initial segment of x, y is purely infinite, and z is
finite. Since x-x1 is infinite it follows that y < x. Also y is an
initial segment of x' hence also an initial segment of x. We claim
now that terms in the representation of exp x contributed by x1 are
mutually cofinal with terms contributed by y. First
exp x1 = exp(y+z) = exp y exp z ~ exp y since exp z is finite. Now
consider an arbitrary expression of the form (u+v) n where u is
positive infinite and v is finite.
If v is negative then certainly (u+v) n u11 u n + 1 . If
v is positive we have:
n
.
.
n
.
n
v /nx l n-i r /nn n-i
rir r / n * n - i i
n+i
), ( . ) u v
< I (.)u v
= u I ) (.)v
I u
1
]
1=0
" i=0 ^
i=0
So in e i t h e r case (u+v) n u n + 1 .
I f we apply this to the above we
/ , xn
(u+v) =

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

obtain

158

exp x ' ( x - x ' ) n ~ exp y ( x - x ' ) n = exp y [ x - ( y + z ) ] n = exp y ( x - y - z ) n

exp y ( x - y )

n+1

1 11

Hence exp x ' U - x )

< exp y ( x - y )

applies to upper i n i t i a l segments.

n+1

A similar argument

We state the f i n a l s i m p l i f i c a t i o n as

theorem 10.5.
Theorem 10.5.

For purely i n f i n i t e x

{0,exp x ' ( x - x ' ) n } | { ^ x j | * } where

\ x ~x /
lower i n i t i a l segments of x and
upper i n i t i a l segments of

is,

has the form

x1

runs through a l l purely i n f i n i t e

runs through a l l purely i n f i n i t e

x.

As an example consider
i n f i n i t e lower segment.

exp x

exp u>.

Therefore

of course, d i f f e r e n t from

is the only purely

exp w = {0,OJ }|<j> = w .

t h i s i l l u s t r a t e s that

Since w

to

as defined

i n chapter f i v e should not be regarded as exponentiation in the sense

that is being developed here.
as the value for

exp (x in w)

We shall see l a t e r what arises naturally

a f t e r we study the function

A similar r e s u l t applies i f

n.

is expressed in a form

which is not necessarily canonical but some caution is required.

that

y e FUG=^|x-y|

of an element in
an element
y-z

is i n f i n i t e .

FUG

In general the purely i n f i n i t e part

need not be in

FUG.

[By the i n f i n i t e part of

is meant the unique purely i n f i n i t e number

is f i n i t e . ]

However, by the uniformity theorem for

s t a r t with using

F and

such that

exp x we can

G and s t i l l do the same s i m p l i f i c a t i o n as in

the canonical representation but instead of taking subsets of

the end we use the i n f i n i t e parts of a l l the elements of
lower terms and of

F|G
Suppose

FUG

at

F for the

G for the upper terms.

In future

x, y, and

w i l l refer to purely i n f i n i t e

numbers unless s p e c i f i c a l l y stated otherwise.

For reference we state two

i n e q u a l i t i e s which are similar to ones we had e a r l i e r .

Assume

and y

are p o s i t i v e ;
,
2n
n n
(x+y)X
> x y
x

n+i n+i v /
n
y
> (x+y)N .

The first is clear since by the binomial theorem

terms in the expansion of (x+y) 2 n .
term

("jxV"""

( )x y is one of the

The second is clear since a typical

in the expansion of (x+y) n

satisfies

EXPONENTIATION

159

("jx 1 ^"" 1 x n + 1 y n + 1 . Although the latter inequality may be strengthened there is not much point to i t for our purpose. ( I t is much like
showing that one can use -JTT instead of TT in analysis!)
Recall that by theorem 10.1, i f x < y, then
exp x[y-x] n < exp y. Hence exp x(y-x) < exp x[y-x] n + 1 < exp y. Thus
we may replace the term [y-x]
by (y-x) n in the inequality stated as
part of the above theorem. Since y-x is infinite, this shows
immediately that x < y + exp x exp y. As we recall, the function OJX
has this property, but there is a crucial distinction, namely for </
this applies to the whole domain, but to exp x this applies only to the
subclass of purely infinite numbers. In fact, i f y-x is finite then
x
y
e ~ e . At any rate, because of the above property the proof of the
addition formula for exp x resembles the proof of the corresponding
formula for
u .
Theorem 10.6.

exp(x+y) = exp x exp y.

Proof. We use induction. Without loss of generality a typical element

in the representation of x+y has the form x + y. Since x is purely
i n f i n i t e , |x-x| is infinite, hence so is |(x+y)-(x+y)|. Therefore in
the computation of exp(x+y) we may use the infinite parts of elements
such as x+y. Since y is purely infinite the infinite part of x +y
is y plus the infinite part of x. Also, as x runs through all
proper i n i t i a l segments of x the infinite parts of x run through all
proper purely infinite i n i t i a l segments and the infinite part of x is
on the same side of x as x . This permits us to write exp(x+y) in
the form {0,exp(x'+y)[(x+y)-(x'+y)] n , exp(x+y')[(x+y)-(x+y')] n }|
exp(x+y")
}
{
,
")()lnr
T
n
[(x +y)-(x+y)]n
[ ( x +y")-( x +y)ln
{0,exp x [ x - x ' ] n } l1{ e x p

and s i m i l a r l y for

y.

Using the inductive

(x"-x) n
hypothesis, a typical lower term simplifies to exp x1 exp y(x-x')

exp y 0 i s , of course, also a lower term.

(x"-x) n
These are terms occurring in the representation of exp x exp y since
0 is a lower element of both x and y. Thus, as in the proof of
theorem 10.4, i t remains to prove that these terms are cofinal with
an upper term to

exp X

and

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

160

respect to the other terms in the formula for the product. If x or y

is 0, there are no other terms so the result is clear. (That case is
trivial anyway and could just as well be eliminated in advance.)
First, we have a lower term such as exp x exp y'(y-y') +
exp y exp x'(x-x') n - exp y'(y-y') m exp x'(x-x') n . Suppose
exp x exp y'(y-y') exp y exp x'(x-x')11. Then
exp x exp y'(y-y') m + exp y exp x ' U - x 1 ) " - exp y'fy-y 1 )" 1 exp x'(x-x') n
exp x exp y'iy-y1) + exp y exp x'(x-x') n 2 exp y exp x'(x-x') n
< exp y exp x 1 (x-x 1 )n+i
A similar result applies if
n
exp y exp x'(x-x') _< exp x exp y'(y-y') m .
Now consider a lower term such as
exp x exp y" exp y exp x" _ e*
We know that
(x"-x) n
(y"-y) m (x"-x) n
(y"-y)
exp x"
exp y" exp x"
> exp x hence

e xp x.
(y"-y) m (x"-x) n
(y"-y)
Similarly
exp y.
(x"-x) r
(y"-y) m (x"-x) n
Therefore the above expression is negative, so that for cofinality it
suffices to take the term zero.
Finally we must consider a term such as
exp x"
exp x exp y'(y-y') m + exp y exp x n - exp y'(y-y')
Now
M
n
(x -x)
(x"-x) n
exp y exp x
Therefore certainly
exp y,
l)m
(x"-x) n
(x n -x) n
exp x exp y'fy-y1)111 + exp y

(xH-x)n
L > \ exp y
^
(x M -x) n ~~ *
(x"-x)

(xM-x) n+i

A similar result applies if we interchange the role of x and

proves the cofinality and completes the proof.
Corollary 10.1. exp(x+y) = exp x exp y

y. This

for all x and y.

Proof. This follows immediately from theorem 10.6 with the help of
theorem 10.4 and the corresponding result for finite x and y.

EXPONENTIATION

161

Corollary 10.2. exp x exp(-x) = 1.

Our next aim is to show that as x runs through all purely
infinite numbers, exp x runs through all numbers of the form u>a where
a >_ 0. (Of course, if a = 0 we obtain exp 0 = 1 = u> so our main
concern is with a > 0.) One direction is almost immediate. The other
direction will be proved by means of the function Jin x.
Theorem 10.7. exp x is a power of

GO.

Proof. We already know that every lower term such as exp x'(x-x)
satisfies exp x'(x-x') exp x and similarly every upper term such
satisfies exp x exD
^ x" Hence every number y of the
(x"-x) n
same order of magnitude as x also satisfies
exp x'(x-x') n < y < e x P x Therefore x is an initial segment of y.
(x"-x) n
So by theorem 5.3 x is a power of w.
The above result is especially striking if one recalls that
purely infinite elements are canonical elements in an additive order of
magnitude. The theorem says that such elements are mapped into canonical
elements in a multiplicative order of magnitude. This is further
heuristic evidence that exp behaves the way an exponential should.
We now define in x. This definition will be made only for
x of the form co .
b"-b
Definition.

n(w ) = {in{m

)+n, n(w

) - OJ

}|{n(w

)-n,

b-b1
n(w

) + a)

where

n runs through a l l positive integers.

Note that lower i n i t i a l segments in

elements in the representation of

n(a) ).

b contribute to upper

This is essential.

