Cristian S.

Calude and Gheorghe Păun, Editors

FINITE VERSUS INFINITE:
CONTRIBUTIONS TO AN ETERNAL DILEMMA
Cristian S. Calude, Gheorghe Păun
Editors

FINITE VERSUS
INFINITE:
CONTRIBUTIONS
TO AN ETERNAL DILEMMA

Springer-Verlag
London, 2000
Cristian S. Calude
Department of Computer Science
The University of Auckland
Private Bag 92019, Auckland, New Zealand
E-mail: cristian@cs.auckland.ac.nz

Gheorghe Păun
Institute of Mathematics of the Romanian Academy
PO Box 1-764, 70700 Bucureşti, Romania
E-mail: gpaun@imar.ro

ISBN, Copyright, etc.

 Preface

The finite – infinite interplay is central in the human thinking, from

ancient philosophers and mathematicians (Zeno, Pythagoras), to modern
mathematics (Cantor, Hilbert) and computer science (Turing, Gödel). Re-
cent developments in mathematics and computer science suggest a) rad-
ically new answers to classical questions (e.g., does infinity exist?, where
does infinity come from?, how to reconcile the finiteness of the human brain
with the infinity of ideas it produces?), b) new questions of debate (e.g.,
what is the role played by randomness?, are computers capable to handle
the infinity through unconventional media of computation?, how can one
approximate efficiently the finite by the infinite and, conversely, the infinite
by finite?).
Distinguished authors from around the world, many of them architects
of the mathematics and computer science for the new century, contribute
to the volume. Papers are as varied as Professor Marcus’ activity, to whom
this volume is dedicated. They range from real analysis to DNA com-
puting, from linguistics to logic, from combinatorics on words to symbolic
dynamics, from automata theory to geography, and so on, plus an incursion
into the old history of conceptions about infinity and a list of philosophical
“open problems”. Mainly mathematical and theoretical computer science
texts, but not all of them purely mathematical. Dealing directly with the
finite – infinite interplay (one proves several times that paths from finite
to infinite and conversely, from infinite to finite, are both inspiring and
useful), or only implicitly (each grammar or automaton is a finite device
meant to describe a possibly infinite object, a language; similarly, an ax-
iomatic system is a finite construction aiming to cover an infinity of truths
– whence the limits of such systems).
We emphasize the precaution taken in the title. The reader can find
here only some contributions to an eternal debate. This is just a step in an
infinite sequence of steps. We hope that this is a step forward. . .
*
Many thanks are due to all contributors, to Bev Ford and Rebecca

v
vi Preface

Mowat, from Springer, London, for the efficient and pleasant cooperation,
as well as to our host institutions, JAIST, Ishikawa, Japan, and TUCS,
Turku, Finland, for providing excellent environments for completing this
book.

Cristian S. Calude
Gheorghe Păun
November 1999
 Tribute to Professor Solomon Marcus

The present book is presented in honour of Professor Solomon Marcus,

on the occasion of his 75th birthday (1st of March, 2000).
The co-editors of this volume have debated for a while which would be
the most appropriate topic for such a work. The decision was, at the same
time, difficult and easy.
Professor Marcus started by being a “pure mathematician”, building
for himself an international reputation in mathematical analysis and re-
lated areas; then, at the end of fifties, he switched to an interesting career
in mathematical linguistics (mainly analytical approaches to phonological,
morphological, syntactic and semantic categories, but also automata and
formal languages).1 Although he was one of the most cited authors in this
area, he has once again enlarged his interests – and definitely, others’ as well
– by founding one more domain, mathematical poetics. Just en passant, in
1968, he introduced (Marcus) contextual grammars, a genuinely new gener-
ative device inspired from analytical models in the study of language.2 This
was not enough for Professor Marcus. In the same way as real analysis and
mathematical linguistics provided to him tools for mathematical poetics,
all these have been applied to semiotics, natural and social sciences.
Then, what to choose for a Festscrift? None of these alone would be
enough, all together are hard to be bound under the covers of a single
book.3
A tempting idea was to look for major trends or meta-ideas in Professor
Marcus oeuvre. Again a difficult choice – at least for the present co-editors,
although both of them were (and psychologically still are!) his students
and had/have the privilege to be his close collaborators. Analogies/bridges
between fields which, at the first sight, look remote? The finesse of study,
1
His book Finite Grammars and Automata, published unfortunately only in Romanian
in 1964, was one of the earliest monographs on this topic.
2
The domain has expanded beyond his expectations (as he confessed somewhere): a
comprehensive book was published recently by Kluwer: Marcus Contextual Grammars.
3
Another book in honour of Professor Marcus, including only papers in mathematical
and computational linguistics, will be published by the Romanian Academy Publishing
House.

vii
viii Tribute to Professor Solomon Marcus

no matter what object of research? The attraction for counter-intuitive,

even paradoxical facts, proven then to be common-frequent-beneficial? The
visible passion for dichotomies? Which one? Artistic versus scientific? Fi-
nite physical existence versus infinite spiritual existence? Old versus young
provocation? Discrete versus continuous? Analytical versus generative?
Local versus global? Finite versus infinite?
That’s it! It fits well with Professor Marcus’ activity and personal-
ity, both of which leave the impression of infinity. . . About four hundred
papers, more than forty authored or co-authored books (published in ten
languages not including Romanian), twenty five edited books, hundreds of
conferences attended, hundreds of lectures at universities, many domains
of direct research, much more domains of general interest, an authoritative
leader of schools, a respected cultural presence, a continuous promoter of
the Romanian mathematical heritage and, at the same time, an energizer
of the first steps of many young researchers, a perfect memory, a rigorous
life style, active scientifically and physically at 75 (at a level which is tiring
for many of his much younger colleagues) – all these have something to do
with in-finity. . .
Antonia, a 10 years old Romanian-Spanish girl from Tarragona, Spain,
had a decisive influence in choosing the topic of the volume. Earlier than
us, she called Professor Marcus El infinito, under circumstances pointing
to his manner of approaching kids. In his opinion, a child should rediscover
Zeno’s paradoxes via a Socrates-type dialogue of the following kind: S.M.:
“Look, Antonia, you have a bread, a usual one, and you eat today half of
it. How much does it remain for tomorrow?” A.: “Half a bread, of course.”
S.M.: “Okay, but tomorrow you eat half of what you have got. Does some
bread remain for after tomorrow?” A.: “Yes, sir, some bread still remains.”
S.M.: “Good, but after tomorrow you again eat half of the bread you have.
Is it true that some bread still remains?” A. (already doubtful): “Yes,
but”. . . After about five iterations, the two parts are always separated: A.,
or any other “victim”, knowing for sure that (s)he can eat a normal bread
in a day or so, hence any continuation of the dialogue is senseless (because
it’s breadless), S.M. looking desperately for the bright light of infinity in
the others’ eyes: eating every day half of the piece of bread you have got
means securing bread for infinitely many days, isn’t it?. . . In many cases,
both sides are disappointed. . . In at least one case, S.M. got a surname: El
Infinito. . .
Happy Birthday, Profesore Infinito!
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

Tribute to Professor Solomon Marcus . . . . . . . . . . . . . . . . . . . . . . vii

Umberto Amato, Dan Tudor Vuza

Smoothing Data: When Finite and Infinite Dimensional
Help Each Other . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Hari Bercovici
Rearrangements of Real Functions Derived from
the Combinatorics of Young Tableaux. . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Massimo Borelli, Andrea Sgarro

A Possibilistic Distance for Sequences of Equal
and Unequal Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

Cristian Calude, Helmut Jürgensen, Shane Legg

Solving Problems with Finite Test Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Cezar Câmpeanu, Kai Salomaa, Sheng Yu

State Complexity of Regular Languages: Finite Versus
Infinite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Gregory J. Chaitin
A Century of Controversy over the Foundations of
Mathematics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .75

Florin Diacu
Finite Versus Infinite in Singularity Dynamics . . . . . . . . . . . . . . . . . . 101

John Ewing, Ciprian Foias

An Interesting Serendipitous Real Number . . . . . . . . . . . . . . . . . . . . . . 119

ix
x Contents

Rudolf Freund, Alexandru Mateescu, Arto Salomaa

Algebraic Representations of Regular Array Languages. . . . . . . . . .127
Salvatore Greco, Benedetto Matarazzo, Roman Slowinski
Rough Set Processing of Vague Information Using Fuzzy
Similarity Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Tom Head
Relativised Code Concepts and Multi-Tube DNA Dictionaries. . .175
Lucian Ilie, Ion Petre, Grzegorz Rozenberg
Uniformly Scattered Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Vincenzo Manca
Splicing Normalization and Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Carlos Martı́n-Vide
Infinitely Many Infinities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Georges Nicolas
Finite and Infinite in Geography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Piergiorgio Odifreddi
Ultrafilters, Dictators, and Gods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Sergiu Rudeanu
On the Behaviour of Semiautomata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Jouko Seppänen
Infinity – An Outline of Conceptions in Mythology,
Cosmology and Natural Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Hava T. Siegelmann
Finite Versus Infinite Neural Computation . . . . . . . . . . . . . . . . . . . . . . 285
Dan Simovici, Szymon Jaroszewicz
On Information-Theoretical Aspects of Relational Databases . . . . 301
Robert M. Solovay
A Version of Ω for Which ZFC Cannot Predict a Single Bit . . . . . 323
Ludwig Staiger
On the Power of Reading the Whole Infinite Input Tape . . . . . . . . 335
Victor Vianu
Finite and Infinite in Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
 Smoothing Data:
When Finite and Infinite Dimensional
Help Each Other
Umberto Amato
Istituto per Applicazioni della Matematica CNR
Via Pietro Castellino 111, 80131 Napoli, Italy
E-mail: amato@iamna.iam.na.cnr.it

Dan Tudor Vuza

Institute of Mathematics of the Romanian Academy
P.O. Box 1-764, RO-70700 Bucharest, Romania
E-mail: dvuza@stoilow.imar.ro

1 Introduction
A huge number of applications require to solve the mathematical prob-
lem of removing noise from data. Therefore the problem has been widely
considered and several methods have been developed for its solution un-
der different points of view, e.g., statistics, approximation theory, inverse
problems, just to cite some in the mathematical literature. Each field uses
its own tools to solve the problem, i.e., nonparametric regression, approxi-
mation, regularization, respectively. The different points of view gave even
rise to different denominations for the problem, e.g., smoothing data, de-
noising (see in this respect the discussion in [13], where distinction between
de-noising and smoothing data is made). In the present paper we shall be
concerned with the problem of removing noise from data in the framework
of inverse problems, where the problem is known as smoothing data, and
indeed it is the simplest prototype.
The dilemma finite-infinite dimensional is classic in mathematics and
obviously it concerns also smoothing data. Who could negate that infinite
dimensional is more desirable? It is the dream of people to have a complete,

1
2 U. Amato, D. T. Vuza

full representation of things, even the simplest (think of people who collect
things, stamps, post-cards, and so on). It is the mankind dream translated
into mathematical language, where it becomes a must. Most finite dimen-
sional quantities rarely have an autonomous life and often live only as an
approximation of infinite dimensional.
This is true also for smoothing data, where in addition we are generally
faced to a real problem with real data. This has two peculiarities: a finite
number of data, N , a finite (strictly greater than zero) noise. This means
two things. First, the problem is finite-dimensional by its very nature, so
that a finite-dimensional approach should be appropriate for its handling.
Second, the type of convergence required is different from the one that
is usually assumed in the general theory of inverse problems: we are not
interested in the case when noise tends to zero, but in the convergence in
some sense (maybe probabilistic) when the size of the sample increases.
Nevertheless, the analysis of the problem in the infinite dimensional
setting has been an approach followed in the literature. There are at least
two important reasons that go beyond the generality of the mathemati-
cians’ dream. First, there is a particular applicative interest when details
finer than obtainable from the data only are required (e.g., when smooth
curves are asked as in CAGD, or zooming in is needed in image processing).
Second, exogenous information has to be introduced in the problem, often
in an infinite dimensional setting, in order to remove the ambiguity due
to the finite size of the sample, as currently done in discrete and semidis-
crete inverse problems. Pioneer in this respect is the celebrated thin-plate
spline, where exogenous information is given by the regularity of the solu-
tion expressed in terms of its p-th derivative. Then the (finite dimensional)
thin-plate spline is the solution of the (infinite dimensional) problem

N
min (ỹi − yiε )2 + λf˜(p) L2 ,
f˜∈H p i=1

where H p is the Sobolev space of regularity p, yiε is a sample of an unknown

function f in a proper grid of size N affected by noise and ỹi is solution of
the regularization problem collocated in the same grid.
Then we can say that, even though finite dimensional is the proper
setting of this problem due to its experimental nature in general, however

infinite dimensional helps finite dimensional.

Recently wavelets have been considered as a tool for solving several

mathematical problems, including noise removal. The interest is due to
Smoothing Data 3

their properties to be localized both in time and frequency; moreover a

fast algorithm can be developed for computing the wavelet transform, also
suitable to be implemented on integrated circuits. What is interesting
for noise removal is their capability to concentrate information in a limited
amount of data; then pieces of not interesting information can be discarded
without affecting too much the whole message whereas in this process a lot
of the accompanying noise is also eliminated. This generates the asked
noise suppression. Let us analyze in more detail this aspect.
Noise and signal are strongly linked and one knows indeed only their
combination; then to reduce one implies to reduce the other too, unfortu-
nately. Luckily, noise and signal have very different characterizations for
the majority of problems; in particular the signal has strong redundancies
and correlations.
Therefore a suitable transformation of the signal gives very few, possible
independent, pieces of information and allows one to discard the redundant
ones. Since the noise is intimately linked to the signal, the former is also
discarded together with the redundant information and since noise does
not exhibit the same redundancy as the signal, the part discarded with the
latter is significant, giving rise to the noise removal. Then finite dimensional
(more exactly, small finite dimensional) is the key for suppressing noise
effectively. The question is how much finite? Of course it depends on the
signal (the more the structures of signal and noise are different, the more
noise removal is better), but in general we can say less, much less than the
size of the sample. The ideal case would be fixed, not depending on the
size of the sample at all. This would give high computational efficiency
and best convergence rate (N −1 ). However this (parametric) approach is
not suitable when no enough information is available on the functional
expression of the solution.
Therefore finite dimensional is something more than simply an obliga-
tion coming from the finite size nature of applications. It is the key for
having effective noise removal. But surprisingly it is also the key for having
effective infinite dimensional solutions. In fact we shall see that the infinite
dimensional solution can be effectively defined only by a limit process of
the finite dimensional problem when the size of the sample increases. An
attempt to deal directly with the infinite dimensional setting gives rise to
lack of convergence or discrepancies. Then we can say that

finite dimensional helps infinite dimensional

and the mutual help finite-infinite dimensional is complete.

4 U. Amato, D. T. Vuza

2 Shrinking Functions and Inverse Problems

Let us introduce the noise removal problem

yiε = yi + εi , 1 ≤ i ≤ N, (1)

where y is a sample of an unknown function f at some points and εi is a

realization of noise having some suitable properties (initially we suppose
white noise with variance σ 2 ). Then y ε is the signal at our disposal, that
is a sample of f affected by noise. In the following we suppose N = 2J
and f sampled at equispaced points. Moreover we choose an orthogonal
wavelet system and we indicate by {Y−1,0 , Yj, , 0 ≤ j < J, 0 ≤  < 2j } the
discrete wavelet transform (DWT) of {2−J/2 yi }N i=1 . For ease of notation
N −1
we mainly indicate the DWT by {Yk }k=0 , with k related to the scale and
translation parameters by the relation k = 0 for j = −1 and k = 2j + 
for 0 ≤ j < J, 0 ≤  < 2j . Yk turns out to be an approximation of the
continuous wavelet transform of f . Such an approximation is usual in all
applications and justified by several papers [12], [20]; it is even better in the
case of Coiflets basis [7]. Analogously Y ε will denote the DWT of y ε . Note
that due to linearity of the model (1) and orthogonality of the transform,
noise is white also in the wavelet domain (with variance N −1 σ 2 ).
Most methods for removing noise from signals based on the wavelet
transform go through the following

Algorithm 2.1
1. Take the wavelet transform of the data 2−J/2 y ε , yielding Y ε .

2. Shrink Y ε according to

Ỹk = s(Ykε ; t)Ykε , 0 ≤ k < N, (2)

where 0 ≤ s(Ykε ; t) ≤ 1 is a (generally nonlinear) shrinking function

depending on a free parameter t.

3. Take the inverse wavelet transform 2J/2 Ỹ , yielding ỹ, which is an

estimate of the cleaned signal.

The reason for which the association between the wavelet transform and
Algorithm 2.1 became so popular in applications is provided by its com-
putational speed and by the possibility to implement Algorithm 2.1 within
integrated circuits. In the literature the number of shrinking functions de-
veloped by researchers is rapidly increasing; here we mention hard, soft [14],
Smoothing Data 5

[15], DKLT [11], regularization [6], [1], firm [16], garrote [8], to cite some.
They were developed under different frameworks (statistics, approximation
theory, inverse problems) with different purposes. All of them are nonlin-
ear, except regularization that relies on tools taken from inverse problems
methods (see the next section). Indeed, even though the shrinking func-
tions were developed under so different frameworks, a unifying approach
for them can be just regularization. In this respect in [2] it is proved that
any shrinking method is the solution of a properly defined regularization
problem: let us consider the prototype regularization problem

min(x − y)2 + 2g(x). (3)

x∈R

Then we have the following

Theorem 2.1 ([2], Th. 2.6) Let r : R → R be an increasing antisym-

metric function such that 0 ≤ r(y) ≤ y for y ∈ [0, ∞), r(y) → ∞. Then
there is a continuous positive solid function g : R → R such that r(y) is the
unique solution of the problem (3) for every y at which r is continuous (a
function f : X → R∪{∞} is called solid if |x| ≤ |y| ⇒ f (x) ≤ f (y)). More-
over the expression of the function g underlying the regularization problem
can be evaluated as
 x
g(x) = (s(t) − t)dt, x ≥ 0,
0

with
s(x) = sup{y|r(y) ≤ x}.

This means that the boundary among the several approaches to problem
(1) broadens quite a lot, and that the approach pursued in the present
paper, based on regularization, could provide a unifying framework for the
problem of removing noise from data.
Finally we mention that convergence properties proved for the methods
can be classified in two types, according to the purposes of the method;
both consider the L2 -loss L(f˜) := f˜− f 2 ỹ − y2 = Ỹ − Y 2 and the
corresponding risk ρ(f˜) := EL(f˜), where E means taking the average value
with respect to the noise distribution function. For regularization conver-
gence holds when L(f˜) → 0 asymptotically in some sense (in probability or
in the average). For thresholding methods the convergence for an estimator
f˜ is given in the framework of the minimax theory: let F[0, 1] be any of
s [0, 1] (the Besov classes) or F s [0, 1] (the Triebel
the function spaces Bp,q p,q
6 U. Amato, D. T. Vuza

classes) and let Fc denote the ball of functions {f : f F ≤ C}; more-

over, let ψ be a wavelet system with basis elements in C R and D vanishing
moments so that 1/p < s < min(R, D) holds. Then

Theorem 2.2 ([13], Th. 1.2) For each ball Fc arising from an F[0, 1] as
above, there is a constant C2 (Fc ; ψ) not depending on N such that for all
N = 2J
sup ρ(f˜) ≤ C2 log N inf sup ρ(fˆ),
f ∈Fc fˆ f ∈Fc

where fˆ is any estimator.

3 Regularization
Let us consider the smoothing data problem (1), with f ∈ H p . In [6],
[1] a wavelet-based regularization method was introduced for solving the
problem. The method was recently rediscovered in [9], [10]. Let fNε be a

function in the span of the first N wavelets and assuming the values ykε at
the design nodes. We consider the following penalization problem

min f˜ − fN L2 + λf˜2H p ,

ε 2
(4)
f˜∈H p

with λ being the regularization parameter.

Considering equivalent norms in the wavelet space (see [18]), problem
(4) can be written in the wavelet domain as


N −1 
N −1
min (Ỹk − Ykε )2 + λ a2p 2
j(k) Ỹk , (5)
Ỹ k=0 k=0

with
aj(k) ≡ ak := 2max{j(k),0} (6)

and Ỹk = 0 for k ≥ N .

The solution of problem (5) is simply given by

Ykε
Ỹk = . (7)
1 + a2p
k λ

Convergence of the regularized solution to the true one was proved accord-
ing to the following
Smoothing Data 7

Theorem 3.1 ([6], Th. 3.1 and [1], Th. 3) Let f ∈ H p , p > 14 . Then
  2p    1 
  σ2 2p+1 2
σ2 2p+1
f 2H p
E f˜ − f 2L2 (0,1) ≤ D(p)f  2p+1
p + + ,
H
N N N 2p

 2p
2p+1
σ2
for λ = C(p) f 2H p N
, where C(p) and D(p) do not depend on N .

Note that the rate of convergence (N −2p/(2p+1) ) is optimal in the frame-

work of nonparametric regression [17].
Two different criteria were considered for the choice of the regularization
parameter based on the available noised data only. In [1] an adaption of the
GCV criterion [21] to the wavelet framework was considered: λ is chosen
as the minimizer of VN (λ), where

(I − RN (λ)) Y ε 2
VN (λ) =  2 ,
1
N −1 Tr (I − RN (λ))

with RN (λ) being a diagonal matrix whose diagonal elements are given by
(1 + a2p −1
k λ) . For this criterion the following asymptotic efficiency theorem
has been proved.

Theorem 3.2 Let f ∈ H p , p > 1/2. Then the choice of λ provided by the
GCV is asymptotically optimal in the sense that if λN is any minimizer of
EVN (λ), then
ρλN (f˜)
lim = 1,
N →∞ minλ≥0 ρλ (f˜)

with ρλ (f˜) being the risk of the regularized solution corresponding to the
regularization parameter λ.

In [6] the regularization parameter is chosen by the Mallows criterion

endowed with an estimate of variance, σMS2 , based on finite differences of
ε
the sample y [19]: the expression to be minimized by λ is
   2 
 1 − a2p  −1
2
σMS 
J−1
j λ +
N

a2p
k λ
Λ̃(λ) = aj (Ykε )2  .
N j=−1 1+ a2p
j λ k=1 1 + a2p
k λ

For this criterion the following convergence theorem has been proved:
8 U. Amato, D. T. Vuza

Theorem 3.3 Let f ∈ H p , p > 1/2. Then


log N
Λ̃(λ) = ρλ (f˜) + .
N 2p
From Theorem 3.1 it is clear that the optimal regularization parameter
goes to zero with N . In the infinite dimensional case λ = 0 and the regu-
larized solution (7) is just the sample affected by noise; of course this does
not converge to the true function (recall that we are interested in conver-
gence when noise is finite). Therefore the infinite dimensional solution of
the regularization problem can be properly defined only as the limit of the
finite dimensional solution when the sample grows. A similar behaviour
also happens in shrinking methods based on thresholds (e.g., hard, soft),
where the asymptotic solution is defined as the limit when the threshold
goes to 0 and it is different from the value corresponding to the threshold
value t = 0.
In [5] it is proved that convergence of the regularized solution endowed
with the GCV criterion also holds in the H q -norm, 0 ≤ q < p, that is

Theorem 3.4 ([5], Proposition 3.2) If f ∈ H p+q \ {0}, 0 ≤ p < q, then

 
lim E f˜ − f 2H p+q → 0.
N

Therefore convergence of the regularized solution to the true function

holds, but in addition the same happens also for some of its derivatives,
which is interesting for those applications where derivatives are useful for
extracting features from a signal.
The wavelet regularization method has been generalized to the case of
correlated noise in [4]. The shrinking function is modified there as
1
s̃(Ykε ; λ) = , 0 ≤ k < N, (8)
1 + a2p 2
k vk λ

with vk2 being the diagonal elements of the wavelet transform of the input
error covariance matrix. In order to estimate the regularization parameter
a modification of the GCV criterion has been introduced in [4], CGCV, as

(I − RN (λ)) Y ε 22

CVN (λ) =  2
1
N Tr [C (I − RN (λ))]

with C being a diagonal matrix whose non null elements are given by vk2 .
Smoothing Data 9

In the case of correlated noise analog theorems on convergence (at the

optimal rate) and asymptotic efficiency of the CGCV criterion have been
stated. In [4] an algorithm that computes vk2 (that is the diagonal part
of the transformed covariance matrix) in O(N ) operations, instead of the
O(N 2 ) required to compute the whole matrix, is introduced.
A further generalization has been also proposed in [3], aimed at mak-
ing the wavelet regularization method fully adaptive. Indeed the shrinking
function should assume high values when the noise affecting a coefficient
is small with respect to the true coefficient itself (Signal to Noise ratio)
and low values in the opposite situation. In the regularization shrinking
function this is accomplished through the regularization parameter λ (that
depends on the overall variance of noise affecting data, see Theorem 3.1),
the coefficients vk2 (that account for noise correlation), and the values a2p k
(that in some sense account for a certain decrease of the true wavelet coeffi-
cients according to the regularity of the underlying function). It is just this
dependence on the true function that is not taken into account completely:
for example the coefficients a2p k only depend on the scale but not on the
translation parameter, so that any dependence of the shrinking function on
the latter is completely missing. Therefore the following shrinking function
has been introduced in [3] that generalizes (7) and (8):
1
s̃(Ykε ; λ) = , 0 ≤ k < N, (9)
1 + a2p
k tk λ

with tk satisfying 0 < c1 ≤ tk ≤ c2 < ∞ for any k and N .

The choice of tk based on the data only is an open problem; the optimal
tk gives rise to the known oracle
1
soracle (Yk ) = vk2 σ 2 /N
(10)
1+ Yk2

that depends on the (true) signal to noise ratio for each coefficient. As
such, Eq. (10) can be used only for theoretical purposes with synthetic
data. However, when the true (unknown) coefficients Yk2 are approximated
in some way, Eq. (10) can yield plenty of shrinking functions, including
known and possibly new ones.

4 Numerical Experiments
In the present section we show the results of numerical experiments based
on synthetic data. We shall pursue the following purposes: to validate
10 U. Amato, D. T. Vuza

the GCV regularization method basing on significant examples; to under-

stand a possible improvement of the method achievable by introducing fully
adaptability of the coefficients; to compare the performance of wavelet reg-
ularization vs. Fourier regularization.

Function Expression
11
Blocks αk (1 + sgn(x − βk ))
k=1 −4
Bumps 11
k=1 γk (1 + ζk |x − ηk |)
Heavysine 4

sin(4πx) − sgn(x − 0.3) − sgn(0.72 − x)
Doppler x(1 − x) sin(2.1π/(x + 0.05))
Ramp x,

x ≤ x0 ; x − 1 otherwise
Cusp |x − x0 |
Sing |x − x0 |−1
HiSine sin(0.6902N πx)
LoSine sin(0.3333N πx)
Leopold δ(x − x0 )
Riemann (IFFT(z)); zi2 = 1/i, 1 ≤ i ≤ N ; zi = 0 otherwise
−70 exp(−(x − 0.5)2 /2σ 2 ), 0 ≤ x ≤ 1/7
−35 exp(−(x − 0.5)2 /2σ 2 ), 1/7 < x ≤ 1/5
−70 exp(−(x − 0.5)2 /2σ 2 ), 1/5 < x ≤ 1/3
−15Bumps(x), 1/3 < x ≤ 1/2
Piece-regular − exp(4x), 1/2 < x ≤ 7/12
− exp(5 − 4x), 7/12 < x ≤ 2/3
25, 11/15 < x ≤ 49/60
exp(4x) − exp(4), 49/60 < x ≤ 403/420
0, otherwise
20(x3 + x2 + 4), 0 ≤ x ≤ 0.05
10, 0.05 < x ≤ 0.15
20(x3 + x2 + 4), 0.15 < x ≤ 0.2
10(0.6 − x)3 + 45, 0.2 < x ≤ 0.4
Piece-polynomial 40(2x3 + x) + 100, 0.4 < x ≤ 0.6
2
16x + 8x + 16, 0.6 < x ≤ 0.8
36 − 20x, 0.8 < x ≤ 0.9
150, 0.9 < x ≤ 0.95
36 − 20x, 0.95 < x ≤ 1
Table 1: List of considered functions, f (x), 0 ≤ x ≤ 1, and corre-
sponding expressions. In the table α, β, γ, ζ, η are suitable vectors,
x0 = 0.37, δ(x) is the Dirac function, IFFT is inverse Discrete Fourier
transform, σ = 6/40.
Smoothing Data 11

It is clear from the paper that wavelet regularization is introduced for

smooth functions (f ∈ H p , p > 1/4). Nevertheless, many interesting sig-
nals in applied sciences do not share such a smoothness feature, in general.
Therefore, in order that regularization be a candidate for removing noise
from real signals, its performance has be to evaluated on non smooth func-
tions. In the present paper we consider several of them, summarized in
Table 1. They have been extensively used in the current literature or im-
plemented in wavelet software.
We sampled each test function in N = 2J equispaced points according
to the model (1), with variance of noise affecting data such that the Signal
to Noise ratio is 7.
In order to evaluate the performance of the regularization method, we
introduce the index I,

 N −1
 (Ỹk − Yk )2
I =  k=0
N −1 2 . (11)
k=0 Yk

Index I is an estimate of the relative (square root) L2 -loss and therefore

it is less sensitive to the particular function. In order to damp influence
of the random noise, 100 different realizations of noise were generated and
index I 2 was averaged over them (estimate of the L2 -risk). Of course for a
fair comparison all sets of tests were run on the same noise sequence.
Another interesting index considered for evaluating performance is the
estimate of the convergence rate, γ, of the L2 -risk, ρ; it is computed by least
squares fitting the experimental estimate of the risk as a function of the
sample size according to the law ρ = αN −γ . Note that for smooth functions
γ should be related to the smoothness index of a function, pf (generally
unknown), as γ = 2pf /(2pf +1), provided that the wavelet system is regular
enough and the actual p used in regularization satisfies p ≥ pf . The fit (and
then convergence rate) is significant only when convergence holds, which,
as we shall see, is not always the case.
Table 2 shows the indexes I for a sample size N = 212 and γ (asymp-
totic value) for the functions of Table 1. The values of I shown correspond
to regularization endowed with the Generalized Cross Validation criterion
for estimating the regularization parameter. This criterion, even though
defined on an asymptotic basis, is highly efficient even for N ≥ 211 ; the
differences between the digits shown and optimal regularization (i.e., when
the regularization parameter is chosen optimally, which is possible in sim-
ulation) are negligible. Moreover p = 2 was used in regularization.
12 U. Amato, D. T. Vuza

REGULARIZATION ORACLE
Function I (N = 1012 ) γ I (N = 1012 ) γ
Blocks 0.077 0.50 0.034 0.74
Bumps 0.076 0.79 0.035 0.82
Heavysine 0.034 0.50 0.016 0.79
Doppler 0.054 0.78 0.023 0.84
Ramp 0.066 0.50 0.015 0.83
Cusp 0.020 0.56 0.013 0.78
Sing N.A. N.A. 0.017 0.83
HiSine N.A. N.A. N.A. N.A.
LoSine N.A. N.A. N.A. N.A.
Leopold N.A. N.A. 0.016 0.84
Riemann 0.12 0.30 0.086 0.30
Piece-regular 0.063 0.52 0.029 0.79
Piece-polynomial 0.083 0.49 0.032 0.77

Table 2: Numerical experiments obtained by wavelet system. For sev-

eral test functions the error index (Eq. (11)) is shown for N = 212 together
with its convergence rate, ρ, when applicable. Lack of convergence is indi-
cated by Not Applicable (N.A.). Values are shown for GCV regularization
and for the oracle solution.

First we note from the table that convergence indeed holds even when
the functions are not smooth, at a convergence rate that strongly depends
on the function itself, of course. Paradoxically, the most interesting indica-
tions come from failures of the method, that occur with highly oscillating
functions (HiSine and LoSine) and highly peaked ones (Sing and Leopold).
A natural question is whether failures depend on the wavelet regularization
method chosen (and accordingly whether the L2 -risk and the convergence
rate can be improved). To this purpose we estimated the analog indexes
I and γ for the oracle solution (Eq. 10, also shown in Table 2). They
are to be intended as a trend in a possible improvement achievable by
wavelet methods for removing noise from data based on other (nonlinear)
shrinking functions (e.g., adaptive regularization, Hard-Thresholding, Soft-
Thresholding). Comparison of left and right columns of Table 2 shows that
this is the case indeed. Convergence now holds also for highly peaked
functions (but again not for highly oscillating functions) and the rate is
much better in general. We also remark the interesting observation that
most functions exhibit a smooth-like convergence rate of about 0.8, corre-
sponding to a smoothness-like parameter of about 2. In some sense it is
Smoothing Data 13

an experimental evidence that to consider regularization with parameter

p = 2 is a good choice (this also applies to spline smoothing).
Wavelet based methods are only a class for solving the problem of noise
removal from data. Before their astonishing success, the methods described
or cited in the present paper were earlier employed in conjunction with
the Fourier transform instead of the wavelet transform. Debate is active
in the scientific community concerning the performance of wavelet based
methods against Fourier based ones and selecting the applications for which
a definite answer concerning the most performing one can be given. In the
present section we attempt an experimental answer, by performing the same
numerical experiments shown in Table 2 by means of a transform in the
Fourier domain. Results are shown in Table 3.

REGULARIZATION ORACLE
Function I (N = 1012 ) γ I (N = 1012 ) γ
Blocks 0.079 0.51 0.066 0.51
Bumps 0.079 0.81 0.064 0.81
Heavysine 0.034 0.51 0.029 0.51
Doppler 0.052 0.81 0.047 0.81
Ramp 0.064 0.52 0.056 0.52
Cusp 0.015 0.68 0.013 0.68
Sing N.A. N.A. N.A. N.A.
HiSine N.A. N.A. 0.047 0.47
LoSine N.A. N.A. 0.047 0.47
Leopold N.A. N.A. N.A. N.A.
Riemann 0.12 0.33 0.040 0.48
Piece-regular 0.064 0.53 0.054 0.53
Piece-polynomial 0.083 0.53 0.067 0.53

Table 3: Numerical experiments obtained by Fourier system. For several

test functions the error index (Eq. (11)) is shown for N = 212 together with
its convergence rate, ρ, when applicable. Lack of convergence is indicated
by Not Applicable (N.A.). Values are shown for GCV regularization and
for the oracle solution.

The first (positive) remark for Fourier regularization is that the method
is very powerful among other nonlinear shrinking functions, since the es-
timated risk approaches the ideal one within 20% in general. Also the
convergence rate is the same. There are no appreciable differences between
wavelet and Fourier regularization (compare the left columns of Tables 2
14 U. Amato, D. T. Vuza

and 3). However comparison of right columns of the same tables clearly
shows that Fourier methods are much less performing than wavelet ones
in removing noise, both in terms of risk and convergence rate. This im-
mediately means that there are significant improvements that are possible
by nonlinear methods (e.g., adaptive regularization, thresholding methods)
only in the wavelet domain. This confirms experimentally that the associa-
tion between the wavelet transform and shrinking functions can be a good
candidate for noise removal.
Finally we note that the Fourier methods are not applicable in the
case of highly peaked functions (where wavelet regularization failed), even
though linear regularization fails in cleaning noise from the signal.
We also remark that several other functions were considered for the
tests, not shown here, mainly concerned with different kinds of chirps. For
them full failure was detected with both wavelet and Fourier methods, so
that specific methods have to be developed for such signals.

5 Conclusions
The present paper dealt with the problem of removing noise from signals.
By its very nature, the problem is strictly finite dimensional. However a
quite intriguing mutual aid finite-infinite dimensional is established in solv-
ing the problem. A wavelet based regularization method that removes noise
from data is discussed in the paper. Convergence properties are given to-
gether with a data driven criterion for estimating the optimal regularization
parameter. Moreover a generalization of the method to the case of corre-
lated noise and to adaptive regularization is presented. Wavelet methods
directly compete with analogous methods that can be developed by relying
on a transform in the Fourier domain. So far we do not know about theoret-
ical results concerning a comparison of accuracy between the two methods
(e.g., examples can be worked out where Fourier regularization works bet-
ter than wavelet regularization). A numerical comparison is shown in the
present paper on the basis of a wide set of test functions. The main con-
clusions from the experiments are that wavelet and Fourier regularization
behave quite the same, as far as both accuracy and convergence rate are
concerned. However, when adaptive methods are considered, wavelet meth-
ods show a much bigger potential for improvement with respect to Fourier
methods, both in accuracy and convergence rate. Moreover wavelet meth-
ods are able to clean noise also from highly peaked signals, whereas on the
contrary Fourier methods are to be preferred for pure trigonometric signals,
Smoothing Data 15

as could be expected.

References
[1] U. Amato, D. T. Vuza, Wavelet approximation of a function from sam-
ples affected by noise, Rev. Roumaine Math. Pures Appl., 42 (1997),
481–493.

[2] U. Amato, D. T. Vuza, Besov regularization, thresholding and wavelets

for smoothing data, Numer. Funct. Anal. and Optimiz., 18 (1997),
461–493.

[3] U. Amato, M. R. Occorsio, D. T. Vuza, Fully adaptive wavelet regu-

larization for smoothing data, Rend. Circ. Matem. Palermo Serie II,
52 (1998), 207–222.

[4] U. Amato, D. T. Vuza, Smoothing data with correlated noise by

wavelet regularization, Rev. Roumaine Math. Pures Appl., 43 (1998),
253–275.

[5] U. Amato, D. T. Vuza, Wavelet simultaneous approximation from sam-

ples affected by noise, Computers & Mathematics with Applications,
36 (1998).

[6] A. Antoniadis, Smoothing noisy data with tapered Coiflet series,

Scand. J. Stat., 23 (1996), 313–330.

[7] G. Beylkin, R. Coifmann, V. Rokhlin, The fast wavelet transform and

numerical algorithms, Comm. Pure Appl. Math., 44 (1991), 141–163.

[8] L. Breiman, Better subset regression using the nonnegative garrote,

Technometrics, 37 (1995), 373–384.

[9] L. T. Dechevsky, S. I. Penev, On Penalized Wavelet Estimation, Re-

port No. S98-15, Department of Statistics, School of Mathematics, The
University of New South Wales, 1998.

[10] L. T. Dechevsky, S. I. Penev, Weak Penalized Least Squares Wavelet

Regression Estimation, Report No. S99-1, Department of Statistics,
School of Mathematics, The University of New South Wales, 1999.

[11] R. A. DeVore, G. Kyriazis, D. Leviatan, V. M. Tikhomirov, Compres-

sion and nonlinear n-widths, J. Adv. Comp. Math., 1 (1993), 197–214.
16 U. Amato, D. T. Vuza

[12] D. L. Donoho, Interpolating Wavelet Transforms, Technical Report

480, Department of Statistics, Stanford University, Stanford, 1992.

[13] D. L. Donoho, De-noising by soft-thresholding, IEEE Trans. Inf. Th.,

41 (1995), 613–627.

[14] D. L. Donoho, I. M. Johnstone, Ideal spatial adaption by wavelet

shrinkage, Biometrika, 81 (1994), 425–455.

[15] D. L. Donoho, I. M. Johnstone, Adapting to unknown smoothness via

wavelet shrinkage, J. Amer. Statist. Assoc., 90 (1995), 1200–1224.

[16] H. Y. Gao, A. G. Bruce, WaveShrink with firm shrinkage, Stat. Sinica,

7 (1997), 855–874.

[17] I. A. Ibragimov, R. Z. Khasminskii, Nonparametric Regression Esti-

mation, Dokl. Akad. Nauk SSSR, 252 (1980), 780–784.

[18] Y. Meyer, Ondelettes et opérateurs, Hermann, Paris, 1990.

[19] H. G. Muller, U. Stadtmuller, Estimation of heteroscedasticity in re-

gression analysis, Ann. Statist., 15 (1987), 610–625.

[20] W. Sweldens, R. Piessens, Quadrature formulae and asymptotic error

expansion for wavelet approximations of smooth functions. SIAM J.
Numer. Anal., 31 (1994), 1240–1264.

[21] G. Wahba, Spline Models for Observational Data, SIAM, Philadelphia,

1990.
 Rearrangements of Real Functions
Derived from the Combinatorics
of Young Tableaux
Hari Bercovici
Mathematics Department
Indiana University
Bloomington, IN 47405, USA
E-mail: bercovic@indiana.edu

Abstract. The relationship between Young tableaux and

longest increasing subsequences is extended to measurable func-
tions on an interval. We also extend Knuth equivalence to this
context, and we show that every function is Knuth equivalent
to a function associated with a continuous Young tableau.

1 Introduction
Young tableaux and their rich combinatorics is related to diverse areas of
mathematics, including representation theory, symmetric functions, the ge-
ometry and topology of Grassmannians, and module theory. More recently
it was noted in [1] that a continuous version of Young tableaux is the key to
the study of certain lattices of invariant subspaces of operators on Hilbert
space. What played a major role in [1] was an extension to continuous
diagrams of the classical Littlewood-Richardson rule. Our purpose in this
paper is to find continuous analogues of other aspects of the combinatorics
of Young tableaux. We will extend to this context Knuth equivalence and a
version of the Schensted bumping algorithm. We will recall in Section 2 the
basics of Young tableaux, and develop our continuous analogue in Section
3. It is worthwhile to note that, while the statements in Section 3 involve
real-valued functions, the proofs are done by reducing to the combinatorial
case. It may be interesting to search for direct proofs.