For

example, l e t

b = 1.

Then

upper terms t h i s would be

definition

in x

in oo = {n(oj) }+n| {n(o))+(o) 1 ) n } .

{n}|(j> = GO.

Since

exp w = w

would not be the inverse f u n c t i o n of

o t h e r hand the given d e f i n i t i o n leads to

does seem more reasonable a priori.

with

exp x.
i_
l

in to = {n}|{u> } = w .

In fact, let us compute

With no
this
On the
This
i_

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

i
= {O}|{-}.

w^ = {0,n}|{u) n }.

Then

Hence by r e s t r i c t i n g ourselves to

7F
exp (to") = {(exp O)a) u) }H ex P (a) 1 ) }.
{ n *]m

i n f i n i t e parts

uo

to .

_L

< (u)n)m = /

for some real number r

i -

infinite i f

(in f a c t - ).
n
H

> a)"

> a).

Now

a) = {n}|t|).

Since

to03 < oi

and

= (to )

(J -^)"

exp(gj n )

Now by the a d d i t i o n formula exp (to ) = (exp to)

11 A 1

Also

The lower terms are

-to J

n
_
simply

162

= to .

So

is

n ^ 1, the conditions of the cofinality theorem are

i_

s a t i s f i e d and

exp(co ) = to.

I n c i d e n t a l l y , the reader may be i n t e r e s t e d in experimenting

w i t h various other computations using the d e f i n i t i o n .

We prefer to

postpone a discussion of e x p l i c i t r e s u l t s u n t i l we have proved some

remarkable r e s u l t s which w i l l
still

s i m p l i f y the computation tremendously.

have a distance to go before t h i s .

know t h a t

in x

Theorem 10.8.

is

We

For example, we d o n ' t y e t even

defined!

in x

is defined.

Furthermore, i f

a > b then
a-b

&n(to ) - n(to )

is positive i n f i n i t e and &n(to ) - Jtn(to ) < u)

a l l positive integers n.

In particular, i f

a > 0

then

for

in{u )

is

positive i n f i n i t e and n(toa) < to11.

Proof.

I t is no longer surprising that we use induction. We now show

that the lower terms are really less than the upper terms. Since
h"
h1
h1
h"
n(to ) - Jtn(to ) is i n f i n i t e , then certainly n(to )+n < n(io )-m.
b-b1
Since

to

is i n f i n i t e then certainly

an(u>

b-b'

)+n < n(a>

) + to m

However, such an inequality is not adequate for our purpose since we must
consider two arbitrary i n i t i a l segments of
(The use of the same
incomplete p r o o f . )

b not necessarily a l i k e .

b-d
b ' , which is a natural mistake, leads to an

So we must show t h a t

n(to )+n < n(oo ) +

m
0)

EXPONENTIATION
where

and

163

are two lower i n i t i a l

segments of

b.

This is

c-d
if

d 2. c.

If

c > d

n(toC) - (tod) < to

we have

trivial

b-d

< to

-n.

Hence

b-d
&n(ioc)+n < ir\{ud)

+ to

Similarly for upper i n i t i a l

segments

and

c-b
d

of

n(toC) - a>

we must show that

< n(to C )-n.

As before this is
c-d

trivial

if

c < d.

If

c > d

we have

c-b

n(to ) - n(w ) < w

< u>

-n

"~
n(<oC) - u> m

hence

< n(to ) - n .
b"-b
bb

I t remains to prove that

n(to
in

By the inductive hypothesis we have

b"-b

Now to 2 n

n.

(o n

(w
) - n(w

n(co

b"-b b-b' 1

= [(o

b"-b

<i[u>

bb-b'
n(<o
< n(<o

) - to

b"-b'

ii

) << w

+ to

Since

to

by the arithmetico-

b -b
r = max(m,n)

/ b \
n (to

we have

/bv
)

in (to

to

b"-b
n
)

b"-b
_< to

+ ui

then

/ b \
in (to

Hence
b"-b
n

so

b-b'
+ to

b-b'
m

^ (0

Hence

is a decreasing function of

for

for a l l

b-b'

+ a) n

geometric inequality which is valid in any ordered field.

b"-b
b"-b
b-b'
x
to

) + to

b-b'
m

/ b \

to

^ &n(to

+ to

We now know that in (to ) is defined. For a > b i t is

immediate from the definition that n(tod) - n(<o ) is infinite i f one
of a and b is an i n i t i a l segment of the other. In the general case
i t suffices to consider the common i n i t i a l segment of x and y since
the sum of two positive infinite numbers is clearly infinite.
For the final inequality let c be the common i n i t i a l
segment of a and b where a < b so that a < c < b. (Again i f
c = a or b the result is t r i v i a l . ) Then from the definition we have
c-a
b-c
n(aj a ) > in(to c ) - to

b-c
*,n(to b )-n(to a ) < a)

n(tob) < &n(o)C) + to

and

c-a
+ to

b-c
.

Now

to

Then

b-a

to

c-a
and

to

b-a

to

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

b-c
wn

Therefore

c-a b-a
+ to n < w n .
b-a

Jln(u) ) - n(ooa) < oa n .

This finally gives us the inequality

This completes the proof.

We would like to show next that

function of

exp x where the domain of

infinite numbers.

Theorem 10.8.
Proof.

Let

n(co )

)+n

be the representation of

We claim that
we get

) - oo

terms.

given by the

c e F =*> (3deF)(d>x+l)

This is easy to see.

d = n(oj

) + (n+1)

If

and i f

and
c has the form
has the form
b"-b

second case

b.

Jin(u) )

b"-b
n(u>

is the inverse

is restricted to the purely

is purely infinite for all

c e G =* (3deG)(d<c-l).
n(u)

Jln(x)

exp

For this purpose we f i r s t need the following theorem.

F|G

definition.

164

then we may choose

d-c

is i n f i n i t e . )

By induction

d = n(u>

) - oo n

(In the

A similar argument applies to upper

1 can be replaced by any positive integer.

This is enough to show that

F < an(o> )+r < G for any real

r.

of its purely infinite part, so

n(oo )+r

Hence

n(a> )

n(a> )

satisfies
is an i n i t i a l segment

must itself be purely infinite.

We are now ready for the theorem which completes the basic
theory.
Theorem 10.9.
Proof.

exp

nn(oo ) = to .

Again we use induction.

,
n(oo ) = F|G

G = {iin(u>

b-b
n

h1

h"

)-n, n(to

c e F = n(to )-c

and

)+n,

Jtn(oob ) - to

"~b
n

and

) + to

c e F = (3deF)(d>c+l)

F = {n(oo b

where

}.

We already know that

c e G => (ideG)(d<c-l).

is infinite

and

d e G=^>d-iin(to )

This implies that

is infinite.

Thus

by our earlier remarks we may use the infinite parts of the elements in
F

and

G to compute

positive power of

exp n(to ).

By theorem 10.8 and the fact that any

to is i n f i n i t e , the infinite parts of the elements in

b"-b
.ii

.i

have the forms

n(oo

and

n(to

) - to

, and s i m i l a r l y those

EXPONENTIATION

165
b-b1

in

G have the form

n(aj

typical lower terms besides

and &n(u>
0

in the formula for

Therefore the
exp[fcn(o) ) ]

have

expUn(u>b ) ] U n ( a j b ) - n U b ) ] n and

the form

bM-b
exp[<w(u

) + w

)-u

b"-b
K

l{nn(u) )-[itn(co

K"

)-u> n

]} , and similarly the upper

b-b'

terms have the form

^ " ^ " ^

and

We use cofinality to show that this gives rise to

is

xp[tn(^)t. "

u>

which

{0,A}|{a)b"s}.

First, we check the betweenness condition.

h1
h
h' n
By the inductive hypothesis exp n(oo )[n(u> )-n(oi
)]
b-b 1
b / b / b -in
b ^ n >.n
b b-b
b
. .. _
= a rLn(a) x)-n(w x) ] < a) (a)
J = w a)
= u) . oS i m i l a r l y
LII

LM

LII

)]

J2
b\

* bA1n

>

2
b -b

b
w

terms are not as hard to deal with as i t may appear.

b"-b

(i

b"-b

)" \

iim

bM-b
- [*n(u)b ) - a) n
expU

)
b"-b

11

)m

s1nCe

b"-b
,n+m

""

b.

^=

^ b -

= l b

Similarly

Fortunately the other

First,

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

b-b 1

b-b'
]

explwi(co ':

to

b<

b-b

b-t

[Un(to b Vto
b-b1
tob'(o)

b-b
la)

*n(.b)] m

'

(0

^t n ( t o b ' ] l+co
b

b-b'
=

(0

0)

This

b-b'

to

- ^r l ( o ) D ]

n ~\

explco
b-b1
r n >

verifies

166

k im

the betweenness cor

The cofinality condition is easier to check since the simpler

lower and upper terms are all we need for the purpose.
h'

o) r

is a typical lower term in the representation of

One of the lower terms found in the representation of

b

expUn(to

)]Un((o -*,n(to

n(o) ) > n(w )

is i n f i n i t e .

)] .

So we may take

is

The f i r s t factor is

and both are purely i n f i n i t e .

n(to )

<o .
Hence ,n(to ) - n(w )

n = 1 and we have an element which is

L I

larger than

o> r

to .

L II

since

i s , of course, f i n i t e .

a typical upper term in the representation of

(o) )
]
b
"

to .
i$

Similarly

o> s is

Here we consider
(Again

1nfinl-tes1mal#

Un(co D )-*n(coD ) ] n
h"

we may use
complete.

n = 1.)

This is smaller than

to s. The proof is now

It is immediate from theorem 10.9 that the function exp x

defined for x purely infinite is onto the class of numbers of the form
w d . By theorem 10.4 and earlier information on exp x for finite x it
follows that the range of exp x where x runs through the class of all
surreal numbers consists of all numbers of the form u> b where b is
the class of all positive finite non-infinitesimal numbers. This is
precisely the class of all numbers of the form
s i n c e t h e l a t t e r can be e x p r e s s e d i n t h e f o r m

w i'bj with
i<a
a
o> ( b o +

b0 > 0

c.
"'b)

l
This is in turn the class of all positive surreals. Thus we have
Corollary 10.3. exp x is onto the class of all positive numbers.
We now know that exp x behaves as it should with respect to
the main properties expected of an exponential function. Thus we have
good philosophical grounds in designating our function as "the

EXPONENTIATION

exponential f u n c t i o n . "

167

Squeamishness about proper classes can be handled

by the techniques of chapter s i x .

With the basics out of the way we are now ready to move
towards the e x o t i c surprises of the exponential f u n c t i o n .

REDUCTION TO THE FUNCTION

The central problem of i n t e r e s t can be expressed as f o l l o w s .

Given

x =

I w "ir-j, what can we say about exp x? I . e . we desire to be

i<a
as e x p l i c i t as p o s s i b l e . We repeat t h a t our i n t e r e s t i s l i m i t e d to the
purely i n f i n i t e case since we already know what goes on i n the f i n i t e
case.

I t turns out there i s much t h a t can be said but the subject is

non-trivial,

i.e.

there i s no single theorem which closes the subject as

r a p i d l y as i n the cases where

i s real or i n f i n i t e s i m a l .

tremendous amount of f u r t h e r s i m p l i f i c a t i o n i s p o s s i b l e ; however, i n

s p i t e of t h i s there are enough complications remaining so t h a t the
subject remains s u b s t a n t i a l .

(This means t h a t we d o n ' t go out of

business by exhausting the subject too r a p i d l y . )

We are f i n a l l y ready f o r the f i r s t of the b e a u t i f u l
Theorem 10.10.

Proof.

If

a > 0

exp(w a )

then

We do not use i n d u c t i o n ! !

has the form

the canonical representation of

we express

as {0,a>a r}|{u)

have i n f i n i t e distance from

the computation of

exp(w ).

<*> .
So

b
ww .

Since we are studying purely

numbers only a t t h i s time, the condition

a > 0

results.

i s understood.

infinite
Hence in

all

terms are non-negative.

s}

all

terms are purely i n f i n i t e and

Thus i f

Therefore we may use these terms for

exp (to

is

a"
{0,(exp Q)(a) ) ,[exp(a) 'r)](u) -a) r) H{ /, a, ^ s)
a \}.
n We now use mutual
(a) S-U) )
a n

a>

exp ( aj

cofinality to simplify. The term 0 is superfluous because of the

second group of lower terms which are of the form w . Now
(o)a-a)a'r)n+1 > ( w a ) n > U a - i / ' r ) n and
(u) s-u) )
> (co ) n > (w s-u> ) . (We may just as well assume that
s _< 1 by cofinality or alternatively we can use the inequality
(u)a ) n + 1 > U

s-w a ) n .) At any rate, the representation for

exp((oa)

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

simplifies

{(to n a ,[exp(to a > r)]to n a } | { e x p ( l s ) } -

to

to

possible value of a 1
a

exp(u) ") > (a) "f)

"!-) =
^

exp(a)

,,s) <
( a s ~\

a)
exp (to

is a

> to ".
-

s).

exp x > exp 0(x ) = x

Thus

Hence t h e u p p e r

So

t e r m s may be

limited

exp(u) ) s i m p l i f i e s f u r t h e r

to

Also we may l i m i t r to integers and s

a'
a'
r
to dyadic f r a c t i o n s .
Hence we may w r i t e exp(w r ) as [exp(o) ) ]
and
exp(w a s) as [exp(to a ) J S .
a1
a"
By theorem 10.7 exp (to ) and exp (to ) are powers of to.
(Note t h a t we are not assuming the stronger f a c t t h a t exp(w ) and
a"
exp(u) ) are i t e r a t e d powers of a> as we would in a t y p i c a l proof by
induction.)

}|{exp(o)

na

s )

,,
na

exp[o) j)
r)]o)

Since in general

n+1

exp(a)

t o terms of the form

{[exp(w

Since

and 1 a possible value of r the first group of

lower terms are superfluous.

we have

na

168

Let

exp (to

s)}.

) = a>

and

exp(u)

) = w

Then

exp((o )

can be expressed as
b'r

na , , b"s

Note t h a t we use the a d d i t i v i t y

justify

the above.

b'r

/n

{0,<O

Since the set

na.., b'V
0)

{b'r+na}

} |{0)

b = {b'r+na}|{b s}.

,n

b'r+na,,,

} = {0,0)

exp x

and

to

has no maximum and b s

to,

b"s,

}|{U)
M

may be expressed as a power of

ll

theorem f o r both

We may j u s t as well w r i t e t h i s as

specifically

}.

has no minimum, this

w

where

I t is redundant to try to show that

b'r+na < b"s

since we began with something which was defined and used mutual
c o f i n a l i t y at every step, so we already know that to
Consider the representation
b'r+na < c < b"s
i s a power of

hence

w.

Let

b = {0,b'r+na}|{b s}.

< to
If

b is an i n i t i a l segment of
b = toC. Then

b~ c
c.

then

Therefore

a = ww .

By theorem 10.10 exp x

induces a function g with domain

g(a)
consisting of a l l positive surreal s such that exp (to ) = tow
. We now
use the above proof to obtain an inductive formula for
purpose we define
v *\*
I u irj then
i

ind a as follows:

ind a = b 0 .

If

g(a).

For this

has the normal form

b 0 can also be characterized as the unique

EXPONENTIATION

surreal

such t h a t

169

a ~ u)X.

We also use the temporary notation

exp(uj ) = /

( x )

f ( 0 ) = 0.

From the above proof we have using t h i s

for

x > 0

so t h a t

f ( a ) = { r f ( a ' ) + n a } | { s f (a 1 1 )}.
case

f(a')

= 0,

by c o f i n a l i t y .
0

f(a) = u>g(a).

x = 0

we set

notation
a1

Since one value of

is

in which

f ( a ) = { 0 , r f (a )+na} | { s f (a 11 )}

we may c e r t a i n l y w r i t e
Now among a l l

If

a1

the elements of the form

we single out

and w r i t e t h i s as
f(a) = {0,na,rf(a')+na}|{sf(a11)}.

Of course, there may be no other

a1

j u s t as there may be no

a"

altogether.
We may now replace
x > 0.
i.e.

So we obtain
a ~ u)C.

Then

equicofinal with

ru

f(a)
Since

f ( a ) = w9

f(x)

by

f(a) = {0,na,ruj9(a
na

or
}

+na}| {su) 9 ( a

is equicofinal with
+ nw

which i s v a l i d for

nwC

}.

and

Let

rou9

which i s in turn e q u i c o f i n a l

we can now determine

is

with

g(a).

In f a c t ,

which i s the same as

{c,g(a )}|{g(a")}

by c o f i n a l i t y .

We have thus ended up w i t h a rather remarkably

simple-looking

However, the reader i s warned t h a t the presence of the term

gives r i s e to t r i c k y phenomena.
prevents

+na

={0)nuc,numax[9(a')'c]}|{soo9(a")}.

g(a) = { c , m a x [ g ( a ' ) , c ] } | { g ( a " ) }

formula.

c = ind a,

from simply being the i d e n t i t y f u n c t i o n as we shall see.

In

can take on negative values as we shall also see. This is

x
c o n s i s t e n t w i t h what we have since w"
is positive i n f i n i t e for a l l

x.

fact,

For example, the presence of

Note t h a t there i s no zero among the lower terms in the expression f o r

g(a).

This i l l u s t r a t e s t h a t in general the i n c l u s i o n of

among the

lower terms must be treated w i t h caution to ensure t h a t i t i s l e g a l .

In view of i t s importance for the r e s t of our work we state
the formula f o r

g(x) as a theorem.

a
u> g ( a )
Theorem 10.11.
I f a i s p o s i t i v e exp(w ) has the form w
where
g(a) = { c , g ( a ' ) } | { g ( a " ) }
and c i s such t h a t a ~ u> . [ I t i s under-

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

stood t h a t

a1.]

i s not a possible value f o r

L e t us a p p l y theorem 1 0 . 1 1 t o o b t a i n
evaluate

exp a>.