17
18 H. Bercovici

2 Knuth Equivalence and Young Tableaux

Our basic reference on Young tableaux is the book [2] where all the results
in this section are proved. Some of the original sources are [3] and [4]. In
this section we will consider words w = w1 w2 · · · wn over the alphabet of real
numbers, i.e., w1 , w2 , . . . , wn ∈ R. As usual, uv indicates the concatenation
of u and v. Knuth equivalence ≡K on words is the smallest equivalence
relation with the following two properties:

(a) if x, y, z ∈ R, x ≤ z < y, and u, v are words, then uxyzv ≡K uyxzv;

(b) if x, y, z ∈ R, z < x ≤ y, and u, v are words, then uxyzv ≡K uxzyv.

A basic result establishes a complete set of representatives of equiva-

lence classes of words relative to Knuth equivalence. These representatives
are words associated with Young tableaux. To explain how these are con-
structed, recall that a diagram is a left-justified collection of boxes placed
in rows with λ1 , λ2 , . . . , λk boxes, where λ1 ≥ λ2 ≥ · · · ≥ λk . A (column
strict) Young tableau is obtained by placing a number in each box of a
diagram in such a way that the rows are increasing and the columns are
strictly increasing. Thus, each number in the tableau is strictly less than
the one above (if any), and at least equal to the one to its left (if any).
Given a tableau T over a diagram with k rows, let us call wj the word
obtained by reading the elements in the jth row from left to right. The
word w(T ) of T is defined as

w(T ) = wk wk−1 · · · w1 .

Theorem 2.1 For every word w there exists a unique Young tableau T
such that w ≡K w(T ).

There are several algorithms for calculating T from w in the above

theorem. One is given by Schensted bumping and it can be done fairly
quickly by hand. Schensted was actually interested in another aspect of
words w, namely in their increasing subsequences. Assume that the tableau
T has row lengths λ1 , λ2 , . . . , λk , and w ≡K w(T ). Then w has increasing
subsequences of length λ1 and no longer increasing subsequences. More
generally, denote by Lj (w) the largest sum of the form n1 + n2 + · · · + nj
such that w has disjoint increasing subsequences of lengths n1 , n2 , . . . , nj .
Then
Lj (w) = λ1 + λ2 + · · · + λj .
Rearrangements of Real Functions 19

This gives a method for calculating the diagram of T . In order to calculate

T itself we need an additional observation. For a fixed real number α
denote by wα the word obtained from w by deleting all the letters which
are greater than or equal to α.

Proposition 2.1 If w ≡K w
, then wα ≡K w
α for every real number α.

This, combined with the fact that wα is a tableau word if w is one,

can now be used to determine the tableau T with w ≡K w(T ). It will be
convenient to set L0 (w) = 0 for all w.

Proposition 2.2 Let w be a word and T a Young tableau such that w ≡K

w(T ), say w(T ) = wk wk−1 · · · w1 , where wj are the row words of T . For
j = 1, 2, . . . , k and real α, wj has precisely Lj (wα ) − Lj−1 (wα ) elements
strictly less than α.

To see how this works, take the word w = 236145 (each digit is a letter)
and note that L1 (w) = 4 and L2 (w) = 6, which gives λ1 = 4 and λ2 = 2.
Look next at w6 = 23145 for which L1 = 4 and L2 = 5. This indicates
that 6 belongs in the last box of the second row of T . Continuing in this
way we find that the top row of T is 1345 and the bottom row 26, so that
w(T ) = 261345 ≡K 236145.
This method of calculating T is not particularly efficient, but I chose to
present it because it is the one which is most easily extended to measurable
functions.

3 Knuth Equivalence of Measurable Functions

In this section the discrete words of the preceding section will be replaced
by ‘continuous’ words. Such a continuous word is simply a real valued,
Lebesgue measurable function f with bounded domain D(f ) ⊂ R. Just
as words are entirely determined by their letters, the relevant aspect of f
will be determined by its values, not by the parametrization. Thus, two
functions f and g will be considered identical if there exists an essentially
invertible, measure preserving map ϕ : D(f ) → D(g) such that f = g ◦
ϕ almost everywhere. Thus any function f can be replaced by another
one defined on the interval (0, ), where  = |D(f )| (we denote by |σ| the
Lebesgue measure of σ). Indeed, the function ϕ : D(f ) → (0, ) defined by
ϕ(t) = |D(f ) ∩ (−∞, t)| is essentially invertible and measure preserving.
20 H. Bercovici

Concatenation is defined in the following way. A function f is the

concatenation of g and h if D(f ) = D(g) ∪ D(h), D(g) lies to the left of
D(h) (i.e., x < y if x ∈ D(g) and y ∈ D(h)), f |D(g) = g, and f |D(h) = h.
We will indicate this by writing f = g  h. Given functions g and h, one
can always form g  h; one might have to shift the domains so that the
domain of h is on the right. More generally, given any family (fi )i∈I of
functions indexed by a totally ordered countable set I, one can form the
 
concatenation i∈I fi provided that i∈I |D(fi )| < ∞. In this construction
D(fi ) must precede D(fj ) if i < j. In this paper we will only use infinite
concatenations indexed by the natural numbers with the reverse ordering,
i.e., concatenations of the form


1
fn = · · ·  fn  · · ·  f2  f1 .
n=∞

We now define the continuous version of Young tableaux. Consider

a sequence T = (fn )∞n=1 of measurable functions, with D(fn ) = (0, λn ),
where λn ≥ 0 for n = 1, 2, . . . Then T will be called a (column strict)
Young tableau if the following conditions are satisfied:

(i) λ1 ≥ λ2 ≥ · · ·;
∞
(ii) n=1 λn < ∞;

(iii) each fn is increasing (not necessarily strictly);

(iv) fn (x) ≤ fn+1 (x), for all x ∈ (0, λn+1 ); and

(v) if x < y < λn+1 , then fn (x) < fn+1 (y).

A few comments are in order. Increasing functions have at most countably

many discontinuities, so it is always possible to assume that the functions
fn in a Young tableau are continuous from the right or from the left. To
visualize conditions (iii) and (iv), let us consider the extended graph of
fn . This is a continuous curve consisting of all points (x, y) such that
x ∈ (0, λn ) and limt↑x fn (t) ≤ y ≤ limt↓x fn (t). If fn is bounded below, one
also adds the points of coordinates (0, y) with y ≤ limt↓0 fn (t). Similarly,
if fn is bounded above, the points (λn , y) with y ≥ limt↑λn fn (t) are added
to the extended graph of fn . (In the language of monotone operators,
the extended graph of fn is a maximal monotone extension of fn with
the smallest possible domain. The range of this extension is the whole
real line.) Condition (iii) indicates that the extended graph of fn+1 lies
Rearrangements of Real Functions 21

above that of fn , while condition (iv) requires that these two graphs have
no horizontal segment in common. These conditions can be expressed in
terms of the distribution functions of the fn . The distribution function µn
of fn is simply defined by µn (α) = |{t ∈ (0, λn ) : fn (t) < α}|. The function
µn is defined on the whole line, it is increasing and continuous from the
left, and its extended graph is the reflection of the extended graph of fn in
the line x = y. Thus the conditions on fn correspond with the fact that
the extended graph of µn lies above the extended graph of µn+1 , and these
extended graphs have no vertical segments in common. In other words,

(a) µn (α) ≥ µn+1 (α) for all α; and

(b) µn (α) ≥ limβ↓α µn+1 (β) for all α.

The alert reader will have noticed that condition (a) is in fact a consequence
of (b) since the functions in question are increasing and continuous from
the left. Similarly, condition (iii) on the functions fn follows from (iv) if
the functions fn are assumed continuous from the left. In the sequel, the
sequence (λn )∞ n=1 will be called the shape of the tableau T .
Given a Young tableau T = (fn )∞ n=1 , one can form the tableau function
1
fT = n=∞ fn . We can always think of fT as being defined on the interval
 ∞
(0, ), with  = ∞ n=1 λn . If we set n = j=n λj , we have fT (t) = fn (t −
n+1 ) for t ∈ (n+1 , n ).
Given a function f and a real number α we will denote by f α the
restriction of f to {t ∈ D(f ) : f (t) < α}. If T = (fn )∞ n=1 , we also set
α α ∞
T = (fn )n=1 . The following result is left as an exercise for the reader.

Lemma 3.1 If T is a Young tableau then T α is also a Young tableau.

Fix now a measurable function f and a natural number n. We will

denote by Ln (f ) the least upper bound of all sums of the form |σ1 | + |σ2 | +
· · · + |σn |, where σ1 , σ2 , . . . , σn are measurable, pairwise disjoint subsets
of D(f ) such that f |σj is increasing for every j. For convenience we set
L0 (f ) = 0. Observe that it is possible that Ln (f ) = 0 for all n, even though
|D(f )| = 0. For instance, f could be strictly decreasing.

Lemma 3.2 If T is a Young tableau of shape (λn )∞

n=1 , then

Ln (fT ) = λ1 + λ2 + · · · + λn .
22 H. Bercovici

Proof. The inequality Ln (fT ) ≥ λ1 + λ2 + · · · + λn is obvious. With the

notation used in the definition of fT , let σ ⊂ (0, ) be a measurable set such
that fT |σ is increasing. Up to a countable set we have
∞
 ∞

σ= [σ ∩ (k+1 , k )] = [σ k + k+1 ],
k=1 k=1

where σ k ⊂ (0, λk ). We claim that the sets σ k are essentially pairwise

disjoint. In fact σ k and σ k cannot have more than one point in common if


k = k
. Indeed, if x, y ∈ σ k ∩ σ k with x < y and k < k
we must have

fk+1 (y) ≤ fk
(y) = fT (y + k
+1 ) ≤ fT (x + k+1 ) = fk (x),

where we used k
≥ k + 1, the definition of fT , the inequality x + k+1 ≥
y + k
+1 , and the fact that fT is increasing on σ. This inequality is not
possible by property (iv) of Young tableaux. It follows that |σ| = |σ
|, where

σ
= ∞

k=1 σ ⊂ (0, λ1 ). Note that σ = u(σ), where u : (0, 1 ) → (0, λ1 ) is

k

defined by u(t) = t−k+1 for t ∈ (k+1 , k ). The above argument shows that
u|σ is essentially one-to-one on σ if fT |σ is increasing. Note furthermore
that u−1 (t) contains exactly j points if t ∈ (λj−1 , λj ).
Consider now pairwise disjoint sets σ1 , σ2 , . . . , σn ⊂ (0, 1 ) such that
f |σj is increasing for j = 1, 2, . . . , n. The above argument shows that

χu(σ1 ) (t) + χu(σ2 ) (t) + · · · + χu(σn ) (t) ≤ j,

for t ∈ (λj−1 , λj ), and therefore

χu(σ1 ) + χu(σ2 ) + · · · + χu(σn ) ≤ χ(0,λ1 ) + χ(0,λ2 ) + · · · + χ(0,λn )

almost everywhere. Therefore


n 
n  
n  
n 
n
|σj | = |u(σj )| = χu(σj ) ≤ χ(0,λj ) = λj ,
j=1 j=1 j=1 j=1 j=1
n
and this implies the desired inequality Ln (f ) ≤ j=1 λj . ✷

It will be useful to know that the definition of Ln (f ) is related with the

corresponding Ln defined for words. The relevant result is as follows.

Lemma 3.3 Let w = w1 w2 · · · wk be a word, and let f be a function defined

on the interval (0, ) such that f (t) = wj for t ∈ ((j − 1)/k, j/k). Then
Ln (f ) = (/k)Ln (w) for every natural number n.
Rearrangements of Real Functions 23

Proof. Any increasing subsequence of length j of w gives rise to set σ

with |σ| = (/k)j such that f |σ is increasing; σ simply consists of a union
of j intervals of length /k. Moreover, disjoint subsequences determine
disjoint subsets. This proves the inequality Ln (f ) ≥ (/k)Ln (w). For the
opposite inequality, assume that σ1 , σ2 , . . . , σn are pairwise disjoint sets
such that the restrictions f |σi are increasing. If a point of one of the
intervals Ij = ((j − 1)/k, j/k) belongs to one of the sets σi , we can add
the whole interval Ij to σi , and f will still be increasing on the resulting
set. In order to keep the sets σi disjoint, the parts that are added to a set σi
must be deleted from the other sets; this does not decrease the sum of the
measures of these sets. Applying this procedure at most k times we arrive
at a new collection σ̃1 , σ̃2 , . . . , σ̃n of pairwise disjoint sets, each one arising
 
from an increasing subsequence of w, and with nj=1 |σ̃j | ≥ nj=1 |σj |. This
proves the opposite inequality Ln (f ) ≥ (/k)Ln (w). ✷

For an arbitrary function f let us denote

λn (f ) = Ln (f ) − Ln−1 (f ), n = 1, 2, . . .

Lemma 3.2 says simply that the numbers λn (fT ) give precisely the shape
of the Tableau T .

Theorem 3.1 For any measurable function f , the sequence (λn (f ))∞
n=1 is
decreasing.

Proof. We need to prove that 2Ln (f ) ≥ Ln−1 (f ) + Ln+1 (f ) for n ≥ 1. Fix

a positive number η and choose sets σ1 , σ2 , . . . , σn−1 , σ1
, σ2
, . . . , σn+1 such
that

(i) the restrictions f |σi and f |σj

are increasing;

(ii) σi ∩ σj = ∅ for i = j, σi
∩ σj
= ∅ for i = j;
n−1 n+1
(iii) i=1 |σi | ≥ Ln−1 (f ) − η and j=1 |σj
| ≥ Ln+1 (f ) − η.
   
n+1

Let us set g = f |σ, where σ = i=1 σi ∪

n−1
j=1 σj . Since Ln (f ) ≥ Ln (g),
while clearly Ln±1 (g) ≥ Ln±1 (f ) − η, we have

2Ln (f ) − Ln−1 (f ) − Ln+1 (f ) ≥ 2Ln (g) − Ln−1 (g) − Ln+1 (g) − 2η.

If the theorem were true for functions of the type of g, the result would
follow for f as well because η is arbitrary. Therefore it suffices to prove the
24 H. Bercovici

theorem for functions f for which there exist measurable sets ω1 , ω2 , . . . , ωm


with D(f ) = m j=1 ωj , and the restrictions f |ωj are increasing. Assume
therefore that f satisfies this property, and observe that the sets ωj can
be
 assumed to be pairwise disjoint; one just needs to replace ωj by ωj \
i<j i for j > 1. Fix again a positive number η and select compact sets
ω
 m
Kj ⊂ ωj such that m j=1 |ωj \ Kj | < η. Define then K = j=1 Kj and h =
f |K. Since |D(f )| − |D(h)| < η it is clear that Lk (h) ≤ Lk (f ) ≤ Lk (h) + η
for all k. Therefore
2Ln (f ) − Ln−1 (f ) − Ln+1 (f ) ≥ 2Ln (h) − Ln−1 (h) − Ln+1 (h) − 2η.
Again, since η is arbitrary, it follows that it suffices to prove the theorem for
functions whose domain is a disjoint union of compact sets K1 , K2 , . . . , Km
such that the restrictions f |Kj are increasing. Assume therefore that f has
this property. There exists then a positive number ε such that |x − y| > ε
whenever x ∈ Ki , y ∈ Kj , and i = j. None of the intervals [qε, (q + 1)ε]
(with q an integer) will intersect more than one of the sets Kj , and a
finite number of these intervals, say I1 , I2 , . . . , Ip , will cover D(f ). We
may assume that Ij = [qj ε, (qj + 1)ε] with q1 < q2 < · · · < qp . If we
set now fj = f |(Ij ∩ D(f )), we see that f1 , f2 , . . . , fp are increasing and
f = f1  f2  . . .  fp . At this point it will be convenient to make a
measure preserving change of variable so that f is defined on an interval
(0, αp ), where 0 = α0 < α1 < · · · < αp and f |(αj−1 , αj ) is increasing for
j = 1, 2, . . . , p. Once more fix a positive number η and set = N = [αp /η].
Define next a function v which is constant and equal to inf t∈((j−1)η,jη) f (t)
on the interval ((j −1)η, jη) for 1 ≤ j ≤ N . Note that at most p−1 of these
intervals contains one of the points αi ; call ω the union of these intervals. If
f |σ is increasing then clearly v|σ \ω is increasing and |σ \ω| ≥ |σ|−(p−1)η.
We conclude that Lk (v) ≥ Lk (f ) − (p − 1)η for all k. If, on the other hand,
σ is a set such that v|σ is increasing, then f |σ is not necessarily increasing,
but f will be increasing on the smaller set obtained by removing from σ \ ω
the intervals of the form ((j − 1)η, jη) which intersect σ, and are closest to
the left and right of one of the points αi , 1 ≤ i < p. This way we obtain the
inequalities Lk (v) ≤ Lk (f ) + 3k(p − 1)η. As in the preceding reductions,
it easy to see now that it suffices to prove the theorem for the function
v. However, v is derived from a word like in Lemma 3.3, and the desired
inequality follows from that lemma and from the facts presented in Section
2. ✷

Let us note that the functions α → Ln (f α ) are increasing and left-

continuous; this follows from the fact that limε↓0 |{t : f (t) ∈ (α−ε, α)}| = 0.
Rearrangements of Real Functions 25

The limits limβ↓α Ln (f β ) can also be calculated. Indeed, if ε > 0 is small,

and σ ⊂ D(f α+ε ) is such that f |σ is increasing then f |σ
, σ
= σ ∩ {t :
f (t) ≤ α}, is also increasing and |σ| − |σ
| ≤ |{t : f (t) ∈ (α, α + ε)}|, and
this quantity tends to zero as ε → 0. We conclude that limβ↓α Ln (f β ) =
Ln (f α+ ), where D(f α+ ) = {t : f (t) ≤ α}.
We can now improve the preceding theorem.

Theorem 3.2 Let f be a measurable function, and set µn (α) = Ln (f α ) −

Ln−1 (f α ) for α ∈ R and n = 1, 2, . . .

(1) For each α, the sequence (µn (α))∞

n=1 is decreasing.

(2) For each n, µn (α) is increasing and continuous from the left.

(3) For each n, limα→−∞ µn (α) = 0 and limα→∞ µn (α) = λn (f ).

(4) For each n and α we have µn (α) ≥ limβ↓α µn+1 (β).

Proof. Statement (1), and the left continuity in (2) follow from the preced-
ing theorem and the discussion above. The equalities in (3) follow easily
from the fact that

lim |{t : f (t) < α}| = lim |{t : f (t) ≥ β}| = 0.

α→−∞ β→∞

To verify (4) and the remainder of (2) fix a positive number η. As in the

proof of the preceding theorem, we can find a restriction g of f which is
a finite concatenation of increasing functions, such that limn→∞ Ln (f ) −
|D(g)| < η. Since Ln (g α ) ≤ Ln (f α ) ≤ Ln (g α ) + η, we must also have
|λn (g α ) − λn (f α )| < 2η for every n and α. Thus it suffices to prove (4) and
(2) in case f is a finite concatenation of increasing functions. Similarly, as
in the conclusion of the proof of Theorem 3.1, we can further reduce to the
case when f is associated with a word as in Lemma 3.3. For these functions
the theorem follows from the results described in Section 2. ✷

Corollary 3.1 For every measurable function f on a bounded set there

exists an essentially unique Young tableau T such that Ln (f α ) = Ln (fTα )
for every α ∈ R and every natural number n.

Proof. Let µn be as in the preceding result and define fn to be the unique

left-continuous function which has distribution function µn . Then T =
26 H. Bercovici

(fn )∞
n=1 is a Young tableau by properties (2) and (4) of the preceding result.
Clearly T satisfies the requirements of the corollary. ✷

Consider the class F of measurable functions with the property that

limn→∞ Ln (f ) = |D(f )|. The natural extension of Knuth equivalence to
the class F is as follows: f, g ∈ F are Knuth equivalent if Ln (f α ) =
Ln (g α ) for every α ∈ R and every n = 1, 2, . . . We indicate this equivalence
by f ≡K g. Observe that the functions f α belong to F if f ∈ F, and
therefore Knuth equivalent functions have the same distribution, that is
|{t : f (t) < α}| = |{t : g(t) < α}|. In particular, if f ∈ F and T is the
tableau associated with f by the preceding corollary, the function fT can
be regarded as a rearrangement of f . As in the combinatorial situation, the
data for performing this rearrangement can be codified in a second Young
tableau (a standard one), thus giving rise to an extension of the Robinson-
Schensted correspondence. We plan to pursue these matters in future work.
Let us just note that a Young tableau T is standard if the tableau function
fT is an essentially invertible, measure preserving transformation of its
domain (0, ).
It is instructive to look at a simple example. Consider the function
f : (0, 2) → R defined by f (t) = t2 for t ∈ (0, 1] and f (t) = t − 1 for
t ∈ (1, 2). This is not a Young tableau and one can calculate L1 (f ) = 5/4,
L2 (f ) = 3/4. The longest set on which f is increasing is (0, 1/2] ∪ [5/4, 2).
Observe that the break points 1/2 and 5/4 have the property that f (1/2) =
f (5/4) and f
(1/2) = f
(5/4); this is not an accident. The reader should
have no difficulty finding the Young tableau associated to f .

References
[1] H. Bercovici, W. S. Li, T. Smotzer, A continuous analogue of the
Littlewood-Ridchardson rule and its application to invariant sub-
spaces, Advances in Mathematics, 134 (1998), 278–293.

[2] W. Fulton, Young Tableaux, With Applications to Representation The-

ory and Geometry, Cambridge University Press, Cambridge, 1997.

[3] D. E. Knuth, Permutations, matrices, and generalized Young tableaux,

Pacific Journal of Mathematics, 34 (1970), 709–727.

[4] C. Schensted, Longest increasing and decreasing subsequences, Cana-

dian Journal of Mathematics, 13 (1961), 179–191.
 A Possibilistic Distance
for Sequences of Equal and Unequal Length
Massimo Borelli1 , Andrea Sgarro2
1,2 Department of Mathematical Sciences
University of Trieste, 34100 Trieste, Italy
1 Scuola
Media Superiore Italiana di Rovigno
Rovinj, Croatia
E-mail: sgarro@univ.trieste.it

Abstract. We generalize Hamming distances to sequences of

possibilistic letters: this corresponds to passing from ordinary
binary logics to an infinite-valued logical setting. Our proposal
is validated on a possibilistic model of a noisy communication
channel, as opposed to the random-noise models which are cur-
rent in information theory; it might prove to be a basis for
“soft decoding” of noisy data. We take into account both syn-
chronous and non-synchronous channels; the latter require a fur-
ther generalization to sequences of unequal length in the spirit
of Levenštejn distance; by so doing, we re-take a problem which
prof. Solomon Marcus had suggested to the second author when
both were even younger than they are nowadays.

1 History and Introduction

It has been very early understood that a binary (black and white) logical
apparatus was wide of the mark if one wanted to cope with the complexity
of human thought; even Aristotelic logics is ternary, having at its disposal
three logical values, false = 0, true = 1, and possible = 12 (the numerical
encoding is of course modern). A much bolder step has been taken in this
century by such giants of modern logics as La ukasiewicz or Moisil; in multi-
valued logics the range of (numerically encoded) logical values can cover

27
28 M. Borelli, A. Sgarro

the whole of the interval [0, 1] and so spreads over a continuum; the ba-
sic logical operations, disjunction, conjunction and negation are (usually)
implemented by the mathematical operations of maximum, minimum and
complementation to 1. Nowadays, quite a variety of “soft logics” are avail-
able, say fuzzy logics, or possibility theory; the relationship between these
approaches is clear at the formal level, but unfortunately formal coinci-
dence or inclusion is not a proof that a similar relationship holds also at
the level of contents, and so philosophers are offered a chance of unend-
ing quarreling. It should be pointed out that these theories are not just
a matter of philosophical discussions: soft (flexible, nuanced) logics have
proved to be extremely advantageous for industrial applications. Actually,
and rather unexpectedly, the motto of these sophisticated tools when used
in the “real world” appears to be: cheap, quick, and effective. The inspir-
ing idea is trying to achieve the “unreasonable success” of human thought,
which is imprecise, vague and ambiguous, but in practice quite reliable and
surprisingly effective. When one of us is driving, one does not solve dif-
ferential equations to avoid crashing onto the guardrail: now, why should
a robot driver be obliged to do so, when he has to perform the very same
task? Therefore, the sophisticated and abstract meditations of L a ukasiewicz
or Moisil have in the end produced a whole range of very concrete objects,
which are supposed to make our life more pleasant (and which do make a
few of us richer).
A compelling generalization of black-and-white Hamming distances be-
tween strings to a “soft” setting was investigated in [1]; the starting point
was a problem in the classification of languages, more precisely Romance
languages, as due to the well-known Croatian romanist Ž. Muljačić, who
stressed the role of the (extinct) Dalmatic language as a “bridge” between
the western group and the eastern group (i.e., Romanian in its four vari-
ants, Dacoromanian, Aromanian, Meglenoromanian and Istroromanian, as
spoken south of Trieste). Now, some of the linguistic features were fuzzy.
At that time prof. Solomon Marcus pointed out to the author of [1] the
interest of further extending fuzzy Hamming distances to strings of unequal
lengths; the thing was put forward so clearly that presumably prof. Marcus
knew very well how to solve the problem he was stating. Unfortunately,
the extension to unequal lengths given by the inexperienced author of [1]
was of a very limited scope only. In this paper we shall re-take professor
Marcus’ request, though we have to frankly acknowledge a deplorable delay
of more than twenty years.
We shall have to extend the notions of [1]; actually the distance in [1]
A Possibilistic Distance 29

involves real (fuzzy) logical values rather than possibilistic letters, as dealt
with below; so, it is again a moot point whether our formal generalization
has also some “philosophical” contents. We shall validate our proposal on
a possibilistic model of a noisy communication channel, as opposed to the
random-noise models which are current in information theory. Actually,
our approach may prove to be the basis for “soft decoding” of noisy data.
A masterly and classical reference to non-classical logics is [2]. As a
standard reference to fuzzy sets we suggest [3], where several similarity
and dissimilarity measures are listed; as a standard reference to possibility
theory cf., e.g., [4]. A comprehensive introduction to “unorthodox” theories
of uncertainty representation and management is [5].

2 Possibilistic Letters and Their Transmission

We recall that a possibility distribution Π over the finite set, or alphabet,
A = {a1 , . . . , aK }, K ≥ 2, is defined by the possibility vector (x1 , . . . , xK )
whose components are the possibilities of the K singletons: Π(ai ) = xi ;
0 ≤ xi ≤ 1, 1 ≤ i ≤ K. For each subset A ⊆ A one sets Π(A) = maxai ∈A xi ,
Π(∅) = 0. Observe that we do not even require Π(A) = 1, i.e., maxi xi =
1. So, we allow for incomplete possibility vectors; incomplete theories of
uncertainty management are quite popular nowadays, but actually go back
to Rényi’s incomplete probabilities; cf. [6]. Nowadays incompleteness is
rather associated with the representation of “self-contradictory” states of
knowledge; cf., e.g., [7,8].
Consider a noisy medium through which letter aj ∈ A is sent. We shall
describe the noisy medium by means of a possibilistic channel. This means
that at the output a possibilistic letter Y is received which is described
through a possibility vector (y1 , . . . , yK ), 0 ≤ yj ≤ 1; we shall write di-
rectly Y = (y1 , . . . , yK ). Also the input letter aj can be thought of as a
possibilistic letter X; just set xj = 1, else xi = 0. Actually, we shall not
in any way distinguish between such a possibilistic letter X and the corre-
sponding alphabet symbol aj ; we shall usually write X = aj rather than
X = (0, . . . , 0, 1, 0, . . . , 0). Should one assume that the possibilistic chan-
nel increases the possibility vector which describes the input letter, then
yj = 1; however, this is never required below. For the moment being we
confine ourselves to observation channels, i.e., to the transmission of single
letters, rather than to the transmission of whole sequences; as for these
cf. Section 4. What we need is a sort of “distance” between possibilistic
letters to be put to good use in minimum distance decoding, as explained
30 M. Borelli, A. Sgarro

in Section 4; we stress that the term “distance” is used in this paper in a

purely heuristic sense.
Example. Let A be the standard stock of printed letters; say capital letter
E has been transmitted and its lower part is erased during transmission;
the possibilistic letter Ξ output by the possibilistic channel has 1 in corre-
spondence to E and F, else zero.
We define some terms. The possibilistic letter X is a deterministic let-
ter, when for a specified letter aj ∈ A one has xj = 1, while xi = 0 whenever
i = j (deterministic letters should not be confused with crisp letters, a more
general notion when all the logical values are either zero or one). In our
context we find it convenient to say that X is a (partial) erasure to mean
that at least two of its logical values are equal to 1 while all the others equal
to zero. In a total erasure X all the logical values are equal to 1; we stress
that a total erasure is located, in the sense that, when it is received at the
output of a possibilistic channel, we know that an input letter must have
been transmitted; in Section 4 we shall also deal with erasures which are
not located. X is a random letter when the sum of its possibilistic values

xi is equal to 1: i xi = 1; the binary case K = 2 will be of a special
interest to us, as dealt with below (we stress that the term random should
be understood in a purely formal sense, our context being not a proba-
bilistic one). Deterministic letters and erasures are normal, or complete, in
the sense that maxi xi is equal to 1. Instead, random letters are usually
incomplete, being complete only in the deterministic case. A possibilistic
letter X is concentrated on ai when xi is the only positive possibility com-
ponent; deterministic letters are both concentrated and complete. More
generally, the support of a possibilistic letter is the subset of the alphabet
A which corresponds to possibility components which are strictly positive.
The extreme case of incompleteness is the all-zero possibilistic letter, whose
support is void: we shall call it the (totally) inconsistent letter.
We define the incompleteness, or inconsistency, ι(X) of the possibilistic
letter X by setting:
ι(X) = 1 − max xi (1)
i
One has: 0 ≤ ι(X) ≤ 1, with ι(X) = 0 iff X is normal, ι(X) = 1 iff X
is totally inconsistent. One defines a partial ordering between possibilistic
letters by setting:
X ≤ Y ⇐⇒ xi ≤ yi for all i. (2)
In this ordering the first element is the totally inconsistent letter for which
all the possibilities are zero, while the last element is the total erasure for
A Possibilistic Distance 31

which all the possibilities are equal to 1. One has:

X ≤ Y =⇒ ι(X) ≥ ι(Y ) (3)

Observe that, in our application to channels, we might have considered

only distances between an input letter X and an output letter Y where the
input X is constrained to be deterministic; this restriction is never used in
the section to follow, however, since it would much hamper the flexibility
of the model.
We need one more technicality. If I is a set (finite in our case), we
shall say that a symmetric function δ : I 2 → + is a pseudometric when
it satisfies, for all x, y, z ∈ I:

0 ≤ δ(x, x) ≤ δ(x, y) ≤ δ(x, z) + δ(z, y)

The inequality on the right is called the triangle inequality. Observe that
we do not even require δ(x, x) = 0. To achieve this without jeopardizing
the triangle inequality, it would be enough to set δ
(x, y) = 0 if x = y, else
δ
(x, y) = δ(x, y); this change would be of no use in our context, though.
Actually, our “distance” as defined below will verify the triangle inequality
only when the “triangulating” element z is constrained to belong to a subset
to be specified.

3 A Possibilistic Non-Metric Distance

Let X = (x1 , . . . , xK ) and Y = (y1 , . . . , yK ) be two possibilistic letters
described by the corresponding possibility vectors over the alphabet A of
K deterministic letters. It is possible that X and Y are equal when there
exists a deterministic letter aj such that it is possible both that X = aj and
Y = aj . Consequently, we tentatively set:

ρ(X, Y ) = max min(xi , yi ) (4)

1≤i≤K

to measure the similarity between the two possibilistic letters X and Y .

We stress the question was not “are the two letters equal?” but rather “is
it possible that they are equal?” Since we need a dissimilarity index, rather
than a similarity one, we (still tentatively) define the possibilistic distance
between the two possibilistic letters X and Y simply as:

d(X, Y ) = 1 − ρ(X, Y )
32 M. Borelli, A. Sgarro

Example. Let X = E be deterministically equal to capital letter E,

Y = F be deterministically equal to capital letter F, let Ξ be the E/F
erasure as in the example above, let Ω be a total erasure as described by
the all-one possibility vector. Then: d(E, F) = 1, d(E, Ξ) = d(F, Ξ) = 0,
d(E, Ω) = d(F, Ω) = d(Ξ, Ω) = 0. All this is as it should be.
As for the formal properties of d(X, Y ), which is obviously symmetric,
we have:
0 ≤ d(X, X) ≤ d(X, Y ) ≤ 1 (5)
as soon checked. The “distance” d(X, Y ) takes on the minimum value 0 iff
xi = yi = 1 for some i, i.e., when X = ai and Y = ai , are both possible
options (which are not necessarily true); d(X, Y ) takes on its maximum
value 1 iff the supports of X and Y are disjoint. In particular d(X, X)
coincides with incompleteness as defined in (1) above:

d(X, X) = ι(X)

and so is zero iff X is normal, and is 1 iff X is totally inconsistent. There-

fore, incomplete letters are characterized by a positive “self-distance”, and
this does make sense. Actually, the lattice structure of possibilistic letters
allows one to always re-express the distance d(X, Y ) as an incompleteness,
more precisely as the incompleteness of the “lattice meet” X ∧ Y , whose
possibilities are equal to min(xi , yi ):

d(X, Y ) = ι(X ∧ Y ) (6)

(recall that the lattice meet is the “highest” element which precedes both
arguments).
The case of the “usual” (deterministic) Hamming distances and the
fuzzy case of [1] are both re-found; the latter, which is rather more complex,
will be dealt with in an addendum at the end of this section.
Deterministic letters. Say X and Y are deterministic: X = ai ,
Y = aj . Then d(X, Y ) is equal to 0 or 1 according whether i = j or i = j,
as in the usual Hamming case. More generally:

X = ai is deterministic =⇒ d(X, Y ) = 1 − yi (7)

Still more generally, if X is concentrated on ai , then d(X, Y ) = 1 −

min(xi , yi ). As for the triangle property, the following example shows that
unfortunately it does not hold even if two of the letters involved are con-
strained to be deterministic.
A Possibilistic Distance 33

Example. Take X = E, Y = F, both deterministic, and Ξ as in the

examples above. Then d(E, Ξ) + d(Ξ, F) = 0 < d(E, F) = 1. Actually, it is
enough to take Ξ concentrated on {E, F} and such that ξE + ξF > 1 to
have d(E, Ξ) + d(Ξ, F) = 2 − (ξE + ξF ) < d(E, F).
However the triangle property does hold when it is the “triangulating”
letter Z which is constrained to be deterministic:

Theorem 3.1 If Z = aj is deterministic then d(X, Z) + d(Z, Y ) ≥

d(X, Y ). A criterion for equality is that at least one of the following con-
ditions i) or ii) holds true: i) d(X, Z) = 0 and d(Y, Z) = d(Y, Y ); ii)
symmetrically, d(Y, Z) = 0 and d(X, Z) = d(X, X).

Proof. Because of (7) the first side is (1−xj )+(1−yj ); so d(X, Z)+d(Z, Y ) ≥
max [(1 − xj ), (1 − yj )] ≥ mini [ max [(1 − xi ), (1 − yi )] = d(X, Y ). To
have equality in the first inequality either xj = 1 or yj = 1, i.e., either
d(X, Z) = 0 or d(Y, Z) = 0. Suppose that d(X, Z) = 0; then to have
equality also in the second inequality one must have yj = maxi yi , i.e.,
d(Y, Z) = d(Y, Y ). ✷

Theorem 3.1 is useful in view of the applications to channels, where one

needs distances of the type d(X, Y ) with X deterministic and variable, Y
fixed. We find it meaningful to determine the set of all possibilistic letters
Z that “universally” verify the triangle inequality; so doing we generalize
the inequality part of Theorem 3.1.

Theorem 3.2 Let ∆ be the set of possibilistic letters Z such that d(X, Z)+
d(Z, Y ) ≥ d(X, Y ), ∀X, ∀Y ; let ∆∗ ⊇ ∆ be the set of possibilistic letters
Z such that d(X, Z) + d(Z, Y ) ≥ d(X, Y ), for all X and Y deterministic;
let ∆∗∗ be the set of possibilistic letters Z whose two highest possibility
components, zM and zm , say, sum at most to 1: zM + zm ≤ 1. Then
∆ = ∆∗ = ∆∗∗ .

Proof. The last example shows that a necessary condition for Z to belong
to ∆∗ , and so to ∆, is Z ∈ ∆∗∗ . We go to sufficiency and prove that
Z ∈ ∆∗∗ implies Z ∈ ∆. In terms of ρ(X, Y ) , which was defined in (4), the
triangle inequality is re-written as ρ(X, Z) + ρ(Z, Y ) ≤ 1 + ρ(X, Y ). We
distinguish the case when d(X, Z) and d(Z, Y ), or, equivalently, ρ(X, Z)
and ρ(Z, Y ), are achieved in the same position and the case when it is
not so. Actually, the triangle inequality always holds when ρ(X, Z) and
ρ(Z, Y ), are achieved in the same position, i, say, since one has ρ(X, Z) +
ρ(Z, Y ) = min(xi , zi ) + min(zi , yi ) ≤ xi + yi = max(xi , yi ) + min(xi , yi ) ≤
34 M. Borelli, A. Sgarro

1+min(xi , yi ) ≤ 1+ρ(X, Y ). Assume now that the positions where ρ(X, Z)

and ρ(Z, Y ) are achieved are distinct, i = j, say. Then ρ(X, Z) + ρ(Z, Y ) =
min(xi , zi ) + min(zj , yj ) ≤ zi + zj ≤ zM + zm ≤ 1 ≤ 1 + ρ(X, Y ). ✷

As a remark, formula (6) soon implies that the triangle inequality always
holds when the three letters involved are ordered; cf. (2). Assume that Z is
the “triangulating” sequence and assume, without real restriction, X ≤ Y .
If X ≤ Y ≤ Z, then the triangle inequality becomes ι(Y ) ≥ 0 and so holds
with equality iff Y is normal (in this case, if Y is normal, then also Z is
such). If X ≤ Z ≤ Y , then the triangle inequality becomes ι(Z) ≥ 0 and
so holds with equality iff Z is normal. If Z ≤ X ≤ Y , then the triangle
inequality becomes 2ι(Z) ≥ ι(X) and so holds with equality iff Z is normal
(use (3) with Z ≤ X). Note that in all three cases a necessary condition to
have equality is that the triangulating letter Z should be normal.

Addendum: Fuzzy Features

Assume K = 2; assume that the second logical value is bound to negate,
i.e., complement, the first, i.e., X = (x, 1 − x), Y = (y, 1 − y). Formally, we
are dealing with a binary random letter, but the meaning is rather pres-
ence/absence of a fuzzy feature. Then d(X, Y ) is precisely the pseudometric
δ(x, y) as given in [1], as we now argue.

Example. Say a Romance language is being studied and X denotes pres-

ence/absence of rotacism in that language; if this linguistic feature is only
weakly present in the language one can set X = ( 12 , 12 ), as the linguist
Muljačić used to do in similar cases.
If x1 , . . . xi , . . . are logical values, then we set:

σ(x1 , . . . , xi , . . .) = min(x1 , . . . , xi , . . . , 1 − x1 , . . . , 1 − xi , . . .).

One has 0 ≤ σ ≤ 12 ; σ = 0 if at least one of the logical values is crisp (i.e.,

either zero or one), σ = 12 if all the logical values are equal to 12 , i.e., if
they are all “totally uncrisp”. In [1] the following pseudo-metric distance
between logical values was investigated:

δ(x, y) = max [ min(x, 1 − y), min(1 − x, y)]

So δ(x, y) is the logical value of the proposition: “either the first logical
value is true and the second is false, or the first logical value is false and
the second is true” (i.e., “the two logical values are different”). Notice that
A Possibilistic Distance 35

δ(x, x) = σ(x), and so is zero iff the logical value x is crisp, as it should

be. The maximum value δ(x, y) = 1 is reached when x and y are crisp and
distinct. The equivalence with the alternative definition (8) as given below
is checked in [1]; (8) stresses the relation between δ(x, y) and is the “usual”
Euclidean distance |x − y|.

δ(x, y) = |x − y| + σ(x, y). (8)

Actually, we also have:

1 + |x − y| − |1 − x − y|
δ(x, y) = . (9)
2
Proof. Because of (8), it is enough to check that:

1 − |x − y| − |1 − x − y|
σ(x, y) = .
2
Assume x = σ(x, y) (else use the transformations x ↔ y and/or x ↔ 1 − x,
y ↔ 1 − y, which leave both sides unchanged). Then also the second side
is equal to x. ✷
Now, for X = (x, 1 − x), Y = (y, 1 − y), ρ(X, Y ) = δ(x, 1 − y) and
d(X, Y ) = 1 − δ(x, 1 − y) = δ(x, y); the last equality is soon checked by
means of (9).

4 Sequence Distances and Possibilistic Channel

Decoding
We now extend possibilistic distances from letters to sequences; our pro-
posal is validated by means of two models of possibilistic channels, the first
synchronous, the second asynchronous.

Synchronous channel (channel with located erasures)

In this case the channel is used n times; n is the length (number of pos-
sibilistic letters) both of the input and the output sequence. Only some
sequences, called codewords, are admissible as input sequences; the out-
put sequence will be decoded to a codeword which is “nearest”. This case
ties up with usual channel decoding when one minimizes the usual (deter-
ministic) Hamming distance; accordingly, we define the Hamming distance
between possibilistic n-sequences as the sum of the Hamming distances of
36 M. Borelli, A. Sgarro

its letters, i.e., as the “number of positions” (in a possibilistic sense) where
they differ: 
d(X, Y ) = d(X , Y ). (10)
1≤≤n

Here X and Y are the -th possibilistic letter of the input n-sequence X
and the output n-sequence Y , respectively. Properties for d(X, Y ) are soon
obtained from (5); in particular: 0 ≤ d(X, Y ) ≤ n. Because of Theorem 3.1,
if all the letters of sequence Z are deterministic, one has d(X, Z)+d(Z, Y ) ≥
d(X, Y ).

The fuzzy Hamming distance of [1] is re-found when the letters

X1 , . . . , Xn which make up the sequence X are all constrained to be bi-
nary and “random”. In this case the possibilistic n-sequence X can re-
interpreted as a fuzzy subset FX ⊆ {1, 2, . . . , n} described through its indi-
cator function (x1 , . . . , xn ), X = (x , 1 − x ). As an example, the linguist
Muljačić used to characterize a language through n linguistic fuzzy fea-
tures, the logical value x telling whether the -th feature, e.g., rotacism,
was present in the language; the Hamming distance between two languages
expressed how far apart the two languages were with respect to the se-
lected features. Notice that the fuzzy Hamming distance d(X, Y ) as in
[1] is nothing else but the fuzzy cardinality |FX " FY | of the symmetric
difference between the corresponding fuzzy subsets, FX and FY .
Example. Say the synchronous channel is used n = 5 times; the possible
codewords to be used as input sequences are all the meaningful English
word of length 5. Let Ξ be the E/F erasure, as above. The output se-
quence ΞΞAST is decoded to FEAST, since EEAST, EFAST, FFAST are
not possible inputs; d(ΞΞAST, FEAST) = 0, d(ΞΞAST, BEAST) = 1,
BEAST being a possible input, i.e., a meaningful English word of length 5.

Asynchronous channel (channel with located and unlocated era-

sures)
In this case the channel is used n times; n is the length of the input se-
quence, while the length of the output sequence can be smaller than n, m
say, because unlocated erasures may have occurred; the channel can erase
letters in such a way that there is no way to know where this has taken
place. This case ties up with the usual Levenštejn distance, rather than
Hamming distance; it is a “bounded” Levenštejn distance, though, n be-
ing the bound. We observe that to come closer to a genuine “unbounded”
Levenštejn distance, we should have considered also unlocated insertions,
A Possibilistic Distance 37

and not only unlocated erasures; we think however that our bounded case
is enough to cover quite a meaningful class of noisy communication media.
We can try to compensate for unlocated erasures by re-insertions of
letters, these re-insertions having their cost, however. We shall take the easy
way out and assume, as it is standard, that each re-insertion has cost one,
the same incurred when the decoder confuses two (distinct) deterministic
letters. To express this fact we find it convenient to add a new “letter”
ℵ which has the property that d(X, ℵ) = d(ℵ, X) = d(ℵ, ℵ) = 1 whatever
the possibilistic letter X may be. Observe that (5) still holds when the
letters X and/or Y are allowed to be equal to ℵ; the triangle inequality
always holds when at least one of the three letters involved is equal to ℵ;
compare with Theorem 3.2. Now, we shall extend the output sequence Y
to a sequencen
U of length n, by inserting n − m times the letter ℵ; this can
be done in m ways. We shall write Y $ U . We generalize (10) and set:

d(X, Y ) = min d(X, U ).

U :Y U

Clearly:
n − m ≤ d(X, Y ) ≤ n.

Moreover, if one assume that X and Z have both length n (are two possible
input sequences, or codewords), and Z is deterministic, then the triangle
inequality d(X, Z) + d(Z, Y ) ≥ d(X, Y ) does hold. A proof of this fact
follows. Say U achieves d(Z, Y ): Y $ U , d(Z, Y ) = d(Z, U ); since U
belongs also to the minimization set in the definition of d(X, Y ) one has
d(X, Y ) ≤ d(X, U ). Then the triangle inequality for sequences of length n
(cf. Theorem 3.1) gives d(X, Z)+d(Z, Y ) = d(X, Z)+d(Z, U ) ≥ d(X, U ) ≥
d(X, Y ).

Example. Say the asynchronous channel is used n = 11 times and the out-
put sequence is ΦΨNVERSY; for the example’s sake, here Φ is a possibilistic
letter which has 1 in correspondence to vowels, and 12 in correspondence to
consonants, while Ψ has 1 in correspondence to consonants, and 12 in corre-
spondence to vowels. There are n − m = 3 unlocated erasures, and so the
distance has to be at least equal to 3. Ω denotes a (located) total erasure,
as above. The decoded sequence is ANNIVERSARY; actually, by taking
U = ΦΨNℵVERSℵℵY one has d(ANNIVERSARY, ΦΨNVERSY) = 3; in-
stead d(CONTROVERSY, ΦΨNVERSY) = 4; take U = ΦΨNℵℵℵVERSY.
38 M. Borelli, A. Sgarro

References
[1] A. Sgarro, A fuzzy Hamming distance, Bulletin Mathématique, 21, 1-2
(1977).

[2] Gr. C. Moisil, Încercări vechi şi noi de logică neclasică, Editura
Ştiinţifică, Bucureşti, 1965.

[3] D. Dubois, H. Prade, Fuzzy Sets and Systems: Theory and Applica-
tions, Academic Press, 1980.

[4] D. Dubois, H. Prade, Possibility Theory, Plenum Press, 1987.

[5] G. J. Klir, T. A. Folger, Fuzzy Sets, Uncertainty and Information,

Prentice-Hall, 1988.

[6] A. Rényi, On measures of entropy anf information, now in Selected

Papers of Alfréd Rényi, 2, Akadémiai Kiadó, Budapest, 1976, 565–
579.

[7] A. Sgarro, An open-frame theory of incomplete interval probabilities,

International Journal of Uncertainty, Fuzziness and Knowledge-Based
Systems, 6, 6 (1998).

[8] P. Castellan, A. Sgarro, Open-frame Dempster conditioning for incom-

plete interval probabilities, Proc. First International Symposium on
Imprecise Probabilities and their Applications, Ghent, Belgium, June
1999.
 Solving Problems with Finite Test Sets1
Cristian S. Calude
Department of Computer Science, The University of Auckland
Private Bag 92019, Auckland, New Zealand
E-mail: cristian@cs.auckland.ac.nz

Helmut Jürgensen
Department of Computer Science, The University of Western Ontario
London, Ontario, Canada N6A 5B7, and
Institut für Informatik, Universiät Potsdam
Am Neuen Palais 10, D-14469, Potsdam, Germany
E-mail: helmut@uwo.ca.

Shane Legg
Department of Mathematics, The University of Auckland
Private Bag 92019, Auckland, New Zealand

Abstract. Every finite and every co-finite set of non-negative

integers is decidable. This is true and it is not, depending on
whether the set is given constructively. A similar constraint
is applicable in language theory and many other fields. The
constraint is usually understood and, hence, omitted.
The phenomenon of a set being finite, but possibly undecidable,
is, of course, a consequence of allowing non-constructive argu-
ments in proofs. In this note we discuss a few ramifications of
this fact. We start out with showing that every number the-
oretic statement that can be expressed in first-order logic can
be reduced to a finite set, to be called a test set. Thus, if one
knew the test set, one could determine the truth of the state-
ment. The crucial point is, of course, that we may not be able
1
The research reported in this paper was partially supported by Auckland University,
Research Grant A18/XXXXX/62090/3414050, and by the Natural Sciences and Engi-
neering Council of Canada, Grant OGP0000243.

39
40 C. S. Calude, H. Jürgensen, S. Legg

to know what the finite test set is. Using problems in the class
Π1 of the arithmetic hierarchy as an example, we establish that
the bound on the size of the test set is Turing-complete and
that it is upper-bounded by the busy-beaver function.
This re-enforces the fact that there is a vast difference between
finiteness and constructive finiteness. In the context of the
present re-opened discussion about the notion of computabil-
ity – possibly extending its realm through new computational
models derived from physics – the constraint of constructivity
of the model itself may add another twist.

1 Introduction
In the early days of decidability theory and also of theoretical computer
science it was not uncommon to find statements like every finite and every
co-finite set of non-negative integers is decidable in the research literature
and in textbooks, and to find “proofs” of this using the argument that
a decision algorithm could use table look-up; moreover, such statements
themselves would be used in proofs of the decidability of other problems
via reduction to finite or co-finite sets.2 Of course every finite or co-finite
set is decidable, but only – as is well-known – if it is given constructively.
Similar constraints are applicable in language theory and many other fields.
The constraint is, of course, usually understood and, hence, omitted. For
example, in the case of the D0L equivalence problem3 it was known for quite
some time that this problem could be reduced to the problem of deciding
whether two regular languages are equal. Unfortunately, this reduction was
not constructive and a constructive one eluded researchers for several years.
The phenomenon of a set being finite, but possibly undecidable, is, of
course, a consequence of allowing non-constructive arguments in proofs. In
this note we discuss a few ramifications of this fact. We start out with
showing that every number theoretic statement that can be expressed in
first-order logic can be reduced to a finite set, to be called a test set. Thus, if
one knew the test set, one could determine the truth of the statement. This
2
We refrain from giving references, because pointing to past mistakes is not the aim
of this paper. However, the interested reader is likely to find such statements by just
perusing a few older books.
3
Given a finitely generated free monoid X ∗ with set X of generators, elements u, v ∈
∗
X and endomorphisms g and h, the DOL equivalence problem is to decide whether the
sets {u, g(u), g(g(u)), . . .} and {v, h(v), h(h(v)), . . .} are equal.
Solving Problems with Finite Test Sets 41

rather simple result models what is sometimes referred to as experimental

mathematics: Simply stated, if the statement is true we don’t need to do
anything and if it is false we find the smallest counter-example by computer.
We then show how several classical problems fall into this category. The
crucial point is, of course, that we may not be able to know what the finite
test set is. Using problems in the class Π1 of the arithmetic hierarchy
as an example, we establish that the bound on the size of the test set is
Turing-complete and that it is upper-bounded by the busy-beaver function.
This re-enforces the fact that there is a vast difference between finite-
ness and constructive finiteness. In the context of the present re-opened
discussion about the notion of computability – possibly extending its realm
through new computational models derived from physics – the constraint
of constructivity of the model itself may add another twist.
Let N denote the set of positive integers, let N0 = N ∪ {0}, and, for
k ∈ N, consider a k-ary predicate P on N, that is, a mapping of Nk into
the set B = {0, 1} of truth values. Consider the formula

f = Q1 n1 Q2 n2 . . . Qk nk P (n1 , n2 , . . . , nk ),

where Q1 , Q2 , . . . , Qk ∈ {∀, ∃} are quantifier symbols. In analogy to the

arithmetic classes, we say that f is in the class Π̂s or Σ̂s if the quantifier
prefix of f starts with ∀ or ∃, respectively, and contains s − 1 alternations
of quantifier symbols. When P is computable, then f is in Πs or Σs ,
respectively.4 It is sufficient to consider only such formulæf in which no
two consecutive quantifier symbols are the same; in the sequel we make this
assumption without special mention. With f as above, one has s = k.
As usual in logic, we write P (n1 , . . . , nk ) instead of P (n1 , . . . , nk ) = 1
when n1 , . . . , nk are elements of N. Thus, ¬P (n1 , . . . , nk ) if and only if
P (n1 , . . . , nk ) = 0. Moreover, since we consider variable symbols only in
the domain N, if f is any formula in first-order logic, we write f is true
instead of f is true in N.
Let Γs be one of the classes Π̂s , Σ̂s , Πs , and Σs . We refer to the task of
proving or refuting a first-order logic formula as a problem and especially,
to problems expressed by formulæ in Γs as Γs -problems.
We say that a problem is being solved if the corresponding formula is
proved or disproved to be true, that is, if the truth value of the formula is
determined. A problem is said to be finitely solvable if it can be solved by
examining finitely many cases.5
4
See [32] for general background on arithmetic classes.
5
A rigorous definition of this notion is given in Section 3 below.
42 C. S. Calude, H. Jürgensen, S. Legg

For example, consider the predicate

!
1, if n is even or n = 1 or n is a prime,
P (n) =
0, otherwise,

that is, P (n) = 0 if and only if n is an odd number greater than 1 which is
not a prime. Then the problem expressed by the formula ∀n P (n) is finitely
solvable;6 indeed, it is sufficient to check all n up to and including 9.
In this paper, we mainly consider Π̂1 -problems and Π1 -problems. For
example, Goldbach’s conjecture is a Π1 -problem. It states that every even
n ∈ N is the sum of two primes.7 To express this in the terminology as
introduced, let PG : N → B be such that
!
1, if n is odd or n is the sum of two primes,
PG (n) =
0, otherwise.

Thus, fG = ∀n PG (n) is true if and only if Goldbach’s conjecture is true.

Similarly, Riemann’s hypothesis is a Π1 problem.8 Consider the complex
function
∞

1 (−1)n−1
ζ(s) = · ,
1 − 21−s n=1 ns
where s = σ + i t, σ, t ∈ R, σ > 0, and s = 1. Riemann conjectured that
all zeroes s0 = σ0 + i t0 of ζ satisfy σ0 = 12 and are simple [30].
By a result of [14], Riemann’s hypothesis can be expressed in terms of
the function δR : N → R defined by
" "
δR (k) = ηR (j),
n<k j≤n

where
!
p, if j = pr for some prime p and some r ∈ N,
ηR (j) =
1, otherwise.
6
This example is based on a folklore joke on induction proofs: To prove that all odd
natural numbers greater than 2 are primes one proceeds as follows: 3 is a prime; 5 is a
prime; 7 is a prime; 9 is a measuring error; 11 is prime; 13 is a prime; this is enough
evidence.
7
The conjecture was stated in 1742 by Goldbach in a letter to Euler [17]. According
to [22], in 1980 the Goldbach conjecture was known to be true for all n ≤ 108 ; in [35] of
December 1994, it is claimed that no counter-example exists up to 2 · 1010 . Hardy states
that the Goldbach problem is “probably as difficult as any of the unsolved problems in
mathematics” [19]. See also [26] and [34].
8
The problem is first proposed in [30]; see also [31].
Solving Problems with Finite Test Sets 43

Riemann’s hypothesis is equivalent with the assertion that

 2
 1 n2
− < 36n3 ,
k≤δR (n)
k 2

for all n ∈ N, see [14].9 Hence, let

#  2
n2
PR (n) = 1, if k≤δR (n)
1
k − 2 < 36n3 ,
0, otherwise.
Thus, fR = ∀n PR (n) is true if and only if the Riemann hypothesis is true.
Clearly, PR is decidable. Therefore, Riemann’s hypothesis is a Π1 -problem.
As in the case of the Goldbach conjecture, also for the Riemann hy-
pothesis huge computations have been performed to search for a counter-
example – or to increase the confidence [3], [4], [5], [6], [25].
Of course, not every mathematical statement is a Π1 -problem. For
instance, the conjecture stating the existence of infinitely many twin primes,
that is, consecutive odd primes such as 857 and 859, is not a Π1 -problem.
With
!
1, m > n and m and m + 2 are primes,
PT (n, m) =
0, otherwise,
this conjecture can be stated as

fT = ∀n∃m PT (n, m).

The formula fT is in the class Π2 . Bennett claims that most mathemati-

cal conjectures can be settled indirectly by proving stronger Π1 -problems,
see [2]. For the twin-prime conjecture such a stronger Π1 -problem is ob-
tained as follows. Consider the predicate
!
1, if there is m with 10n−1 ≤ m ≤ 10n , m and m + 2 primes,
PT
(n) =
0, otherwise.
Let fT
= ∀nPT
(n). Thus, fT
gives rise to a Π1 -problem and, if fT
is true,
then also fT is true.
In this paper we discuss the fact – surprising (only) at first thought
– that every Π̂s -problem and every Σ̂s -problem has a finite test set. Of
course, there cannot be a constructive proof of this statement. Moreover,
already for s = 1 the size of the test sets behaves as badly as the busy
beaver.
9
For another proof see [23], pp. 117–122.
44 C. S. Calude, H. Jürgensen, S. Legg

2 Notation and Basic Notions

In this section we briefly review some basic notions and introduce some
notation. Let X be a non-trivial alphabet, that is, a non-empty, finite set
with at least 2 elements. Then X ∗ is the set of all words over X. A (formal)
language over X is a subset of X ∗ .
We assume that the reader is familiar with the theory of computable
functions on integers or strings (see [32], [8]). If U is a universal Turing
machine which maps strings over X to non-negative integers, and π is a
program for U then U (π) denotes the result of applying U to π and an
empty input tape. In particular, we write U (π) = ∞ when U does not halt
on π.

3 Finite Solvability
For s ∈ N, let Γ̂s denote any of Π̂s and Σ̂s , and let Γs denote any of Πs
and Σs .

Definition 3.1 Let

f = Q1 n1 Q2 n2 . . . Qs ns P (n1 , n2 , . . . , ns ),

with s ∈ N, where Q1 , Q2 , . . . , Qs are alternating quantifier symbols.

1. A test set for f is set T ⊆ Ns such that f is true in Ns if and only

if it is true in T .

2. The problem of f is finitely solvable if there is a finite test set for f .

Theorem 3.1 Let s ∈ N. Every f ∈ Γ̂s is finitely solvable.

Proof. Let
f = Q1 n1 Q2 n2 . . . Qs ns P (n1 , n2 , . . . , ns ),
with s ∈ N, where Q1 , Q2 , . . . , Qs are alternating quantifier symbols.
We determine a sequence N1 , N2 , . . . , Ns of finite sets with Ni ⊆ Ni
such that the problem posed by f can be solved by checking all s-tuples
(n1 , n2 , . . . , ns ) ∈ Ns .
We define the sets Ni by induction on i. For this purpose, let

fi (m1 , . . . , mi−1 ) = Qi ni . . . Qs ns P (m1 , . . . , mi−1 , ni , . . . , ns ),

Solving Problems with Finite Test Sets 45

where m1 , . . . , mi−1 ∈ N. In particular, f1 ( ) = f and fs+1 (m1 , . . . , ms ) =

P (m1 , . . . , ms ).
For i = 1, if Q1 = ∀, let

ν1 = 1 if f = f1 ( ) is true,

and
ν1 = min{m1 | m1 ∈ N, ¬f2 (m1 )} otherwise;
if Q1 = ∃, let
ν1 = 1 if f = f1 ( ) is not true,
and
ν1 = min{m1 | m1 ∈ N, f2 (m1 )} otherwise.
Let N1 = {(m1 ) | m1 ∈ N, m1 ≤ ν1 }.
Now, suppose Ni−1 has been defined and i ≤ s. For each (m1 , . . . ,
mi−1 ) ∈ Ni−1 , define νi (m1 , . . . , mi−1 ) ∈ N0 as follows. If Qi = ∀, let

νi (m1 , . . . , mi−1 ) = 1 if fi (m1 , . . . , mi−1 ) is true,

and

νi (m1 , . . . , mi−1 ) = min{mi | mi ∈ N, ¬fi+1 (m1 , . . . , mi−1 , mi )} otherwise;

if Qi = ∃, let

νi (m1 , . . . , mi−1 ) = 1 if fi (m1 , . . . , mi−1 ) is not true,

and

νi (m1 , . . . , mi−1 ) = min{mi | mi ∈ N, fi+1 (m1 , . . . , mi−1 , mi )} otherwise.

Let

Ni = {(m1 , . . . , mi ) | (m1 , . . . , mi−1 ) ∈ Ni−1 , mi ∈ N,

mi ≤ ν(m1 , . . . , mi−1 )}.

We now prove,10 by induction on i, that each set Ti = Ni × Ns−i is a

test set for f . Then, in particular, Ns is a finite test set for f .
10
We decided to include this rather straight-forward proof as it was only by this proof
that we discovered some subtle traps in the construction of the test sets.
46 C. S. Calude, H. Jürgensen, S. Legg

Consider i = 1. Suppose first that Q1 = ∀. The set N1 is {(1)} and,

clearly, the set T1 is a test set11 for f . When f is false the set N1 consist
of all positive integers up to the first counter-example for the first variable
of P . Hence, again, T1 is a test set for f . On the other hand, suppose that
Q1 = ∃. Then N1 = {(1)} when f is false. Clearly T1 is a test set12 for f .
When f is true the set N1 consists of all positive integers up to the first
witness for the first variable of P . Again T1 is a test set for f .
Now consider i > 1 and assume that Ti−1 is a test set for f . First
suppose that Qi = ∀. Consider (m1 , . . . , mi−1 ) ∈ Ni−1 . If fi (m1 , . . . , mi−1 )
is true then νi (m1 , . . . , mi−1 ) = 1. As Ti−1 is a test set for f , to test
whether f is true on {(m1 , . . . , mi−1 )} × Ns−i+1 it suffices to test on
{(m1 , . . . , mi1 , 1)} × N s−i , and (m1 , . . . , mi−1 , 1) ∈ Ni . If fi (m1 , . . . , mi−1 )
is false, then Ni contains all the i-tuples (m1 , . . . , mi−1 , mi ) with mi rang-
ing from 1 to the smallest counter-example. Hence, as Ti−1 is a test set for
f so is Ti .
Now suppose that Qi = ∃. If fi (m1 , . . . , mi−1 ) is false then
νi (m1 , . . . , mi−1 ) = 1. As Ti−1 is a test set for f , to test whether f is true on
{(m1 , . . . , mi−1 )} × Ns−i+1 it suffices to test on {(m1 , . . . , mi1 , 1)} × N s−i ,
and (m1 , . . . , mi−1 , 1) ∈ Ni . If fi (m1 , . . . , mi−1 ) is true then Ni contains
all the i-tuples (m1 , . . . , mi−1 , mi ) with mi ranging from 1 to the smallest
witness. Hence, as Ti−1 is a test set for f so is Ti . ✷

The proof of Theorem 3.1 is non-constructive and this remains so even

when P is decidable. Thus, from this proof we do not learn anything about
the number of cases one needs to check in order to prove or disprove the
trueity of f . It is clear from the theories of arithmetic classes and degrees
of unsolvability that, in general, finite test sets cannot be constructed for
this type of problems even when the predicate is computable. We try to
shed some light, from a different perspective, on some of the reasons why
this cannot be done.
The proof of Theorem 3.1 highlights a typical pitfall in proofs in com-
putability theory when the reasoning of classical logic is used. The proof
and the statement proved are computationally meaningless as neither helps
with actually solving the Γ̂s -problem. The “construction” of the sets Ni in
the proof disguises the fact that none of these finite sets may be computable.
11
In fact, the empty set would be a test set for f . However, if one uses this idea, that
is sets ν1 to 0 rather than 1 – and similarly for νi in general – then the ‘construction’
seems to break down.
12
Again the empty set could have been used were it not for problems with the subse-
quent steps of the ‘construction’.
Solving Problems with Finite Test Sets 47

See, for example, the formula fG expressing Goldbach’s conjecture.

The statement of Theorem 3.1 has some similarity with the Test Set
Theorem in formal language theory. This theorem, originally known as
Ehrenfeucht’s conjecture, can be stated as follows: Let X and Y be alpha-
bets, and let L ⊆ X ∗ . There exists a finite subset F of L, a test set, such
that, for any two morphisms f, g from X ∗ to Y ∗ , f (u) = g(u) for all u ∈ L
whenever f (u) = g(u) for all u ∈ F . This was proved independently in [1]
and [18].13 In [7] and also [11] it is pointed out that the existence of the test
sets is not constructive.14 In the statement of the Test Set Theorem for lan-
guages, the order of the quantifiers, that is, ∀L∃F ∀f ∀g, is very important.
The modified order ∀L∀f ∀g∃F results in a far simpler statement, for which
a proof can be given using the same ideas as in the proof of Theorem 3.1.
In the sequel, for f ∈ Γ̂s , let N (f ) = Ns with Ns as in the proof of
Theorem 3.1. In particular, when s = 1, then N (f ) is the set {(n1 ) | n1 ∈
N, n1 ≤ ν1 }. For this case, we define ν(f ) = ν1 .

4 Π1 -Problems
In this section, we analyze the case of Π1 -problems in greater detail. Let X
be an arbitrary but fixed alphabet. We use X as the alphabet for programs
of universal Turing machines. We also fix a computable bijective function
( , ) : X ∗ × N0 → X ∗ . Consider f = ∀n P (n) where P is a computable
predicate on N. We assume that P is given as a program for an arbitrary,
but fixed universal Turing machine U . Thus P is given as a word πP ∈ X ∗
such that U ((πP , n)) = P (n) for all n ∈ N. One can, therefore, consider ν
as a partial function of X ∗ into N0 , that is, ν(πP ) = ν(f ) with f as above.
We first determine an upper bound on ν(f ) for f ∈ Π1 .
The busy beaver function σ : N → N ([29]; see also [15], [16], Chap-
ter 39) is defined as follows:

σ(n) = max{U (x) | x is a program of length n for U

and U (x) halts on x}.

Let P be a computable unary predicate on N, let f = ∀n P (n), hence

f ∈ Π1 . Consider a program pf for U such that

U (pf ) = min{n | ¬P (n)},

13
Explanations of the proofs are given in [27] and [33]. For further information see [13].
14
Under special assumptions on L like regularity, test sets can be effectively constructed
[20], [21]; see also [13].
48 C. S. Calude, H. Jürgensen, S. Legg

if f is not true, and such that U runs forever on pf if f is true. Such a

program always exists because the program, which tries P (1), P (2), . . .
and halts with the first n such that ¬P (n), has the required properties.
Let mf = |pf |. If f is not true, then U halts on pf with output ν(f ).
Hence ν(f ) ≤ σ(mf ). If f is true, then ν(f ) = 0. This proves the following
statement.

Proposition 4.1 For every f ∈ Π1 , ν(f ) ≤ σ(mf ).

By Theorem 4.1, to solve the problem of f one only needs to check

the truth value of P (n) for all n not exceeding σ(mf ). This could be very
useful if σ were computable. However, σ grows faster than any computable
function. Hence, the bound ν(f ) ≤ σ(mf ) does not help in the actual
solution of the problem of f . In fact, no computable bound exists! Here is
the argument. For any π ∈ X ∗ , define the predicate Pπ on N by
!
1, U (π) does not halt within n steps,
Pπ (n) =
0, otherwise.

Clearly, the predicate is computable. Let fπ = ∀n Pπ (n). Then fπ is

true if and only if U (π) does not halt.
Assume now that there is a program to compute an upper bound of ν(f )
for any f ∈ Π1 ; this program takes, as input, a program ρ computing the
predicate P ρ and computes as output an integer ν
(ρ) such that ν(f ρ ) ≤
ν
(ρ), where f ρ = ∀n P ρ (n). We show that this assumption implies the
existence of an algorithm deciding the halting problem for Turing machines.
Indeed, consider π ∈ X ∗ . To decide whether U (π) halts, first compute a
program pπ computing Pπ . Next compute ν
(pπ ). As fπ = f pπ , one has
ν(fπ ) ≤ ν
(pπ ). Hence, to determine whether fπ is true, it is sufficient
to determine whether Pπ (n) for all n ≤ ν
(pπ ). If so, then U (π) halts;
otherwise it doesn’t.

Theorem 4.1 The upper bound ν is Turing-complete.

Proof. We already showed that an oracle for ν or an upper bound on

ν allows one to decide the halting problem. The conversely follows from
Proposition 4.1. ✷

Corollary 4.1 There is no constructive proof showing that every f ∈ Π1

has a finite test set.
Solving Problems with Finite Test Sets 49

With appropriate modifications, a statement similar to Corollary 4.1

can be proved for Σ1 . In fact, for any s ∈ N and any Γs , there is no
constructive proof of the fact that every f ∈ Γs has a finite test set.

5 Conclusions

Many true Π1 -problems are undecidable, hence independent with respect to

a given sufficiently rich, sound, and computably axiomatizable theory. The
analysis above can help us in understanding this phenomenon. Knowing
that P is false can be used to get a proof that “P is false”: we keep
computing P (n), for large enough n until we get an n such that ¬P (n).
But this situation is not symmetric: if we know that P is true we might not
be able to prove that “P is true”, and this case is quite frequent [10]. Indeed,
even when we “have” the proof, that is, we have successfully checked that
P (n) = 0, for all n ≤ ν((∀n)P (n)), we might not be able to “realise” that
we have achieved the necessary bound.
The correspondence P → ν((∀n)P (n)) exists and is perfectly legitimate
from a classical point of view, but has no constructive “meaning”. To a
large extent the mathematical activity can be regarded as a gigantic, col-
lective effort to compute individual instances of the function ν((∀n)P (n)).
This point of view is consistent with Post’s description of mathematical
creativity [28]: “Every symbolic logic is incomplete and extendible relative
to the class of propositions constituting K0 . The conclusion is inescapable
that even for such a fixed, well defined body of mathematical propositions,
mathematical thinking is, and must remain, essentially creative.”15 It also
gives support to the “quasi-empirical” view of mathematics, which sustains
that although mathematics and physics are different, it is more a matter
of degree than black and white [12, 9]; see also [24].
In essence, the seemingly paradoxical situation arises from the fact that,
in classical logic, it may happen that only finite resources are needed for
defining a finite object while finite resources will not suffice to determine
the same object constructively. The finite “character” of a problem may
nevertheless rule out – in a very fundamental way – that its solution can
be obtained by finite means.

15
As usual, K0 means the halting problem in this quote.
50 C. S. Calude, H. Jürgensen, S. Legg

Acknowledgment
We thank Douglas Bridges, Greg Chaitin and Solomon Marcus for stimu-
lating discussions.

References
[1] M. H. Albert, J. Lawrence, A proof of Ehrenfeucht’s conjecture, The-
oret. Comput. Sci., 41 (1985), 121–123.

[2] C. H. Bennett, Chaitin’s Omega, in Fractal Music, Hypercards, and

More . . . (M. Gardner, ed.), W. H. Freeman, New York, 1992, 307–319.

[3] R. P. Brent, J. v. d. Lune, H. J. J. t. Riele, D. T. Winter, On the zeros

of the Riemann zeta function in the critical strip I, Math. Comp., 33
(1979), 1361–1372.

[4] R. P. Brent, J. v. d. Lune, H. J. J. t. Riele, D. T. Winter, On the zeros

of the Riemann zeta function in the critical strip II, Math. Comp., 39
(1982), 681–688.

[5] R. P. Brent, J. v. d. Lune, H. J. J. t. Riele, D. T. Winter, On the zeros

of the Riemann zeta function in the critical strip III, Math. Comp.,
41 (1983), 759–767.

[6] R. P. Brent, J. v. d. Lune, H. J. J. t. Riele, D. T. Winter, On the zeros

of the Riemann zeta function in the critical strip IV, Math. Comp., 46
(1986), 667–681.

[7] C. Calude, Note on Ehrenfeucht’s conjecture and Hilbert’s basis theo-

rem, Bull. EATCS, 29 (1986), 18–22.

[8] C. Calude, Theories of Computational Complexities, North-Holland,

Amsterdam, 1988.

[9] C. S. Calude, G. J. Chaitin, Randomness everywhere. Nature, 400, 22

July (1999), 319–320.

[10] C. Calude, H. Jürgensen, M. Zimand, Is independence an exception?

Applied Mathematics and Computation, 66 (1994), 63–76.

[11] C. Calude, D. Vaida, Ehrenfeucht test set theorem and Hilbert ba-
sis theorem: A constructive glimpse, in Mathematical Foundations of
Solving Problems with Finite Test Sets 51

Computer Science, 1989 (A. Kreczmar, G. Mirkowska, eds.), Lecture

Notes in Computer Science 379, Springer-Verlag, Berlin, 1989, 177–
184.

[12] G. J. Chaitin, The Unknowable, Springer-Verlag, Singapore, 1999.

[13] C. Choffrut, J. Karhumäki, Combinatorics on words, in Handbook of

Formal Language Theory (G. Rozenberg, A. Salomaa, eds.), Vol. 1,
Springer-Verlag, Berlin, 1987, 329–438.

[14] M. Davis, Y. V. Matijasevič, J. Robinson, Hilbert’s tenth prob-

lem. Diophantine equations: Positive aspects of a negative solution,
in Mathematical Developments Arising from Hilbert Problems (F. E.
Browder, ed.), American Mathematical Society, Providence, RI, 1976,
323–378.

[15] A. K. Dewdney, A computer trap for the busy beaver, the hardest-
working Turing machine, Scientific American, 251(8) (1984), 19–23.

[16] A. K. Dewdney, The New Turing Omnibus, Computer Science Press,

New York, 1993.

[17] L. E. Dickson, History of the Theory of Numbers, Carnegie Institute,

Washington, 1919, 1920, 1923, 3 volumes.

[18] V. S. Guba, The equivalence of infinite systems of equations in free

groups and semigroups, Mat. Zametki, 40 (1986), 321–324 (in Rus-
sian).

[19] G. H. Hardy, Goldbach’s theorem, Mat. Tid. B, 1 (1922), 1–16.

Reprinted in Collected Papers of G. H. Hardy, vol. 1, Oxford Uni-
versity Press, Oxford, 1966, 545–560.

[20] J. Karhumäki, W. Rytter, S. Jarominek, Efficient constructions of test

sets for regular and context-free languages, Theoret. Comput. Sci., 116
(1993), 305–316.

[21] J. Karhumäki, W. Plandowski, W. Rytter, Polynomial size test sets

for context-free languages, J. Comput. System Sci., 50 (1995), 11–19.

[22] W. A. Light, T. J. Forres, N. Hammond, S. Roe, A note on the Gold-

bach conjecture, BIT, 20 (1980), 525.
52 C. S. Calude, H. Jürgensen, S. Legg

[23] Y. V. Matijasevič, Hilbert’s Tenth Problem, MIT Press, Cambridge,

MA, 1993, 117–122.

[24] S. Marcus, Bridging linguistics and computer science, via mathematics,

in People and Ideas in Theoretical Computer Science (C. S. Calude,
ed.), Springer-Verlag, Singapore, 1998, 163–176.

[25] A. M. Odlyzko, Tables of zeros of the Riemann zeta function,

http://www. research.att.com/~amo/zeta tables/index.html.

[26] C.-T. Pan, Goldbach Conjecture, Science Press, Beijing, 1992.

[27] D. Perrin, On the solution of Ehrenfeucht’s conjecture, Bull. EATCS,

27 (1985), 68–70.

[28] E. L. Post, Recursively enumerable sets of positive integers and their

decision problems, Bull. (New Series) Amer. Math. Soc., 50 (1944),
284–316.

[29] T. Rado, On non-computable numbers, Bell System Tech. J., 3 (1962),

977–884.

[30] B. Riemann, Über die Anzahl der Primzahlen unter einer gegebenen
Größe, in Gesammelte mathematische Werke und, wissenchaftlicher
Nachlaß, Springer-Verlag, Berlin, 1990, 177–185.

[31] H. Riesel, Prime Numbers and Computer Methods for Factorization,

Birkhäuser, Boston, second ed., 1994.

[32] H. Rogers, Theory of Recursive Functions and Effective Computabil-

ity, McGraw-Hill, New York, 1967.

[33] A. Salomaa, The Ehrenfeucht conjecture: A proof for language theo-

rists, Bull. EATCS, 27 (1985), 71–82.

[34] W. Yuan, ed., Goldbach Conjecture, World Scientific, Singapore, 1984.

[35] Names of large numbers & unsolved problems, http://www.

smartpages.com/faqs/sci-math-faq/unsolvedproblems/faq.html,
December 1994.
 State Complexity of Regular Languages:
Finite versus Infinite1
Cezar Câmpeanu
Faculty of Mathematics
University of Bucharest
Str. Academiei 14, 70109 Bucureşti, Romania
E-mail: cezar@funinf.math.unibuc.ro

Kai Salomaa, Sheng Yu

Department of Computer Science
University of Western Ontario
London, Ontario, Canada N6A 5B7
E-mail: {ksalomaa, syu}@csd.uwo.ca

Abstract. We consider the state complexity of regular lan-

guages and their operations. Especially, we compare the state
complexity results on finite languages and general regular lan-
guages. The similarity relation ∼L and the equivalence relation
≡L over Σ∗ are also compared. Their applications on minimiza-
tion of deterministic finite cover automata and deterministic
finite automata, respectively, are investigated.

1 Introduction
There are many ways to measure the complexity of a deterministic finite
automaton (DFA): (1) the number of states, (2) the number of transitions,
or (3) both the number of states and the number of transitions. Although
(3) gives more complete information, (1) is both simpler in presentation
and cleaner and purer for investigation. The number of states also gives a
1
This research is supported by the Natural Sciences and Engineering Research Council
of Canada grants OGP0041630 and OGP0147224.

53
54 C. Câmpeanu, K. Salomaa, S. Yu

linear bound on the number of transitions. In the case of a complete DFA,

i.e., a DFA whose transition function is defined for each state and each
letter in the alphabet, the number of transitions is totally determined by
the number of states when the alphabet is given. Therefore, the number of
states is a natural complexity measure for a DFA as well as for the language
it accepts. In the following, by a DFA we always mean a complete DFA.
A regular language is accepted by infinitely many different DFA. How-
ever, there is one that has the minimal number of states and it is unique
up to a renaming of the states. We use the number of states of the min-
imal automaton to measure the complexity of the given language. Thus,
by the state complexity of a regular language L, denoted C(L), we mean
the number of states in the minimal DFA that accepts L. By the state
complexity of a class L of regular languages, denoted C(L), we mean the
maximum among all C(L), L ∈ L. When we speak about the state com-
plexity of an operation on regular languages, we mean the state complexity
of the resulting languages from the operation. For example, we say that
the state complexity of the intersection of an m-state DFA language, i.e.,
a language accepted by an m-state complete DFA, and an n-state DFA
language is exactly mn. This means that mn is the state complexity of
the class of languages each of which is the intersection of an m-state DFA
language and an n-state DFA language. In other words, there exist two reg-
ular languages that are accepted by an m-state DFA and an n-state DFA,
respectively, such that the intersection of them is accepted by a minimal
DFA of mn states, and this is the worst case. So, in a certain sense, state
complexity is a worst-case complexity.
The state complexity of a regular-language operation gives a lower-
bound for the space as well as the time complexity of the same operation. In
many cases, the bounds given are tight. For example, the state complexity
of the union of an m-state DFA language and an n-state DFA language
is exactly mn. This also gives both the space and time complexity of the
union operation within a constant factor.
State complexity is a complexity measure only for regular languages.
However, it can be extended to cover other families of languages as well. For
example, the automaticity studied by Shallit et al. [17] can be considered
as an extension of the state complexity. We will not consider any extension
of state complexity in this article.
Examining the state complexity results on the basic operations (e.g.,
catenation, union, intersection, and complementation) on regular languages
in [19], one would notice that all the worst cases are given by using infinite
State Complexity of Regular Languages 55

languages only. This observation raises the question: Are finite languages
significantly different from (infinite) regular languages in state complexity
of their operations? For example, would the state complexity of the union
of two finite languages accepted by an m-state and an n-state DFA, respec-
tively, be O(m + n) instead of mn? We will investigate these questions in
this article.
Finite languages are, perhaps, one of the most often used but least
studied classes of languages. Finite languages are exactly the languages
accepted by acyclic finite automata. It has been shown that there is a
linear (time) algorithm for the minimization of an acyclic DFA by Revuz in
1992 [13]. However, for the minimization of a general DFA, the best known
algorithm has a time complexity O(n log n) by Hopcroft in 1971 [6].
In this article, we compare the state complexity results for finite and
infinite regular languages. We first consider the relatively simple cases,
i.e., the operations on languages over a one-letter alphabet. Then we con-
sider the general cases. In the one-letter cases, most of the operations on
finite languages have a much lower state complexity than the correspond-
ing operations on regular languages. However, in the general cases, only
the catenation of two finite languages, when the first language is accepted
by a DFA with a constant number of final states, has a much lower state
complexity than its regular language counterpart.
Due to the not-so-positive results in the general cases, we resort to a
different concept to try to reduce the number of states for DFA accepting
finite languages. The concept of cover automata for finite languages is
described in the last section of this article. In many cases, cover automata
are a much more concise representation than DFA for finite languages.