Since

w = w

c = 0.

I n t h i s case we r e g a r d

and

i s separated o u t .

we have
A

Hence

and

170

a = 1.
A"

as

exp(u) W ).
Since

F i r s t we

w = 1

<>
j since

g ( l ) = {0}|<}> = 1 .

we have

1 = {0}|<j>

Thus

g(l) = 1.

Therefore

exp u = w

the earlier methods.

g() = .

=w

which agrees with the result we obtained by

We use induction to obtain

g().

Assume

Then

= q f { 0 } | { ( 4 : ) } l = lind ( I - ) } | { g ( i - n = {0} I { - U = 4 r r r .
fore

g(^) = g [ { 0 } | { - ~ } ] = { i n d ^ K g ^ ) }

= {-1>|{|^> = 0.

There-

Hence

o
expdo^) = OJW =w

power of

which again agrees with what we obtained earlier.

a
We show next a kind of converse; namely that in{^
) is a
co for a l l a. In a way this is redundant since i t w i l l also

follow from later results.

However, the proof is of interest since i t

leads to an e x p l i c i t inductive formula for the inverse of

g.

For this

we need a lemma which states that the uniformity theorem is valid for the
function

n x.

This result probably deserves to be a theorem but since

we have no lemmas in this chapter so far and since we are thus already
top heavy with theorems we leave this as a lemma.
interested primarily in the function

Besides, we are

exp x, so that the function

is introduced primarily to help us prove results about

exp x

is onto the set of a l l positive surreals.

Lemma 10.1. The uniformity theorem is valid for

Proof.

n x

exp x, e.g.

Recall that this means that we can obtain

representation of

in the form

to the canonical representation.

a = F|G

n x.
n(a) )

by using any

instead of being restricted

As in a l l proofs of such theorems in

which we use the inverse cofinality theorem followed by the cofinality

theorem we need inequalities of a certain kind.
K

1
x

/b"

) + n , n(co
If

b' < x < b

b"-b
n 1

) - oo
then

i r /

b " x

In this case we have

, b \

b-b1
n i

}|{n(a)
) - n , n(w )+io
x
b1
n(w )+n > n(w )+n and

}.

EXPONENTIATION
b-x
n ,

/ xx

icn v a) ; + 03

j + 03

b < y < b"

y-b
n

/ Vv

b-b 1
n

^ ion 103

If

n(co

, b\

171

>

n(o3

)+n

and

b"-b
n

/ b"

03

Jtn(o3y)+n <_ Z n U

then

- 03

The first inequalities of each pair are obvious since &n(o3X)

is an increasing function of x. The proofs of the second ones in each
b-b'
b-x x-b1
b-b'
pair are s i m i l a r .

03 n

First

- 03 n

> 03

since

b-x
03 n

higher order of magnitude than e i t h e r

or 03 n

x-b'
03

> n(oj )

- n(o3

b"-b
Similarly

O 3

).

- o >

>

Jcnio3

Hence

> o 3

n(o3 ) + a>

b-b'
< nn(o3

) + 03

n(o3b ) - Jtn(o3 y ).

>

Hence

b"-b
n

- 03

For a l l

a,

a
>m(o3W )

is a power of

03.

The lemma allows us to apply the formula in the d e f i n i t i o n to

the r e p r e s e n t a t i o n

{0,w

}|{u>

Hence
e

} o f 03 .

a"
a'
a

^ )

11
c 5\

03

j-n,

/ 0 \
x,n{03 j+03

an integer we have

n(o3W

03

We now simplify.
0, thus getting

)+n,

- U

a"
a
r

z' w
x,n[o3

v >
' j + 03

- co

a
-03

As a s t a r t we can get r i d of

Jin(03 ) , which

as a lower term. Since we may assume that

a1
a'
n(o)W r ) = rn(o3W ) . Note that the addition

formula for the function

for

Also

b"-y

/ bMx

03

Theorem 10.12.

Proof.

y-b

y-b
n

/ Y\

is

has a

b-x

>inio3

03

x-b'

in x

since we already know that i t is the inverse function.

a'
x > 0 &n(o3X) is positive i n f i n i t e .
Hence inl^
) is
a1

i n f i n i t e so

rn(o3

a'
)

is

follows from the corresponding property

exp x

that i f

i s mutually cofinal with

rJm(o3W

)+n.

Note

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

Now consider

172

a" _ ^a
which occurs as an exponent.

a"
a" a
a
dearly - ^ < " *"" - Since i t n ^ s) < u n
this shows that all
the lower terms in the third group are negative and hence can be discarded because of the presence of n. So the set of lower terms have
a
been simplified to {n,rn(a)a) )}. Similarly for the upper terms we
may assume that s is dyadic and drop n so that the first group of
a1
terms simplify to sln^u* ). The second group of terms have the form
a
a a
a a
^n~
- 'r
'r -a
a) . Now consider -^
which occurs as an exponent. -p- >
a1
W r
Since n(a)
) > 0 we may discard the third troup of terms by
cofinality because of the second group. Thus the set of upper terms
a
a
a
II

GJ__

0)

0)

simplify to { s J l n ^ ), w n } . Again, since w11 < sw n we may again

by c o f i n a l i t y s l i g h t l y unsimplify the set of upper terms to
a

The final representation now exhibits

of the form
then

</. In fact, if we define

a
h(a) = {O,h(a')} |{h(a"), ^ } .

a
n(u)W ) as a surreal

h(x) so that in(u* ) = w h ( x )

Note that h(a) > 0 for all a. This shows that the range
of g consists of the class of all surreal numbers. Thus exp x
induces a map from the class of positive surreals onto the class of all
surreals! This sounds like the opposite of what an exponential function
does but recall that the map goes from exponents to iterated exponents so
that there is nothing unreasonable about that, although the existence of
such a g is a surprising phenomenon.
The uniformity theorem is valid for g and h. This is
clear since x _< y * ind x _< ind y. We are primarily interested in g
and it is thus convenient to know that we may use any representation of
b to compute g(b) from the formula given by theorem 10.11.
We next prove that exp x has a kind of generalized
linearity property. This will enable us to reduce the study of exp x
for purely infinite x entirely to the study of the function g.

EXPONENTIATION
Theorem 10.13.
exp

If

173

a-,- > 0 for a l l

J w i'rj = coY where

y = a>

then

i r-j.

i
Proof.

Since

exp(x+y) = exp x exp y

clear for f i n i t e sums and r a t i o n a l

and u>

r-j.

= w O3J the r e s u l t is

We shall prove f i r s t that the

r e s u l t is v a l i d for monomials for a r b i t r a r y real

rj

and then prove the

r e s u l t for a r b i t r a r y sums.
war

Consider
expressed as

where

{oo r'}|{o3 r"}

is real.

toar

Then

may be

where a l l elements of the form

r1

and

are dyadic, in p a r t i c u l a r r a t i o n a l , so that the r e s u l t is known for

r"
u> r 1

exp(o>ar) = {0,exp(a)V ) U a r - a ) V )"} | { e5

5 p (a} Pn}.
a
r"-aj
r)
u)
r"
Since the set of elements of the form r 1 is non-empty we can eliminate

and

0.

wV.

Since

Then

( A V r

) n + 2 > O3a(n+1) > U a r - u > V ) n

we may simplify the

set of lower elements by mutual c o f i n a l i t y and similarly for the upper

elements so that we obtain
r'

= {0,03

na11 r a)
} | {{o3
o

03
0
3

r"

}|{o3

-nai

03

g(a)
naJU^
r"9(a)
Now

exp(oj a ) > 03na,

general.
tt9(a)

i.e.

> (J-.ja,

i.e.

g(a)
r
O3W

{or

/ v
o ) 9 ( a j > na

thus

in

r'}|{o3

. g ( a ) r - O3 9(a) r' > na

Similarly

hence

u> g(a) r M - na > O3 9(a) r.

This proves

s a t i s f i e s the betweenness condition with respect to the

representation of
1 a

na}.

Since this is true for a l l positive integers we also have

co g ( a ) r' + na < O3 9(a) r.

that

> O3na

Now 039

exp(o3 r ) .

r"}.

can be expressed as

Since the lower terms have no maximum and the

uppper terms no minimum we have

03
03

Certainly

g(a)
u

g(a)
Y

= {0,

03

g(a)

r ' + na >_ u 9 ( a l ) r '

ry ' " | r O 3

g(a)
'

{03

and

g(a)
u

v l

r " - na < u g ( a ) r "

so

c o f i n a l i t y is immediate, completing the proof in the case for monomials.

For arbitrary sums we use induction on

a.

For non-limit

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

174

o r d i n a l s the r e s u l t f o l l o w s immediately from the a d d i t i v e p r o p e r t i e s of

the f u n c t i o n s

exp x

ordinary a d d i t i o n .

and

03

since the problem reduces to one of

For l i m i t o r d i n a l s we need an argument which i s

roughly s i m i l a r to the one used before f o r monomials.

where

y < a

and

is positive.