2 Preliminaries
A deterministic finite automaton (DFA) is denoted by a quintuple
(Q, Σ, δ, q0 , F ), where Q is the finite set of states, Σ is the finite alpha-
bet, δ : Q × Σ → Q is the transition function, q0 ∈ Q is the start state, and
F ⊆ Q is the set of final states. In this paper, all the DFAs are assumed to
be complete DFAs. By a complete DFA we mean that there is a transition
defined for each letter of the alphabet from each state, i.e., δ is a total
function. In contrast, a DFA is called an incomplete DFA if its transition
function is a partial function.
For any x ∈ Σ∗ , we use #(x) to denote the length of x and #a (x) for
some a ∈ Σ to denote the number of appearances of a in x. The empty
56 C. Câmpeanu, K. Salomaa, S. Yu

word is denoted by ε.
The transition function δ of a DFA is extended to δ̂ : Q × Σ∗ → Q by
setting δ̂(q, ε) = q and δ̂(q, ax) = δ̂(δ(q, a), x) for q ∈ Q, a ∈ Σ, and x ∈ Σ∗ .
In the following, we simply use δ to denote δ̂ if there is no confusion.
A word w ∈ Σ∗ is accepted by a DFA A = (Q, Σ, δ, q0 , F ) if δ(q0 , w) ∈ F .
The language accepted by A, denoted L(A), is the set {w ∈ Σ∗ | δ(q0 , w) ∈
F }. Two DFA are said to be equivalent if they accept the same language.
An incomplete DFA can be transformed to an equivalent complete DFA
by adding a ‘sink state’ and transitions, which were undefined before, to the
‘sink state’, as well as transitions from the ‘sink state’ to the ‘sink state’.
Let A = (Q, Σ, δ, s, F ) be a DFA. Then
a) a state q is said to be accessible if there exists w ∈ Σ∗ such that
δ(s, w) = q;
b) a state q is said to be useful if there exists w ∈ Σ∗ such that δ(q, w) ∈
F.
It is clear that for every DFA A there exists an equivalent DFA A
such
that every state of A
is accessible and at most one state is useless (the
‘sink state’). A DFA A
as above is called a reduced DFA. We will use only
reduced DFA in the following.
A nondeterministic finite automaton (NFA) is denoted also by a quin-
tuple (Q, Σ, η, q0 , F ) where η ⊆ Q × (Σ ∪ {ε}) × Q is a transition relation
rather than a function, and Q, Σ, q0 , and F are defined similarly as in a
DFA. The words and languages accepted by NFA are defined similarly as
for DFA.
For a set s, we use |s| to denote the cardinality of s. For a language L,

l
we define L≤l = Li .
i=0
For L ⊆ Σ∗ , we define a relation ≡L ⊆ Σ∗ × Σ∗ by

x ≡L y iff xz ∈ L ⇔ yz ∈ L for all z ∈ Σ∗ .

Clearly, ≡L is an equivalence relation, which partitions Σ∗ into equivalence

classes. The number of equivalence classes of ≡L is called the index of ≡L .
The Myhill-Nerode Theorem [7] states that L is regular if and only if ≡L
has a finite index and the minimal number of states of a complete DFA
that accepts L is equal to the index of ≡L .
For a rather complete background knowledge in automata theory, the
reader may refer to [7], [15].
The following lemmas will be used in the subsequent sections. They
can be proved rather easily. Thus, we omit the proofs to concentrate on
State Complexity of Regular Languages 57

our main results.

Lemma 2.1 Let R ⊆ Σ∗ be a regular language. If there exists an integer

n such that
max{#(w) | w ∈ Σ∗ & w ∈ R} = n,

then any DFA accepting R needs at least n + 2 states. In particular, if Σ

is a singleton, the minimal DFA accepting R uses exactly n + 2 states.

Lemma 2.2 Let m, n > 0 be two arbitrary integers such that (m, n) = 1
(m and n are relatively prime).

(i) The largest integer that cannot be presented as cm + dn for any inte-
gers c, d > 0 is mn.

(ii) The largest integer that cannot be presented as cm + dn for any inte-
gers c > 0 and d ≥ 0 is (m − 1)n.

(iii) The largest integer that cannot be presented as cm + dn for any inte-
gers c, d ≥ 0 is mn − (m + n).

3 Finite Versus Regular Languages Over a One-

Letter Alphabet
As we have mentioned in the introduction, we start our comparison of the
state complexity of operations on regular and finite languages from the
relatively easy cases, i.e., the languages over a one-letter alphabet.
We first list the basic results below and then give detailed explanations
for some of the operations.
We assume that L1 is an m-state DFA language and L2 an n-state DFA
language, Σ = {a}, and m, n > 1.

Finite Regular
L1 ∪ L2 max(m, n) mn, for (m, n) = 1
L1 ∩ L2 min(m, n) mn, for (m, n) = 1
Σ∗ − L1 m m
L1 L2 m+n−1 mn, for (m, n) = 1
LR
1 m m
L∗1 m2 − 7m + 13, for m > 4 (m − 1)2 − 1
58 C. Câmpeanu, K. Salomaa, S. Yu

Note that for finite languages, the state complexity for each of the union,
intersection, and catenation operations is linear, while it is quadratic for
infinite regular languages.
In the above table, all results for finite languages are relatively trivial
except for L∗1 . We give an informal proof in the following. Let L1 be
accepted by an m-state DFA A1 and A is a minimal DFA accepting L∗1 .
It is clear that the length of the longest word accepted by A1 is m − 2.
(Note that the m states include a ‘sink state’.) We consider the following
three cases: (1) A1 has one final state; (2) A1 has two final states; or (3)
A1 has three or more final states. If (1), then A has m − 1 states. For (2),
the worst case is given by L = {am−2 , am−3 }. By (iii) of Lemma 2.2, the
length of the longest word that is not in L∗1 is

(m − 2)(m − 3) − (2m − 5) = m2 − 7m + 11.

Then A has exactly m2 − 7m + 13 states. In case (3), it is easy to see that

A cannot have more than m2 − 7m + 13 states.
For regular languages, we give a more detailed discussion below.
For the union operation, it is clear that mn states are sufficient for the
resulting minimal DFA. To show that mn states are necessary, it suffices to
show that there are at least mn distinct equivalence classes of the relation
≡L1 ∪L2 . Let L1 = (am )∗ and L2 = (an )∗ , m, n > 1 and (m, n) = 1. For
positive integers p and q, let mp = p mod m, np = p mod n, mq = q mod m,
and nq = q mod n, 0 ≤ mp , mq < m, 0 ≤ np , nq < n. It turns out that if
mp = mq or np = nq , then ap and aq are not equivalent. However, this is
not immediately clear for some cases. For example, let mp = −2 mod m,
np = −1 mod n, mq = −1 mod m, and nq = −2 mod n. (In order to
explain easily, we use the negative numbers.) Then neither mp = mq nor
np = nq , but mp = nq and np = mq . So, both ap a2 , aq a2 ∈ L1 ∪ L2 and
ap a, aq a ∈ L1 ∪ L2 . It then appears that ap ≡L1 ∪L2 aq . However, this is not
true because it can be proved that ap a2+m ∈ L1 ∪ L2 , but aq a2+m ∈ L1 ∪ L2
assuming that m < n.
The state complexity result for the intersection of two regular languages
can be similarly proved.
The result for the catenation of two regular languages is more involved.
We outline a proof in the following. A more detailed proof can be found in
[19].
We first give a general example of an m-state DFA language and an n-
state DFA language, (m, n) = 1, such that mn states are necessary for any
State Complexity of Regular Languages 59

DFA that accepts the catenation of the two languages. Let L1 = am−1 (am )∗
and L2 = an−1 (an )∗ . Obviously, L1 and L2 can be accepted by an m-
state DFA and an n-state DFA, respectively, and L = L1 L2 = {ai | i =
(m−1)+(n−1)+cm+dn for some integers c, d ≥ 0}. By (iii) of Lemma 2.2,
for (m, n) = 1, the largest number that cannot be represented by cm + dn,
c, d ≥ 0, is mn − (m + n). Then the largest i such that ai ∈ L is mn − 2.
So, the minimal DFA that accepts L has at least mn states. We show that
mn states are sufficient in the following theorem.

Theorem 3.1 For any integers m, n ≥ 1, let A and B be an m-state DFA

and an n-state DFA, respectively, over a one-letter alphabet. Then there
exists a DFA of at most mn states that accepts L(A)L(B).

Proof. The cases when m = 1 or n = 1 are trivial. We assume

that m, n ≥ 2 in the following. Let A = (QA , {a}, δA , sA , FA ) and
B = (QB , {a}, δB , sB , FB ). By a variation of the subset construction, we
know that L(A)L(B) is accepted by the DFA C = (QC , {a}, δC , sC , FC )
where
QC = {< q, P > | q ∈ QA & P ⊆ QB };
sC =< sA , ∅ > if sA ∈ FA and sC =< sA , {sB } > if sA ∈ FA ;
δC (< q, P >, a) =< q
, P
> where q
= δA (q, a) and P
= δB (P, a) ∪
{sB } if q
∈ FA , P
= δB (P, a) otherwise;
and FC = {< q, P > | P ∩ FB = ∅}.
Now we show that at most mn states of QC are reachable from sC .
First we assume that in A there is a final state f in the loop of A’s
transition diagram. Then δA (sA , at ) = f and δA (f, al ) = f for some non-
negative integers t < m and l ≤ m. Let j1 , . . . , jr , 0 < j1 < . . . < jr < l,
be all the integers such that δA (f, aji ) ∈ FA for each 1 ≤ i ≤ r. Denote
P0 = {sB },
P1 = {δB (sB , al ), δB (sB , al−j1 ), . . . , δB (sB , al−jr )},
and for i ≥ 2 we define
Pi = δB (Pi−1 , al ).
Let δC (sC , at ) =< f, S >. Denote S0 = S − {sB } and Si = δB (Si−1 , al ) for
each i ≥ 1. Then we have the following state transition sequence of C:

sC -tC < f, P0 ∪ S0 >

-lC < f, P0 ∪ P1 ∪ S1 >
............
-lC < f, P0 ∪ P1 ∪ . . . ∪ Pn−1 ∪ Sn−1 >
-lC < f, P0 ∪ P1 ∪ . . . ∪ Pn ∪ Sn >
60 C. Câmpeanu, K. Salomaa, S. Yu

Here p -kC q stands for δC (p, ak ) = q. Denote P0 ∪ . . . ∪ Pi by Pi , i ≥ 0. Let

i be the smallest integer such that Pi−1 = Pi . It is clear that i ≤ n since
B has n states. If i = n, then Pn−1 = QB and

< f, Pn−1 ∪ Sn−1 >=< f, Pn ∪ Sn >=< f, QB > .

Therefore, C needs at most m + l(n − 1) ≤ m + m(n − 1) = mn states. If

i < n, consider the set Si−1 = Si−1 − Pi−1 . Note that every state in Si−1

is in the loop of the transition diagram of B. If for each element r of Si−1 ,

there exists j, 0 ≤ j ≤ n − i, such that δB (r, ajl ) ∈ Pi−1 (i.e., Pn−1 ), then
the proof is concluded as above. Otherwise, there is an element r0 of Si−1

and a transition sequence

r0 -lB r1 -lB . . . -lB rn−i

such that, for some j, k ≤ n − i and j < k, rj = rk . (There are at most

n − i states not in Pi−1 .) Then it is easy to verify that Si−1+j = Si−1+k .
Therefore, < f, Pi−1+j ∪ Si−1+j >=< f, Pi−1+k ∪ Si−1+k >. Thus, the
number of states that are reachable from sC is at most t + 1 + l(n − 1) ≤
(m − 1) + 1 + m(n − 1) = mn.
Finally we consider the case when no final states of A are in the loop.
Let QA = {0, . . . , m − 1} where sA = 0 and δA (0, ai ) = i for 0 ≤ i ≤ m − 1.
We can assume that m−2 is a final state and m−1 loops to itself. Otherwise,
L(A) can be accepted by a complete DFA with less than m states. Consider
the following m + n − 1 transition steps of C

sC -m−2
C < m − 2, T > -C < m − 1, T0 > -C < m − 1, T1 >

-C . . . -C < m − 1, Tn > .
Let the state δB (sB , ai+1 ) be ti , for each i ≥ 0. Note that sB ∈ T and
ti is in Ti . It is clear that there exist j, k such that 0 ≤ j < k ≤ n and
tj = tk . Then it is not difficult to see that < m − 1, Tj >=< m − 1, Tk >.
Therefore, at most m + n states are necessary for C. (We have m + n < mn
for m, n ≥ 2.) ✷

For the union, intersection, and catenation operations, we have consid-

ered only the cases when (m, n) = 1. For (m, n) = t > 1, we have not
obtained exact formulas for those cases. Note that neither mn/(m, n) nor
lcm(m, n) (the least common multiple of m and n) is the solution. For
example, a(a5 )∗ and (a9 )∗ are accepted by a 6-state and a 9-state DFA,
respectively, but the union of them needs at least 45 states rather than
6 × 9/3 = 18 states.
State Complexity of Regular Languages 61

We now consider the last operation in the table, i.e., the star operation
on infinite regular languages. We find that the proofs for both directions
are interesting.

Theorem 3.2 The number of states which is sufficient and necessary in

the worst case for a DFA to accept the star of an n-state DFA language,
n > 1, over a one-letter alphabet is (n − 1)2 + 1.

Proof. For n = 2, the necessity is shown by a 2-state DFA which accepts

(aa)∗ . For each n > 2, the necessary condition can be shown by the DFA
A = ({0, . . . , n − 1}, {a}, δ, 0, {n − 1}) where δ(i, a) = i + 1 mod n for each
i, 0 ≤ i ≤ n − 1. The star of L(A) is the language {ai | i = c(n − 1) +
dn, for some integers c > 0 and d ≥ 0, or i = 0}. By (ii) of Lemma 2.2,
the largest i such that ai ∈ L(A)∗ is (n − 2)n. So, the minimal DFA that
accepts (L(A))∗ has (n − 2)n + 2, i.e., (n − 1)2 + 1, states.
The proof for showing that (n − 1)2 + 1 states are sufficient is more
interesting. Let A = (Q, {a}, δ, s, F ) be an arbitrary n-state DFA, n > 1,
and R = L(A). If s is the only final state of A, then R∗ = R. So, we
assume that there is at least one final state f such that f = s. Clearly,
R∗ (excluding ε if s ∈ F ) is accepted by the NFA A
= (Q, {a}, δ
, s, F )
where δ
= δ ∪ {(q, ε, s) | q ∈ F }. For any X ⊆ Q, denote by closure(X) the
set X ∪ {q ∈ Q | (p, ε, q) ∈ δ
for some p ∈ X}. Now we follow the subset
construction approach to build a DFA B = (P, {a}, η, {s}, FP ) from A
to
accept R∗ such that P ⊆ 2Q , η(X, a) = closure({q ∈ Q | there exists p ∈
X such that (p, a, q) ∈ δ
}), and FP = {X ∈ P | X ∩ F = ∅ or X = {s}}.
Let f be the first final state from s in A and at is the shortest word such that
δ(s, at ) = f . Then η({s}, at ) = {s, f }. Denote by pki the state η({s}, ait )
in P , i ≥ 0, which is a subset of Q.
We claim that pki ⊇ pki−1 for all i ≥ 1. It is true for i = 1, because
η({s}, at ) = {s, f }, and also true for i > 1 since

pki = η({s}, ait ) = η({s, f }, a(i−1)t ) = η({s}, a(i−1)t ) ∪ η({f }, a(i−1)t )

= pki−1 ∪ η({f }, a(i−1)t ).

Then one of the following must be true:

(1) pki = pki−1 for some i ≤ n − 1;
(2) pkn−1 = Q.
This is because if (1) is false, then pkn−1 contains at least n states and,
therefore, (2) is true. Note that if (2) is true, then η(pkn−1 , a) = pkn−1 . In
62 C. Câmpeanu, K. Salomaa, S. Yu

any of the cases, the number of states of B is no more than t(n − 1) + 1

which is at most (n − 1)2 + 1. ✷

For the transformation from an n-state NFA to a DFA, it is clear in

the case of finite languages over a one-letter alphabet that at most n states
are needed. However, in the case of an infinite regular language over a
one-letter alphabet, the problem is still open.

4 Finite Versus Regular Languages Over an Ar-

bitrary Alphabet
For the one-letter alphabet case which we have discussed in the previous
section, the state complexities for most operations on finite languages are
of a lower order than their counterpart for regular languages. However,
this is no longer true in the case when the size of the alphabet is arbitrary.
Although none of the operations on finite languages, except the comple-
mentation, can reach the exact bound for regular languages, most of them
have a complexity that is of the same order as the corresponding operation
on regular languages.
We list the state complexity of the basic operations for both finite and
regular languages over an arbitrary alphabet below. All the results for
regular languages are given as exact numbers. However, we use the big
“O” notation for most of the results for finite languages due to the fact
that either the formulas we have obtained are acutely nonintuitive or we
do not have an exact result, yet. More detailed explanations follow the
table.
We assume that L1 and L2 are accepted by an m-state DFA A1 =
(Q1 , Σ, δ1 , s1 , F1 ) and an n-state DFA A2 = (Q2 , Σ, δ2 , s2 , F2 ), respectively,
and m, n > 1. We use t to denote the number of final states in A1 .

Finite Regular
L1 ∪ L2 O(mn) mn
L1 ∩ L2 O(mn) mn
Σ∗ − L1 m m
L1 L2 O(mnt−1 + nt ) (2m − 1)2n−1
LR1 O(2m/2 ), for |Σ| = 2 2m
L∗1 2m−3 + 2m−4 , for m ≥ 4 2m−1 + 2m−2

For the union and the intersection of finite languages, it was expected
that their state complexities would be linear, more specifically O(m + n),
State Complexity of Regular Languages 63

but it turns out that both of them are of the order of mn although neither
of them can reach the exact bound mn.
It is easy to show that mn states are sufficient for both union and
intersection by the following simple argument. We can construct a DFA
A = (Q, Σ, δ, s, F ) which is the cross-product of A1 and A2 , i.e., Q =
Q1 × Q2 , δ = δ1 × δ2 that is δ((q1 , q2 ), a) = (δ(q1 , a), δ2 (q2 , a)), s = (s1 , s2 ).
For the union operation, F = {(q1 , q2 ) | q1 ∈ F1 or q2 ∈ F2 } and for the
intersection operation, F = {(q1 , q2 ) | q1 ∈ F1 and q2 ∈ F2 }.
Note that the pairs of the form (s1 , q2 ) where q2 = s2 and (q1 , s2 ) where
q1 = s1 are never reached from (s1 , s2 ), and therefore, are useless. So, mn−
(m + n − 2) states are sufficient for both the union and intersection of two
finite languages accepted by an m-state and an n-state DFA, respectively.
However, this is a very rough upper bound. Much tighter upper bounds
for the union and intersection of finite languages are given in [2], which
unfortunately are in a very complicated and highly incomprehensible form.
Thus, we will not quote them in this paper.
It is more interesting to show that the state complexities of those two
operations are indeed of the order of mn but not lower. The following
examples were originally given by Shallit [16]. Automaton-based examples
are given in [2], which give better lower bounds than the examples be-
low. We choose to present the following examples due to their clarity and
intuitiveness.
For the intersection of two finite languages, consider the following ex-
ample. Let Σ = {a, b} and

L1 = {w ∈ Σ∗ | #a (w) + #b (w) = 2n},

L2 = {w ∈ Σ∗ | #a (w) + 2#b (w) = 3n}.

Clearly, L1 is accepted by a DFA with 2n + 2 states and L2 by a DFA with

3n + 2 states. The intersection L = L1 ∩ L2 is

{w ∈ Σ∗ | #a (w) = #b (w) = n}

One can prove that any DFA accepting L needs at least n2 states by the
Myhill-Nerode Theorem [7].
For the union of two finite languages, the example is slightly more com-
plicated. Let Σ = {a, b} and

L1 = {w ∈ Σ∗ | #(w) ≤ 3t and #a (w) + #b (w) = 2t},

L2 = {w ∈ Σ∗ | #a (w) + 2#b (w) < 3t}.
64 C. Câmpeanu, K. Salomaa, S. Yu

It is clear that L1 ∪ L2 includes all words in Σ∗ of length less than or equal

to 3t except those words w such that #(w) = 2t and #b (w) ≥ t. One
can prove that any DFA accepting L1 ∪ L2 needs more than t2 states by
checking the number of the equivalence classes of ≡L1 ∪L2 .
We now consider the catenation operation. Notice that for the finite
language case, if the number of final states in A1 is a constant, then the
state complexity of catenation is a polynomial in terms of m and n. In
particular, if A1 has only one final state, then the state complexity is linear,
i.e., m + n. In contrast, for infinite regular languages, there are examples in
which A1 has only one final state but any DFA accepting the catenation of
the two languages needs at least (2m − 1)2n−1 states [19], [18]. This is one
of a few cases in which the state complexities for finite and infinite regular
languages, respectively, are in different orders.
We now give the proof for the finite language case. For the general case
for the catenation of regular languages, the reader may refer to [19] or [18].
Without loss of generality, we assume that all the DFA we are con-
sidering are reduced and ordered. A DFA A = (Q, Σ, δ, 0, F ) with Q =
{0, 1, . . . , n} is called an ordered DFA if, for any p, q ∈ Q, the condition
δ(p, a) = q implies that p ≤ q.
For convenience, we introduce the following notation:
   
n 
i
n
= .
≤i j=0
j

Theorem 4.1 Let Ai = (Qi , Σ, δi , 0, Fi ), i = 1, 2, be two DFA accepting

finite languages Li , i = 1, 2, respectively, and #Q1 = m, #Q2 = n, #Σ =
k, and #F1 = t. There exists a DFA A = (Q, Σ, δ, s, F ) such that L(A) =
L(A1 )L(A2 ) and
#    $

m−2
n−2 n−2
#Q ≤ min k , i
,
i=0
≤i ≤t−1
#  $
m−1 n−2
+ min k , . (∗)
≤t

Proof. The DFA A is constructed in two steps. First, an NFA A

is con-
structed from A1 and A2 by adding a λ-transition from each final state in
F1 to the starting state 0 of A2 . Then, we construct a DFA A from the
NFA A
by the standard subset construction. Again, we assume that A is
reduced and ordered.
State Complexity of Regular Languages 65

It is clear that we can view each q ∈ Q as a pair (q1 , P2 ), where q1 ∈ Q1

and P2 ⊆ Q2 . The starting state of A is s = (0, ∅) if 0 ∈ F1 and s = (0, {0})
if 0 ∈ F1 . Let us consider all states q ∈ Q such that q = (i, P ) for a
particular state i ∈ Q1 − {m − 1} and some set P ⊆ Q2 . Since A1 is
ordered and acyclic, the number ofsuch states  in Q is restricted  by the
n − 2 n − 2
following three bounds: (1) k i , (2) , and (3) . We
≤i ≤t−1
explain these bounds below informally.
We have (1) as a bound since all states of the form q = (i, P ) are at a
level ≤ i, which have at most k i−1 predecessors. By saying that a state p
is at level i we mean that the length of the longest path from the starting
state to q is i.
We now consider (2). Notice that if q, q
∈ Q such that δ(q, a) = q
, q =
(q1 , P2 ) and q
= (q1
, P2
), then δ1 (q1 , a) = q1
and P2
= {δ2 (p, a) | p ∈ P2 }
if q1
∈ F1 and P2
= {0} ∪ {δ2 (p, a) | p ∈ P2 } if q1
∈ F1 . So, #P2
> #P2 is
possible only when q1
∈ F1 . Therefore, for q = (i, P ), #P ≤ i if i ∈ F1 and
#P ≤i + 1 if  i ∈ F1 . In both cases, the maximum number of distinct sets
n−2
P is . The number n − 2 comes from the exclusion of the sink
≤i
state n − 1 and starting state 0 of A2 . Note that, for a fixed i, either 0 ∈ P
for all (i, P ) ∈ Q or 0 is not in any set P such that (i, P ) ∈ Q.
(3) is a bound since for each state i ∈ Q1 − {m − 1}, there are at most
t − 1 final states on the path from the starting state to i (not including i).
For the second term of (∗), it suffices to explain that for each (m−1, P ),
P ⊆ Q2 , #P is bounded by the total number of final states in F1 . ✷

Corollary 4.1 Let Ai = (Qi , Σ, δi , 0, Fi ), i = 1, 2, be two DFA accept-

ing finite languages Li , i = 1, 2, respectively, and #Q1 = m, #Q2 = n,
and #F1 = t, where t > 0 is a constant. Then there exists a DFA
A = (Q, Σ, δ, s, F ) of O(mnt−1 + nt ) states such that L(A) = L(A1 )L(A2 ).

It has been shown in [1] that the bound given in Theorem 4.1 can be
reached in the case |Σ| = 2.
About the state complexity of the reversal of an m-state DFA language,
one may easily have a misconception. Many thought, without any hesita-
tion, that it should be linear (in terms of m), especially in the case of finite
languages. In fact, it is not even polynomial for both finite and infinite
regular languages. We break the misconception by giving two examples
in the following: one for finite languages and the other for infinite regular
66 C. Câmpeanu, K. Salomaa, S. Yu

languages. Note that a nontrivial proof for a tight upper bound on the
state complexity of the reversal of finite languages can be found in [1].
Example 4.1 Let m = 2n + 2 and L = {a, b}n a{a, b}≤n , where {a, b}≤n
denotes
λ ∪ {a, b} ∪ {a, b}2 ∪ · · · ∪ {a, b}n .
It is clear that L is a finite language accepted by an m-state DFA. One can
prove that any DFA accepting LR needs at least 2n states.
Example 4.2 An n-state DFA that accepts an infinite regular language is
shown in Figure 1. A proof showing that any DFA accepting the reversal
of this language requires at least 2n states can found in [19].

b ✟ b, c
☛ ☛✟

✏
a ✏
❄ ❄
a, c 1 ✙ 2
✒✑ ✒✑
✗ I a
✏✎ ...
✲ 0n b, c
✒✑ ❖
✏
a ✲
n-1 ✲ ...
✒✑a
✻
✍✌
b, c
Figure 1: An n-state DFA such that L(A)R requires 2n states

For the star operation, the difference between the state complexity for
finite and infinite regular languages is that the latter is 4 times the former.

a,
✎
b, c
b
✬ ✩

✗✔ ✗✔ ✛✘
✗✔ ✛✘ ❄
a a, b, ✗✔
c a, b
✗✔
a, b, c a, b a, b,
✗✔ ✗✔ ❄
✲ 0 ✲ 1 ✲ 2 ✲ 3 ✲ . . . ✲ n-3 ✲ n-2 a, b,✲c n-1
c
✖✕ ✖✕ ✖✕ ✖✕ ✖✕ ✚✙
✚✙ ✖✕ ✖✕
c ✻✻
c ✖ ✕
✫ ✪

Figure 2: An n-state DFA such that L(A)∗ needs 2n−3 + 2n−4 states
State Complexity of Regular Languages 67

Both bounds have been shown to be tight [1], [19]. Here we only give
two examples to demonstrate that the bound given on the table can be
reached.

Example 4.3 The n-state DFA A shown in Figure 2 accepts a finite lan-
guage. It is shown in [1] that any DFA accepting L(A)∗ needs at least
2n−3 + 2n−4 states (assuming that n is even).

Example 4.4 Let L be the language accepted by the DFA shown in Figure
3. It is shown in [18] that any DFA accepting L∗ requires at least 2n−1 +2n−2
states.
✏ a, b ✏
b 1 z
a ✿✒✑ 2
✒✑ a, b
✏
N
✲ 0 ❄
✒✑ ...
❑ ✏
✎
a, b n-1 ... ✮
✍✌
✒✑ y
a, b
Figure 3: An n-state DFA such that L(A)∗ requires 2n−1 + 2n−2 states

5 Cover-Automata for Finite Languages

Let L be a finite language. The number of states of a deterministic finite
automaton accepting L can, in the worst case, be exponential in the max-
imal length of words appearing in L, maxlen(L). The number of states of
the minimal deterministic automaton for L can be significantly reduced if
we require only that the automaton gives the “correct answer” for words of
length at most maxlen(L), and the automaton may accept also words hav-
ing length greater than maxlen(L). The more relaxed definition is natural
in many applications since the system can keep track of the input word
length separately. In a high-level programming language environment the
length-control function can be readily implemented by an integer variable.
The above idea leads to the definition of cover-automata [3]. A sim-
ilar notion called automaticity is used in [17] as a descriptive complexity
measure for arbitrary languages.
Let L be a finite language over Σ and maxlen(L) = l. A DFA A =
(Q, Σ, δ, q0 , F ) is said to be a deterministic finite cover-automaton (DFCA)
68 C. Câmpeanu, K. Salomaa, S. Yu

of L if L(A) ∩ Σ≤l = L. We say that A is a minimal DFCA of L if A has

the smallest number of states among all the cover-automata of L.
It is clear that a minimal cover-automaton of a finite language L may
require a smaller number of states than the minimal DFA accepting L. For
instance, the language {a, a3 , a5 } has a minimal cover-automaton with two
states whereas any DFA accepting it requires seven states.
It is well known that the minimal DFA accepting a regular language L
is unique up to isomorphism [7], [15], [18], the number of states is given by
the index of the Nerode congruence ≡L . The cover-automata do not enjoy
this uniqueness property. For instance let Σ = {a} and consider the finite
language L1 = {ε, a, a3 }. Since a2 ∈ L1 it is easy to see that any DFCA of
L1 needs at least three states. Thus

A = ({0, 1, 2}, Σ, δ, 0, {0, 1}),

where δ(i, a) = (i + 1 mod 3) is a minimal DFCA of L1 . On the other

hand, also A
= ({0, 1, 2}, Σ, δ
, 0, {0, 1}) is a minimal DFCA of L1 where
δ
(i, a) = δ(i, a), i = 0, 1, and δ
(2, a) = 1.
However, it turns out that for a finite language L the number of states
of any minimal DFCA of L is unique [3]. In order to show this we consider
an L-similarity relation on Σ∗ and a corresponding similarity relation de-
fined on the set of states of a cover-automaton. The notion of L-similarity
generalizes the Nerode congruence ≡L and it has first been introduced in
[8].
Let L ⊆ Σ∗ be a finite language and denote l = maxlen(L). The L-
similarity relation ∼L ⊆ Σ∗ ×Σ∗ is defined by setting x ∼L y iff the following
condition holds:

for all z ∈ Σ∗ such that #(xz) ≤ l and #(yz) ≤ l, xz ∈ L iff yz ∈ L.

The relation ∼L is reflexive and symmetric, but not transitive. For

instance if L = {a, b, ab} then a ∼L ab, ab ∼L b and a ∼L b. However, the
L-similarity relation has some “transitive-like” properties when we impose
additional conditions on the lengths of the words. The properties are listed
in the below lemma, the proof of which is straightforward.

Lemma 5.1 Let L ⊆ Σ∗ be a finite language and x, y, z ∈ Σ∗ , #(x) ≤

#(y) ≤ #(z). The following statements hold:

(i) If x ∼L y, x ∼L z, then y ∼L z.

(ii) If x ∼L y, y ∼L z, then x ∼L z.
State Complexity of Regular Languages 69

(iii) If x ∼L y, y∼L z, then x∼L z.

The relation ∼L is not an equivalence relation. However we can consider

L-similarity sets where all elements are pairwise similar. Then following [3]
we can introduce canonical dissimilar sequences of L where the elements are
pairwise dissimilar and the sequence contains an element from a maximal
number of L-similarity sets. The length of canonical dissimilar sequences
of L turns out to give the number of states of a minimal DFCA of L. These
notions are formalized in the below definition.

Definition 5.1 Let L ⊆ Σ∗ be a finite language.

(i) A set S ⊆ Σ∗ is called an L-similarity set if x ∼L y for any x, y ∈ S.

(ii) A sequence of words [x1 , . . . , xn ] over Σ is called a dissimilar sequence

of L if xi ∼L xj for each pair i, j, 1 ≤ i < j ≤ n.

(iii) A dissimilar sequence [x1 , . . . , xn ] is called a canonical dissimilar se-

quence of L if there exists a partition π = {S1 , . . . , Sn } of Σ∗ such
that for each i, 1 ≤ i ≤ n, xi ∈ Si , and Si is an L-similarity set.

(iv) A dissimilar sequence [x1 , . . . , xn ] of L is called a maximal dissimilar

sequence of L if for any dissimilar sequence [y1 , . . . , ym ] of L, m ≤ n.

Note that maximal dissimilar sequences always exist since all words
w ∈ Σ∗ such that #(w) > maxlen(L) are pairwise L-similar.

Theorem 5.1 Let L be a finite language. A dissimilar sequence of L is a

canonical dissimilar sequence of L if and only if it is a maximal dissimilar
sequence of L.

Proof. Let [x1 , . . . , xn ] be a canonical dissimilar sequence of L and π =

{S1 , . . . , Sn } the corresponding partition of Σ∗ . Let [y1 , . . . , ym ] be an ar-
bitrary dissimilar sequence of L. Assume that m > n. Then there are yi
and yj , i = j, such that yi , yj ∈ Sk for some k, 1 ≤ k ≤ n. Since Sk is an
L-similarity set, yi ∼L yj . This is impossible.
Conversely, let [x1 , . . . , xn ] be a maximal dissimilar sequence of L.
Without loss of generality we can suppose that #(x1 ) ≤ . . . ≤ #(xn ).
For i = 1, . . . , n, define

Xi = {y ∈ Σ∗ | y ∼L xi and y ∈ Xj for j < i}.

70 C. Câmpeanu, K. Salomaa, S. Yu

Note that for each y ∈ Σ∗ , y ∼L xi for at least one i, 1 ≤ i ≤ n, since

[x1 , . . . , xn ] is a maximal dissimilar sequence. Thus, π = {X1 , . . . , Xn } is a
partition of Σ∗ . The remaining task is to show that each Xi , 1 ≤ i ≤ n, is
a similarity set.
For the sake of contradiction assume that for some i, 1 ≤ i ≤ n, there
exist y, z ∈ Xi such that y∼L z. We know that xi ∼L y and xi ∼L z by the
definition of Xi . We have the following three cases: (1) #(xi ) < #(y), #(z),
(2) #(y) ≤ #(xi ) ≤ #(z) (or #(z) ≤ #(xi ) ≤ #(y)), and (3) #(xi ) >
#(y), #(z). If (1) or (2), then y ∼L z by Lemma 5.1 which is impossible.
In case (3) it is easy to prove that y ∼L xj and z ∼L xj , for all j = i, using
Lemma 5.1 and the definition of Xi . Then we can replace xi by both y
and z to obtain a longer dissimilar sequence [x1 , . . . , xi−1 , y, z, xi+1 , . . . , xn ].
This contradicts the fact that [x1 , . . . , xi−1 , xi , xi+1 , . . . , xn ] is a maximal
dissimilar sequence of L. Hence, y ∼L z and Xi is a similarity set. ✷

The above theorem implies that the number of elements in any canonical
dissimilar sequence of L is fixed. We denote this number by N (L).
It turns out that any minimal DFCA of a finite language L will have
exactly N (L) states. In the following, if A = (Q, Σ, δ, q0 , F ) is a DFA, in
order to simplify the notation, we always assume that Q = {0, 1, . . . , |Q|}
and q0 = 0. First we define the level of states and a similarity relation on
the states of a DFCA.

Definition 5.2 Let A = (Q, Σ, δ, 0, F ) be a DFA. We define, for each state

q ∈ Q,
level(q) = min{#(w) | δ(0, w) = q},
i.e., level(q) is the length of the shortest path from the initial state to q.

Definition 5.3 Let A = (Q, Σ, δ, 0, F ) be a DFCA of a finite language L

and l = maxlen(L). Let level(p) = i and level(q) = j, m = max{i, j}. We
say that p ∼A q (state p is L-similar to q in A) if for every w ∈ Σ≤l−m ,
δ(p, w) ∈ F iff δ(q, w) ∈ F .

Again ∼A is a weakened version of the Nerode equivalence relation

defined on states of a DFA. Note that, strictly speaking, the relation ∼A
depends also on L (and not just on A). It follows from the definition that
A is a minimal DFCA of L if and only if no distinct pair of states of A is
in the L-similarity relation.
If A = (Q, Σ, δ, 0, F ) is a DFA, for each q ∈ Q, we denote xA (q) =
min{w | δ(0, w) = q}, where the minimum is taken according to the quasi-
State Complexity of Regular Languages 71

lexicographical order on Σ∗ . When the automaton A is understood, we

write xq instead of xA (q).

Lemma 5.2 Let A = (Q, Σ, δ, 0, F ) be a DFCA of a finite language L. Let

x, y ∈ Σ∗ be such that δ(0, x) = p and δ(0, y) = q. If p ∼A q, then x ∼L y.

Proof. Let level(p) = i and level(q) = j, m = max{i, j}, and p ∼A q.

Choose an arbitrary w ∈ Σ∗ such that #(xw) ≤ l and #(yw) ≤ l. Because
i ≤ #(x) and j ≤ #(y) it follows that #(w) ≤ l − m. Since p ∼A q we
have that δ(p, w) ∈ F iff δ(q, w) ∈ F , that is, δ(0, xw) ∈ F iff δ(0, yw) ∈ F .
This implies that xw ∈ L(A) iff yw ∈ L(A). Hence x ∼L y. ✷

Lemma 5.3 Let A = (Q, Σ, δ, 0, F ) be DFCA of a finite language L. Let

level(p) = i and level(q) = j, m = max{i, j}, and x ∈ Σi , y ∈ Σj be such
that δ(0, x) = p and δ(0, y) = q. If x ∼L y, then p ∼A q.

Proof. Let x ∼L y and w ∈ Σ≤l−m . If δ(p, w) ∈ F , then δ(0, xw) ∈ F .

Because x ∼L y, it follows that δ(0, yw) ∈ F , so δ(q, w) ∈ F . Using
symmetry we get p ∼A q. ✷

Now we can prove the main result of this section.

Theorem 5.2 Let L ⊆ Σ∗ be finite and A = (Q, Σ, δ, 0, F ) be a minimal

DFCA of L. Then A has exactly N (L) states.

Proof. First assume that |Q| < N (L). Let [y1 , . . . , yN (L) ] be a canonical
dissimilar sequence of L. Then there exist i, j, 1 ≤ i < j ≤ N (L), such that
δ(0, yi ) = δ(0, yj ) = q for some q ∈ Q. Using Lemma 5.2 and reflexivity of
the relation ∼A we get yi ∼L yj . This is a contradiction.
Secondly, consider the possibility that N (L) < |Q|. By Theorem 5.1
there exist p, q ∈ Q, p = q, such that xp ∼L xq . By the definition of
the words xp , xq we have #(xp ) = level(p) and #(xq ) = level(q). Now
Lemma 5.3 gives that p ∼A q. Without loss of generality we can assume
that level(p) ≤ level(q). We define a DFA A
= (Q
, Σ, δ
, 0, F
) where
Q
= Q − {q}, F
= F − {q}, and
#

δ(s, a) if δ(s, a) = q,
δ (s, a) =
p δ(s, a) = q
72 C. Câmpeanu, K. Salomaa, S. Yu

for each s ∈ Q
and a ∈ Σ. It can be verified that L(A
) ∩ Σ≤(maxlen(L)) = L
and thus A
is a DFCA of L. This contradicts the minimality of A. ✷

In general, a minimal DFCA of a finite language L is smaller than

the minimal DFA accepting L. In [3] it is shown that if A is a minimal
DFCA of L then A is minimal also as a DFA, i.e., it is the smallest DFA
accepting the language L(A). (Note that L(A) and L may differ on words
of length greater than maxlen(L).) The paper [3] also gives an algorithm
for constructing the L-similarity relation of a DFCA A, and using it a
minimization algorithm for cover-automata.

References
[1] C. Câmpeanu, K. Culik II, K. Salomaa, S. Yu, State complexity of ba-
sic operations on finite languages, Proc. Fourth International Work-
shop on Implementing Automata, WIA’99, to appear.
[2] C. Câmpeanu, K. Culik II, S. Yu, Finite languages and cover au-
tomata, in preparation.
[3] C. Câmpeanu, N. Sântean, S. Yu, Minimal cover automata for fi-
nite languages, Proc. Third International Workshop on Implement-
ing Automata, WIA’98, Lecture Notes in Computer Science, 1660,
Springer-Verlag, Berlin, 1999, 33–42.
[4] A. K. Chandra, D. C. Kozen, L. J. Stockmeyer, Alternation, Journal
of the ACM, 28 (1981), 114–133.
[5] A. Fellah, H. Jürgensen, S. Yu, Constructions for alternating finite
automata, Intern. J. Computer Math., 35 (1990) 117–132.
[6] J. E. Hopcroft, An n log n algorithm for minimizing the states in
a finite automaton, The Theory of Machines and Computations (Z.
Kohavi, ed.), Academic Press, New York, 189–196.
[7] J. E. Hopcroft, J. D. Ullman, Introduction to Automata Theory, Lan-
guages, and Computation, Addison Wesley, Reading, Mass., 1979.
[8] J. Kaneps, R. Freivalds, Minimal nontrivial space complexity of prob-
abilistic one-way Turing machines, Proc. Mathematical Foundations
of Computer Science, Banská Bystryca, Czechoslovakia, August 1990,
Lecture Notes in Computer Science, 452, Springer-Verlag, Berlin,
1990, 355–361.
State Complexity of Regular Languages 73

[9] E. Leiss, Succinct representation of regular languages by boolean au-

tomata, Theoretical Computer Science, 13 (1981), 323–330.

[10] A. R. Meyer, M. J. Fischer. Economy of description by automata,

grammars, and formal systems, FOCS, 12 (1971), 188–191.

[11] B. Ravikumar, Some applications of a technique of Sakoda and Sipser,

SIGACT News, 21 (1990), 73–77.

[12] B. Ravikumar, O. H. Ibarra, Relating the type of ambiguity of fi-

nite automata to the succinctness of their representation, SIAM J.
Comput., 18, 6 (1989), 1263–1282.

[13] D. Revuz, Minimisation of acyclic deterministic automata in linear

time, Theoretical Computer Science, 92 (1992), 181–189.

[14] T. Jiang, B. Ravikumar, Minimal NFA problems are hard, Proc. 18th
ICALP, Lecture Notes in Computer Science, 510, Springer-Verlag,
Berlin, 1991, 629–640.

[15] A. Salomaa, Theory of Automata, Pergamon Press, Oxford, 1969.

[16] J. Shallit, Private communication, Nov. 1996.

[17] J. Shallit and Y. Breitbart, Automaticity I: Properties of a Measure of

Descriptional Complexity, Journal of Computer and System Sciences,
53 (1996), 10–25.

[18] S. Yu, Regular languages, in vol. 2 of Handbook of Formal Languages

(G. Rozenberg, A. Salomaa, eds.), Springer-Verlag, Berlin, 1997.

[19] S. Yu, Q. Zhuang, K. Salomaa, The state complexity of some basic

operations on regular languages, Theoretical Computer Science, 125
(1994), 315–328.
 A Century of Controversy
over the Foundations of Mathematics1
Gregory J. Chaitin
IBM T. J. Watson Research Center
30 Saw Mil River Road
Hawthorne, NY 10532 USA
E-mail: chaitin@watson.ibm.com

1 Introduction
I would like to talk about some crazy stuff. The general idea is that some-
times ideas are very powerful. I’d like to talk about theory, about the
computer as a concept, a philosophical concept.
We all know that the computer is a very practical thing out there in
the real world! It pays for a lot of our salaries, right? But what people
don’t remember as much is that really—I’m going to exaggerate, but I’ll
say it—the computer was invented in order to help to clarify a question
about the foundations of mathematics, a philosophical question about the
foundations of mathematics.
Now that sounds absurd, but there’s some truth in it. There are actually
lots of threads that led to the computer, to computer technology, which
come from mathematical logic and from philosophical questions about the
limits and the power of mathematics.
The computer pioneer Turing was inspired by these questions. Turing
was trying to settle a question of Hilbert’s having to do with the philosophy
of mathematics, when he invented a thing called the Turing machine, which
is a mathematical model of a toy computer. Turing did this before there
were any real computers, and then he went on to actually build computers.
The first computers in England were built by Turing.
And von Neumann, who was instrumental in encouraging the creation
of computers as a technology in the United States (unfortunately as part
1
Lecture given Friday 30 April 1999 at UMass-Lowell. The lecture was videotaped;
this is an edited transcript.

75
76 G. J. Chaitin

of the war effort, as part of the effort to build the atom bomb), he knew
Turing’s work very well. I learned of Turing by reading von Neumann
talking about the importance of Turing’s work.
So what I said about the origin of the computer isn’t a complete lie, but
it is a forgotten piece of intellectual history. In fact, let me start off with
the final conclusion of this talk. . . In a way, a lot of this came from work
of Hilbert. Hilbert, who was a very well-known German mathematician
around the beginning of this century, had proposed formalizing completely
all of mathematics, all of mathematical reasoning—deduction. And this
proposal of his is a tremendous, glorious failure!
In a way, it’s a spectacular failure. Because it turned out that you
couldn’t formalize mathematical reasoning. That’s a famous result of
Gödel’s that I’ll tell you about, done in 1931.
But in another way, Hilbert was really right, because formalism has been
the biggest success of this century. Not for reasoning, not for deduction, but
for programming, for calculating, for computing, that’s where formalism
has been a tremendous success. If you look at work by logicians at the
beginning of this century, they were talking about formal languages for
reasoning and deduction, for doing mathematics and symbolic logic, but
they also invented some early versions of programming languages. And
these are the formalisms that we all live with and work with now all the
time! They’re a tremendously important technology.
So formalism for reasoning did not work. Mathematicians don’t rea-
son in formal languages. But formalism for computing, programming lan-
guages, are, in a way, what was right in the formalistic vision that goes
back to Hilbert at the beginning of this century, which was intended to
clarify epistemological, philosophical questions about mathematics.
So I’m going to tell you this story, which has a very surprising outcome.
I’m going to tell you this surprising piece of intellectual history.

2 The Crisis in Set Theory

So let me start roughly a hundred years ago, with Georg Cantor.
The point is this. Normally you think that pure mathematics is static,
unchanging, perfect, absolutely correct, absolute truth. . . Right? Physics
may be tentative, but math, things are certain there! Well, it turns out
that’s not exactly the case.
In this century, in this past century there was a lot of controversy over
the foundations of mathematics, and how you should do math, and what’s
A Century of Controversy over the Foundations of Mathematics 77

right and what isn’t right, and what’s a valid proof. Blood was almost shed
over this. . . People had terrible fights and ended up in insane asylums over
this. It was a fairly serious controversy. This isn’t well known, but I think
it’s an interesting piece of intellectual history.
More people are aware of the controversy over relativity theory. Einstein
was very controversial at first. And then of the controversy over quantum
mechanics. . . These were the two revolutions in the physics of this century.
But what’s less well known is that there were tremendous revolutions and
controversies in pure mathematics too. I’d like to tell you about this. It
really all starts in a way from Cantor.
What Cantor did was to invent a theory of infinite sets.
He did it about a hundred years ago; it’s really a little more than a
hundred years ago. And it was a tremendously revolutionary theory, it was
extremely adventurous. Let me tell you why.
Cantor said, let’s take 1, 2, 3, . . . We’ve all seen these numbers, right?!
And he said, well, let’s add an infinite number after this.

1, 2, 3, . . . ω

He called it ω, lowercase Greek omega. And then he said, well, why stop
here? Let’s go on and keep extending the number series.

1, 2, 3, . . . ω, ω + 1, ω + 2, . . .

Omega plus one, omega plus two, then you go on for an infinite amount of
time. And what do you put afterwards? Well, two omega? (Actually, it’s
omega times two for technical reasons.)

1, 2, 3, . . . ω . . . 2ω

Then two omega plus one, two omega plus two, two omega plus three, two
omega plus four. . .

1, 2, 3, . . . 2ω, 2ω + 1, 2ω + 2, 2ω + 3, 2ω + 4, . . .

Then you have what? Three omega, four omega, five omega, six omega,. . .

1, 2, 3, . . . , 3ω, . . . , 4ω, . . . , 5ω, . . . , 6ω, . . .

Well, what will come after all of these? Omega squared! Then you
keep going, omega squared plus one, omega squared plus six omega plus
eight. . . Okay, you keep going for a long time, and the next interesting thing
78 G. J. Chaitin

after omega squared will be? Omega cubed! And then you have omega to
the fourth, omega to the fifth, and much later?

1, 2, 3, . . . , ω, . . . , ω 2 , . . . , ω 3 , . . . , ω 4 , . . . , ω 5

Omega to the omega!

1, 2, 3, . . . , ω, . . . , ω 2 , . . . , ω ω

And then much later it’s omega to the omega to the omega an infinite
number of times!
...
ωω
1, 2, 3, . . . , ω, . . . , ω 2 , . . . , ω ω , . . . , ω ω

I think this is usually called epsilon nought.

ω ...
ωω
ε0 = ω

It’s a pretty mind-boggling number! After this point things get a little
complicated. . .
And this was just one little thing that Cantor did as a warm-up exercise
for his main stuff, which was measuring the size of infinite sets! It was
spectacularly imaginative, and the reactions were extreme. Some people
loved what Cantor was doing, and some people thought that he should be
put in an insane asylum! In fact he had a nervous breakdown as a result
of those criticisms. Cantor’s work was very influential, leading to point-
set topology and other abstract fields in the mathematics of the twentieth
century. But it was also very controversial. Some people said, it’s theology,
it’s not real, it’s a fantasy world, it has nothing to do with serious math!
And Cantor never got a good position and he spent his entire life at a
second-rate institution.

3 Bertrand Russell’s Logical Paradoxes

Then things got even worse, due mainly, I think, to Bertrand Russell, one
of my childhood heroes.
Bertrand Russell was a British philosopher who wrote beautiful essays,
very individualistic essays, and I think he got the Nobel prize in literature
for his wonderful essays. Bertrand Russell started off as a mathematician
and then degenerated into a philosopher and finally into a humanist; he
went downhill rapidly! Anyway, Bertrand Russell discovered a whole bunch
A Century of Controversy over the Foundations of Mathematics 79

of disturbing paradoxes, first in Cantor’s theory, then in logic itself. He

found cases where reasoning that seemed to be okay led to contradictions.
And I think that Bertrand Russell was tremendously influential in
spreading the idea that there was a serious crisis and that these contradic-
tions had to be resolved somehow. The paradoxes that Russell discovered
attracted a great deal of attention, but strangely enough only one of them
ended up with Russell’s name on it! For example, one of these paradoxes
is called the Burali-Forti paradox, because when Russell published it he
stated in a footnote that it had been suggested to him by reading a paper
by Burali-Forti. But if you look at the paper by Burali-Forti, you don’t see
the paradox!
But I think that the realization that something was seriously wrong,
that something was rotten in the state of Denmark, that reasoning was
bankrupt and something had to be done about it pronto, is due princi-
pally to Russell. Alejandro Garciadiego, a Mexican historian of math, has
written a book which suggests that Bertrand Russell really played a much
bigger role in this than is usually realized: Russell played a key role in
formulating not only the Russell paradox, which bears his name, but also
the Burali-Forti paradox and the Berry paradox, which don’t. Russell was
instrumental in discovering them and in realizing their significance. He
told everyone that they were important, that they were not just childish
word-play.
Anyway, the best known of these paradoxes is called the Russell para-
dox nowadays. You consider the set of all sets that are not members of
themselves. And then you ask, “Is this set a member of itself or not?” If
it is a member of itself, then it shouldn’t be, and vice versa! It’s like the
barber in a small, remote town who shaves all the men in the town who
don’t shave themselves. That seems pretty reasonable, until you ask “Does
the barber shave himself?” He shaves himself if and only if he doesn’t shave
himself, so he can’t apply that rule to himself!
Now you may say, “Who cares about this barber!” It was a silly rule
anyway, and there are always exceptions to the rule! But when you’re
dealing with a set, with a mathematical concept, it’s not so easy to dismiss
the problem. Then it’s not so easy to shrug when reasoning that seems to
be okay gets you into trouble!
By the way, the Russell paradox is a set-theoretic echo of an earlier
paradox, one that was known to the ancient Greeks and is called the Epi-
menides paradox by some philosophers. That’s the paradox of the liar:
“This statement is false!” “What I’m now saying is false, it’s a lie.” Well,
80 G. J. Chaitin

is it false? If it’s false, if something is false, then it doesn’t correspond with

reality. So if I’m saying this statement is false, that means that it’s not
false—which means that it must be true. But if it’s true, and I’m saying
it’s false, then it must be false! So whatever you do you’re in trouble!
So you can’t get a definite logical truth value, everything flip flops, it’s
neither true nor false. And you might dismiss this and say that these are
just meaningless word games, that it’s not serious. But Kurt Gödel later
built his work on these paradoxes, and he had a very different opinion.
He said that Bertrand Russell made the amazing discovery that our log-
ical intuitions, our mathematical intuitions, are self-contradictory, they’re
inconsistent! So Gödel took Russell very seriously, he didn’t think that it
was all a big joke.
Now I’d like to move on and tell you about David Hilbert’s rescue plan
for dealing with the crisis provoked by Cantor’s set theory and by Russell’s
paradoxes.

4 David Hilbert to the Rescue with Formal Ax-

iomatic Theories
One of the reactions to the crisis provoked by Cantor’s theory of infinite
sets was the suggestion to escape into formalism. If we get into trouble
with reasoning that seems okay, then one solution is to use symbolic logic,
to create an artificial language where we’re going to be very careful and
say what the rules of the game are, and make sure that we don’t get the
contradictions. Because here’s a piece of reasoning that looks okay but it
leads to a contradiction, we’d like to get rid of that. Natural language is
ambiguous—you never know what a pronoun refers to. So let’s create an
artificial language and make things very, very precise and make sure that
we get rid of all the contradictions! So this was the notion of formalism.
Now I don’t think that Hilbert actually intended that mathematicians
should work in such a perfect artificial language. It would sort of be like a
programming language, but for reasoning, for doing mathematics, for de-
duction, not for computing, that was Hilbert’s idea. But he never expressed
it that way, because there were no programming languages back then.
So what are the ideas here? First of all, Hilbert stressed the importance
of the axiomatic method.
The notion of doing mathematics that way goes back to the ancient
Greeks and particularly to Euclidean geometry, which is a beautifully clear
mathematical system. But that’s not enough; Hilbert was also saying that
A Century of Controversy over the Foundations of Mathematics 81

we should use symbolic logic.

And symbolic logic also has a long history: Leibniz, Boole, Frege,
Peano. . . These mathematicians wanted to make reasoning like algebra.
Here’s how Leibniz put it: He talked about avoiding disputes—and he
was probably thinking of political disputes and religious disputes—by cal-
culating who was right instead of arguing about it! Instead of fighting, you
should be able to sit down at a table and say, “Gentleman, let us compute!”
What a beautiful fantasy!. . .
So the idea was that mathematical logic should be like arithmetic and
you should be able to just grind out a conclusion, no uncertainty, no ques-
tions of interpretation. By using an artificial math language with a symbolic
logic you should be able to achieve perfect rigor. The idea is that an argu-
ment is either completely correct or else it’s total nonsense, with nothing
in between. And a proof that is formulated in a formal axiomatic system
should be absolutely clear, it should be completely sharp!
In other words, Hilbert’s idea was that we should be completely precise
about what the rules of the game are, and about the definitions, the ele-
mentary concepts, and the grammar and the language—all the rules of the
game—so that we can all agree on how mathematics should be done. In
practice it would be too much work to use such a formal axiomatic system,
but it would be philosophically significant because it would settle once and
for all the question of whether a piece of mathematical reasoning is correct
or incorrect.
Hilbert’s idea seemed fairly straightforward. He was just following the
axiomatic and the formal traditions in mathematics. Formal as in formal-
ism, as in using formulas, as in calculating! He wanted to go all the way,
to the very end, and formalize all of mathematics, but it seemed like a
fairly reasonable plan. Hilbert wasn’t a revolutionary, he was a conserva-
tive. . . The amazing thing, as I said before, was that it turned out that
Hilbert’s rescue plan could not work, that it couldn’t be done, that it was
impossible to make it work!
Hilbert was just following the whole mathematics tradition up to that
point: the axiomatic method, symbolic logic, formalism. . . He wanted to
avoid the paradoxes by being absolutely precise, by creating a completely
formal axiomatic system, an artificial language, that avoided the paradoxes,
that made them impossible, that outlawed them! And most mathematicians
probably thought that Hilbert was right, that of course you could do this—
it’s just the notion that in mathematics things are absolutely clear, black
or white, true or false.
82 G. J. Chaitin

So Hilbert’s idea was just an extreme, an exaggerated version of the

normal notion of what mathematics is all about: the idea that we can decide
and agree on the rules of the game, all of them, once and for all. The big
surprise is that it turned out that this could not be done. Hilbert turned
out to be wrong, but wrong in a tremendously fruitful way, because he had
asked a very good question. In fact, by asking this question he actually
created an entirely new field of mathematics called metamathematics.
Metamathematics is mathematics turned inward, it’s an introspective
field of math in which you study what mathematics can achieve or can’t
achieve.

5 What is Metamathematics?
That’s my field—metamathematics! In it you look at mathematics from
above, and you use mathematical reasoning to discuss what mathemat-
ical reasoning can or cannot achieve. The basic idea is this: Once you
entomb mathematics in an artificial language à la Hilbert, once you set
up a completely formal axiomatic system, then you can forget that it has
any meaning and just look at it as a game that you play with marks on
paper that enables you to deduce theorems from axioms. You can forget
about the meaning of this game, the game of mathematical reasoning, it’s
just combinatorial play with symbols! There are certain rules, and you can
study these rules and forget that they have any meaning!
What things do you look at when you study a formal axiomatic system
from above, from the outside? What kind of questions do you ask?
Well, one question you can ask is if you can prove that “0 equals 1”?
Hopefully you can’t, but how can you be sure? It’s hard to be sure!
And for any question A, for any affirmation A, you can ask if it’s possible
to settle the matter by either proving A or the opposite of A, not A. That’s
called completeness.
A formal axiomatic system is complete if you can settle any question A,
either by proving it (A), or by proving that it’s false (¬A). That would be
nice! Another interesting question is if you can prove an assertion (A) and
you can also prove the contrary assertion (¬A). That’s called inconsistency.
So what Hilbert did was to have the remarkable idea of creating a
new field of mathematics whose subject would be mathematics itself. But
you can’t do this until you have a completely formal axiomatic system.
Because as long as any “meaning” is involved in mathematical reasoning,
it’s all subjective. Of course, the reason we do mathematics is because it has
A Century of Controversy over the Foundations of Mathematics 83

meaning, right? But if you want to be able to study mathematics, the power
of mathematics, using mathematical methods, you have to “desiccate” it
to “crystallize out” the meaning and just be left with an artificial language
with completely precise rules, in fact, with one that has a mechanical proof-
checking algorithm.
The key idea that Hilbert had was to envision this perfectly desiccated
or crystallized axiomatic system for all of mathematics, in which the rules
would be so precise that if someone had a proof there would be a referee,
there would be a mechanical procedure, which would either say “This proof
obeys the rules” or “This proof is wrong; it’s breaking the rules”. That’s
how you get the criterion for mathematical truth to be completely objective
and not to depend on meaning or subjective understanding: by reducing it
all to calculation. Somebody says “This is a proof”, and instead of having
to submit it to a human referee who takes two years to decide if the paper is
correct, instead you just give it to a machine. And the machine eventually
says “This obeys the rules” or “On line 4 there’s a misspelling” or “This
thing on line 4 that supposedly follows from line 3, actually doesn’t”. And
that would be the end, no appeal!
The idea was not that mathematics should actually be done this way.
I think that that’s calumny, that’s a false accusation. I don’t think that
Hilbert really wanted to turn mathematicians into machines. But the idea
was that if you could take mathematics and do it this way, then you could
use mathematics to study the power of mathematics. And that is the im-
portant new thing that Hilbert came up with. Hilbert wanted to do this
in order to reaffirm the traditional view of mathematics, in order to justify
himself. . .
He proposed having one set of axioms and this formal language, this
formal system, which would include all of mathematical reasoning, that we
could all agree on, and that would be perfect! We’d then know all the rules
of the game. And he just wanted to use metamathematics to show that this
formal axiomatic system was good—that it was consistent and that it was
complete—in order to convince people to accept it. This would have settled
once and for all the philosophical questions “When is a proof correct?” and
“What is mathematical truth?” Like this everyone could agree on whether
a mathematical proof is correct or not. And in fact we used to think that
this was an objective thing.
In other words, Hilbert’s just saying, if it’s really objective, if there’s
no subjective element, and a mathematical proof is either true or false,
well, then there should be certain rules for deciding that and it shouldn’t
84 G. J. Chaitin

depend, if you fill in all the details, it shouldn’t depend on interpretation.

It’s important to fill in all the details—that’s the idea of mathematical logic,
to “atomize” mathematical reasoning into such tiny steps that nothing is
left to the imagination, nothing is left out! And if nothing is left out, then
a proof can be checked automatically, that was Hilbert’s point, that’s really
what symbolic logic is all about.
And Hilbert thought that he was actually going to be able to do this.
He was going to formalize all of mathematics, and we were all going to agree
that these were in fact the rules of the game. Then there’d be just one ver-
sion of mathematical truth, not many variations. We don’t want to have a
German mathematics and a French mathematics and a Swedish mathemat-
ics and an American mathematics, no, we want a universal mathematics,
one universal criterion for mathematical truth! Then a paper that is done
by a mathematician in one country can be understood by a mathematician
in another country. Doesn’t that sound reasonable?! So you can imagine
just how very, very shocking it was in 1931 when Kurt Gödel showed that
it wasn’t at all reasonable, that it could never be done!

6 Kurt Gödel Discovers Incompleteness

Gödel did this is Vienna, but he was from the Czech Republic, from the
city of Brünn or Brno. And later he was at the Institute for Advanced
Study in Princeton,
Gödel sort of exploded this whole view of what mathematics is all about.
He came up with a famous incompleteness result, “Gödel’s incompleteness
theorem”.
There’s a lovely book explaining the way Gödel originally did it. It’s
by Nagel and Newman, and it’s called Gödel’s Proof. I read it when I was
a child, and forty years later it’s still in print!
What is this amazing result of Gödel’s? Gödel’s amazing discovery is
that Hilbert was wrong, that it cannot be done, that there’s no way to take
all of mathematical truth and to agree on a set of rules and to have a formal
axiomatic system for all of mathematics in which it is crystal clear whether
something is correct or not!
More precisely, what Gödel discovered was that if you just try to deal
with elementary arithmetic, with 0, 1, 2, 3, 4,. . . and with addition and
multiplication
+ × 0, 1, 2, 3, 4, . . .
(this is “elementary number theory” or “arithmetic”) and you just try
A Century of Controversy over the Foundations of Mathematics 85

to have a set of axioms for this—the usual axioms are called Peano
arithmetic—even this can’t be done! Any set of axioms that tries to have
the whole truth and nothing but the truth about addition, multiplication,
and 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,. . . will have to be incomplete. More pre-
cisely, it’ll either be inconsistent or it’ll be incomplete. So if you assume
that it only tells the truth, then it won’t tell the whole truth. There’s no
way to capture all the truth about addition, multiplication, and 0, 1, 2, 3,
4,. . . ! In particular, if you assume that the axioms don’t allow you to prove
false theorems, then it’ll be incomplete, there’ll be true theorems that you
cannot prove from these axioms!
This is an absolutely devastating result, and all of traditional mathe-
matical philosophy ends up in a heap on the floor! At the time this was
considered to be absolutely devastating. However you may notice that in
1931 there were also a few other problems to worry about. The situation
in Europe was bad. There was a major depression, and a war was brew-
ing. I agree, not all problems are mathematical! There’s more to life than
epistemology! But you begin to wonder, well, if the traditional view of
mathematics isn’t correct, then what is correct? Gödel’s incompleteness
theorem was very surprising and a terrible shock.
How did Gödel do it? Well, Gödel’s proof is very clever. It almost looks
crazy, it’s very paradoxical. Gödel starts with the paradox of the liar, “I’m
false!”, which is neither true nor false.

“This statement is false!”

And what Gödel does is to construct a statement that says of itself “I’m
unprovable!”
“This statement is unprovable!”
Now if you can construct such a statement in elementary number theory,
in arithmetic, a mathematical statement—I don’t know how you make a
mathematical statement say it’s unprovable, you’ve got to be very clever—
but if you can do it, it’s easy to see that you’re in trouble. Just think
about it a little bit. It’s easy to see that you’re in trouble. Because if it’s
provable, it’s false, right? So you’re in trouble, you’re proving false results.
And if it’s unprovable and it says that it’s unprovable, then it’s true, and
mathematics is incomplete. So either way, you’re in trouble! Big trouble!
And Gödel’s original proof is very, very clever and hard to understand.
There are a lot of complicated technical details. But if you look at his
original paper, it seems to me that there’s a lot of LISP programming in
it, or at least something that looks a lot like LISP programming. Anyway,
86 G. J. Chaitin

now we’d call it LISP programming. Gödel’s proof involves defining a

great many functions recursively, and these are functions dealing with lists,
which is precisely what LISP is all about. So even though there were
no programming languages in 1931, with the benefit of hindsight you can
clearly see a programming language in Gödel’s original paper. And the
programming language I know that’s closest to it is LISP, pure LISP, LISP
without side-effects, interestingly enough—that’s the heart of LISP.
So this was a very, very shocking result, and people didn’t really know
what to make of it.
Now the next major step forward comes only five years later, in 1936,
and it’s by Alan Turing.

7 Alan Turing Discovers Uncomputability

Turing’s approach to all these questions is completely different from
Gödel’s, and much deeper. Because Turing brings it out of the closet! What
he brings out of the closet is the computer! The computer was implicit in
Gödel’s paper, but this was really not visible to any ordinary mortal, not
at that time, only with hindsight. And Turing really brings it out in the
open.
Hilbert had said that there should be a “mechanical procedure” to de-
cide if a proof obeys the rules or not. And Hilbert never clarified what
he meant by a mechanical procedure, it was all words. But, Turing said,
what you really mean is a machine, and a machine of a kind that we now
call a Turing machine—but it wasn’t called that in Turing’s original paper.
In fact, Turing’s original paper contains a programming language, just like
Gödel’s paper does, what we would now call a programming language. But
the two programming languages are very different. Turing’s programming
language isn’t a high-level language like LISP, it’s more like a machine lan-
guage. In fact, it’s a horrible machine language, one that nobody would
want to use today, because it’s too simple.
But Turing makes the point that even though Turing machines are very
simple, even though their machine language is rather primitive, they’re
very flexible, very general-purpose machines. In fact, he claims, any com-
putation that a human being can perform, should be possible to do using
such a machine. Turing’s train of thought now takes a very dramatic turn.
What, he asks, is impossible for such a machine? What can’t it do? And he
immediately finds a question that no Turing machine can settle, a problem
that no Turing machine can solve. That’s the halting problem, the problem
A Century of Controversy over the Foundations of Mathematics 87

of deciding in advance if a Turing machine or a computer program will

eventually halt.
So the shocking thing about this 1936 paper is that first of all he comes
up with the notion of a general-purpose or universal computer, with a
machine that’s flexible, that can do what any machine can do. One cal-
culating machine that can do any calculation, which is, we now say, a
general-purpose computer. And then he immediately shows that there are
limits to what such a machine can do. And how does he find something
that cannot be done by any such machine? Well, it’s very simple! It’s the
question of whether a computer program will eventually halt, with no time
limit.
If you put a time limit, it’s very easy. If you want to know if a program
halts in a year, you just run it for a year, and either it halted or doesn’t.
What Turing showed is that you get in terrible trouble if there’s no time
limit. Now you may say, “What good is a computer program that takes
more than a year, that takes more than a thousand years?! There’s always
a time limit!” I agree, this is pure math, this is not the real world. You
only get in trouble with infinity! But Turing shows that if you put no time
limit, then you’re in real difficulties.
So this is called the halting problem. And what Turing showed is that
there’s no way to decide in advance if a program will eventually halt.
If it does halt, by running it you can eventually discover that, if you’re
just patient. The problem is you don’t know when to give up. And Turing
was able to show with a very simple argument which is just Cantor’s diago-
nal argument—coming from Cantor’s theory of infinite sets, by the way—I
don’t have time to explain all this—with a very simple argument Turing
was able to show that this problem cannot be solved.
No computer program can tell you in advance if another computer pro-
gram will eventually halt or not. And the problem is the ones that don’t
halt, that’s really the problem. The problem is knowing when to give up.
So now the interesting thing about this is that Turing immediately
deduces as a corollary that if there’s no way to decide in advance by a
calculation if a program will halt or not, then there cannot be any way to
deduce it in advance using reasoning either. No formal axiomatic system
can enable you to deduce in advance whether a program will halt or not.
Because if you can use a formal axiomatic system to always deduce
whether a program will halt or not, that will give you a way to calculate
in advance whether a program will halt or not. You simply run through
all possible deductions—you can’t do this in practice—but in principle you
88 G. J. Chaitin

can run through all possible proofs in size order, checking which ones are
correct, until either you find a proof that the program will halt eventually
or you find a proof that it’s never going to halt.
This is using the idea of a completely formal axiomatic system where
you don’t need a mathematician—you just run through this calculation on
a computer—it’s mechanical to check if a proof is correct or not. So if there
were a formal axiomatic system which always would enable you to prove,
to deduce, whether a program will halt or not, that would give you a way
to calculate in advance whether a program will halt or not. And that’s
impossible, because you get into a paradox like “This statement is false!”
You get a program that halts if and only if it doesn’t halt, that’s basically
the problem. You use an argument having the same flavor as the Russell
paradox.
So Turing went more deeply into these questions than Gödel. As a
student I read Gödel’s proof, and I could follow it step by step. But I
couldn’t really feel that I was coming to grips with Gödel’s proof, that I
could really understand it. The whole thing seemed too delicate, it seemed
too fragile, it seemed too superficial. . . And there’s this business in the
closet about computing, that’s there in Gödel, but it’s hidden, it’s not in
the open, we’re not really coming to terms with it.
Now Turing is really going, I think, much deeper into this whole matter.
And he’s showing, by the way, that it’s not just one particular axiomatic
system, the one that Gödel studied, that can’t work, but that no formal
axiomatic system can work. But it’s in a slightly different context. Gödel
was really looking at 0, 1, 2, 3, 4. . . and addition and multiplication, and
Turing is looking at a rather strange mathematical question, which is does
a program halt or not. It’s a mathematical question that did not exist at the
time of Gödel’s original paper. So you see, Turing worked with completely
new concepts. . .
But Gödel’s paper is not only tremendously clever, he had to have the
courage to imagine that Hilbert might be wrong. There’s another famous
mathematician of that time, von Neumann. Von Neumann was probably
as clever as Gödel or anyone else, but it never occurred to him that Hilbert
could be wrong. And the moment that he heard Gödel explain his result,
von Neumann immediately appreciated it and immediately started deduc-
ing consequences. But von Neumann said, “I missed it, I missed the boat,
I didn’t get it right!” And Gödel did, so he was much more profound. . .
Now Turing’s paper is also full of technical details, like Gödel’s paper,
because there is a programming language in Turing’s paper, and Turing
A Century of Controversy over the Foundations of Mathematics 89

also gives a rather large program, which of course has bugs, because he
wasn’t able to run it and debug it—it’s the program for a universal Turing
machine. But the basic thing is the ideas, and the new ideas in Turing’s
work are just breathtaking! So I think that Turing went beyond Gödel, but
you have to recognize that Gödel took the first step, and the first step is
historically the most difficult one and takes the most courage. To imagine
that Hilbert could be wrong, which never occurred to von Neumann, that
was something!

8 I Discover Randomness in Pure Mathematics

Okay, so then what happened? Then World War II begins. Turing starts
working on cryptography, von Neumann starts working on how to calcu-
late atom bomb detonations, and people forget about incompleteness for a
while.
This is where I show up on the scene. The generation of mathematicians
who were concerned with these questions basically passes from the scene
with World War II. And I’m a kid in the 1950s in the United States reading
the original article by Nagel and Newman in Scientific American in 1956
that became their book.
And I didn’t realize that mathematicians really preferred to forget about
Gödel and go on working on their favorite problems. I’m fascinated by
incompleteness and I want to understand it. Gödel’s incompleteness result
fascinates me, but I can’t really understand it, I think there’s something
fishy. . . As for Turing’s approach, I think it goes much deeper, but I’m still
not satisfied, I want to understand it better.
And I get a funny idea about randomness. . . I was reading a lot of
discussions of another famous intellectual issue when I was a kid—not the
question of the foundations of mathematics, the question of the foundations
of physics! These were discussions about relativity theory and cosmology
and even more often about quantum mechanics, about what happens in the
atom. It seems that when things are very small the physical world behaves
in a completely crazy way that is totally unlike how objects behave in real
life. In fact things are random—intrinsically unpredictable—in the atom.
Einstein hated this. Einstein said that “God doesn’t play dice!” By the
way, Einstein and Gödel were friends at Princeton.
I was reading about all of this, and I began to wonder—in the back of
my head I began to ask myself—could it be that there was also randomness
in pure mathematics?
90 G. J. Chaitin

The idea in quantum mechanics is that randomness is fundamental, it’s

a basic part of the universe. In normal, everyday life we know that things
are unpredictable, but in theory, in Newtonian physics and even in Ein-
stein’s relativity theory—that’s all called classical as opposed to quantum
physics—in theory in classical physics you can predict the future. The
equations are deterministic, not probabilistic. If you know the initial con-
ditions exactly, with infinite precision, you apply the equations and you
can predict with infinite precision any future time and even in the past,
because the equations work either way, in either direction. The equations
don’t care about the direction of time. . .
This is that wonderful thing sometimes referred to as Laplacian de-
terminism. I think that it’s called that because of Laplace’s Essai
Philosophique sur les Probabilités, a book that was published almost two
centuries ago. At the beginning of this book Laplace explains that by
applying Newton’s laws, in principle a demon could predict the future ar-
bitrarily far, or the past arbitrarily far, if it knew the exact conditions at
the current moment. This is not the type of world where you talk about
free will and moral responsibility, but if you’re doing physics calculations
it’s a great world, because you can calculate everything!
But in the 1920s with quantum mechanics it began to look like God
plays dice in the atom, because the basic equation of quantum mechanics
is the Schrödinger equation, and the Schrödinger equation is an equation
that talks about the probability that an electron will do something. The
basic quantity is a probability and it’s a wave equation saying how a prob-
ability wave interferes with itself. So it’s a completely different kind of
equation, because in Newtonian physics you can calculate the precise tra-
jectory of a particle and know exactly how it’s going to behave. But in
quantum mechanics the fundamental equation is an equation dealing with
probabilities!
You can’t know exactly where an electron is and what its velocity vector
is—exactly what direction and how fast it’s going. It doesn’t have a specific
state that’s known with infinite precision the way it is in classical physics.
If you know very accurately where an electron is, then its velocity—its
momentum—turns out to be wildly uncertain. And if you know exactly
in which direction and at what speed it’s going, then its position becomes
infinitely uncertain. That’s the infamous Heisenberg uncertainty principle,
there’s a trade-off, that seems to be the way the physical universe works. . .
It’s an interesting historical fact that before people used to hate this—
Einstein hated it—but now people think that they can use it! There’s a
A Century of Controversy over the Foundations of Mathematics 91

crazy new field called quantum computing where the idea is that maybe
you can make a brand new technology using something called quantum
parallelism. If a quantum computer is uncertain, maybe you can have it
uncertainly do many computations at the same time! So instead of fighting
it, the idea is to use it, which is a great idea.
But when I was a kid people were still arguing over this. Even though
he had helped to create quantum mechanics, Einstein was still fighting it,
and people were saying, “Poor guy, he’s obviously past his prime!”
I began to think that maybe there’s also randomness in pure mathemat-
ics. I began to suspect that maybe that’s the real reason for incompleteness.
A case in point is elementary number theory, where there are some very
difficult questions. Take a look at the prime numbers. Individual prime
numbers behave in a very unpredictable way, if you’re interested in their de-
tailed structure. It’s true that there are statistical patterns. There’s a thing
called the prime number theorem that predicts fairly accurately the over-all
average distribution of the primes. But as for the detailed distribution of
individual prime numbers, that looks pretty random!
So I began to think about randomness. . . I began to think that maybe
that’s what’s really going on, maybe that’s a deeper reason for all this
incompleteness. So in the 1960s I, and independently some other people,
came up with some new ideas. And I like to call this new set of ideas
algorithmic information theory.
That name makes it sound very impressive, but the basic idea is just
to look at the size of computer programs. You see, it’s just a complexity
measure, it’s just a kind of computational complexity. . .
I think that one of the first places that I heard about the idea of com-
putational complexity was from von Neumann. Turing came up with the
idea of a computer as a mathematical concept—it’s a perfect computer, one
that never makes mistakes, one that has as much time and space as it needs
to work—it’s always finite, but the calculation can go on as long as it has
to. After Turing comes up with this idea, the next logical step for a math-
ematician is to study the time, the work needed to do a calculation—its
complexity. And in fact I think that around 1950 von Neumann suggested
somewhere that there should be a new field which looks at the time com-
plexity of computations, and that’s now a very well-developed field. So of
course if most people are doing that, then I’m going to try something else!
My idea was not to look at the time, even though from a practical
point of view time is very important. My idea was to look at the size of
computer programs, at the amount of information that you have to give a
92 G. J. Chaitin

computer to get it to perform a given task. From a practical point of view,

the amount of information required isn’t as interesting as the running time,
because of course it’s very important for computers to do things as fast as
possible. . . But it turns out that from a conceptual point of view, it’s not
that way at all. I believe that from a fundamental philosophical point of
view, the right question is to look at the size of computer programs, not at
the time. The reason is because program-size complexity connects with a
lot of fundamental stuff in physics.
You see, in physics there’s a notion called entropy, which is how disor-
dered a system is. Entropy played a particularly crucial role in the work of
the famous 19th century physicist Boltzmann, and it comes up in the field
of statistical mechanics and in thermodynamics. Entropy measures how
disordered, how chaotic, a physical system is. A crystal has low entropy,
and a gas at high temperature has high entropy. It’s the amount of chaos
or disorder, and it’s a notion of randomness that physicists like.
And entropy is connected with some fundamental philosophical
questions—it’s connected with the question of the arrow of time, which
is another famous controversy. When Boltzmann invented this wonderful
thing called statistical mechanics—his theory is now considered to be one
of the masterpieces of 19th century physics, and all physics is now statisti-
cal physics—he ended up by committing suicide, because people said that
his theory was obviously wrong! Why was it obviously wrong? Because
in Boltzmann’s theory entropy has got to increase and so there’s an arrow
of time. But if you look at the equations of Newtonian physics, they’re
time reversible. There’s no difference between predicting the future and
predicting the past. If you know at one instant exactly how everything is,
you can go in either direction, the equations don’t care, there’s no direction
of time, backward is the same as forward.
But in everyday life and in Boltzmann statistical mechanics, there is a
difference between going backward and forward. Glasses break, but they
don’t reassemble spontaneously! And in Boltzmann’s theory entropy has
got to increase, the system has to get more and more disordered. But
people said, “You can’t deduce that from Newtonian physics!” Boltzmann
was pretending to. He was looking at a gas. The atoms of a gas bounce
around like billiard balls, it’s a billiard ball model of how a gas works. And
each interaction is reversible. If you run the movie backwards, it looks the
same. If you look at a small portion of a gas for a small amount of time,
you can’t tell whether you’re seeing the movie in the right direction or the
wrong direction.
A Century of Controversy over the Foundations of Mathematics 93

But Boltzmann gas theory says that there is an arrow of time—a sys-
tem will start off in an ordered state and will end up in a very mixed up
disordered state. There’s even a scary expression in German, heat death.
People said that according to Boltzmann’s theory the universe is going to
end up in a horrible ugly state of maximum entropy or heat death! This
was the dire prediction! So there was a lot of controversy about his theory,
and maybe that was one of the reasons that Boltzmann killed himself.
And there is a connection between my ideas and Boltzmann’s, because
looking at the size of computer programs is very similar to this notion of
the degree of disorder of a physical system. A gas takes a large program
to say where all its atoms are, but a crystal doesn’t take as big a program,
because of its regular structure. Entropy and program-size complexity are
closely related. . .
This idea of program-size complexity is also connected with the philos-
ophy of the scientific method. You’ve heard of Occam’s razor, of the idea
that the simplest theory is best? Well, what’s a theory? It’s a computer
program for predicting observations. And the idea that the simplest theory
is best translates into saying that a concise computer program is the best
theory. What if there is no concise theory, what if the most concise pro-
gram or the best theory for reproducing a given set of experimental data
is the same size as the data? Then the theory is no good, it’s cooked up,
and the data is incomprehensible, it’s random. In that case the theory isn’t
doing a useful job. A theory is good to the extent that it compresses the
data into a much smaller set of theoretical assumptions. The greater the
compression, the better!—That’s the idea. . .
So this idea of program size has a lot of philosophical resonances, and
you can define randomness or maximum entropy as something that cannot
be compressed at all. It’s an object with the property that basically the
only way you can describe it to someone is to say “this is it” and show
it to them. Because it has no structure or pattern, there is no concise
description, and the thing has to be understood as “a thing in itself”, it’s
irreducible.
Randomness = Incompressibility

The other extreme is an object that has a very regular pattern so you
can just say that it’s “a million 0s” or “half a million repetitions of 01”,
pairs 01, 01, 01 repeated half a million times. These are very long objects
with a very concise description. Another long object with a concise descrip-
tion is an ephemeris, I think it’s called that, it’s a table giving the positions
of the planets as seen in sky, daily, for a year. You can compress all this
94 G. J. Chaitin

astronomical information into a small FORTRAN program that uses New-

tonian physics to calculate where the planets will be seen in the sky every
night.
But if you look at how a roulette wheel behaves, then there is no pattern,
the series of outcomes cannot be compressed. Because if there were a
pattern, then people could use it to win, and having a casino wouldn’t be
such a good business! The fact that casinos make lots of money shows
that there is no way to predict what a roulette wheel will do, there is no
pattern—the casinos make it their job to ensure that!
So I had this new idea, which was to use program-size complexity to
define randomness. And when you start looking at the size of computer
programs—when you begin to think about this notion of program-size or
information complexity instead of run-time complexity—then the inter-
esting thing that happens is that everywhere you turn you immediately
find incompleteness! You immediately find things that escape the power
of mathematical reasoning, things that escape the power of any computer
program. It turns out that they’re everywhere!
It’s very dramatic! In only three steps we went from Gödel, where
it’s very surprising that there are limits to reasoning, to Turing, where it
looks much more natural, and then when you start looking at program size,
well, incompleteness, the limits of mathematics, it just hits you in the face!
Why?! Well, the very first question that you ask in my theory gets you
into trouble. What’s that? Well, in my theory I measure the complexity of
something by the size of the smallest computer program for calculating it.
But how can I be sure that I have the smallest computer program?
Let’s say that I have a particular calculation, a particular output, that
I’m interested in, and that I have this nice, small computer program that
calculates it, and I think that it’s the smallest possible program, the most
concise one that produces this output. Maybe a few friends of mine and I
were trying to do it, and this was the best program that we came up with;
nobody did any better. But how can you be sure? Well, the answer is that
you can’t be sure. It turns out you can never be sure! You can never be sure
that a computer program is what I like to call elegant, namely that it’s the
most concise one that produces the output that it produces. Never ever!
This escapes the power of mathematical reasoning, amazingly enough.
But for any computational task, once you fix the computer program-
ming language, once you decide on the computer programming language,
and if you have in mind a particular output, there’s got to be at least one
program that is the smallest possible. There may be a tie, there may be
A Century of Controversy over the Foundations of Mathematics 95

several, right?, but there’s got to be at least one that’s smaller than all the
others. But you can never be sure that you’ve found it!
And the precise result, which is one of my favorite incompleteness re-
sults, is that if you have N bits of axioms, you can never prove that a
program is elegant—smallest possible—if the program is more than N bits
long. That’s basically how it works. So any given set of mathematical
axioms, any formal axiomatic system in Hilbert’s style, can only prove that
finitely many programs are elegant, are the most concise possible for their
output.
To be more precise, you get into trouble with an elegant program if it’s
larger than a computerized version of the axioms—It’s really the size of the
proof-checking program for your axioms. In fact, it’s the size of the program
that runs through all possible proofs producing all possible theorems. If you
have in mind a particular programming language, and you need a program
of a certain size to implement a formal axiomatic system, that is to say,
to write the proof-checking algorithm and to write the program that runs
through all possible proofs filtering out all the theorems, if that program
is a certain size in a language, and if you look at programs in that same
language that are larger, then you can never be sure that such a program
is elegant, you can never prove that such a program is elegant using the
axioms that are implemented in the same language by a smaller program.
That’s basically how it works.
So there are an infinity of elegant programs out there. For any compu-
tational task there’s got to be at least one elegant program, and there may
be several, but you can never be sure except in a finite number of cases.
So it turns out that you can’t calculate the program-size complexity,
you can never be sure what the program-size complexity of anything is.
Because to determine the program-size complexity of something is to know
the size of the most concise program that calculates it—but that means—
it’s essentially the same problem—then I would know that this program
is the most concise possible, I would know that it’s an elegant program,
and you can’t do that if the program is larger than the axioms. So if it’s
N bits of axioms, you can never determine the program-size complexity of
anything that has more than N bits of complexity, which means almost
everything, because almost everything has more than N bits of complexity.
Almost everything has more complexity than the axioms that you’re using.
Why do I say that? The reason for using axioms is because they’re
simple and believable. So the sets of axioms that mathematicians normally
use are fairly concise, otherwise no one would believe in them! Which
96 G. J. Chaitin

means that in practice there’s this vast world of mathematical truth out
there, which is an infinite amount of information, but any given set of
axioms only captures a tiny finite amount of this information! And that’s
why we’re in trouble, that’s my bottom line, that’s my final conclusion,
that’s the real dilemma.
So in summary, I have two ways to explain why I think Gödel incom-
pleteness is natural and inevitable rather than mysterious and surprising.
The two ways are—that the idea of randomness in physics, that some things
make no sense, also happens in pure mathematics, is one way to say it. But
a better way to say it, is that mathematical truth is an infinite amount of
information, but any particular set of axioms just has a finite amount of
information, because there are only going to be a finite number of prin-
ciples that you’ve agreed on as the rules of the game. And whenever any
statement, any mathematical assertion, involves more information than the
amount in those axioms, then it’s very natural that it will escape the ability
of those axioms.
So you see, the way that mathematics progresses is you trivialize ev-
erything! The way it progresses is that you take a result that originally
required an immense effort, and you reduce it to a trivial corollary of a
more general theory!
Let me give an example involving Fermat’s “last theorem”, namely the
assertion that
xn + y n = z n
has no solutions in positive integers x, y, z, and n with n greater than 2.
Andrew Wiles’s recent proof of this is hundreds of pages long, but, probably,
a century or two from now there will be a one-page proof! But that one-page
proof will require a whole book inventing a theory with concepts that are
the natural concepts for thinking about Fermat’s last theorem. And when
you work with those concepts it’ll appear immediately obvious—Wiles’s
proof will be a trivial afterthought—because you’ll have imbedded it in the
appropriate theoretical context.
And the same thing is happening with incompleteness.
Gödel’s result, like any very fundamental basic result, starts off by being
very mysterious and complicated, with a long impenetrable proof. People
said about Gödel’s original paper the same thing that they said about
Einstein’s theory of relativity, which is that there are less than five people
on this entire planet who understand it.
So in 1931 Gödel’s proof was like that. If you look at his original paper,
it’s very complicated. The details are programming details we would say
A Century of Controversy over the Foundations of Mathematics 97

now—really it’s a kind of complication that we all know how to handle

now—but at the time it looked very mysterious. This was a 1931 math-
ematics paper, and all of a sudden you’re doing what amounts to LISP
programming, thirty years before LISP was invented! And there weren’t
even any computers then!
But when you get to Turing, he makes Gödel’s result seem much
more natural. And I think that my idea of program-size complexity and
information—really, algorithmic information content—makes Gödel’s result
seem more than natural, it makes it seem, I’d say, obvious, inevitable. But
of course that’s the way it works, that’s how we progress.