Hence

expH u ai H) = {0, expH u ^ - i A o t / V ) " } I

{exp( I 03 irf+u)
a

)(o3 V ' ) ~ n }

2,0) T r j
Y<i<a

e1

is such that

and s i m i l a r l y f o r

are i n f i n i t e s i m a l .

g ( a

where

. . a .

03=o)I+
Ie-e"I

g ( a

i .

e".

|e-e'|

and

Also
g(a

Y}

9(ai)

9(a Y )

Since the lower terms have no maximum and the upper terms no minimum we

have

03

= {0,03 } | {03 }

terms and

where

stands f o r the set of lower

G f o r the set of upper terms.

We now use c o f i n a l i t y to show t h a t the representation of

exp( I 03 "ir-j)

does give w

First we verify the betweenness condition.

ity terms such as
ur

> na

Y > na

(03 e ' ) may be replaced by

for all integers

n. Since

Y
03

u
where

By mutual cofinalNow we know that

is not infinitesimal we have

for all integers n. Now a typical term among the lower

d

terms of

exp( I 03 1 r 1 -)

is

d *Y

exp( [ u ^ j - m

9( an -)
9(aY )
1 CJ rj-03
e + naY

ii d *\/

eju

This is of the form

y =

and addition theorem for the function

using the inductive hypothesis

exp x.

This in turn is smaller

EXPONENTIATION

175

then
g(a.)
g(a Y )
g(a Y )e
), a) ' r^-a)
e + a)
y
v

the lexicographical order.

.
which i s less than

g(a.)
\ w ! rj

This i s exactly what we need.

by

A similar

argument applies to the upper terms.

form

C o f i n a l i t y i s immediate since a typical term of w has the

a.
aY
exp( I u> V j - u ) Ye) by the inductive hypothesis, thus a l l we

need i s the f a c t that

terms of the form

to

w.

>_ 1.

The s i t u a t i o n i s similar for

This completes the proof.

The study of the function

study of

g.

exp x

Recall from theorem 10.11 that

{ind(a),g(a')}|{g(a")}

where

has now been reduced to the

g(a) =

ind a i s the unique

c such that

a ~ </

Thus, as has been promised e a r l i e r , the subject has in one respect become
greatly s i m p l i f i e d .

On the other hand,

g behaves in t r i c k y i n t e r e s t i n g

ways so that the subject i s s t i l l f a r from t r i v i a l .

D

PROPERTIES OF g AND EXPLICIT RESULTS

We shall f i r s t determine

that

g ( l ) = 1.

g for ordinals.

I t i s easy to see that

We know already

g(2) = 2, g(3) = 3 , . . . , g ( w ) = o>,

g(u)2) = co2, g{u^) = u/"3, e t c . We thus have an excellent p h y s i c i s t ' s proof

that

g(x) = x

f o r a l l ordinals which i s similar but somewhat superior

to the p h y s i c i s t ' s proof that a l l odd numbers are prime!

At least in

t h i s case the f i r s t counter-example i s not a f i n i t e number such as 9.

Before keeping the reader i n too much suspense l e t us
consider an epsilon number
However,
Hence

e.

By d e f i n i t i o n

</ = e hence

ind e = e.

ind e appears as a lower element in the formula for g ( e ) .

g(e) > e.

This can also be seen from the formula for

exp x.

F i r s t , we have

a>e = e and w03 = e.

Now exp(x) > x n

for positive
e

purely i n f i n i t e numbers. So certainly exp e > e, i.e. exp(we) > o)W .

Thus the subject of the exponential function has taken on a
new twist.

We shall see that epsilon numbers and numbers related to them

play an important part in the determination of an explicit formula for

g.

Incidentally, i t is impossible to give an explicit definition

of "explicit."
definition.

The reader may claim that theorem 10.11 already gives a

However, we shall see that we can obtain results which are

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

much more e x p l i c i t ,

although unlike the e a r l i e r p a r t of the theory in

t h i s chapter there i s a f e e l i n g of lack of f i n a l i t y

results.

176

in some of the

For example, what we do not do in general i s obtain the sign

sequence of

g(x)

d i r e c t l y from the sign sequence of

though in a

sense we approximate t h i s in some respects to an extent which is

conveniently t r a c t a b l e and i n t e r e s t i n g .
F i r s t , we obtain a formula for
a

g(a)

i n the special case where

i s an o r d i n a l .

Theorem 10.14.

If

the i n e q u a l i t y

e _< a < e + w

i s an o r d i n a l , then

g(a) = a

unless

f o r some epsilon number

satisfies

e, in which case

g(a) = a+1.

Proof.
a"

If

i s an ordinal

i s empty.

Hence

g(a)

g(a)

F i r s t suppose t h a t
number.

Then

ind a < a.

{ i n d a, a'}|<j>.

Since

Now l e t

{ i n d a, g(a')}|<j>

i s less than the f i r s t

this is

since

We now use i n d u c t i o n .

By the inductive hypothesis

ind a < a
a

has the form

i s an o r d i n a l .

epsilon

g(a) =

by the c o f i n a l i t y

be the f i r s t epsilon number.

theorem.

Then g(a) =

{ i n d a, a'}|<j> = {a,a'}|<j> = a+1.

We now determine
p o s i t i v e integer
ind(a+n+l) = a.

g(a+n)

Hence

we obtain

Next we determine

have so f a r .
a _< e a

where

ea

i s the

is a
Now

In f a c t , g(a+w) = {ind(a+w),

g(x) = x

from now on u n t i l we get

The general argument is s i m i l a r to what we

In f a c t , suppose the theorem i s v a l i d for a l l

where

Thus we get back to the e q u a l i t y

I t i s now easy to see t h a t

to the next epsilon number.

by c o f i n a l i t y .

{a,a+n+l} |<)> = a+n+2.

g(a+w).

g(a+n)}|<j> = { a , a+n+l}|<|> = a+a>.

g(x) = x.

i n d u c t i v e l y on

g(a+n+l) = { i n d ( a + n + l ) , g(a+n)}|<|>

ath epsilon number.

such t h a t

Then by the same argument

as f o r the f i r s t epsilon number we have by induction on

that

g(e a +n+l) = | i n d ( e a + n + l ) , g(ea+nHj> = { e a , e a + n + l } \$ = ea+n+2

and

g(ea+o>) = { e a , e a + n + l } <
| J> = ea+u>.

then we

have
ea+x

If

x > u> but

g(e a +x) = { i n d ( e a + x ) , ( e a + x ) ' } |<j>.

by c o f i n a l i t y .

ea+x < e a + 1

Since ind(e a +x) < ea+x

Then g ( e a + 1 ) = { i n d ( e a + 1 ) , ( e a + 1 ) ' } |

To complete the proof we s t i l l

must consider

this

is

= ea+i lea

where

i s a l i m i t o r d i n a l , but t h i s i s also more or less s i m i l a r to what we had

previously.

In f a c t , by c o f i n a l i t y

ea = {ea'}|<J>, hence

EXPONENTIATION

177

g(e a ) = (1nd ea,g(ea')}|4> = {e a a l+ l} l<fr = e a + l since certainly

ea+l < e a for any a 1 .
The above theorem is the first example of a phenomenon which
occurs in the study of g. Essentially we have a kind of "singularity"
in the neighborhood of an epsilon number. This helps to make the study
of g somewhat tricky.
We now consider the special case where a consists of a plus
followed by a sequence of minuses. These sequences lead to arbitrarily
small positive surreals and is thus a natural class of surreals to
consider after the ordinals, which lead to arbitrarily large surreals.
Actually we already computed the result when the number of
minuses is finite or exactly u>. In fact, we have g ( ) = and
g() = 0. We note here that the pattern

g(x) = x is broken at in

to

co

a different way than at epsilon numbers.

Now by the sign sequence formula 2 GO where n is a
positive integer and b is an ordinal consists of a plus followed by
cob+n minuses. Thus we are interested in a formula for g(2 a> ) .
In the sequel we expect the reader to have some facility in
computing expressions F|G from the sign sequences of elements in F
and G. This is an essential skill for the study of the function g.
Theorem 10.15.

g ( 2 ~ V b ) = -b + 2~ n .

Proof. We use induction on b and induction on n for fixed

Suppose g l 2 ~ V b ) = -b + 2~n. Consider 2 ~ n " V b . This is

b.