9 Where Do We Go from Here?!

I should say, though, that if this were really true, if it were that simple,
then that would be the end of the field of metamathematics. It would be
a sad thing, because it would mean that this whole subject is dead. But I
don’t think that it is!
I’we been giving versions of this talk for many years. In these talks
I like to give examples of things that might escape the power of normal
mathematical reasoning. And my favorite examples were Fermat’s last
theorem, the Riemann hypothesis, and the four-color conjecture. When I
was a kid these were the three most outstanding open questions in all of
mathematics.
But a funny thing happened. First the four-color conjecture was settled
by a computer proof, and recently the proof has been greatly improved.
The latest version has more ideas and less computation, so that’s a big
step forward. And then Wiles settled Fermat’s last theorem. There was a
misstep, but now everyone’s convinced that the new proof is correct.
Fortunately the Riemann hypothesis is still open at this point, as far as
I know!
But I was using Fermat’s last theorem as a possible example of incom-
pleteness, as an example of something that might be beyond the power
of the normal mathematical methods. I needed a good example, because
people used to say to me, “Well, this is all very well and good, Algorithmic
Information Theory is a nice theory, but give me an example of a specific
mathematical result that you think escapes the power of the usual axioms.”
And I would say, well, maybe Fermat’s last theorem!
So there’s a problem. Algorithmic information theory is very nice and
shows that there are lots of things that you can’t prove, but what about
98 G. J. Chaitin

individual mathematical questions? How about a natural mathematical

question? Can these methods be applied? Well, the answer is no, my
methods are not as general as they sound. There are technical limitations.
I can’t analyze Fermat’s last theorem with these methods. Fortunately!
Because if I had announced that my methods show that Fermat’s last the-
orem can’t be settled, then it’s very embarrassing when someone settles
it!
So now the question is, how come in spite of these negative results,
mathematicians are making so much progress? How come mathematics
works so well in spite of incompleteness? You know, I’m not a pessimist,
but my results have the wrong kind of feeling about them, they’re much
too pessimistic!
So I think that a very interesting question now is to look for positive
results. . . There are already too many negative results! If you take them
at face value, it would seem that there’s no way to do mathematics, that
mathematics is impossible. Fortunately for those of us who do mathemat-
ics, that doesn’t seem to be the case. So I think that now we should look
for positive results. . . The fundamental questions, like the questions of phi-
losophy, they’re great, because you never exhaust them. Every generation
takes a few steps forward. . . So I think there’s a lot more interesting work
to be done in this area.
And here’s another very interesting question: Program size is a com-
plexity measure, and we know that it works great in metamathematics,
but does it have anything to do with complexity in the real world? For
example, what about the complexity of biological organisms? What about
a theory of evolution?
Von Neumann talked about a general theory of the evolution of life. He
said that the first step was to define complexity. Well, here’s a definition of
complexity, but it doesn’t seem to be the correct one to use in theoretical
biology. And there is no such thing as theoretical biology, not yet!
As a mathematician, I would love it if somebody would prove a general
result saying that under very general circumstances life has to evolve. But
I don’t know how you define life in a general mathematical setting. We
know it when we see it, right? But as a mathematician I don’t know how
to tell the difference between a beautiful deer running across the road and
the pile of garbage that my neighbor left out in the street! Well, actually
that garbage is connected with life, it’s the debris produced by life. . .
So let’s compare a deer with a rock instead. Well, the rock is harder,
but that doesn’t seem to go to the essential difference that the deer is alive
A Century of Controversy over the Foundations of Mathematics 99

and the rock is a pretty passive object. It’s certainly very easy for us to
tell the difference in practice, but what is the fundamental difference? Can
one grasp that mathematically?
So what von Neumann was asking for was a general mathematical the-
ory. Von Neumann used to like to invent new mathematical theories. He’d
invent one before breakfast every day: the theory of games, the theory of
self-reproducing automata, the Hilbert space formulation of quantum me-
chanics. . . Von Neumann wrote a book on quantum mechanics using Hilbert
spaces—that was done by von Neumann, who had studied under Hilbert,
and who said that this was the right mathematical framework for doing
quantum mechanics.
Von Neumann was always inventing new fields of mathematics, and
since he was a childhood hero of mine, and since he talked about Gödel
and Turing, well, I said to myself, if von Neumann could do it, I think I’ll
give it a try. Von Neumann even suggested that there should be a theory of
the complexity of computations. He never took any steps in that direction,
but I think that you can find someplace where he said that this has got to
be an interesting new area to develop, and he was certainly right.
Von Neumann also said that we ought to have a general mathematical
theory of the evolution of life. . . But we want it to be a very general theory,
we don’t want to get involved in low-level questions like biochemistry or
geology. . . He insisted that we should do things in a more general way,
because von Neumann believed, and I guess I do too, that if Darwin is
right, then it’s probably a very general thing.
For example, there is the idea of genetic programming, that’s a computer
version of this. Instead of writing a program to do something, you sort of
evolve it by trial and error. And it seems to work remarkably well, but can
you prove that this has got to be the case? Or take a look at Tom Ray’s
Tierra. . . Some of these computer models of biology almost seem to work
too well—the problem is that there’s no theoretical understanding why
they work so well. If you run Ray’s model on the computer you get these
parasites and hyperparasites, you get a whole ecology. That’s just terrific,
but as a pure mathematician I’m looking for theoretical understanding, I’m
looking for a general theory that starts by defining what an organism is and
how you measure its complexity, and that proves that organisms have to
evolve and increase in complexity. That’s what I want, wouldn’t that be
nice?
And if you could do that, it might shed some light on how general the
phenomenon of evolution is, and whether there’s likely to be life elsewhere
100 G. J. Chaitin

in the universe. Of course, even if mathematicians never come up with

such a theory, we’ll probably find out by visiting other places and seeing
if there’s life there. . . But anyway, von Neumann had proposed this as an
interesting question, and at one point in my deluded youth I thought that
maybe program-size complexity had something to do with evolution. . . But
I don’t think so anymore, because I was never able to get anywhere with
this idea. . .
So I think that there’s a lot of interesting work to be done! And I
think that we live in exciting times. In fact, sometimes I think that maybe
they’re even a little bit too exciting!. . . And I hope that if this talk were
being given a century from now, in 2099, there would be another century of
exciting controversy about the foundations of mathematics to summarize,
one with different concerns and preoccupations. . . It would be interesting
to hear what that talk would be like a hundred years from now!

Further Reading
[1] G. J. Chaitin, The Unknowable, Springer-Verlag, 1999.
[2] G. J. Chaitin, The Limits of Mathematics, Springer-Verlag, 1998.
 Finite Versus Infinite in Singularity Dynamics
Florin Diacu1
Department of Mathematics and Statistics
University of Victoria, P.O. Box 3045
Victoria, B.C., Canada, V8W 3P4
E-mail: diacu@math.uvic.ca

Abstract. We study the dynamics of particle systems near

solution singularities for different types of potentials. Solution
singularities appear in finite time, but they can be blown up
and removed to infinity with the help of certain transforma-
tions, thus creating the framework of a dynamical system. In
this paper we present several examples of solution singularities
and various ways of removing them by the blow-up technique.
These will put into the evidence the benefits of blow-up transfor-
mations towards understanding the motion near singularities.

1 Introduction
The beginnings of the study of singularities in the dynamics of particles and
systems are difficult to trace in the scattered mathematical and astronom-
ical literature. It is clear, however, that this problem was first considered
in connection with the Newtonian n-body problem of celestial mechanics,
characterized by Whittaker [16] as “the most celebrated of all dynamical
problems.” The first significant study appeared in the last chapter of [15],
a treatise based on Painlevé’s lectures delivered in 1895 at the University
of Stockholm upon the invitation of King Oscar II of Sweden and Norway.
Details on the origin, content, and history of these lectures can be found
in [5] and [9].
The understanding of the qualitative behavior of solutions near singu-
larities is an important but difficult task. It is important because motion
near singularities is unpredictable and it can drastically change the “fate”
1
Supported in part by Grant OGP0122045 of the NSERC of Canada.

101
102 F. Diacu

not only of an orbit but of the whole phase-space picture and consequently
of the global dynamical behavior. It is difficult because most methods fail
to provide significant results. Exact techniques are powerless and numerical
ones prove far from reliable for anything but short time intervals. A quali-
tative method that offers some insight into the matter is that of blow up,
introduced in celestial mechanics in the context of regularization techniques
by the Italian mathematician Tullio Levi-Civita (see [12] and [13]). In 1974
Richard McGehee of the University of Minnesota (see [14]) improved this
method and applied it to understand the triple collision in the rectilinear
3-body problem. The advancement of the field would have been hard to
imagine without McGehee’s seminal contribution. The blow-up techniques
we present below are based on McGehee’s idea.
The goal of this paper is to discuss the qualitative blow-up technique
for solution singularities and see how it applies to different situations. We
will describe it, emphasize its advantages for understanding the behavior
of the solutions near a singularity, and make the connection between the
finite and the infinite aspects of it.
In Section 2 we define the notion of solution singularity. We will show
that a singularity is either due to a physical collision or to a so-called
pseudocollision, which may occur if the motion reaches no limit position
while becoming unbounded in finite time.
In Section 3 we will discuss collisions for the Manev potential, which
is an approximation of general relativity in the framework of classical me-
chanics, as proved in [10]. More precisely, the Manev gravitational law
is the analogue of the Schwarzschild solution to Einstein’s field equations,
assuming that the study is restricted to almost circular orbits, for example
to planetary motion.
In Section 4 we consider the anisotropic case of the Manev problem,
proposed by the present author in 1995. Unlike the classical case, the
anisotropic Manev problem is not integrable. This problem stays at the
intersection between classical mechanics, quantum mechanics, and general
relativity, and can be also seen as an approximation of special relativity. It
therefore presents interest for understanding the possible connections and
for lying bridges between these fields.
In Section 5 we discuss the blow-up technique for 2-body problems with
drag or thrust. Such problems appear in astronomy and astrophysics when
studying the motion of microcosmic particles near the gravitational field of
a celestial body. The drag or the thrust can be due to the friction with the
atmosphere, to magnetic interactions, to radiation pressure, to the solar
Finite Versus Infinite in Singularity Dynamics 103

wind, or to other nongravitational forces that act in the neighborhood of

planets or stars.
In Section 6 we consider the Mücket-Treder gravitational law, which
unlike the previous models is given by a potential that is neither homo-
geneous nor quasihomogeneous. This model was proposed in 1977 by the
German astronomers J.P. Mücket and H.-J. Treder (see [7]). We will show
that the blow-up technique can be also adapted to such situations and we
will apply to the 3-body problem.
In Section 7 we draw conclusions regarding the importance of blow-up
techniques in particle-systems dynamics. We mention the interplay between
finite and infinite for understanding the qualitative behavior of motion near
singularities and the constructive and generative role of this dichotomy in
mathematics. This paper illustrates one of the various aspects of this role.

2 Solution Singularities
In this paper we will study systems of the form
!
q̇ = M−1 p
(1)
ṗ = ∇U (q),

where q = (q1 , q2 , . . . , qn ) is the configuration of an n-particle sys-

tems, qi = (qi1 , qi2 , qi3 ), i = 1, 2, . . . n, are the position vectors,
p = Mq̇ is the momentum, M is the 3n-dimensional matrix of the
masses, having 0 everywhere except on the main diagonal which is
m1 , m1 , m1 , m2 , m2 , m2 , . . . , mn , mn , mn , where mi > 0 for each i =
1, 2, . . . , n. U is a real function of q called potential function, which signi-
fies the negative of the potential energy, ∇ is the gradient operator, and
the upper dot denotes differentiation with respect to the independent time-
variable t. From the physical point of view such a system describes the
motion of n point masses under a force law defined by the potential func-
tion U . For example, the potential
 mi mj
U (q) = G , (2)
1≤i<j≤n
|qi − qj |

where G is the gravitational constant, defines the Newtonian n-body prob-

lem; for
 mi mj 3G2  mi mj (mi + mj )
U (q) = G + 2 , (3)
1≤i<j≤n
|qi − qj | 2c 1≤i<j≤n
|qi − qj |2
104 F. Diacu

equations (1) are those of the Manev n-body problem of celestial mechanics
(see [2], [4], [6], or [10]). The Lennard-Jones potential (see e.g. [11])
 1  1
U (q) = − + (4)
2≤i<j≤n
|qi − qj |6
1≤i<j≤n
|qi − qj |12

and the equations (1) with m1 = m2 = . . . = mn = 1 are used in physical

chemistry to explain crystal formation. The Liboff potential
 1  1
U (q) = + (5)
1≤i<j≤n
|qi − qj |3
1≤i<j≤n
|qi − qj |4

appears in electromagnetics.
Notice that all the above potentials are quasihomogeneous, i.e., sums
of homogeneous functions. (In fact the potential (2) is homogeneous
but (3), (4), and (5) are not.) For U quasihomogeneous, standard results of
the theory of differential equations ensure the existence and the uniqueness
of an analytic solution for the initial value problem given by system (1)
with initial conditions (q, p)(0) ∈ (IR3n \ ∆) × IR3n , where

∆= {q ∈ IR3n |qi = qj } (6)
1≤i<j≤n

is the collision set. This solution can be extended analytically to its maxi-
mal domain [0, t∗ ). If t∗ = ∞, the solution is called regular. If t∗ < ∞, the
solution is called singular.
In his 1895-Stockholm lectures, Painlevé proved a result concerning
singularities for the Newtonian potential (see [3]). Its extension to quasi-
homogeneous potentials is straightforward. If (q, p) is an analytic solution
of system (1) given by a quasihomogeneous potential U and if this solution
is defined on a maximal interval [0, t∗ ), then t∗ is a singularity if an only
if limt→t∗ min(q(t)) = 0, where min(q(t)) is the minimum of the mutual
distances between particles at the given time t.
This result allows us to differentiate between two types of solution sin-
gularities, those that have a limit and those that don’t. We can formulate
this as follows. A singularity t∗ is due to a collision if q has a definite limit
when t → t∗ . The time instant t∗ is called a collision singularity. If q
has no limit or becomes unbounded when t → t∗ , we call t∗ a noncollision
singularity or a pseudocollision.
This definition has the following interpretation. If a solution reaches the
set ∆, the denominator of ∇U (q) cancels and system (1) loses its meaning.
Finite Versus Infinite in Singularity Dynamics 105

The set ∆, however, can be approached either with asymptotic phase (i.e.,
the solution tends to a certain element of ∆) or without asymptotic phase
(i.e., the solution tends to ∆ but oscillates among its elements without
reaching any of them). In the former case we have a collision singularity
and in the latter a noncollision singularity.
The dynamics near ∆ is very different for each of the above potentials.
For the Newtonian one, the set of initial conditions leading to collisions
has measure zero and is of the first Baire category. For n = 4 the set of
initial conditions leading to pseudocollisions has measure zero and is also of
the first Baire category. But the size of this set is unknown in the general
case. For the Manev potential the set of initial data leading to collisions
has positive measure. The same is true for the Liboff potential. In case of
the Lennard-Jones potential this set is empty because when particles get
close, the repelling force overcomes the attractive one, so collisions cannot
occur. For more details regarding this type of results see [3].
In his Stockholm lectures Painlevé proved that pseudocollisions do not
exist in the Newtonian 2- and 3-body problems and conjectured that they
can appear in the Newtonian n-body problem for n ≥ 4. This problem is
notoriously difficult and it took a lot of energy and many attempts to answer
it. In 1992, almost a century after the conjecture was stated, Zihong Xia
from Northwestern University produced in his Ph.D. thesis the first example
of a pseudocollision in the spatial 5-body problem (see [17] and [5]). Shortly
after, Joe Gerver from Rutgers University came up with the first example
of a pseudocollision for particles moving in a plane (see [3]) for an n-body
problem with n large and which exhibits certain symmetries. In principle,
Xia’s example can be generalized to any n ≥ 5 but not to n = 4. So
the existence of pseudocollisions in the 4-body problem is still an open
problem. It is important to mention that Xia’s example makes extensive
use of blow-up techniques, which we will discuss below, first in the context
of the Manev problem.

3 The Manev Potential

Let us now consider system (1) with potential (3) for n = 2, called the
Manev problem after the Bulgarian physicist G. Manev, who proposed it
in the 1920s as a classical alternative to general relativity. It is interesting
that although Manev’s physical derivation of this law was initially based
on shaky physical grounds (like ether, for example), his law explains the
perihelion advance of the inner planets with the same accuracy as Ein-
106 F. Diacu

stein’s theory. In fact, as we proved in [10], there is a strong similarity

between the Schwarzschild solution of Einstein’s field equations and the
Manev potential.
The Manev problem can be written in Hamiltonian form with the help
of the Hamiltonian function
1 1 k
H(p, q) = (|p1 |2 + |p2 |2 ) − − , (7)
2 |q1 − q2 | |q1 − q2 |2
2
where k = 3G2c2
(see (3)). In this setting, the McGehee transformations for
removing the singularity of the equations occurring when q1 = q2 , reduce
to first write the Hamiltonian in polar coordinates r > 0, θ ∈ S 1 , where S 1
is the segment [0, 2π] with the end points identified. Thus the Hamiltonian
takes the form

H(pr , pθ , r) = (1/2)(p2r + p2θ /r2 ) − 1/r − k/r2 , (8)

where pr , pθ are the new polar variables of the momenta.

The next step consists of blowing up the collision singularity that now
occurs at r = 0. For this we can formally multiply the energy integral by
r2 , which takes the form

(1/2)r2 p2r + p2θ − r − k = hr2 . (9)

Further introducing the transformations v = rpr and u = pθ and scaling

the time variable by using dt = r2 dτ , we complete the sequence of blow-up
transformations. The equations given by the Hamiltonian (8) become




r = rv


v = r(1 + 2hr)
(10)

 θ
= u


u = 0,

where the prime denotes differentiation with respect to the fictitious time
variable τ . The energy relation (9) gets transformed into

v 2 + u2 − 2r − 2hr2 = 2k. (11)

We define the collision manifold as the set of solutions given by rela-

tion (11) when r = 0. Notice that, geometrically, the collision manifold
is a cylinder in the 3-dimensional space of coordinates (u, θ, v), and since
θ ∈ [0, 2π], it follows that this cylinder can be identified with a torus. In
Finite Versus Infinite in Singularity Dynamics 107

fact, the 2-dimensional torus representing the collision manifold is embed-

ded in the 4-dimensional phase space of the coordinates (r, u, θ, v). The
equations (10) show that the flow on the collision manifold is formed al-
most exclusively by periodic orbits, except
√ the upper and lower circles of
the torus given by r = 0, u = 0, v = ± 2k, which consist of equilibrium
points (see Figure 1a).

Figure 1: (a) The collision manifold embedded in the 4-

dimensional phase space, (b) The collision manifold in the re-
duced phase space

Figure 2: The flow in the reduced phase space for negative

energy

Since θ does not appear explicitly in the equations (10) or in the energy
relation (11), we can further reduce the 4-dimensional phase space to a
108 F. Diacu

3-dimensional one by factorizing the flow to S 1 . Exploiting this symmetry

we can obtain clear pictures of the global flow in phase space. In fact the
global description can be reduced to different energy levels by regarding
the energy constant h as a parameter. Factorizing the collision manifold
to S 1 , the torus becomes a circle (see Figure 1b). The points M and N
on this circle correspond to the circles of equilibria on the torus, while all
the other points correspond to the periodic orbits on the torus. In case of
negative energy h < 0, for example, the flow in the reduced phase space
looks as in Figure 2.
Through the above blow-up technique we have removed the collision sin-
gularity by shifting it to infinity in time, so we can now study the properties
of solutions passing close to this type of collision. The complete qualitative
analysis of these solutions for the Manev potential in the particular case
n = 2 is done in [2] and generalized in [10].

4 The Anisotropic Manev Potential

The Manev problem is integrable but its relative, the anisotropic Manev
problem, is not. Its flow is also far from fully understood. The first prob-
lem of this type, the anisotropic Kepler problem, was proposed by Martin
Gutzwiller in the late 1960s in an attempt to find a bridge between classi-
cal and quantum mechanics (for references see [1]). The anisotropic Manev
problem, suggested by the present author in 1995, attempts even more. It
is a bridge between classical mechanics, quantum mechanics, and general
relativity. As we have shown in [1], this problem exhibits properties from
all three fields.
The anisotropic Manev problem is given by system (1) with potential

1 b
U (q) = ) + ,
µq12 + q22 µq12 + q22

where q = (q1 , q2 ), and µ > 0 and b > 0 are constants. These equations
define the motion of two particles of unit mass in an anisotropic space, i.e.,
one in which the attraction forces act differently in every direction. The
above potential defines the anisotropy of the space as a function of the
parameter µ. If µ < 1, the attraction is the weakest in the direction of the
q1 -axis and the strongest in that of the q2 -axis, the situation being reversed
if µ > 1. If µ = 1, the space is isotropic and we are in the case of the
classical Manev problem discussed above. Since both remaining cases have
Finite Versus Infinite in Singularity Dynamics 109

a weakest-force and a strongest-force direction, we can assume, without loss

of generality, that µ > 1.
The Hamiltonian function of the system is given by
1
H(p(t), q(t)) = |p(t)|2 − U (q(t)),
2
the sum of kinetic and potential energies, which yields the energy integral

H(p(t), q(t)) = h.

However, since the force ∇U is not central, the angular momentum L(t) =
p(t) × q(t) is not an integral of the system, as it is in the classical Manev
problem.
To perform the McGehee transformations of the second kind we proceed
as follows. We first change the dependent variables using the formulas

 r = |q|

 θ = arctan(q /q )
2 1

 y = ṙ = (q 1 1 + q2 p2 )/|q|
p

x = rθ̇ = (q1 p2 − q2 p1 )/|q|,
and !
v = ry
u = rx,
and then change the independent variable using

dτ = r−2 dt.

Composing these transformations, which are analytic diffeomorphisms in

their respective domains, the energy relation becomes

u2 + v 2 − 2rδ −1/2 − 2bδ −1 = 2r2 h,

and the equations of motion take the form



 r = rv



v = 2r2 h + rδ −1/2

 θ
= u


u = (1/2)(µ − 1)(rδ −3/2 + 2bδ −2 ) sin 2θ,

where δ = µ cos2 θ + sin2 θ. The new variables (r, v, θ, u) ∈ (0, ∞) × IR ×

S 1 × IR depend on the fictitious time τ , so the prime denotes here differ-
entiation with respect to the new independent variable τ . Note that the
110 F. Diacu

new equations extend analytically to r = 0, an invariant manifold which

physically corresponds to binary collisions.
Notice that the sets {(r, v, θ, u) | r = 0} and {(r, v, θ, u) | r > 0} are
invariant manifolds for the new system. The set
C = {(r, v, θ, u) | r = 0 and the energy relation holds}
is the collision manifold. It replaces the set of singularities {(q, p) | q = 0}
of the original system with a 2-dimensional manifold in the space of the
new variables. This 2-dimensional manifold is embedded in IR3 × S 1 and is
given by the equations
r=0 and u2 + v 2 = 2bδ −1 .
This shows that C is homeomorphic to a torus (see Figure 3).

Figure 3: The flow on the collision manifold in the anisotropic

Manev problem

A qualitative analysis of the flow near the collision manifold and the
physical interpretation of the results are given in [1]. As in the case of
the Manev problem, we have transformed a finite-in-time phenomenon to
an infinite-in-time one in order to study the behavior of solutions near the
singularity.

5 Potentials with Drag or Thrust

The blow-up technique can be used in a larger context. For example, it
also applies to systems in which additional forces appear. The 2-body
Finite Versus Infinite in Singularity Dynamics 111

problem with drag or thrust is such an example, rooted in the astronomical

literature of the beginning of the 20th century. It models the motion of
2 point masses, one of negligible mass, subject to Newtonian gravitation
and a drag force that involves the velocity (due to atmospheric friction,
radiation pressure, solar wind, magnetic interaction, etc.).
Assume that the motion of two point masses (a particle and a primary)
is subject to a perturbation force of the type F = −r−2 (αvr + βvt ), where r
is the distance between bodies, vr and vt are the radial and the tangential
velocity, respectively, and α, β are real, nonzero constants. With respect
to a frame whose origin is in the primary, the equations of motion in polar
coordinates (r, θ) are given by the system
#
r̈ − rθ̇2 = −µr−2 − αṙr−2
(12)
dt (r θ̇) = −β θ̇,
d 2

where µ combines several physical constants (mass, gravitational constant,

charge, etc.). Unlike in the absence of drag or thrust, the total energy and
the angular momentum are not conserved quantities.
All possible choices of signs for α, β, and µ present practical interest,
from the one modeling the motion of small satellites near the Earth to the
one involving microcosmic particles under radiation pressure. Therefore we
will further discuss the qualitative aspects of each case. Notice that µ > 0
corresponds to an attractive force, µ < 0 represents a repelling one, while
α > 0 implies the presence of a radial drag and α < 0 that of a radial
thrust.
Using the substitutions ṙ = u and θ̇ = ϕ, we transform equations (12)
into the first-order system

 ṙ = u

 θ̇ = ϕ

 u̇ = rϕ2 − µr−2 − αur−2
 −2 −1
ϕ̇ = −βϕr − 2uϕr .

Since the vector field of this system is independent of θ, we can drop the
second equation and thus obtain the simpler form

 ṙ = u
u̇ = rϕ2 − µr−2 − αur−2

ϕ̇ = −βϕr−2 − 2uϕr−1 ,

which under the time-rescaling analytic diffeomorphism dτ = r−2 dt is con-

112 F. Diacu

verted into the system


2
r = r u
u
= −αu + r3 ϕ2 − µ (13)


ϕ = −βϕ − 2uϕr.

Equations (13) are defined for (r, u, ϕ) ∈ [0, ∞) × IR2 and

represents the
derivative with respect to the new fictitious time-variable τ . Notice that
though equations (12) were undefined at r = 0, system (13) extends to this
set, i.e., the plane r = 0 is pasted to the 3-dimensional phase space of the
variables (r, u, ϕ). The collision plane r = 0 is thus an invariant manifold
for the equations (13). (Recall that an invariant manifold is a union of
orbits, i.e.‘ initial data in the set constrain the whole orbit to the set.)
Since system (13) has only one equilibrium solution at (0, − αµ , 0), we
will shift the origin of the frame into it with the help of the transformation
v = u + αµ . Thus system (13) becomes


µ
 r = r (v − α )
2

v
= −αv + r3 ϕ2 (14)

 ϕ
= −βϕ − 2(v − µ )rϕ,
α

and will be the main object of our further qualitative investigations. Sys-
tem (14) has the unique equilibrium solution (r, v, ϕ) = (0, 0, 0). Due to
the embedding of the collision manifold into the phase space, system (14)
is well suited for understanding the behavior of orbits near collision.
We will also study an alternative system, suitable for understanding the
behavior of orbits at infinity. To obtain it, we transform system (14) via
the change of variable
1
x= ,
r
which is an analytic diffeomorphism that brings infinity into the phase
space. The new equations are

µ
 x = −(v − α )
v
= −αv + x−3 ϕ2 (15)


ϕ = −βϕ − 2(v − αµ )x−1 ϕ.

The time-rescaling transformation

ds = x−3 dτ,
Finite Versus Infinite in Singularity Dynamics 113

which is an analytic diffeomorphism, changes system (15) into the regular-

ized system 
 µ
 x = −x (v − α )
, 3

v , = x3 − αx3 v + ϕ2 (16)

 ϕ, = −βx3 ϕ − 2(v − µ )x2 ϕ,
α

where , denotes differentiation with respect to the new fictitious time vari-
able s. The invariant plane x = 0, called the infinity manifold, represents
the physical points at infinity. In other words, a solution tending to this
plane reaches infinity in physical space. On the other hand a solution of sys-
tem (16) that tends to infinity in phase space reaches a collision in physical
space. Equations (16) form the alternative system to be studied.
Unlike in previous cases, we can now use the blow-up technique not
only for the study of motion near the singularity but also at infinity. We
pasted an infinity manifold to the phase space, which we can use to draw
conclusions about the qualitative behavior of solutions at infinity. The
analysis of the above aspects of the motion is done in [8].

6 Mücket-Treder Potentials
All the above cases involve quasihomogeneous potentials. Blow-up tech-
niques, however, go beyond the class of quasihomogeneous functions. In
1977 the German astronomers J. P. Mücket and H.-J. Treder proposed a
gravitational law, which for the 3-body problem is given by a potential of
the form
U = V + W, (17)
where the functions V and W are
 m i mj
V (q) = G
1≤i<j≤3
|qi − qj |

and
 αmi mj (1 + ln |qi − qj |)
W (q) = .
1≤i<j≤3
|qi − qj |

Here G is the gravitational constant and α is a small negative constant.

If α > 0, singularities do not occur and the equations (1) are globally de-
fined (see [7]). In our case, α < 0, noncollision singularities are excluded
but collisions can take place. In what follows we will blow up the singu-
larity that appears due to triple collisions. For this we first consider the
114 F. Diacu

transformation  1

 r = (qT M q) 2


s = r−1 q (18)

 y = pT s


x = s − yM s.
Notice that the following relations take place:

sT M s = 1, sT x = 0
V (q) = r−1 V (s), ∇V (q) = r−2 ∇V (s)
W (q = r−1 W s) + αr−1 (ln r)V (s)
∇W (q) = r−2 ∇W (s) + αr−2 (ln r)∇V (s).

Since V is a homogeneous function of degree −1, Euler’s relation yields

qT ∇V (q) = −V (q).

Using the transformation (18), equations (1) given by the potential (17)
take the form

 ṙ = y

 −1 T −1 −2 T −2
 ẏ = r x M x + r s ∇W (s) − (1 + α ln r)r V (s)

ṡ = r−1 M −1 x (19)





ẋ = −r−1 yx − r−1 (xM −1 x + ∇V (s)] + (1 + α ln r)−2 [V (s)M s+
−2r −2 T
∇W (s)] − r [s ∇W (s)]M s.
For r positive and small, which implies that the 3 particles are close to-
gether, we further consider the change of variables
#
v = r 2 (− ln r)− 2 y
1 1

(20)
u = r 2 (− ln r)− 2 x,
1 1

which transforms equations (19) into


ṙ = r− 2 (− ln r) 2 v
1 1





 v̇ = r− 2 (− ln r) 2 [ v2 (1 − ln1r ) + uT M −1 u − s ∇W
3 1 2 T (s)

 ln r +

 1

 (α + ln r )V (s)]

ṡ = r− 2 (− ln r) 2 M−1 u
3 1
(21)

 u̇ = −r− 2 (− ln r) 2 (1 + ln1r ) uv
3 1
 T −1 uM s+

 2 + (u M



 (α + ln1r )[V (s)M s + ∇V (s)] + ln1r [∇W (s)+

 

 (s ∇W (s))M s] .
T
Finite Versus Infinite in Singularity Dynamics 115

To eliminate the negative powers of r and − ln r in the above equations, we

consider the time rescaling transformation

dτ = r− 2 (− ln r) 2 dt,
3 1

which introduces the fictitious time variable τ . With this transformation

system (21) becomes



 r = rv

v
= v2 (1 − ln1r ) + uT M −1 u − s ∇W
T


2 (s)
+ (α + ln1r )V (s)


ln r
s = M u −1
(22)

 u
= − (1 + ln1r ) uv T −1 uM s + (α + 1 )[V (s)M s+

 2 + (u M ln
r


 ∇V (s)] + 1
[∇W (s) + (s ∇W (s))M s] ,
T
ln r

where the prime denotes differentiation with respect to the independent

variable τ . With the above transformations the energy relation takes the
form
1 V (s) + W (s) hr
(uM −1 u + v 2 ) + + αV (s) = − .
2 ln r ln r
With the transformation
1
ρ=−
ln r
system (22) becomes



 ρ = ρ2 v

 sT ∇W (s)

 v
= (1 + ρ) v2 + uT M −1 u + ρ ln r
2
+ (α − ρ)V (s)

s =M u −1 (23)



 u
= (ρ − 1) uv + uT M −1 uM s − (ρ − α)[V (s)M s + ∇V (s)]+

 2
ρ[∇W (s) + (s ∇W (s))M s],
T

and the energy relation changes to

1 −1
(uM −1 u + v 2 ) − ρ[V (s) + W (s] + αV (s) = hρe ρ .
2
Since the triple collision takes place at ρ = 0, system (23) is now defined
to include this singularity. However, notice that the energy relation is still
undefined at ρ = 0. To avoid this difficulty, we extend the energy relation
to
1
(uM −1 u + v 2 ) − ρ[V (s) + W (s] + αV (s) = hg(ρ),
2
where #
− ρ1
g(ρ) = ρe , if ρ = 0
0, if ρ = 0.
116 F. Diacu

The function g is analytic for ρ = 0 and of class C ∞ at ρ = 0. Notice

that this is the only way to extend g to 0 in the class C ∞ . Now both the
equations of motion and the energy relation make sense at ρ = 0. Therefore
we can define the triple collision set to be

1
C = {(ρ, y, s, u)|ρ = 0 and (uM −1 u + v 2 ) + αV (s) = 0},
2
which is a C ∞ -manifold. The flow of this manifold provides information on
the dynamics near triple collision. The study of this flow is done in [7].

7 Conclusions
As we have seen above, the blow-up techniques connect finite and infinite
time phenomena in the dynamics of particles. Some finite-time phenom-
ena are better understood through infinite-time techniques, whereas some
asymptotic properties at infinity can be better grasped with the help of
finite-space investigations. This can be done for various classes of poten-
tials. We can thus see that the strong finite-versus-infinite relationship is
not restricted to discrete mathematics. It leaves its imprint in all branches
of the field, including the qualitative theory of differential equations and
dynamical systems, and it can be successfully applied to many practical
problems.

References
[1] S. Craig, F. Diacu, E.A. Lacomba, E. Pérez, The anisotropic Manev
problem, Journ. Math. Phys., 40, 3 (1999), 1359–1375.

[2] J. Delgado, F. Diacu, E.A. Lacomba, A. Mingarelli, V. Mioc, E. Pérez,

C. Stoica, The global flow of the Manev problem, Journ. Math. Phys.,
37, 6 (1996), 2748–2761.

[3] F. Diacu, Singularities of the N -Body Problem–An Introduction to Ce-

lestial Mechanics, Les Publications CRM, Montréal, 1992.

[4] F. Diacu, The planar isosceles problem for Maneff’s gravitational law,
Journ. Math. Phys., 34, 12 (1993), 5671–5690.

[5] F. Diacu, Painlevé’s conjecture, Math. Intelligencer, 15 (1993), 6–12.

Finite Versus Infinite in Singularity Dynamics 117

[6] F. Diacu, Near-collision dynamics for particle systems with quasihomo-

geneous potentials, Journ. Differential Equations, 128 (1996), 58–77.

[7] F. Diacu, On the Mücket-Treder gravitational law, in New Trends

for Hamiltonian Systems and Celestial Mechanics (E. A. Lacomba,
J. Llibre, eds.), World Scientific, London, 1996, 127–139.

[8] F. Diacu, Two-body problems with drag or thrust: qualitative results,

Celestial Mechanics (to appear).

[9] F. Diacu, P. Holmes, Celestial Encounters–The Origins of Chaos and

Stability, Princeton Univ. Press, Princeton, N.J., 1996.

[10] F. Diacu, V. Mioc, C. Stoica, Phase-space structure and regularization

of Manev-type problems, Nonlinear Analysis (to appear).

[11] M. R. Hoare, P. Pal, Statics and stability of small cluster nuclei, Nature
(Physical Science), 230 (1971), 5–8.

[12] T. Levi-Civita, Sulla regolarizzazione del problema piano dei tre corpi,
Rendiconti dell’Accademia Nazionale dei Lincei, 24, 5 (1915), 61–75.

[13] T. Levi-Civita, Sur la régularisation du problème des trois corps, Acta

Mathematica, 42 (1918), 92–144.

[14] R. McGehee, Triple collision in the collinear three-body problem, In-

vent. Math., 27 (1974), 191–227.

[15] P. Painlevé, Leçons sur la théorie analytique des équations differen-

tielles, Hermann, Paris, 1897.

[16] E.T. Whittaker, A Treatise on the Analytical Dynamics of Particles

and Rigid Bodies, 4th edition, Cambridge Univ. Press, 1937.

[17] Z. Xia, The existence of noncollision singularities in the N -body prob-

lem, Annals of Math., 135 (1992), 411–468.
 An Interesting Serendipitous Real Number
John Ewing, Ciprian Foias
Mathematics Department
Indiana University
Bloomington, IN 47405, USA
E-mail: foias@indiana.edu, ewing@ams.org

1 Introduction
This is the story of a remarkable real number, the discovery of which was
due to a misprint. Namely, in the midseventies, while Ciprian was at the
University of Bucharest, one of his former students1 approached him with
the following question:
(Q) If x1 > 0 and xn+1 = (1 + 1 n
xn ) (n = 1, 2, . . .), can xn → ∞?
This was listed in a fall issue of the “Gazeta Matematică” as one of
the problems given at the previous summer admission examination for
prospective freshmen in the Department of Mathematics at the University
of Bucharest. Ciprian found the answer in about one day, but considered
that the problem was even above the sophomore level. He also found that
(Q) is a misprinted version of the following question (given by Professor N.
Boboc):
(Q
) If x1 > 0 and xn+1 = (1 + 1 xn
xn ) (n = 1, 2, . . .), can xn → ∞?
This was an appropriate exam question since the answer is clearly “No.”
Years later, in the 1980s, Ciprian told the story to Professor P. Halmos,
who in turn told the story to John, but mischieveously did not mention
at all Ciprian’s answer to (Q), so a day later John also found the answer.
This answer is given by the following
1
If by chance that gentleman reads this article, we would appreciate if he will let us
know his name.

119
120 J. Ewing, C. Foias

Theorem 1.1 There exists exactly one real number2 a ∼ 1.187 such that
if x1 = a then xn → ∞. Moreover in this case
ln n
xn → 1 for n → ∞. (1)
n
Relation (1) can be rewritten as
xn
lim =1 (2)
n→∞ π(n)

where π(n) is the number of primes less than n. However after many at-
tempts to establish a deeper connection with the Prime Number Theorem,
we came to believe that relation (2) is fortuitous. A strong argument for
this opinion is provided by Theorem 3.1 below in which we show that the
estimate for the error
ln n
xn −1
n
differs from its analog in the Prime Number Theorem.
We dedicate this short note to Professor Solomon Marcus, for his 75th
birthday. Among many things Ciprian learned from him, was also the
interest in the anecdotical aspects of mathematics which explains the little
story above.

2 Proof of Theorem 1.1

To start we recall that on (0, ∞), (1 + 1/β)β increases from 1 to e and
(1 + 1/β)β+1 decreases from ∞ to e. It follows

xn+1 < en/xn and xn+1 > en/(xn +1) , (3)

for all n ≥ 1.

Lemma 2.1 Let n satisfy

ln(n + 3) 1
≤ (4)
n+2 1+e
and
n
xn ≤ − 1. (5)
ln(n + 1)
Then
xn+2k → 1, xn+2k+1 → 0 for k → ∞. (6)
2
As computed by Professor Clay C. Ross a = 1.187452351126501054595480158396 . . ..
An Interesting Real Number 121

Proof. We have, using (3) with n replaced by n + 1,

 n+1
1 n+1
xn+2 = 1 + < e xn+1
xn+1
whence, due to (3) and (5),
n+1
< (n + 1)e−n/(xn +1) < 1.
xn+1
It follows xn+2 < e. Relation (4) now yields
n+2
xn+2 ≤ − 1,
ln(n + 3)
and therefore xn+4 ≤ 2. Now note that if (4) holds, so does
1 ln(n + k + 1)
≥ for all k ≥ 2.
1+e n+k
Thus xn+2k < e for all k = 1, 2, . . . This in turn implies

xn+2k+1 > en/(xn+2k +1) > en/(e+1) → ∞ for k → ∞

and consequently

xn+2k+2 < en/xn+2k+1 → 1 for k → ∞.

Lemma 2.2 If n satisfies (4) and

n n+1
xn ≥ , where yn = − 1, (7)
ln yn ln(n + 2)
then
xn+2k → ∞, xn+2k+1 → 1 f or k → ∞. (8)

Proof. If (4) holds, so does

ln(n + 4) 1
≤ . (9)
n+3 1+e
Moreover
n n+1
xn+1 < e xn ≤ −1
ln(n + 2)
122 J. Ewing, C. Foias

so that we apply Lemma 2.1 with n + 1 replacing n, and obtain (8). ✷

For convenience, we denote

 n
1
fn (x) = 1 + , x ∈ (0, ∞), n = 1, 2, . . . (10)
x
Thus if x1 = x, the sequence (xn )∞
n=1 is given by

xn+1 = (fn ◦ fn−1 ◦ · · · ◦ f1 )(x) for n = 1, 2, . . . (11)

Clearly fn is decreasing, f ((0, ∞)) = (1, ∞) and fn ((1, ∞)) = (1, 2n ), n =
2, 3, . . .. For n = 1, 2, . . . we define the disjoint subsets An and Bn of (0, ∞)
by
An = {x : xn+m+1 = (fm+n ◦ · · · ◦ fn )(x)(m = 1, 2, . . .) satisfies (6)}.
(12)
Bn = {x : xn+m+1 = (fm+n ◦ · · · ◦ fn )(x)(m = 1, 2, . . .) satisfies (8)}.
(13)
Lemma 2.3 Let x ∈ An and y ∈ Bn . Then
fn (x) ∈ Bn+1 and (0, x] ⊂ An , (14)
fn (y) ∈ An+1 and [y, ∞) ⊂ Bn , (15)
and
x < y. (16)
Proof. The first relations in (14), (15) are direct consequences of the def-
initions (12), (13); these together with the monotonicity property of the
functions fn also directly imply the second properties in (14), (15). These
properties then yield (16). ✷

As a corollary, if we set
an = sup An , bn = inf Bn (n = 1, 2, . . .), (17)
then
fn (an ) = bn+1 , fn (bn ) = an+1 and an ≤ bn . (18)
Moreover the proofs of Lemmas 2.1 and 2.2 also show that if n satisfies (4)
then
n n
− 1 ≤ an , bn ≤ . (19)
ln(n + 1) ln yn
Thus
ln(n + 1)
0 ≤ εn = (bn − an ) → 0 for n → ∞. (20)
n
An Interesting Real Number 123

Lemma 2.4 For all n = 1, 2, . . .

min{|fn
(x)| : an ≤ x ≤ bn } = |fn
(bn )|. (21)

Moreover
|fn
(bn )| → ∞ for n → ∞. (22)

Proof. Relation (21) follows from

* n−1 +
d
d 1 1
|fn (x)| = n 1+ < 0.
dx dx x x2

For (22), we observe first that for n large enough

*  n−1 + * +

1 1 (ln yn )2 bn−1
ln |f (bn )| = ln n 1 + ≥ ln e n +1 =
bn b2n n
  −1
1 1
≥ 2 ln ln yn − ln n + 1 − (ln yn ) 1 + ln yn
n n
= 2 ln ln yn + ln(yn /n) + αn = ln(yn (ln yn )2 /n) + αn =
 
(ln yn )2
= ln + βn + αn ,
ln(n + 2)

where αn , βn → 0 for n → ∞. Since

ln yn = ln(n + 1) − ln ln(n + 2) + γn ,

with γn → 0, the lower bound of ln |f

(bn )| established above goes to ∞ as
ln ln(n + 1) for n → ∞. This establishes relation (22). ✷

We can now conclude the proof of Theorem 1.1. Indeed we have, for
some n0 large enough,

|fn
(bn )| ≥ 2, for all n ≥ n0

and consequently (using (21))

|f (bn ) − f (an )| ≥ 2|bn − an |, for n ≥ n0 .

The first two equalities in (18) imply

|bn+1 − an+1 | ≥ 2|bn − an |, for n ≥ n0 . (23)

124 J. Ewing, C. Foias

It is obvious that (20) and (23) are compatible only if bn0 = an0 and
consequently only if
an = bn for n = 1, 2, . . .
Let a = a1 = b1 . Then, due to (18), the sequence

x1 = a, xn+1 = (fn ◦ · · · ◦ f1 )(a) for n = 1, 2, . . .

satisfies
xn = an = bn for all n = 1, 2, . . . . (24)
By virtue of (19),
x n
− 1 ≤ xn ≤ , (25)
ln(n + 1) ln yn
which obviously implies (1). Moreover, since

(0, ∞)\{a} = A1 ∪ B1

it follows that for any x = a the sequence

 n
1
x1 = x, xn+1 = 1 + for n = 1, 2, . . .
xn

will satisfy
lim inf xn = 1, lim sup xn = ∞.
n→∞ n→∞

This concludes the proof of Theorem 1.1.

3 The “Disconnection” from the Prime Number

Theorem
For our purpose referred in the Introduction, it suffices to recall la Vallée
Poussin’s version of the Prime Number Theorem3 Namely,

π(n) = Li(n)(1 + ε(n)) (26)

where for n large enough

√
|ε(n)| ≤ Ce− c ln n
(27)
3
See H. M. Edwards, Riemann’s Zeta Function, Academic Press, 1974, Ch. 5.
An Interesting Real Number 125

and c and C are adequate positive constants. The function Li(x) is defined
for x > 1 by
 x  1−ε  x
dt dt dt
Li(x) = (p.v. at 1) = lim + .
0 ln t ε↓0 0 ln t 1+; ln t

Clearly

n n n n
Li(n) = + 2
+ · · · + (k − 1)! k
+O .
ln n (ln n) (ln n) (ln n)k+1

In particular, 
ln n 1 1
π(n) −1= +O ,
n ln n (ln n)2
so that 
ln n
π(n) − 1 ln n → 1 for n → ∞. (28)
n
This is in dire contrast with the behavior of our sequence in Theorem 1.1
as displayed in the next theorem.

Theorem 3.1 For the sequence (xn )∞ n=1 considered in Theorem 1.1, we
have 
ln n
xn − 1 ln n − ln ln n → 0, for n → ∞. (29)
n

Proof. We start by noticing that (see (3))

n n
− 1 < xn < . (30)
ln xn+1 ln xn+1

Writing (30) with n + 1 in place of n, and introducing the result back into
(30), we obtain
n n
− 1 < xn <  
ln(n + 1) − ln ln xn+2 ln ln xn+2 − 1
n+1

n
=
ln(n + 1) − ln ln xn+2 + ln(1 − (ln xn+2 )/(n + 1))

Using (1), it follows


xn ln(n + 1) 1 ln n
= ln ln xn+2
+O . (31)
n 1 − ln(n+1) n
126 J. Ewing, C. Foias

Writing (31) with n + 2 in place of n, one obtains


ln ln n ln ln n
ln ln xn+2 = ln ln n − +O .
ln n (ln n)2

Introducing this into (31) yields


xn ln n 1 ln ln n
= +O ,
n 1 − ln n
ln ln n (ln n)2

whence

ln n
xn − 1 ln n − ln ln n =
n
 
ln ln n ln ln n ln ln n
+O = O ,
ln n − ln ln n (ln n)2 ln n

which establishes (29). ✷

 Algebraic Representations
of Regular Array Languages
Rudolf Freund
Institut für Computersprachen, Technische Universität Wien
Karlsplatz 13, A-1040 Wien, Austria
E-mail: rudi@logic.at

Alexandru Mateescu, Arto Salomaa

Turku Center for Computer Science
Lemminkäisenkatu 14 A, 20520 Turku, Finland
E-mail: mateescu, asalomaa@utu.fi

Abstract. We consider formal sequences of d-dimensional vec-

tors and symbols representing d-dimensional arrays and intro-
duce the operations of catenation as well as iterated catena-
tion of these formal sequences (d-dimensional formal arrays).
Together with the usual set union, these operations allow us
to define d-dimensional regular array expressions and thus to
develop an algebraic representation of regular array languages
generated by specific d-dimensional array grammars. In that
way, specific infinite regular array languages allow for a finite
representation as regular array expressions. Whereas, in gen-
eral, it is undecidable whether the array language generated
by a given regular array grammar is empty, finite or infinite,
for these specific regular array grammars, these questions are
decidable.

1 Introduction
The generalization of one-dimensional strings to d-dimensional arrays has
been studied during the last two decades; many interesting results could be
obtained not only in the theoretical modelling of generating and accepting

127
128 R. Freund, A. Mateescu, A. Salomaa

devices, but also with respect to applications in pattern recognition (e.g.,

see [1], [5], [9], [12]). Yet so far no algebraic approach to characterize
specific d-dimensional array languages is known. In this paper we start
the investigation of these topics by introducing specific operations on d--
dimensional arrays which can be seen as the d-dimensional generalization
of string catenation and its iterated variant. Moreover, using these d-di-
mensional catenation operations we define specific d-dimensional regular
array expressions and show that they characterize specific subclasses of the
family of d-dimensional regular array languages.
After recalling some well-known notions on string grammars and d-di-
mensional array grammars, in the third section we introduce the notion
of d-dimensional formal arrays and the operations of catenation of d-di-
mensional formal arrays and of iterated catenation of d-dimensional formal
arrays. Using these operations and the set union, we then define d-dimen-
sional regular array expressions. In the fifth section we focus on the al-
gebraic representation of specific d-dimensional regular array languages; in
particular, we exhibit the correspondence between d-dimensional array lan-
guages generated by specific d-dimensional regular array grammars and the
corresponding d-dimensional regular array expressions; we also show how a
d-dimensional regular array expression describing the array language gen-
erated by a specific regular array grammar can be obtained as the fixpoint
solution of a linear system of equations directly constructed from the given
regular array grammar. Furthermore, both fixed and general membership
for d-dimensional regular array grammars, even in the unary case, is NP-
complete, while it is undecidable whether the array language generated by
a given d-dimensional regular array grammar is empty, finite or infinite, if
d ≥ 2 (see [3], [8]). These results correspond with similar results observed
in other picture describing formalisms (e.g., see [6], [11]). Yet for specific
classes of d-dimensional regular array languages allowing for a characteri-
zation by d-dimensional regular array expressions, these questions can be
shown to be decidable as for the case of regular string grammars. In the
last section, we provide a list of further research topics.

2 Preliminaries
First, we recall some basic notions from the theory of formal languages (for
more details, the reader is referred to [10]).
For an alphabet V , by V ∗ we denote the free monoid generated by V
under the operation of catenation; the empty string is denoted by λ, and
Algebraic Representations of Regular Array Languages 129

V ∗ − {λ} is denoted by V + . Any subset of V ∗ is called a (string) language.

A (string) grammar is a quadruple G = (VN , VT , P, S), where VN and
VT are finite sets of non-terminal and terminal symbols, respectively, with
VN ∩VT = ∅, P is a finite set of productions α → β with α ∈ V + and β ∈ V ∗ ,
where V = VN ∪ VT , and S ∈ VN is the start symbol. For x, y ∈ V ∗ we say
that y is directly derivable from x in G, denoted by x =⇒G y, if and only if
for some α → β in P and u, v ∈ V ∗ we get x = uαv and y = uβv. Denoting
the reflexive and transitive closure of the derivation relation =⇒G by =⇒∗G ,
the (string) language generated by G is L(G) = {w ∈ VT∗ | S =⇒∗G w}.
A grammar is called regular, if every production in P is of the form
α → β with α ∈ VN and β ∈ VT ∪ VT VN or of the form S → λ; yet
if S → λ ∈ P , then S must not appear on the right-hand side β of a
production α → β ∈ P . The families of (string) languages generated by
arbitrary and regular grammars with the terminal alphabet VT are denoted
by L(enum(VT )) and L(reg(VT )), respectively.
In the second part of this section, we recall the definitions and notations
for d-dimensional arrays and d-dimensional array grammars (e.g., see [1],
[5], [9], [12]).
Let Z denote the set of integers, let N denote the set of positive inte-
gers, N = {1, 2, . . .}, and let d ∈ N . Then a d-dimensional array A over
an alphabet V is a function A : Z d → V ∪ {#t}, where shape(A) = {v ∈
Z d | A(v) = #} is finite and # ∈ / V is called the background or blank
symbol. We usually shall write A = {(v, A(v)) | v ∈ shape(A)}.
The set of all d-dimensional arrays over V is denoted by V ∗d . The
empty array in V ∗d with empty shape is denoted by Λd . Moreover, we
define V +d = V ∗d − {Λd }. Any subset of V ∗d is called a d-dimensional
array language.
Let v ∈ Z d . Then the translation τv : Z d → Z d is defined by τv (w) =
w + v for all w ∈ Z d , and for any array A ∈ V ∗d we define τv (A), the
corresponding d-dimensional array translated by v, by (τv (A))(w) = A(w −
v) for all w ∈ Z d . The vector (0, . . . , 0) ∈ Z d is denoted by Ωd , and
Z0d = Z d − {Ωd }.
The equivalence class with respect to linear translations [A] of an array
A ∈ V ∗d is defined by [A] = {B ∈ V ∗d | B = τv (A) for some v ∈ Z d }. The
set of all equivalence classes of d-dimensional arrays over V with respect to
linear translations is denoted by [V ∗d ] etc.
A d-dimensional array production p over V is a triple (W, A1 , A2 ), where
W ⊆ Z d is a finite set and A1 and A2 are mappings from W to V ∪ {#};
p is called Λ-free if shape(A2 ) = ∅, where we define shape(Ai ) = {v ∈ W |
130 R. Freund, A. Mateescu, A. Salomaa

Ai (v) = #}, 1 ≤ i ≤ 2. We say that the array C2 ∈ V ∗d is directly derivable

from the array C1 ∈ V ∗d by the d-dimensional array production (W, A1 , A2 )
if and only if there exists a vector v ∈ Z d such that C1 (w) = C2 (w) for all
w ∈ Z d − τv (W ) as well as C1 (w) = A1 (τ−v (w)) and C2 (w) = A2 (τ−v (w))
for all w ∈ τv (W ), i.e., the subarray of C1 corresponding to A1 is replaced
by A2 , thus yielding C2 ; we also write C1 =⇒p C2 . In the sequel we will
represent the d-dimensional array production (W, A1 , A2 ) also by writing
A1 → A2 , i.e., {(v, A1 (v)) | v ∈ W } → {(v, A2 (v)) | v ∈ W }.
A d-dimensional array grammar is a septuple

G = (d, VN , VT , #, P, v0 , S),

where VN is the alphabet of non-terminal symbols, VT is the alphabet of

terminal symbols, VN ∩ VT = ∅, # ∈ / VN ∪ VT ; P is a finite set of d-dimen-
sional array productions over VN ∪ VT , v0 is the start vector, and S is the
start symbol; {(v0 , S)} is called the start array (axiom).
We say that the array B2 ∈ V ∗d is directly derivable from the array
B1 ∈ V ∗d in G, denoted B1 =⇒G B2 , if and only if there exists a d-dimen-
sional array production p = (W, A1 , A2 ) in P such that B1 =⇒p B2 . Let
=⇒∗G be the reflexive transitive closure of =⇒G . Then the (d-dimensio-
nal) array language generated by G, L(G), is defined by L(G) = {A | A ∈
VT∗d , {(v0 , S)} =⇒∗G A}.
A d-dimensional array grammar G, G = (d, VN , VT , #, P, v0 , S), is
said to be #-context-free, if for every d-dimensional array production
p = (W, A1 , A2 ) in P we have card(shape(A1 )) = 1. G is called regular, if:
d
1. W = {Ωd , v} for some v ∈ Ud , where Ud = {(i1 , . . . , id ) | k=1 |ik | =
1},
and A1 = {(Ωd , A), (v, #t)}, A2 = {(Ωd , a), (v, B)}, with A, B ∈ VN
and a ∈ VT , – we also write A → avB –, or

2. W = {Ωd }, A1 = {(Ωd , A)}, A2 = {(Ωd , a)}, with A ∈ VN and a ∈ VT

– we also write A → a –, or

3. W = {Ωd }, A1 = {(Ωd , S)}, A2 = {(Ωd , #)} – we also write S → #;

yet this array production S → # is only allowed if for any other array
production A → avB in P we have B = S.

The families of d-dimensional array languages generated by arbitrary,

#-context-free, and regular d-dimensional array grammars with the termi-
nal alphabet VT are denoted by L(X(VT )), X = d-enum, d-#-cf, d-reg.
Algebraic Representations of Regular Array Languages 131

An interesting feature of d-dimensional array grammars is the fact that

even regular array productions make use of some special context, namely
the context of blank symbols #. This blank-sensing ability induces a rela-
tively high generating power for even regular two-dimensional array gram-
mars and yields some rather astonishing results, e.g., the set of all solid
squares can be generated by a regular two-dimensional array grammar [13].
Moreover, both fixed and general membership for regular d-dimensional
array grammars, even in the unary case, is NP-complete, while the empti-
ness problem is undecidable for d-dimensional regular array grammars, if
d ≥ 2 (see [3], [8]). Moreover, every recursively enumerable string language
can be represented by an at least two-dimensional #-context-free array
grammar (see [3]), so that the membership problem for this grammar and
language type is undecidable.

3 Catenation of Arrays and Iterated Catenation

The formal concepts we use to define the catenation operations on arrays
are elaborated in this section.

Definition 3.1 Let V be a finite alphabet, k ∈ N, ai ∈ V, and vi ∈ Z d

for 1 ≤ i ≤ k; then the sequence (v1 a1 . . . vk ak ) (we also write (vi ai )ki=1 ) is
called a (well-defined) d-dimensional formal array, if for arbitrary i and

i 
m
m with i = m and 1 ≤ i, m ≤ k we have vj = vj . The for-
j=1 j=1
mal array (vi ai )ki=1 represents the d-dimensional array A with shape(A) =

i 
i
{ vj | 1 ≤ i ≤ k} and A( vj ) = ai for 1 ≤ i ≤ k; we also write
j=1 j=1  
array((vi ai )ki=1 ) = A. The corresponding formal sequence (vi ai )ki=1 rep-
resents the equivalence class of d-dimensional arrays [A]. The empty formal
arrays (∅d ) and [(∅d )] represent the empty array (we also write (vi ai )0i=1 and
[(vi ai )0i=1 ]).
The set of all well-defined formal arrays of the form (vi ai )ki=1 with k ∈
N, ai ∈ V, and vi ∈ Z d for 1 ≤ i ≤ k such that for arbitrary i and m

i 
m
with i = m and 1 ≤ i, m ≤ k we have vj = vj , is denoted by (V +d ),
j=1
 j=1  
the corresponding set of formal arrays (vi ai )i=1
k is denoted by (V +d ) .
   
Moreover, we define (V ∗d ) = (V +d ) ∪ {(∅d )} and (V ∗d ) = (V +d ) ∪
{[(∅d )]}.
132 R. Freund, A. Mateescu, A. Salomaa

 
Any subset L of (V ∗d ) ( (V ∗d ) ) is called a language of d-dimensional
formal arrays; if L does not contain the empty array, then L is called Λ-free.

Obviously, the d-dimensional array assigned to a d-dimensional formal

array is uniquely defined. On the other hand, for a d-dimensional array
there may be a finite number of d-dimensional formal arrays representing
this d-dimensional array. The mapping array defines a function from (V ∗d )
to V ∗d as defined above with array((∅d )) = Λd .
The concept of formal arrays now allows us to introduce catenation
operations 5v with v ∈ Z0d to define the catenation of two such formal
arrays:

Definition 3.2 Let V be a finite alphabet, k, l ∈ N, ai , bj ∈ V, and vi , wj ∈

Z d for 1 ≤ i ≤ k and 1 ≤ j ≤ l; then for every w ∈ Z0d the formal array

(v1 a1 . . . vk ak wb1 . . . wl bl ),

if well-defined, denotes the 5w -catenation

(v1 a1 . . . vk ak ) 5w (w1 b1 . . . wl bl )

of the two well-defined formal arrays (v1 a1 . . . vk ak ) and (w1 b1 . . . wl bl ).

The formal array
 
(vi ai )ki=1 5w (wj bj )lj=1

represents the 5w -catenation

   
(vi ai )ki=1 5w (wj bj )lj=1
   
of the two formal arrays (vi ai )ki=1 and (wj bj )lj=1 .
In addition, for the empty (formal) arrays (∅d ) and [(∅d )] , respectively,
we define

(vi ai )ki=1 5w (∅d ) = (vi ai )ki=1 and (∅d ) 5w (vi ai )ki=1 = (vi ai )ki=1

as well as
       
(vi ai )ki=1 5w [(∅d )] = (vi ai )ki=1 and [(∅d )] 5w (vi ai )ki=1 = (vi ai )ki=1 ,

i.e., the empty formal arrays (∅d ) and [(∅d )] , respectively, are unit elements
with respect to each catenation operation 5w .
Algebraic Representations of Regular Array Languages 133

The associativity of the catenation operations is a direct consequence

of the definitions:
 
Lemma 3.1 Let x, y, z ∈ (V ∗d ) or x, y, z ∈ (V ∗d ) , respectively, and
v, w ∈ Z0d . Then, if all catenations are well-defined,

(x 5v y) 5w z = x 5v (y 5w z).

For example, any (well-defined) formal array (vi ai )ki=1 ∈ (V +d ) can also
be written as (v1 a1 ) 5v2 (Ωd a2 ) . . . 5vk (Ωd ak ).
Obviously, the definitions concerning the catenation operations 5w can
be extended to sets of formal arrays and non-empty sets C ⊆ Z0d in the
usual natural way, e.g., for two sets of formal arrays L1 and L2 we have

L1 5C L2 = {z | z = x 5w y is well-defined, x ∈ L1 , y ∈ L2 , and w ∈ C}.

We now define the iterated versions of the catenation operations 5w :

Definition 3.3 Let V be a finite alphabet, C be a (non-empty) subset of

Z0d , and L ⊆ (V ∗d ). Then the iterations of L with respect to C are defined
in the following way:

• LC = {(∅d )},

0

m+1 
• LC (L 5w LC ) for all m ≥ 0,
m
=
w∈C
+ 
• LC = (LC ),
m

m≥1

∗ +
• LC = LC ∪ {(∅d )}.

by obvious extensions of these definitions, we obtain [L]C =

m

 Moreover,

LC for m ∈ N ∪ {0, ∗, +}.
m

∗ d
Example 3.1 For example, (V ∗d )
= {v0 a0 | v0 ∈ a0 ∈ V } Z0 ; observe
Z d,
that there is no completely finite representation of (V ∗d ), i.e., if M is a
∗
finite subset of (V ∗d ) and C is a finite subset of Z0d , then M C ⊂ (V ∗d ).

As many results for d-dimensional arrays for a special d can be carried

over immediately for higher dimensions, we introduce the following notion.
134 R. Freund, A. Mateescu, A. Salomaa

Let n, m ∈ N with n ≤ m. For n < m, the natural embedding in,m :

Zn → Z m is defined by in,m (v) = (v, Ωm−n ) for all v ∈ Z n ; for n = m we
define in,n : Z n → Z n by in,n (v) = v for all v ∈ Z n . To an n-dimensional
array A ∈ V ∗n with A = {(v, A(v)) | v ∈ shape(A)} we assign the m-di-
mensional array in,m (A) = {(in,m (v), A(v)) | v ∈ shape(A)}. In a similar
way, the natural embedding of an n-dimensional formal array (vi ai )ki=1 is
defined by (in,m (vi )ai )ki=1 .

4 Regular Array Sets and Regular Array Gram-

mars
In this section we consider specific regular array grammars and the cor-
responding specific regular array sets. The special representation of the
(formal) arrays in these specific array sets is based on the following defini-
tions:

Definition 4.1 Let C ⊆ Z0d and u ∈ Z d . A (C, u)-representation of a d-

dimensional array A from V +d is a formal array (v0 a0 . . . vk ak ) such that
array((v0 a0 . . . vk ak )) = A and v0 = u as well as vi ∈ C, 1 ≤ i ≤ k;
moreover, (∅d ) is a (C, u)-representation of Λd . A formal array (vi ai )ki=1
is called a C-representation of the array [A] if and only if (Ωd a0 . . . vk ak )
is a (C, Ωd )-representation of B for some B ∈ [A] . The set of all formal
arrays that are (C, u)-representations of d-dimensional arrays A from V ∗d
is denoted by (V ∗d (C, u)).

Remark 4.1 In contrast to (V ∗d ), which, as already mentioned in Ex-

∗
ample 3.1, has no finite representation of the form M C for some finite
M ⊂ (V ∗d ) and some finite C ⊂ Z0d , for (V ∗d (C, u)) we obtain such a
finite representation by taking M = {(ua) | a ∈ V }, i.e., (V ∗d (C, u)) =
∗
{(ua) | a ∈ V }C .

The following result is an immediate consequence of the definitions:

Lemma 4.1 If card(C) = 1, then – if it exists – the (C, u)-representation

(C-representation) of the array A ([A]) is uniquely determined.

Definition 4.2 For any well-defined formal array (vi ai )ki=1 the string im-
age of (vi ai )ki=1 is a1 . . . ak , and we write string((vi ai )ki=1 ) = a1 . . . ak ; the
string image of  the empty  array Λd is the empty word λ. Moreover, we
define string( (vi ai )i=1 ) = a1 . . . ak as well as string([(∅d )]) = λ.
k
Algebraic Representations of Regular Array Languages 135

Example 4.1 Consider the one-dimensional array {((0), a), ((1), b),
((2), c)} ∈ {a, b, c}∗1 . This array has the unique ({(1), (−1)}, (0))-
representation ((0)a(1)b(1)c), whereas we find two ({(1), (−1), (2)}, (0))-
representations ((0)a(1)b(1)c) and ((0)a(2)c(−1)b). The corresponding
string images are abc and acb.

According to Lemma 4.1, each connected one-dimensional array [A] (for

the notion of connectedness, e.g., see [5]) has a uniquely determined ({(1)})-
representation and the corresponding string image is a natural representa-
for card(C) = 1 with C = {v}, v ∈ Z0 , and
tion of the array [A]. In general, d

∗d ∗d
each array A ∈ V ([A] ∈ V ) that has a – uniquely determined – (C, u)-
representation (ua1 . . . vak ) (C-representation [(Ωd a1 . . . vak )]), a1 . . . ak is
the string representation of A ([A]), which allows us to recover the (C, u)-
representation (ua1 . . . vak ) (C-representation [(Ωd a1 . . . vak )]) as well as
the array A ([A]) itself. Obviously, this possibility of recovering the orig-
inal arrays is not given any more if card(C) > 1. On the other hand,
Lemma 4.1 can be extended in the following way:

Lemma 4.2 Let C be a positive subset of Z0d , i.e., every component of each
vector v ∈ C is non-negative, and let u ∈ Z d . Then – if it exists – the
 (C,  u)-
∗d
representation (C-representation) of the array A ∈ V ([A] ∈ V ) is ∗d

uniquely determined.

Proof. On the contrary, let us assume that (v1 a1 . . . vk ak ) and

(v1
a
1 . . . vk
a
k ) with v1 = v1
= u and a1 = a
1 are two different formal
arrays which both are (C, u)-representations of the same array A ∈ V +d .
Hence,

m 
m
A = {( vj , am ) | 1 ≤ m ≤ k} = {( vj
, a
m ) | 1 ≤ m ≤ k}.
j=1 j=1


m 
m
Both sequences of vectors ( vj )km=1 and ( vj
)km=1 are strictly increas-
j=1 j=1
ing, and obviously,

m 
m
{ vj | 1 ≤ m ≤ k} = { vj
| 1 ≤ m ≤ k}.
j=1 j=1

As (v1 a1 . . . vk ak ) and (v1

a
1 . . . vk
a
k ) are two different formal arrays both
representing the same array A, there must be some n with 1 ≤ n ≤ k such
136 R. Freund, A. Mateescu, A. Salomaa

that (vj , aj ) = (vj

, a
j ) for 1 ≤ j ≤ n − 1 and (vN , aN ) = (vN

, a
); a = a

N N N
but vN = vN
is impossible, because in that case the array A would have

n

for
different symbols at the position vj , hence, we must have vN = vN
j=0
some n ≥ 2.

m 
m
As { vj | 1 ≤ m ≤ k} = { vj
| 1 ≤ m ≤ k},
j=1 j=1


n 
m 
n 
m
vj ∈ { vj | 1 ≤ m ≤ k} and vj
∈{ vj | 1 ≤ m ≤ k};
j=1 j=1 j=1 j=1


n 
n−1
moreover, vj
= (
. Now, without loss of generality, let us
vj ) + vN
j=1 j=1
assume

n−1 
n−1

( vj ) + vN < ( vj ) + vN .
j=1 j=1


n−1 
n−1
As ( vj ) < ( vj ) + vN , in the strictly increasing sequence of vectors
j=1 j=1

m 
n
( vj
)km=1 the vector vj could not appear any more, which is a contra-
j=1 j=1

m 
m
diction to the fact that { vj | 1 ≤ m ≤ k} = { vj
| 1 ≤ m ≤ k}. Hence,
j=1 j=1
we conclude that the (C, u)-representation of any array A ∈ V +d is uniquely
determined. The observation that Λd has a unique (C, u)-representation
completes the proof. ✷

Remark 4.2 As an immediate consequence of the preceding lemma we ob-

serve the following. Let C be a positive subset of Z0d and let u ∈ Z d ; if
the (C, u)-representation B of the array A ∈ V ∗d exists, then there exists
exactly one string image of A with resprect to (C, u), i.e., string(B).

We now define the specific classes of regular array grammars to be

considered in this paper:

Definition 4.3 ((C, u)-regular array grammar) Let C be a finite subset of

Z0d and u ∈ Z d . A (C, u)-regular array grammar is a d-dimensional array
grammar G, G = (d, VN , VT , #, P, v0 , S), such that v0 = u and every array
production in P is of the form A → a with A ∈ VN , a ∈ VT , or of the
form {(Ωd , A), (v, #)} → {(Ωd , a), (v, B)} with A, B ∈ VN , a ∈ VT , and
Algebraic Representations of Regular Array Languages 137

v ∈ C (we also write A → avB); finally, we also allow the array production
S → #, if for any array production in P of the form A → avB, we have
B = S. If, in addition, for every array production A → avB in P with
A, B ∈ VN , a ∈ VT , and v ∈ C we have A → av
B ∈ P for every v
∈ C,
then G is called maximal.

The d-dimensional regular array grammars G, G = (d, VN , VT , #,

P, v0 , S), usually considered in the literature (e.g., see [5]), can be seen as
d-dimensional (Ud , v0 )-regular array grammars, where Ud = {(i1 , . . . , id ) |
d
k=1 |ik | = 1}. In contrast to the standard definition of a derivation
in a d-dimensional array grammar with the array productions working
on d-dimensional arrays, for a (C, u)-regular array grammar G, G =
(d, VN , VT , #, P, u, S), we now define a terminal derivation as a sequence
(Ai )ni=0 of formal arrays Ai such that

1. A0 = (uS) = (v0 X0 ),

2. for every i with 0 < i < n,

Ai = (v0 a0 . . . vi−1 ai−1 vi Xi ) is a well-defined formal array and
Xi−1 → ai−1 vi Xi ∈ P as well as vi ∈ C, ai−1 ∈ VT and Xi ∈ VN ;
finally,

3. AN = (v0 a0 . . . vn−1 an−1 ) (is a well-defined formal array), Xn−1 →

an−1 ∈ P and an−1 ∈ VT .

If S → # ∈ P , then ((uS), (∅d )) is a terminal derivation, too, which

derives the empty array (observe that in this case the non-terminal symbol
S cannot appear in another terminal derivation (Ai )ni=0 in an Ai with i ≥ 1).
Any subsequence (Ai )m i=0 of a terminal derivation (Ai )i=0 , 0 ≤ m ≤ n, is
n

called a derivation in G. Moreover, from the definitions we immediately

infer that Ai ∈ ((VN ∪ VT )∗d (C, u)), 0 ≤ i ≤ n − 1, and AN ∈ (VT∗d (C, u)).
Obviously, the terminal derivation of formal arrays (Ai )ni=0 corresponds
with the sequence of arrays (array(Ai ))ni=0 , which represents the usual
derivation of the terminal array array(AN ) ∈ VT∗d , i.e.,

array(A0 ) =⇒G array(A1 ) =⇒G . . . =⇒G array(AN ).

Hence, we restrict ourselves to consider the set of formal arrays that can
be generated by G and thus we define

L(G) = {A | A ∈ (VT∗d (C, u)), (Ai )ni=0 is a terminal derivation in G,

and AN = A}.
138 R. Freund, A. Mateescu, A. Salomaa

The family of (array) languages that can be generated by (maximal) d-

dimensional (C, u)-regular array grammars with the terminal alphabet
VT now is denoted by L(d-(C, u)-reg(VT )) (and L(d-(C, u)-regmax (VT )));
moreover, we define [L(d-C-reg(VT ))] = [L(d-(C, Ωd )-reg(VT ))] as well as
[L(d-C-regmax (VT ))] = [L(d-(C, Ωd )-regmax (VT ))].
We now define the specific classes of regular array sets that correspond
with some special classes of specific (C, u)-regular (C-regular) array lan-
guages.

Definition 4.4 (C-regular

 array
 sets) Let C be a finite subset of Z0d . A set
∗d
of formal arrays R ⊆ (V ) is called a C-regular array set over V , if it
can be obtained in a finite number of steps using the following rules:

1. ∅ is C-regular.

2. {[(∅d )]} is C-regular.

3. {[(Ωd a)]} is C-regular for every a ∈ V .

4. If X and Y are C-regular, then X ∪ Y is C-regular.

5. If X and Y are C-regular, then X 5d Y is C-regular for every non-

empty D ⊆ C.
∗
6. If X is C-regular, then X d is C-regular for every non-empty D ⊆ C.

If we restrict the catenation operations to D = C, then the set of formal

arrays R is called maximal. The set of all (maximal) C-regular array sets
over V is denoted by [REGAd (V, C)] ([REGAd,max (V, C)]).

For subsets
 of (V ∗d ) the situation is a little bit more difficult than for
subsets of (V ∗d ) , because in this case all the arrays have to start from
the same origin:

Definition 4.5 ((C, u)-regular array sets) Let C be a finite subset of Z0d and
let u ∈ Z d . A set of formal arrays R ⊆ (V ∗d ) is called a (C, u)-regular
array set over V , if it can be obtained in a finite number of steps using the
following rules:

1. ∅ is (C, u)-regular.

2. {(∅d )} is (C, u)-regular.

Algebraic Representations of Regular Array Languages 139

3. {(ua)} is (C, u)-regular for every a ∈ V .

4. If X and Y are (C, u)-regular, then X ∪ Y is (C, u)-regular.

5. If X and Y are (C, u)-regular, then X 5d Y is (C, u)-regular for every

non-empty D ⊆ C.
∗
6. If X is (C, u)-regular, then X d is (C, u)-regular for every non-empty
D ⊆ C.

If we restrict the catenation operations to D = C, then the set of formal

arrays R is called maximal. The set of all (maximal) (C, u)-regular array
sets over V is denoted by REGAd (V, C, u) (REGAd,max (V, C, u)).

According to the construction rules given in the definitions above,

(C, u)-regular array sets can be represented as specific finite regular ex-
pressions using the operations union, catenation, and iterated catenation
of sets of formal arrays. In the following we shall not distinguish between
the notions (C, u)-regular array set and (C, u)-regular array expression.

Remark 4.3 Obviously, REGAd (V, C, u) ⊂ (VT∗d (C, u)); moreover,

[REGAd (V, C)] = [REGAd (V, C, Ωd )] and

[REGAd,max (V, C)] = [REGAd,max (V, C, Ωd )] .

If C = {w} for some w ∈ Z0d , then, by definition, all C-regular and (C, u)-
regular array sets are maximal:

REGAd (V, {w}, u) = REGAd,max (V, {w}, u) and

[REGAd (V, {w})] = [REGAd,max (V, {w})] .

Finally, L(d-(∅, u)-reg(VT )), REGAd (V, ∅, u), L(d-(∅, u)-regmax (VT )), and
REGAd,max (V, ∅, u) characterize the finite family of subsets of {(ua) | a ∈
VT } ∪ {(∅d )}, and [L(d-∅-reg(VT ))], REGAd (V, ∅), L(d-∅-regmax (VT )), and
REGAd,max (V, ∅) characterize the finite family of subsets of {[(Ωd a)] | a ∈
VT } ∪ {[(∅d )]}.

More interesting features of specific families of C-regular and (C, u)-

regular array sets will be established in the succeeding section.
140 R. Freund, A. Mateescu, A. Salomaa

5 Regular Array Sets with Algebraic Representa-

tion
As already mentioned earlier in this paper, for Ud -regular array grammars
with d ≥ 2 it is undecidable whether the array language generated by a
given d-dimensional Ud -regular array grammar is empty, finite or infinite.
This follows from the constructions used in [5], where it was shown how ev-
ery recursively enumerable string language can be represented by an at least
two-dimensional #-context-free array grammar. Using this construction we
might show that for any recursively enumerable string language L there ex-
ists a U2 -regular array grammar G such that card(L) = card(L(G)). For
d > 2 we may consider i2,d (G) yielding i2,d (L(G)). In sum, we obtain the
following result (we have to omit the proof which would go much beyond
the scope of this paper).