{0}|{2 GO } by the sign sequence formula and cofinality. (This can

also be seen by other methods from chapter five, but i t is easiest to
quote the sign sequence formula.) Hence
g ( 2 ' n " V b ) = lind 2 " n - V b m g ( 2 - V b ) } = {-b} | |-b+2~n} = -b+2""- 1 .
g(uTb) = {ind(uT b }||g(2~V b 'j} = {-b} | {-b'+2~n} = -b+1 = -b+2*"0. So the
formula is valid in this case.
Remark. Note that -b+1 is simply the negative of an ordinal when b
is a non-limit ordinal but is different when b is a limit ordinal. In
fact, i f b is a non-limit ordinal -b+1 consists of b-1 minuses,
whereas i f b is a limit ordinal -b+1 consists of b minuses followed

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

178

by a plus. In either case the computation {-b}|{-b'+2 } = -b+1 is

valid. (Note that -b consists of a sequence of b minuses and
-b'+2 n consists of a sequence of less than b minuses followed by a
plus and a finite number of minuses.)
The next case of interest is an ordinary e number followed
by a sequence of minuses. This case has some resemblance to the case
just considered. Here we feel that it is more instructive to use a
somewhat more informal approach.
Now clearly ind b is a monotonic function of b. As we add
on more and more minuses to e, ind b remains constant until we reach a
certain point at which the value is decreased. We first investigate the
pattern while ind b is constant and then determine when ind b changes
and the effect of the change on g(b).
We know that g(e) = e+1 which is e followed by a plus. If
b consists of e followed by a sequence of a minuses then the lower
elements of g(b) have the form {ind b,x}. Since ind b = e and
x < e we may simply replace this by e. The upper elements of g(b)
have the form g(x) where x consists of e followed by 3 minuses
where 3 < a. It now follows trivially by induction that g(b) consists
of e followed by a plus and a minuses. This is an example of the
convenient fact which we shall see in a more general form later that as
long as ind b does not change, then roughly speaking g(b) continues
the same way as b. Complications are caused by changes in ind b
although not all changes in ind b cause problems. In fact, we have
already seen that for ordinals tricky changes occur at epsilon numbers
although indices change at many other ordinals. (For example, although
ind b changes at w 2 , this causes no complication in the computation of
gU2).)
We now study the variation in ind b as the number of
minuses is increased. For this we need the sign sequence formula. In
fact, if c is e followed by a minuses, then w is e followed by
eaja minuses and more generally 2 w is e followed by ewe + en
minuses. This is enough information for the determination of ind b. In
fact, if b is e followed by 3 minuses, then ind b is e followed
by a minuses where a is the quotient obtained by dividing 3 by ew
using the division algorithm; i.e. informally speaking, as minuses are
added to b, minuses are added to ind b at intervals of length eu.

179
So the above description of g(b) applies if the number of
minuses after e is less than eco. If there are exactly ew minuses
after e, then ind b is e followed by a minus. The set of lower
elements now have the form {e-l,x}. The upper elements all consist of
e followed by a plus and a sequence of minuses. It is clear by
cofinality that g(b) = e, i.e. g(we l) = e. We now have a setup for a
double induction.
We lead the reader by the hand for a short distance. If b
has ew+1 minuses after e then we may express g(b) as {e-l}|{e}
by cofinality. Hence g(b) consists of e followed by a minus and a
plus, e-1 remains as a lower element until we obtain e w 2 minuses;
hence as we noted earlier g(b) continues the same way as b. Hence if
b has ew+r minuses after e where r is finite, then g(b) is e
followed by a minus and plus followed by r-1 minuses. If r is
infinite then g(b) also has a tail of r minuses since the shortage
by one gets wiped out at ew + w.
When b has eu>-2 minuses after e, e-1 no longer occurs
as a lower element and we thus easily obtain g(b) = e-1. The situation
for eu>-2 + r is similar to what we had before, the only difference being
that g(b) has two minuses after e rather than just one. This
continues in a similar way up to any b which consists of e followed
by ewn + r minuses. g(b) consists of n minuses after e followed
by a plus and the contribution of r.
We now write 3 = ewa + r. The induction on a works the
same way for all non-limit ordinals. In fact, if 3 = ewa then g(b)
consists of e followed by a-1 minuses. If 3 = ewa + r for finite
r then g(b) consists of e followed by a minuses, a plus, and r-1
minuses. If 3 = ewa + r for infinite r then g(b) consists of e
followed by a minuses, a plus, and r minuses. A slight difference
occurs if a is a limit ordinal. The case 3 = ew 2 is typical. In
this case ind b which is e followed by u> minuses is cofinal in the
set of lower elements. For the upper elements we have the cofinal set
consisting of the elements of the form e followed by n minuses.
Hence g(b) consists of e followed by w minuses and a plus. If
3 = ew 2 + r then g(b) consists of e followed by GO minuses, a plus,
and r minuses. Note that for finite r this is different from the
case where a is a non-limit ordinal because of the start at r = 0. In

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

180

general, if 3 = ewa + r where a is a limit ordinal g(b) consists

of e followed by a minuses, a plus, and r minuses.
Note that if b is e followed by ecoa + em + r minuses
then b has the form 2 or - r where y is e followed by a
minuses. The r minuses in the tail of
b contribute -r to the
value of b. On the other hand, the corresponding minuses in g(b)
contribute something entirely different. For example, if g(b) consists
of e followed by a plus and r minuses then g(b) = e+y where y is
the sequence consisting of a plus followed by r minuses. We already
noted just before the proof of theorem 10.15 that y is of the form
0 -n -a
2 a) .
The above computation does not appear to be too interesting
in general. It seems that the most succinct way to look at the behavior
of g(b) is to describe b in terms of a normal form followed by a
sequence of pluses and minuses. We shall see that some nice general
results exist if b is expressed in this form. On the other hand,
attempts to deal with b described purely in algebraic form or purely as
a sign sequence lead rapidly to messy computations.
We now turn to a study of more general b. We shall see that
the special cases considered are basic building blocks for such a study.
First, we have a rather general inequality.
Theorem 10.16.

If b >_ 1 then

g(b) _> b.

Proof. By theorem 10.14 this is true if b is an ordinal. (In fact,

the theorem gives more precise information concerning g(b).) We use
induction on the length of b. Since we know the result for b = 1 we
may assume that b > 1. Hence b begins with at least two pluses.
Therefore in the canonical representation of b all lower terms b1
begin with at least two pluses unless b1 = 0 or 1. Since 0 is
discarded in the computation of g(b), we may thus assume that b1 ^ 1,
hence g(b') ^ b 1 . Certainly every upper term b" satisfies b" 2. 1,
hence g(b") b".
Now g(b) = {ind b, g(b')>|{g(b")>. Of course we have
b = {b'}|{b"}. We now apply theorem 2.5. b < b" _< g(b"). Also
b1 g(b') < g(b) if b1 * 0. If b1 = 0 we have 0 <_ g(l) < g(b),
hence b < g(b).

EXPONENTIATION
Remarks.

181

Although the c o f i n a l i t y theorems have been frequently used

throughout the book t h i s is our f i r s t and only use of theorem 2.5.

It

happens to be convenient to use t h i s theorem here since we are proving an

inequality.
There is a technicality to watch for.
is no

g(b').

In our proof we used

lower element of

b.

false proof that

g(b) >_ b for a l l

If

b1 = 0

then there

g ( l ) , which is legal,since

1 is a

I f one is not cautious about t h i s , one can get a

The requirement

b.

b _> 1 can be s l i g h t l y weakened.

We shall

sharpen theorem 10.16 l a t e r . At any rate, we know that g() = 0 so

w
1
g ( b ) < b i f b < . This gives us a l i m i t a t i o n as to how much the
w

hypothesis can be weakened. Actually, we may replace b j> 1 by b _> 2

for a l l positive integers n thanks to the fact that g(2 ) = 2 . On
the other hand, this does not give the ultimate in sharpness since there
is s t i l l a question concerning infinitesimals which are larger than
general

b.

We now generalize part of theorem 10.14 from ordinals to more

Theorem 10.17. If ea+cu <_ b _< a < ea+i for some epsilon number
some ordinal a then g(b) = b. The same conclusion applies if
1 _< b a < e 0 .

e a and

Note. The inequality b _< a < e a + 1 cannot be replaced by b < e a + 1 # In

fact, if b has the form e-1 then we already know that g(b) > b.
Proof. Since we already have the result when b is an ordinal we may
assume that b is not an ordinal, so ea+u) < b < a. We use induction.
Let 3 be the least ordinal larger than b. Then ea+u) < b < 3. Since
b is not an ordinal, it follows from the definition of 3 that b
consists of 3 followed by a minus and possibly other terms. So 3 is
an upper element in the canonical representation of b and all other
upper elements are less than 3. For the lower elements we may assume
by cofinality that they are all at least e a + w . Hence all numbers in the
representation of b satisfy the hypothesis of the theorem so that by
the inductive hypothesis g(b) = {ind b,g(b*)}|{g(b")} = {ind b,b'}|{b"}.
(As a technical point, we may reinsert all ordinals which are less than

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

182

e a +w into the set of all b' so that the lower set {ind b,b'} has its
usual meaning.)
Now 3 is not an epsilon number since it is strictly between
e a and eOL+l. Hence ind 3 is an ordinal less than 3. Therefore
ind b _< ind 3 < b. Therefore {ind b, b'}|{b"} = b by cofinality. This
completes the proof.
The same argument applies to the second part. As we noted
earlier, as a technical point we need the element one as a lower element
so that we can reintroduce zero by cofinality. Formally speaking, if one
is picayune and use the notation B' for the set of all b 1 , the set of
lower elements should be expressed first as {ind b},(B'-{0}) which
simplifies to {ind b},{B'} by cofinality. We prefer a slight abuse of
notation for the sake of readability.
The above theorem shows that the equality g(b) = b is par
for the course, i.e in a rough sense "most" numbers are covered by the
theorem. We now consider numbers which are "close" to epsilon numbers.
Theorem 10.18. If e _< b _< e+n for some epsilon number e and integer
n then g(b) is obtained from b as follows. If the sign sequence of
b is e followed by the sequence S then g(b) is the juxtaposition
e followed by a plus and then S.
Note. By a careful look at the sign sequence formula one can show that
g(b) = b+1, as was stated as part of theorem 10.14 in the special case
where b is an ordinal. We prefer to deemphasize this point of view
since in general a simple operation such as inserting a plus into a
sequence can have a complicated effect on the normal form. This is
because the existence of the plus affects the contribution of the
subsequent signs. Although the general problem concerning the effect of
insertion of signs on the algebraic value of a surreal number is of
intrinsic interest, it is far removed from the theory of exponentiation.
For our purpose it suffices to attempt to express the function g in
what appears to be the most tractable way.
Proof. The result is known for ordinals. We may thus assume that for
some integer n>0 we have b = e a +n+c where - K c < 0 , i.e. b begins
with e+n pluses followed by a minus. Hence a cofinal lower set in the

EXPONENTIATION

183

canonical representation of
which begin w i t h

pluses and except f o r

b"

satisfy

a lower element of

e+n

then

g(b')

and

lexicographical

i t s e l f continue w i t h a minus.

g(b")

order

but

Also, since

Thus

e+n
b1

i s superfluous by c o f i n a l i t y .

also has t h i s

and
Since

g [ e + n - l ] = e+n

is
Hence

the elements of

have a plus inserted a f t e r

g(b)

e.

By the

property.

one of the gaps in our study of

strengthen theorem 10.17

Theorem 10.19.

ind b = e.

By the inductive hypothesis a l l

We now f i l l

Note.

A l l upper elements begin with

g ( b ) , ind b = e

g(b) = { g ( b * ) } { g ( b " ) } .

e+n _< b

may be obtained by r e s t r i c t i n g to numbers

the hypothesis of the theorem so we can use i n d u c t i o n .

ind e = ind(e+w) = e

the form

e + ( n - l ) pluses.

and

slightly.

I f for some epsilon number

b < e+w, then

and a l l

integers

n,

g(b) = b.

This takes care of one of the gaps l e f t by theorems 10.17 and

10.18.

Proof.

Such a

begins with

g(b) = { i n d b g ( b ' ) } | { g ( b " ) } .

lower element,

e+co

followed by a minus.

ind b = e

i s superfluous.

and since

Hence

Assume f i r s t t h a t a f t e r the

g(e+n) = e+n+1

g(b) = { g ( b ' ) } | { g ( b " ) } .

e+w

pluses and minus there are

no f u r t h e r changes in s i g n .

We then set up an easy i n d u c t i o n .

i s only one minus a f t e r the

e+w

g(b) = { g ( b ' ) } | { g ( b M ) }

is a

If

there

pluses then

= {e+n+1} |{e+w} = {e+n}|{e+w} = b.

An easy

i n d u c t i o n gives the r e s u l t in t h i s case.

If
elements b

there are f u r t h e r changes in sign then we have lower

g ( b ' ) = b1

which s a t i s f y

by the previous case.

i n d u c t i o n we have a c o f i n a l set of lower elements

g(b')

= b1,

so

g(b) = { g ( b ' ) } | { g ( b " ) }

= { b ' } | { b n } = b.

There i s only one "border" area l e f t .

one where

begins w i t h

by a minus, i . e .
b _< 3.

b < e

Namely, t h i s is the

pluses for an epsilon number

but for no ordinal

$ < e

does

This turns out to be the most complicated case.

numbers less than

followed by a minus.

So using

satisfying

i n t o t h i s case, i . e .

e
b

followed
satisfy

We include

numbers beginning with a plus

Note t h a t a l l other numbers have been handled by

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

184

earlier theorems. These two classes of numbers behave in a similar

manner. It seems strange to regard 1 as an epsilon number, yet when it
comes to the study of g, 1 behaves enough like an epsilon number so
that it is convenient for the sake of efficiency to study the classes
simultaneously. So in future when we refer to an epsilon number followed
by minuses it is understood that this includes the case where we have one
plus followed by minuses.
In theorem 10.15 and in the discussion afterwards we already
saw what happens when the continuation consists solely of minuses. The
main characteristic of this case is that ind b changes as the number of
minuses is increased at various intervals so that g(b) changes in an
abrupt way rather than continuing in the same manner as b.
Suppose that the number of minuses after e before the next
plus is expressed in the form ewa+r where r < eu>. We now break this
up into several subcases, only one of which is especially tricky.
Case 1. r > w. Here we obtain an upper element d in the canonical
representation of b by taking e followed by ewa+w minus and a lower
element c by taking e followed by ewa+r minuses. Now ind x is
constant for c <_ x d. Furthermore
g(d) has the form p followed
by w minuses and g(c) has the form p followed by r minuses. (In
this case p consists of e followed by a minuses followed by a plus.
However, since the subsequent reasoning is independent of the specific
nature of p it is convenient for the sake of generality to ignore this
fact now.) Any time we have such a situation we can easily determine
g(x) by induction for any x satisfying c <_ x d by a type of
argument we have already used several times. First, since
ind x = ind c < g(c) and c is a lower element in the canonical
representation of x, it is superfluous to have ind x as a lower
element by cofinality. By a further use of cofinality we may limit the
upper and lower elements used in the representation to numbers y such
that c _< y _< d. Then by an obvious induction and use of the lexicographical order we see that if any number x satisfying c _< x d
is expressed as the juxtaposition consisting of e followed by ewa+w
minuses followed by q then g(x) is the juxtaposition p followed by
CJ minuses followed by q.

EXPONENTIATION

185

Case 2. r > 0 and a is a limit ordinal. Then the same reasoning

applies exactly providing w is replaced by the empty sequence.
Case 3. r = u> and a is a non-limit ordinal. In this case we replace
u) by any finite n larger than 0. We note that if d is the
juxtaposition of e followed by ewa+n minuses then the last string of
minuses in g(d) contains only n-1 minuses. However, since n runs
through the set of all positive integers in obtaining upper sums, this
makes no essential difference.
Case 4. r > 1 and a is a non-limit ordinal. We can now replace u>
by 1. This is otherwise like case 1 except that the r minuses in x
give rise to r-1 minuses in g(x).
Case 5. r = 1 and a is a non-limit ordinal. Here we let d = e
followed by ewa minuses and c = e followed by ewa+1 minuses. This
is essentially the same as the other cases, the only difference being
that the last minus in c contributes a minus followed by a plus in
g(c). For c _< x <_ d the sign sequence in g(x) continues as in x.
Note only that if there is a finite string of n pluses after the
ewa+1 minuses in x, g(x) will have a string of n+1 pluses because
of the extra plus at the beginning.
At any rate all these cases in which r > 0 are essentially
the same, differing only in very minor ways. In all cases the sign
sequence of g(x) continues as in x. The cases where r = 0 are more
challenging. Roughly speaking, we are then closer to the "border" where
ind x changes.
Case 6. r = 0 and a is a non-limit ordinal. This case also turns out
not to be too complicated. Recall that if c is e followed by ewa
minuses then g(c) is e followed by a-1 minuses; and if dr is e
followed by eco(a-l)+r minuses where r is less than ew and non-zero,
then g(d r ) is e followed by a-1 minuses followed by a plus and r-1
or r minuses depending on whether r is finite or not. Here we have
c
x i. d r f r a<ll r < e w
Ind x is n t fixed in this interval.
However, ind x _< ind dr = g(c), so if x > c then ind x is
superfluous in the representation of g(x).

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

186

Now let b = c followed by a plus. Then it is easy to see

that g(b) = {g(c)}|{g(d r } which is e followed by a-1 minuses
followed by a plus and ew minuses. In other words, the final plus
contributes a plus followed by ew minuses. It is now easy to see that
for any x satisfying c _< x <_ dr the sign sequence of g(x) for x
beyond e, the ewa minuses, and plus continues as in x. So this case
is much like the cases where r > 0, the main difference being the
existence of a plus in x which contributes a plus followed by eco
minuses in g(x).
Case 7. r = 0 and a is a limit ordinal. This is the only case left
to consider and the only case which is really fundamentally different.
Let c = e followed by ewa minuses and dg = e followed by ea)3
minuses where 3 < a. g(c) is e followed by a minuses and a plus.
Vie may assume by cofinality that 3 is a non-limit ordinal in which case
g(dg) is e followed by 3-1 minuses. We are now interested in the
interval c <_ x _< de for all 3 < a.
As in the last case let b = c followed by a plus. Then
g(b) = {ind b, g(c)}|{g(d 6 )}. It is immediate that this is e followed
by a minuses and two pluses. (Note that we use the fact that
ind b = ind c since an extra plus cannot change the value of ind.)
So far this case looks easier than the previous one since the
extra plus contributes simply a plus. One can almost say that for an
arbitrary x satisfying c _< x <_ d$ the sign sequence for the rest of
g(x) continues as in x. However, we must beware of the contribution of
ind x.
In fact, suppose that after e and the ewa minuses we have
only pluses. The variation in ind x as the number of pluses keeps
increasing can be determined from the sign sequence formula. When the
number of plus reaches ew then ind x increases by the juxtaposition
of a single plus. Since this is precisely g(c) it can be discarded by
cofinality. However, consider what happens when the number of plus
reaches the first epsilon number above e. The continuation of the sign
sequence in x is precisely ind x. We almost have a generalized
epsilon number as discussed in chapter nine. What is missing is that we
are not assuming that a absorbs ew multiplicatively. At any rate, at
that point ind x can no longer be discarded, and, in fact, is cofinal
by itself as a lower element in the computation of g(x). So g(x) gets

EXPONENTIATION

187

an extra plus.
This is similar to what happens in the simple case for
ordinals at the first epsilon number. In fact, for any x such that
c
Ji x JS d$ the same reasoning we used so far including the splitting up
into various cases can be applied to the tail of c in x in order to
determine the sign sequence of the rest of g(x). Again we are left with
only one case to consider further. Specifically, the case consists of an
element beginning with c followed by e pluses for some epsilon number
e greater than e and then followed by e^ft minuses for some limit
ordinal 3.
We now have a set up for a "grand" induction. At every stage
we are left with one case to consider further so it appears that we will
never be done. The pattern is as follows. We have strings of pluses and
minuses, the ith string of pluses has length e. where (e.) is a
strictly increasing sequence of epsilon numbers, and the ith string of
minuses has length e.uxx. where a. is a limit ordinal for all i.
What we do next is to consider a sequence b such as the one
described above where i runs through the set of all positive integers.
By the induction described above we know the value of g for every
proper initial segment of b. Now g(b) = {ind b, g(b')}|{g(b")}. By
cofinality we may limit ourselves to those initial segments b of b
which consist of full strings. Hence by the inductive hypothesis g(b)
has the following form: the ith string of pluses has length e., the
ith string of minuses has length a. and a plus is added at the end.
(This is valid whether the last string in b consists of pluses or
minuses.) Also, ind b has the same form but, of course, with no plus at
the end. By mutual cofinality we can ignore the plus at the end in
g(b). This is easy to see since the lower element obtained by stopping
after the ith string of minuses then followed by a plus is less than
the element obtained by stopping after the (i+l)st string of minuses,
and the upper element obtained by stopping after the ith string of
pluses is greater than the element obtained by stopping after the
(i+l)st string of minuses then followed by a plus.
Hence g(b) simplifies to {ind b, ind b 1 }|{Ind b"}. If we
apply the remark concerning restricting oneself to full strings by
cofinality to ind b, it follows that {ind b'}|{ind b'1} is precisely
ind b.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

188

We have shown that g(b) has the form {ind b, F}|{G} where
F|G = ind b and F and G are initial segments of ind b. It follows
easily that g(b) is ind b followed by a plus. We see this directly
as follows. If we denote ind b followed by a plus by c then clearly
F < ind b < c. Since G is an initial segment of ind b and
ind b < G, c < G by the lexicographical order. Hence F < c < G. Since
ind b < c, we see that c satisfies the betweenness condition. On the
other hand, suppose F < x < G and ind b < x. Since F|G = ind b it
follows that ind b is an initial segment of x. Since ind b < x it
follows that x must begin with ind b followed by a plus, i.e. x has
c as an initial segment.
It is now easy to see that the situation after the co
strings of pluses and minuses behaves the same way as the situation after
e discussed earlier. Furthermore, the same argument used above for to
strings of pluses and minuses applies to a strings of pluses and
minuses where a is any limit ordinal.
Thus our discussion takes care of the value of g for an
arbitrary sequence of pluses and minuses. Although the final result
could be expressed as a formal theorem we feel that this would make the
pattern look more complicated than it is, i.e. we feel that it is most
lucid to express the procedure in the somewhat informal manner which we
used.
However, we shall end this book by a formal theorem which
b
describes precisely when g(b) = b, i.e. exp w b = ^ . This result
should be of interest since it is a natural culmination of earlier
theorems.
Theorem 10.20.
two forms.

g(b) = b if and only if

b has either of the following

(1) b is less than some ordinal a < e 0 where

epsilon number and < b for all integers n.

e0

is the first

(2) b begins with at least e 0 pluses and the first string in the
sequence for b such that the initial segment of b which terminates at
the end of the string is not a generalized epsilon number is a string of
pluses. Furthermore, if a is the number of pluses then choose e to
be the largest number of pluses such that e < a and the sequence

EXPONENTIATION

189

obtained by replacing the final string of a pluses by e pluses is a

generalized epsilon number. Then b > e+n for all integers n. (Such an
e exists since the I.u.b. of e numbers is an e number.)
Proof.

The proof is easier than the statement of the theorem!

Essentially it follows directly from the preceding discussion. First suppose that b is less than some ordinal a < e 0 . The
case b ^ 1 is taken care of by theorem 10.17. The result for b =
has been mentioned earlier. This is the case of a plus followed by a
finite number of minuses. Now the earlier discussion dealing with an
epsilon number e followed by a sequence of minuses applies to a single
plus followed by a sequence of minuses, as was also noted earlier. Hence
if b begins with a plus followed by a finite number of minuses, g(b)
continues the same way as b so g(b) = b. Since
g(^) = 0, g(b) * b for b <_.
We still must consider the subcase where b begins with a
plus followed by u> minuses, and a plus. The earlier discussion dealing
with e followed by ewa minuses followed by a plus where a is a
non-limit ordinal actually applies to this case. Since we regard e as
1, a) = euxx with a = 1. The plus followed by a> minuses contributes
the empty sequence to g(b) (we know that g() = 0). The next plus
contributes a plus followed by w minuses and from then on g(b)
continues the same way as b. Note that we obtain g(b) resembling b
for a strange reason: i.e. the plus at the end contributes precisely
what has been previously discarded. g(b), however, does not have the
extra plus in b. It is thus clear that g(b) = b precisely if the
string of pluses following the plus and w minuses contains at least u>
members. This is precisely the condition that < b for all integers
GO

n.
Now suppose that for no ordinal a < e 0 does b satisfy
b _< a. This is equivalent to the condition that b begins with at least
e 0 pluses.
Now assume that b is not a generalized epsilon number but
b consists of c followed by a string of minuses and c is a generalized epsilon number. Then ind b consists of c followed by a
string of minuses of length smaller than the corresponding last string in
c. Hence ind b > b. Since g(b) > ind b it follows that g(b) * b.

AN INTRODUCTION TO THE THEORY OF SURREAL NUMBERS

190

This argument extends immediately to any sequence having b

as an initial segment. A very weak use of the sign sequence formula is
all that is needed.
If b is a generalized epsilon number then certainly
g(b) > ind b = b. We know in fact that g(b) consists of b followed
by a plus. Now let b be a generalized epsilon number followed by a
string of a pluses. Then g(b) begins with c followed by a string
of 1+a pluses. Then the situation resembles the one in the proof of
theorem 10.17. The distinction between a and 1+a is wiped out
precisely when a is infinite in which case g(x) = x. This completes
the proof.
In concluding this book we hope that the reader is convinced
that the class of surreal numbers is a fascinating class of objects found
in nature. Since the study of these numbers is still at an early stage,
we are confident that there are many exciting results which are just
waiting to be discovered by an alert reader. For this reason we expect
and even hope that much of the presentation here will be improved as new
insights are gained.

191

REFERENCES

[1]

Conway, J . H . ( 1 9 7 6 ) . On Numbers and Games, Academic Press,

London & New York

[2]

Knuth, D.E. ( 1 9 7 4 ) . Surreal Numbers, Addison-Wesley,

Reading, Massachusetts

192
INDEX

Abelian group
addition
, natural

17
13
41

betweenness property

15

canonical representation
cofinal
c o f i n a l i t y theorem
, inverse
commutative ring
convex

11
9, 10
10
12

19
54

Dedekind cut
diagonalization
dyadic f r a c t i o n

32
121
28a, 29, 34ff

element, lower
, upper
epsilon numbers
exp y F
exponentiation

17
17
121, 175
154ff
143, 145

FIG
field
fixed points

7, 10, 11
22, 39
124ff

groupoid

24

Hensel's lemma

73

, natural
non-standard analysis
normal form

, common
integer
length
logarithm

1
39, 48
3
7
111
3
161

48, 116
52 , 58, 62

order
order of magnitude
, higher
ordinals

3
54, 55
54
41

power series, generalized

purely i n f i n i t e number
real number
reciprocal
semigroup
sequent
shuffle
sign sequence
square root
standard part
subshuffle
surreal numbers

71
156
33
21, 37
52
41
104
76 , 85, 94
24
116
104
1, 3
3
60
59

tail
tai1 property
transfinite sequence
uniformity theorems

impartial games
infinitesimal
i n i t i a l segment

17
41

multiplication

15.

18

valuation theory

118

[x]n
n

144

55