Proposition 5.1 Let d ≥ 2, C ⊇ Ud be a finite subset of Z0d , and u ∈ Z d .

Then it is undecidable whether for a given d-dimensional (C, u)-regular
array grammar G the language of formal arrays L(G) is empty, finite or
infinite.

In the one-dimensional case we obtain the following results, which

mostly are consequences of well-known results (e.g., see [4]):

Proposition 5.2 For any terminal alphabet VT ,

[REGA1 (VT , {(1)})] = [REGA1 (VT , {(−1)})] =

[L(1-{(1)}-reg(VT ))] = [L(1-{(−1)}-reg(VT ))];

moreover, string([L(1-{(1)}-reg(VT ))]) = L(reg(VT )).

We now prove these results in a more general context for arbitrary

singleton sets C ⊂ Z0d :

Theorem 5.1 For any terminal alphabet VT as well as w ∈ Z0d and u ∈ Z d ,

1. REGAd (VT , {w}, u) = L(d-({w}, u)-reg(VT )) =

REGAd,max (VT , {w}, u) = L(d-({w}, u)-regmax (VT )),

2. [REGAd (VT , {w})] = [L(d-{w}-reg(VT ))] =

[REGAd,max (VT , {w})] = [L(d-{w}-regmax (VT ))]
Algebraic Representations of Regular Array Languages 141

3. string(L(d-({w}, u)-reg(VT ))) = L(reg(VT )).

4. [REGAd (VT , {−w})] = [L(d-{−w}-reg(VT ))] =

[L(d-{w}-reg(VT ))] = [REGAd,max (VT , {−w})] =
[L(d-{−w}-regmax (VT ))]

Proof. According to Remark 4.3, ({w}, u)-regular array sets are already
maximal. Moreover, obviously any ({w}, u)-regular array grammar is max-
imal, too. Therefore we need not take care of the feature of maximality
throughout the rest of the proof.
REGAd,max (VT , {w}, u) = L(d-({w}, u)-regmax (VT )) will be shown
later in Theorem 5.2, hence, we obtain 1 and 2.
Let G, G = (d, VN , VT , #, P, u, S), be a d-dimensional ({w}, u)-
regular array grammar. Then consider the string grammar G
, G
=
(VN , VT , P
, S), with P
= P1
∪ P2
∪ P3
and

P1
= {A → a | A → a ∈ P, A ∈ VN , a ∈ VT },
P2
= {A → aB | A → awB ∈ P, A, B ∈ VN , a ∈ VT },
P3
= {S → λ | S → # ∈ P }.

Clearly, string(L(G)) = L(G

). On the other hand, because of the sim-
ple structure of the formal arrays it contains, L(G) is easily reconstructible
from L(G
), i.e., given the regular string grammar G
, G
= (VN , VT , P
, S),
we immediately obtain the corresponding d-dimensional ({w}, u)-regular
array grammar G, G = (d, VN , VT , #, P, u, S), with string(L(G)) = L(G
)
by taking P = P1 ∪ P2 ∪ P3 and

P1 = {A → a | A → a ∈ P
, A ∈ VN , a ∈ VT },
P2 = {A → awB | A → aB ∈ P
, A, B ∈ VN , a ∈ VT },
P3 = {S → # | S → λ ∈ P
}.

Therefore we conclude that (3) is valid.

Finally, for any string language L, let Lr denote the mirror image of
L. Let G be a d-dimensional ({−w}, u)-regular array grammar. Then
consider the string grammar Gs for string(L(G)). L(reg(VT )) is closed
under mirror image, i.e., from Gs we can effectively construct a string
grammar Gr such that L(Gr ) = (L(Gs ))r . According to the construc-
tion above, from Gr we obtain a ({w}, u)-regular array grammar G
with
string(L(G
)) = (string(L(G)))r , and moreover, [L(G
)] = [L(G)]. Hence,
142 R. Freund, A. Mateescu, A. Salomaa

we conclude [L(d-{−w}-reg(VT ))] = [L(d-{w}-reg(VT ))], which completes

the proof. ✷

Remark 5.1 In addition to the results proved above we should like

to mention that for any ({w, −w}, u)-regular array grammar G, G =
(d, VN , VT , #, P, u, S), we can construct a ({w}, u)-regular array grammar
G
such that [L(G
)] = [L(G)], i.e.,

[L(d-{w, −w}-reg(VT ))] = [L(d-{w}-reg(VT ))].

This follows from the fact that [L(G)] simply is the union [L(G)] =
[L(G1 )] ∪ [L(G2 )] of a language [L(G1 )] ∈ [L(d-{w}-reg(VT ))] and a lan-
guage [L(G2 )] ∈ [L(d-{−w}-reg(VT ))], where Gi = (d, VN , VT , #, Pi , u, S)
with

Pi = P − {A → a((−1)2−i w)B | A → a((−1)2−i w)B ∈ P,

A, B ∈ VN , a ∈ VT }, i ∈ {1, 2}.

The correspondence of array languages in L(d-({w}, u)-reg(VT )) with

regular string languages immediately yields the following decidability re-
sults.

Corollary 5.1 Let w ∈ Z0d and u ∈ Z d . Then emptiness and finiteness of

L(G) are decidable for any ({w}, u)-regular array grammar G.

Proof. According to the proof of Theorem 5.1, the string grammar

for string(L(G)) can effectively be constructed from G, and L(G) is
empty/finite if and only if string(L(G)) is empty/finite. ✷

Corollary 5.2 Let C be a finite subset of Z0d and u ∈ Z d . Then emptiness

and finiteness of L(G) are decidable for any maximal (C, u)-regular array
grammar G.

Proof. Let G, G = (d, VN , VT , #, P, u, S), be a d-dimensional maxi-

mal (C, u)-regular array grammar and w ∈ C. Then consider the d-
dimensional maximal ({w}, u)-regular array grammar Gw , Gw = (d, VN , VT ,
#, Pw , u, S), where

Pw = P − {A → avB | A → avB ∈ P and v = w}.

Algebraic Representations of Regular Array Languages 143

Obviously, L(Gw ) ⊆ L(G) and L(G) is empty/finite if and only if L(Gw ) is

empty/finite, which is decidable by Corollary 5.1. ✷

In the second part of this section we elaborate the equivalence of char-

acterizing specific d-dimensional regular array languages by d-dimensio-
nal maximal (C, u)-regular array grammars and by d-dimensional maxi-
mal (C, u)-regular expressions. We also show how d-dimensional maximal
(C, u)-regular array sets can be obtained as the fixpoint solution of a linear
system of equations directly constructed from the corresponding d-dimen-
sional maximal (C, u)-regular array grammar.
Our main result can be formulated in the following way:

Theorem 5.2 For any terminal alphabet VT , u ∈ Z d , and any non-empty

finite subset C of Z0d ,

REGAd,max (VT , C, u) = L(d-(C, u)-regmax (VT )).

The inclusion ⊆ can be proved by showing an even more general result:

Lemma 5.1 For any terminal alphabet VT , u ∈ Z d , and any non-empty

finite subset C of Z0d ,

1. REGAd (VT , C, u) ⊆ L(d-(C, u)-reg(VT )) and

2. [REGAd (VT , C)] ⊆ [L(d-C-reg(VT ))] as well as

3. REGAd,max (VT , C, u) ⊆ L(d-(C, u)-regmax (VT )) and

4. [REGAd,max (VT , C)] ⊆ [L(d-C-regmax (VT ))].

Proof. Let L ∈ REGAd (VT , C, u).

If L ∈ {∅, {(∅d )}} ∪ {(ua) | a ∈ VT }, then obviously L can be generated
by a d-dimensional even maximal (C, u)-regular array grammar.
We now show that L(d-(C, u)-reg(VT )) and L(d-(C, u)-regmax (VT )) are
closed under union as well as catenation and iterated catenation with re-
spect to C. Let Li ∈ L(d-(C, u)-reg(VT )), i ∈ {1, 2}, i.e., there are two
(i)
d-dimensional (C, u)-regular array grammars Gi = (d, VN , VT , #, Pi , u, Si ),
such that L(Gi ) = Li , i ∈ {1, 2}. Without loss of generality we may assume
(1) (2)
that VN and VN are disjoint sets.
144 R. Freund, A. Mateescu, A. Salomaa

1. If L is obtained as the union L1 ∪L2 , we define the new d-dimensional

(C, u)-regular array grammar G, G = (d, VN , VT , #, P, u, S), such that
(1) (2)
S is a new symbol, VN = VN ∪ VN ∪ {S}, P = ((P1 ∪ P2 ) − {S1 →
#, S2 → #}) ∪ P
∪ P

, where P
= {S → α | S1 → α ∈ P1 } ∪ {S →
β | S2 → β ∈ P2 } and P

= {S → #} if either S1 → # ∈ P1 or
S2 → # ∈ P2 and P

= ∅ otherwise. Clearly, G is a d-dimensional

(C, u)-regular array grammar and L(G) = L1 ∪ L2 . If G1 and G2 are
maximal, then G is maximal, too.

2. Now let us consider L = L1 5D L2 for some D ⊆ C; moreover, assume

that the empty d-dimensional formal array (∅d ) does not belong to
(1) (2)
L1 ∪L2 . Then we define G = (d, VN , VT , #, P, u, S1 ), VN = VN ∪VN ,
P = (P1 − {X → a | X → a ∈ P1 }) ∪ P2 ∪ P
, where P
= {X →
awS2 | X → a ∈ P1 and w ∈ D}. One can easily verify that G is a
d-dimensional (C, u)-regular array grammar and L(G) = L1 5D L2 .
If G1 and G2 are maximal and D = C, then G is maximal, too.
The cases where the empty d-dimensional formal array (∅d ) is in L1
or in L2 or in both of them are left to the reader.

3. Let L = L 1
D
for some D ⊆ C. Then we define the d-dimensional
(C, u)-regular array grammar G = (d, VN ∪ {S}, VT , #, P1
, u, S) such
(1)

that S is a new symbol, P1

= (P1 − {S1 → #}) ∪ P
∪ P

∪ {S → #},
where P
= {S → α | S1 → α ∈ P1 } and P

= {X → awS1 | X →
a ∈ P1 and w ∈ D}. Obviously, G is a d-dimensional (C, u)-regular
array grammar and L(G) = L D
1 . If G1 is maximal and D = C, then
G is maximal, too.

All the closure properties proved above for L(d-(C, u)-reg(VT ))

and L(d-(C, u)-regmax (VT )) are valid for the corresponding families
[L(d-C-reg(VT ))] and [L(d-C-regmax (VT ))], too, i.e., in these cases we sim-
ply may consider d-dimensional (C, Ωd )-regular array grammars. ✷

In order to prove the inclusion ⊇ in Theorem 5.2, we need some addi-

tional notations and results (e.g., see [7]).

Remark 5.2 For any terminal alphabet VT , u ∈ Z d , and any non-empty

finite subset C of Z0d , let A(d, VT , C, u) and [A(d, VT , C)] denote the pow-
erset of (VT∗d (C, u)) and [(VT∗d (C))] (= [(VT∗d (C, u))]), respectively. Then
(A(d, VT , C, u), ∪, 5C , ∅, {(∅d )}) and (A(d, VT , C), ∪, 5C , ∅, {[(∅d )]}) is an
ω-complete semiring with first element ∅.
Algebraic Representations of Regular Array Languages 145

Remark 5.3 Let S = (S, +, 5, 0, 1) be an ω-complete semiring with first

element 0. Let Y = {Yi | 1 ≤ i ≤ m} be a set of variables and let E be a
rational system of equations over S, i.e., E is a set of equations of the form
Yi = αi1 5 Y1 + . . . 5 αim 5 Ym + βi ,
where αij ∈ S, βi ∈ S, 1 ≤ i ≤ m, 1 ≤ j ≤ m.
It is well-known (see [7]) that the system E has a solution, and moreover,
each component of the solution of E is a rational set in S.
Example 5.1 For a terminal alphabet VT , u ∈ Z d , and a non-empty finite
subset C of Z0d , consider the equation Y = α 5C Y ∪ β, where α and β
are subsets of (V ∗d (C, u)) or [(V ∗d (C))]. Then the minimal fixpoint of this
∗
equation is αC 5C β according to Kleene’s theorem.
For any d-dimensional maximal (C, u)-regular array grammar we now
define a rational system of equations EG such that L(G) is a component of
the (minimal) fixpoint solution of EG .
Definition 5.1 Let G = (d, VN , VT , #, P, u, Y1 ) be a d-dimensional max-
imal (C, u)-regular array grammar. The rational system of equations EG
associated with G is defined as follows:
1. The elements in VN , VN = {Yi | 1 ≤ i ≤ m}, are the variables of the
system EG .
2. The equations forming EG are
Yi = αi1 5C Y1 ∪ . . . αim 5C Ym ∪ βi ,
where αij = {(ua) | Yi → avYj ∈ P for all v ∈ C}, 1 ≤ j ≤ m,
1 ≤ i ≤ m, and
βi = {(ua) | Yi → a ∈ P }, 1 ≤ i ≤ m; if Y1 → # ∈ P, then we also
add (∅d ) to β1 .
Using these notations (similar notations can be introduced for C-regular
array grammars in an obvious way) we state the following important result:
Theorem 5.3 Let G = (d, VN , VT , #, P, u, Y1 ) be a d-dimensional maximal
(C, u)-regular array grammar and let EG be the rational system of equations
associated with G as defined in Definition 5.1. Let LYi be the i-th component
(Yi -component) of the minimal fixpoint solution of EG . Then,
LYi = L(GYi ), where GYi = (d, VN , VT , #, P, u, Yi ),
and therefore L(G) = LY1 .
146 R. Freund, A. Mateescu, A. Salomaa

Proof. First we show that if the formal array A is the result of a ter-
minal derivation in GYi starting from (uYi ), then it is an element of the
Yi -component LYi of the solution of EG , too, which can be shown by in-
duction with respect to the length of terminal derivations in the grammars
GYi .
If the terminal derivation is ((uY1 ), (∅d )), then by definition, Y1 → # ∈
P and therefore (∅d ) ∈ β1 . If the terminal derivation is ((uYi ), (ua)) for
some a ∈ VT , then by definition, Yi → a ∈ P and therefore (ua) ∈ βi .
Hence, in any case, for a terminal derivation starting with (uYi ) of length 1,
the result of a terminal derivation in GYi is an element of the Yi -component
LYi of the solution of EG , too. Now assume that the sequence (Ak )nk=0 of
formal arrays Ak is a terminal derivation in GYi with n > 1 such that:

1. A0 = (uYi ) = (v0 X0 ),

2. for every k with 0 < k < n, Ak = (v0 a0 . . . vk−1 ak−1 vk Xk ) is a well-

defined formal array and Xk−1 → ak−1 vk Xk ∈ P as well as vk ∈ C,
ak−1 ∈ VT , and Xk ∈ VN ; finally,

3. AN = (v0 a0 . . . vn−1 an−1 ) (is a well-defined formal array), Xn−1 →

an−1 ∈ P and an−1 ∈ VT .

Then we consider the sequence (Bk )nk=1 with:

1. B1 = (uX1 ) = (v1
X1 );

2. for every k with 1 < k < n, Bk = (v1

a1 . . . vk−1

ak−1 vk
Xk ), v1
= u,

vl = vl , 1 < l ≤ k;

3. BN = (v1
a1 . . . vn−1

an−1 ) = (v1 a1 . . . vn−1 an−1 ).

The formal arrays Bk , 1 ≤ k ≤ n, are well-defined, and (Bk )nk=1

is a terminal derivation in GX1 of length n − 1. By the induction hy-
pothesis, BN ∈ LX1 . According to the construction of the system EG ,
X0 → a0 v1 X1 ∈ P implies that (ua0 ) ∈ αij , where j is the index
such that X1 = Yj . Hence, by the definition of the fixpoint solution,
AN = (ua0 ) 5v1 BN and AN ∈ LYi .
For the converse inclusion, the proof is by induction with respect to the
number of Kleene-operations necessary to obtain a formal array A as an
element of the component LYi :
If A belongs to the first iteration, this means that A has to be an
element of βi . According to the definition of EG , A then is the result of
Algebraic Representations of Regular Array Languages 147

a terminal derivation of length 1 in GYi . Now assume that for all formal
arrays coming up in the (n − 1)-th Kleene-iteration, n > 1, there exists a
terminal derivation in GYi of length (n − 1) yielding A. Now consider a
formal array A coming up in the n-th Kleene-iteration of the component
Yi , i.e., A = (ua0 ) 5v1 B, where B is a formal array having come up in
the (n − 1)-th Kleene-iteration of the component Yj for some index j and
(ua0 ) ∈ αij . By the induction hypothesis, B is the result of a terminal
derivation (Bk )nk=1 of length n − 1 in GYj . Moreover, by the construction of
the system EG , Yi → a0 v1 Yj ∈ P ; therefore, the terminal derivation (Bk )nk=1
of B can be extended to a terminal derivation (Ak )nk=0 of A of length n in
GYi with A0 = (uYi ) and Ak = (ua0 ) 5v1 Bk , 1 ≤ k ≤ n; consequently,
A ∈ L(GYi ). ✷

The proof of Theorem 5.2 is now complete, it follows from Lemma 5.1
and Theorem 5.3. The results of Theorem 5.2 and its consequences can be
summed up in the following way:
Corollary 5.3 For any terminal alphabet VT , u ∈ Z d , and any (non-
empty) finite subset C of Z0d ,
1. REGAd,max (VT , C, u) = L(d-(C, u)-regmax (VT ));
2. [REGAd,max (VT , C)] = [L(d-C-regmax (VT ))];
3. L(d-(C, u)-regmax (VT )) ([L(d-C-regmax (VT ))]) coincides with the
family of rational languages in A(d, VT , C, u) (in A(d, VT , C)).

6 Future Research Topics

This paper can be seen as a starting point for further investigations in the
field of algebraic representations of d-dimensional array languages. One
major research topic will be the investigation of possible algebraic repre-
sentations of d-dimensional regular array languages generated by d-dimen-
sional (C, u)-regular array grammars that are not maximal as well as of
possible algebraic representations of d-dimensional array languages gener-
ated by specific (#-)context-free d-dimensional array grammars. Moreover,
all the notions introduced in this paper can also be introduced in the frame-
work of Cayley grids as proposed in [2], where the underlying grid is based
on the more general structure of Cayley graphs of finitely generated groups;
many of the results obtained in this paper can be carried over into this more
general framework, although some of them are not valid any more in the
form presented in this paper.
148 R. Freund, A. Mateescu, A. Salomaa

References
[1] C. R. Cook, P. S.-P. Wang, A Chomsky hierarchy of isotonic array
grammars and languages, Computer Graphics and Image Processing,
8 (1978), 144–152.
[2] E. Csuhaj-Varjú, R. Freund, V. Mitrana, Arrays on Cayley Grids,
TUCS Report, 1998.
[3] H. Fernau, R. Freund, M. Holzer, The generative power of d-dimensio-
nal #-context-free array grammars, Proceedings MCU’98, Vol. II (M.
Margenstern, ed.), Metz, 1998, 43–56.
[4] R. Freund, Gh. Păun, One-dimensional matrix array grammars, J.
Inform. Process. Cybernet., EIK, 29, 6 (1993), 1–18.
[5] R. Freund, Control mechanisms on #-context-free array grammars, in
Mathematical Aspects of Natural and Formal Languages (Gh. Păun,
ed.), World Scientific Publ., Singapore, 1994, 97–136.
[6] C. Kim, I. H. Sudborough, The membership and equivalence problems
for picture languages, Theoretical Computer Science, 52 (1987), 177–
191.
[7] W. Kuich, A. Salomaa, Semirings, Automata, Languages, Springer-
Verlag, Berlin, 1986.
[8] K. Morita, Y. Yamamoto, K. Sugata, The complexity of some de-
cision problems about two-dimensional array grammars, Information
Sciences, 30 (1983), 241–262.
[9] A. Rosenfeld, Picture Languages, Academic Press, Reading, MA, 1979.
[10] A. Salomaa, Formal Languages, Academic Press, Reading, MA, 1973.
[11] I. H. Sudborough, E. Welzl, Complexity and decidability for chain code
picture languages, Theoretical Computer Science, 36 (1985), 173–202.
[12] P. S.-P. Wang, Some new results on isotonic array grammars, Infor-
mation Processing Letters, 10 (1980), 129–131.
[13] Y. Yamamoto, K. Morita, K. Sugata, Context-sensitivity of two-di-
mensional regular array grammars, in Array Grammars, Patterns and
Recognizers (P. S.-P. Wang, ed.), WSP Series in Computer Science,
Vol. 18, World Scientific Publ., Singapore, 1989, 17–41.
Rough Set Processing of Vague Information
Using Fuzzy Similarity Relations
Salvatore Greco, Benedetto Matarazzo
Faculty of Economics, University of Catania
Corso Italia 55, 95129 Catania, Italy
E-mail: salgreco@vm.unict.it

Roman Slowinski
Institute of Computing Science, Poznan University of Technology
Piotrowo 3a, 60-965 Poznan, Poland
E-mail: slowinsk@sol.put.poznan.pl

Abstract. The rough sets theory has proved to be a very use-

ful tool for analysis of information tables describing objects by
means of disjoint subsets of condition and decision attributes.
The key idea of rough sets is approximation of knowledge ex-
pressed by decision attributes using knowledge expressed by
condition attributes. From a formal point of view, the rough
sets theory was originally founded on the idea of approximating
a given set represented by objects having the same description
in terms of decision attributes, by means of an indiscernibility
binary relation linking pairs of objects having the same descrip-
tion by condition attributes. The indiscernibility relation is
an equivalence binary relation (reflexive, symmetric and tran-
sitive) and implies an impossibility to distinguish two objects
having the same description in terms of the condition attributes.
It produces crisp granules of knowledge that are used to built
approximations. In reality, due to vagueness of available in-
formation about objects, small differences are not considered
significant. This situation may be formally modelled by simi-
larity or tolerance relations instead of the indiscernibility rela-
tion. We are using a similarity relation which is only reflexive,
relaxing therefore the properties of symmetry and transitivity.

149
150 S. Greco, B. Matarazzo, R. Slowinski

Moreover, the credibility of our similarity relation is gradual,

i.e., it is a fuzzy similarity relation. In consequence, the gran-
ules of knowledge produced by this relation are fuzzy and their
credibility can change gradually from any finite degree to an
infinitely small degree, thus giving different credibility of rough
approximations.

1 Introduction
The rough sets theory proposed by Pawlak [11], [12] was originally founded
on the idea of approximating a given set by means of an indiscernibility
binary relation which was assumed to be an equivalence binary relation
(reflexive, symmetric and transitive). It produces crisp granules of knowl-
edge that are used to built approximations. With respect to the basic
rough set idea, three main theoretical developments have been proposed
for handling different types of the available information:

1. extensions to a fuzzy description of objects (e.g., [2], [3], [18], [19],

[21]);

2. extensions of the indiscernibility relation to more general binary re-

lations modelling similarity between objects (e.g., [7], [8], [14], [16],
[23], [24], [25]);

3. extensions to preference-ordered scales of attributes (criteria) and to

multicriteria decision problems, using dominance relations and pair-
wise comparison tables (e.g., [5]).

In this paper, we put together extensions 1) and 2), considering the ap-
proach proposed in [23], [24] within a fuzzy context. More specifically, we
propose to approximate a given fuzzy set by means of reflexive fuzzy simi-
larity relations. This extension is motivated by vagueness of the available
information about objects. In consequence of considering fuzzy similar-
ity relations instead of the classical indiscernibility relation, the granules
of knowledge produced by this relation are fuzzy and their credibility can
change gradually from any finite degree to an infinitely small degree, thus
giving different credibility of rough approximations.
The paper is structured as follows. In Section 2, we recall some basic
concepts concerning rough sets, fuzzy sets, approximations by similarity
relations and rough fuzzy sets. In Section 3, we introduce the rough ap-
proximation by fuzzy similarity relations. In Section 4, we discuss the rule
Rough Set Processing of Vague Information 151

induction from rough approximations by fuzzy similarity relations. In Sec-

tion 5, we present an application of the proposed approach to an exemplary
problem. Section 6 groups conclusions.

2 Basic Elements of Fuzzy Sets and Rough Sets

Theories
2.1 Elements of the Rough Sets Theory
The rough set concept proposed by Pawlak [11], [12] is founded on the
assumption that with every object of the universe of discourse there is
associated some information (data, knowledge). For example, if objects
are firms submitted to a bankruptcy risk evaluation, their financial, eco-
nomic and technical characteristics form information (description) about
the firms. Objects characterized by the same description are indiscernible
in view of available information about them. The indiscernibility relation
generated in this way is the mathematical basis of the rough sets theory.
Any set of indiscernible objects is called an elementary set and forms
a crisp granule (atom) of knowledge about the universe. Any subset Y of
the universe can either be expressed precisely, in terms of the granules, or
roughly only. In the latter case, subset Y can be characterised by two ordi-
nary sets called lower and upper approximations. The two approximations
define the rough set. The lower approximation of Y consists of all elemen-
tary sets included in Y , whereas the upper approximation of Y consists of
all elementary sets having a non-empty intersection with Y . Obviously, the
difference between the upper and the lower approximation constitutes the
boundary region including objects which cannot be properly classified as
belonging or not to Y , using the available information. Cardinality of the
boundary region says, moreover, how exactly we can describe Y in terms
of available information.
For algorithmic reasons, knowledge about objects will be represented
in the form of an information table. The rows of the table are labelled by
objects, whereas columns are labelled by attributes and entries of the table
are attribute-values. Formally, by an information table we understand the
4-tuple S = (U, Q, V, f ), where U is a finite set of objects, Q is a finite

set of attributes, V = q∈Q Vq and Vq is a domain of the attribute q, and
f : U × Q −→ V is a total function such that f (x, q) ∈ Vq for every
q ∈ Q, x ∈ U , called an information function (see [12]).
An information table can be seen as decision table assuming that the
152 S. Greco, B. Matarazzo, R. Slowinski

set of attributes Q is equal to C ∪ D such that C ∩ D = ∅, where set C

contains so-called condition attributes, and D, decision attributes.
From the decision table a set of decision rules can be induced and
expressed as logical statements “if. . . then. . . ” relating descriptions of con-
dition and decision classes. The rules are exact or approximate depending
whether a description of a condition class corresponds to a unique decision
class or not. Different procedures for derivation of decision rules have been
presented (see, e.g., [6], [27], [10], [22]).

2.2 Elements of the Fuzzy Sets Theory

We use the following basic concepts with respect to fuzzy sets theory. For a
more extensive review see, e.g., [1], [4]. Each set X in a universe of discourse
U is represented by its membership function µX : U −→ [0, 1], where µX (x)
is the grade of membership of x in X, from full nonmembership (µX (x) = 0)
to full membership (µX (x) = 1), through all intermediate values. Within
fuzzy logic the conjunction operator “and” is represented by a triangular
norm or t-norm T (x, y) while the disjunction operator “or” is represented
by a triangular conorm or t-conorm T ∗ (x, y) ([9], [15], [26]). More precisely,
• a t-norm is a function T : [0, 1]×[0, 1] −→ [0, 1] satisfying the following
properties:
T (1, x) = x, for all x ∈ [0, 1],
T (x, y) = T (y, x), for all x, y ∈ [0, 1],
T (x, y) ≤ T (u, v), for all x ≤ u, y ≤ v,
T (x, T (y, z)) = T (T (x, y), z), for all x, y, z ∈ [0, 1];
• a t-conorm is a function T ∗ : [0, 1] × [0, 1] −→ [0, 1] satisfying the
following properties:
T ∗ (0, x) = x, for all x ∈ [0, 1],
T ∗ (x, y) = T ∗ (y, x), for all x, y ∈ [0, 1],
T ∗ (x, y) ≤ T ∗ (u, v), for all x ≤ u, y ≤ v,
T ∗ (x, T ∗ (y, z)) = T ∗ (T ∗ (x, y), z), for all x, y, z ≤ [0, 1].
The powers of the function T and T ∗ are defined as follows:
T 1 (x1 , x2 ) = T (x1 , x2 ),
T ∗1 (x1 , x2 ) = T ∗ (x1 , x2 ),
for all x1 , x2 ∈ [0, 1], and
T n (x1 , . . . , xn+1 ) = T (T n−1 (x1 , . . . , xn ), xn+1 ),
T ∗n (x1 , . . . , xn+1 ) = T ∗ (T ∗n−1 (x1 , . . . , xn ), xn+1 ),
Rough Set Processing of Vague Information 153

for all x1 , . . . , x2 ∈ [0, 1] and for all n ≥ 2.

We adopt the following notation:

T n−1 (x1 , . . . , xn ) = T (x1 , . . . , xn ) = Ti=1

n
xi = Tx∈A x,
T ∗n−1 (x1 , . . . , xn ) = T ∗ (x1 , . . . , xn ) = Ti=1
∗n ∗
xi = Tx∈A x,

where A = {x1 , . . . , xn } and x1 , . . . , xn ∈ [0, 1].

Let us observe that for each t-norm T and for each t-conorm T ∗ the
following conditions hold for all x, y ∈ [0, 1]

T (x, y) ≤ min x, y, (1)

∗
T (x, y) ≥ max x, y. (2)

In the context of fuzzy logic the negation operator N (·) is a nonincreas-

ing function N : [0, 1] −→ [0, 1] such that N (0) = 1 and N (1) = 0. A
negation is strict if N (·) is a strictly decreasing continuous function. Often
the negation N is required to satisfy also the involutory condition

N (N (x)) = x, for all x ∈ [0, 1]. (3)

Given a fuzzy set X in universe U , whose membership function is µX (x)

the complement of X, denoted by U − X, is the fuzzy set on U having the
membership function µU −X (x) = N (µX (x)).
Given a t-norm T , a t-conorm T ∗ and a strict negation N , (T, T ∗ , N )
is a De Morgan triplet iff N (T ∗ (x, y)) = T (N (x), N (y)).
In fuzzy logic several implication operators have been proposed. Let us
remember that an implication is a function I → (x, y) : [0, 1] × [0, 1] −→ [0, 1]
satisfying the following properties ([4]):

I → (x, y) ≥ I → (z, y), for all x ≤ z and for all y,

I → (x, y) ≥ I → (t, y), for all x ≤ t and for all y,
I → (0, x) = 1, for all x,
I → (x, 1) = 1, for all x,
I → (1, 0) = 0.

In the paper we consider a T ∗ -implication IT→∗ ,N , i.e., a function IT→∗ ,N :

[0, 1] × [0, 1] −→ [0, 1] associated with a t-conorm T ∗ and a negation N
defined by IT→∗ ,N (x, y) = T ∗ (N (x), y)).
A fuzzy binary relation is a function R : U × U −→ [0, 1]. It is

• reflexive if R(x, x) = 1, for each x ∈ U,

154 S. Greco, B. Matarazzo, R. Slowinski

• symmetric if R(x, y) = R(y, x), for all x, y ∈ U,

• T -transitive if T (R(x, y), R(y, z)) ≤ R(x, z), for all x, y, z ∈ U.
A fuzzy binary relation which is reflexive, symmetric and T -transitive is
a fuzzy equivalence binary relation. In fuzzy sets theory, a fuzzy equivalence
relation is called fuzzy similarity, however, we prefer to keep the term of
fuzzy similarity for fuzzy binary relations that are only reflexive.

2.3 Fuzzy Rough Sets

Let X ⊆ U . Let also R be a binary equivalence relation defined in U , and
[x]R the equivalence class including x ∈ U . The lower approximation of X,
denoted by R(X), and the upper approximation of X, denoted by R(X),
are defined as
R(X) = {x ∈ U | [x]R ⊆ X}, (4)
R(X) = {x ∈ U | [x]R ∩ X = ∅}. (5)
Let X be a fuzzy set defined in a finite universe U, µX (x) the membership
function of X, R an equivalence relation defined on U , [x]R the equivalence
class including x ∈ U . The lower approximation of X can be defined as a
fuzzy set whose membership function associated to each x ∈ U is equal to
the credibility that “each y ∈ [x]R belongs to X”, i.e.,
µ(x, R(X)) = Ty∈[x]R µX (y). (6)
Analogously, the upper approximations of X can be defined as a fuzzy set
whose membership function associates to each x ∈ U the credibility that
“there is at least one y ∈ [x]R belonging to X”, i.e.,
∗
µ(x, R(X)) = Ty∈[x] µ (y).
R X
(7)
Let X ⊆ U , and let R be an equivalence fuzzy binary relation defined
on U . In this case the lower approximation of X can be defined as a fuzzy
set whose membership function associates to each x ∈ U the credibility
that “for each y ∈ X, y is not in the relation R with x”, i.e.,
µ(x, R(X)) = Ty∈X
/ N (R(y, x)). (8)
Analogously the upper approximations of X can be defined as a fuzzy set
whose membership function associates to each x ∈ U the credibility that
“there is at least one y ∈ X which is in the relation R with x”, i.e.,
∗
µ(x, R(X)) = Ty∈X R(y, x). (9)
Rough Set Processing of Vague Information 155

Finally, let X be a fuzzy set defined in a finite universe U . Moreover, let

µX (x) be the membership function of X and R be an equivalence fuzzy
binary relation defined on U . In this case, the lower approximation of X
can be defined as a fuzzy sets whose membership function associates to
each x ∈ U the credibility that “for each y ∈ U , y is not in the relation R
with x and/or y belongs to X”, i.e.,

µ(x, R(X)) = Ty∈U T ∗ (N (R(y, x), µX (y))). (10)

The upper approximation of X can be defined in turn as a fuzzy set whose

membership function associates to each x ∈ U the credibility that “there
is at least one y ∈ U which is in the relation R with x and which belongs
to X”, i.e.,
∗
µ(x, R(X)) = Ty∈U T (R(y, x), µX (y)). (11)
Let us observe that

a) definitions (5) and (6) are special cases of definitions (10) and (11),
when the fuzzy equivalence binary relation R becomes a crisp equiv-
alence binary relation,

b) definitions (8) and (9) are special cases of definitions (10) and (11),
when the fuzzy set X becomes a crisp set,

c) the classical definitions (5) and (5) represent special cases of defini-
tions (10) and (11), when the fuzzy equivalence binary relation R
becomes a crisp equivalence binary relation and the fuzzy set X be-
comes a crisp set.

2.4 Similarity
As observed above, indiscernibility implies an impossibility to distinguish
two objects of U having the same description in terms of the attributes from
Q. This relation induces equivalence classes on U , which constitute the
crisp granules of knowledge. In reality, due to uncertainty and imprecision
of data describing the objects (vague information), small differences are
often not considered significant for the purpose of discrimination. This
situation may be formally modelled by considering similarity or tolerance
relations (see e.g., [7], [8], [14], [16], [23], [24], [25]).
In general, the similarity relations R do not generate partitions of U ;
the information regarding similarity may be represented using similarity
156 S. Greco, B. Matarazzo, R. Slowinski

classes for each object x ∈ U . Precisely, the similarity class of x, denoted

by R(x), consists of the set of objects which are similar to x:

R(x) = {y ∈ U | yRx}.

It is obvious that an object y ∈ R(x) may be similar to another object

z ∈ U , and z ∈ / R(x). The similarity relation is of course reflexive (each
object is similar to itself). In [23], [24] one proposes a similarity relation
which is only reflexive, relaxing therefore the properties of symmetry and
transitivity. The abandon of the transitivity requirement is easily justifi-
able, remembering – for example – Luce’s paradox of the cups of tea. As
for the symmetry, one should notice that yRx, which means “y is similar
to x”, is directional; there is a subject y and a referent x, and in general
this is not equivalent to the proposition “x is similar to y”. This is quite
immediate when the similarity relation is defined in terms of a percent-
age difference between evaluations of the objects compared on a numerical
attribute, calculated with respect to the evaluation of the referent object.
Therefore, the symmetry of the similarity relation should not be imposed
and it makes sense to consider the inverse relation of R, denoted by R−1 ,
where xR−1 y means again “y is similar to x”; R−1 (x), x ∈ U, is then the
class of referent objects to which x is similar:

R−1 (x) = {y ∈ U | xRy}.

Given a subset X ⊆ U and a similarity relation R on U , an object x ∈ U

is said to be non-ambiguous in each of the two following cases:
• x clearly belongs to X, that is x ∈ X and R−1 (x) ⊆ X; such objects
are called positive;

• x clearly does not belong to X, that is x ∈ U −X and R−1 (x) ⊆ U −X

(or R−1 (x) ∩ X = ∅); such objects are called negative.
The objects which are neither positive nor negative are said to be am-
biguous.
A more general definition of lower and upper approximation may thus
be offered (see [24]). Let X ⊆ U and R be a reflexive binary relation
defined on U ; the lower approximation of X, denoted by R(X), and the
upper approximation of X, denoted by R(X), are defined, respectively, as:

R(X) = {x ∈ U | R−1 (x) ⊆ X},


R(X) = R(x).
x∈X
Rough Set Processing of Vague Information 157

It may be demonstrated that the key property R(X) ⊆ X ⊆ R(X) still

holds and that

R(X) = U − R(U − X) (complementarity property) and

R(X) = {x ∈ U | R−1 (x) ∩ X = ∅}.

Moreover, the definitions proposed are the only ones which correctly
characterize the set of positive objects (lower approximation) and the set
of positive or ambiguous objects (upper approximation) when a similarity
relation is reflexive, but not necessarily symmetric nor transitive.
Using similarity relation one is able to induce decision rules from a
decision table. The syntax of a rule is the following:

“if f (x, q1 ) is similar to rq1 and f (x, q2 ) is similar to rq2 and

. . . f (x, qp ) is similar to rqp , then x belongs to Yj1 or Yj2 or
. . . Yjk ”,

where {q1 , q2 , . . . , qp } ⊆ C, (rq1 , rq2 , . . . , rqp ) ∈ Vq1 × Vq2 × . . . × Vqp and

Yj−1 , Yj2 , . . . , Yjk are some classes of the considered classification (D-
elementary sets). If k = 1, then the rule is exact, otherwise it is approximate
or uncertain. Procedures for generation of decision rules adapt the general
induction scheme.

3 Rough Approximations by Reflexive Fuzzy

Similarity Relations
3.1 Basic Concepts
Let U be a finite non empty set of objects, called universe, and R a fuzzy
reflexive binary relation defined on U , which represents a certain form of
similarity. Let X be a fuzzy set in U and let also µX : U −→ [0, 1] be the
membership function of X. Given x ∈ U we say that:

1. the membership degree of x to the set of positive objects with respect

to X, denoted by P os(x, X), is the credibility that “for each y ∈ U ,
y is not similar to x and/or y belongs to X”, i.e.,

P os(x, X) = Ty∈U T ∗ (N (R(x, y)), µX (y));

2. the membership degree of x to the set of negative objects with respect

to X, denoted by N eg(x, X), is the credibility that “for each y ∈ U ,
158 S. Greco, B. Matarazzo, R. Slowinski

y is not similar to x and/or y does not belong to X”, i.e.,

N eg(x, X) = Ty∈U T ∗ (N (R(x, y)), N (µX (y))).

Let us remark that, remembering the definition of T ∗ -implication, we

can write

P os(x, X) = Ty∈U IT→∗ ,N (R(x, y), µX (y)), (12)

N eg(x, X) = Ty∈U IT→∗ ,N (N (R(x, y)), N (µX (y))). (13)

On the basis of (12), P os(x, X) can be seen as the credibility that

“for each y ∈ U the similarity of x to y implies that y belongs to X”.
Analogously, from (13), N eg(x, X) can be seen as the credibility that “for
each y ∈ U the ‘non similarity’ of x to y implies that y does not belong to
X”.
Considering a subset X ⊆ U and a reflexive binary relation R defined on
U , the lower approximation of X, denoted by R(X), and the upper approx-
imation of X, denoted by R(X), are fuzzy subsets of U whose membership
functions are respectively defined as

µ(x, R(X)) = Ty∈U T ∗ (N (R(x, y)), µX (y))),

∗
µ(x, R(X)) = Ty∈U T (R(x, y), µX (y)).

Theorem 3.1 For each x ∈ U we have: µ(x, R(X)) ≤ µX (x) ≤

µ(x, R(X)).

Proof. From property (1) we have for each x ∈ U

µ(x, R(X)) = Ty∈U T ∗ (N (R(x, y)), µX (y))

≤ min T ∗ (N (R(x, y)), µX (y)). (14)
y∈U

Since minT ∗ (N (R(x, y)), µX (y)) ≤ T ∗ (N (R(x, x)), µX (x)) and from the
y∈U
reflexivity of R we have

T ∗ (N (R(x, x)), µX (x)) = T ∗ (N (1), µX (x)) = T ∗ (0, µX (x)) = µX (x),

from (14) we obtain

µ(x, R(X)) ≤ µX (x).
Rough Set Processing of Vague Information 159

Analogously, from property (2), we have: µ(x, R(X)) = Ty∈U∗ T (R(x, y),

µX (y)) ≥ maxT (R(x, y), µX (y)) ≥ T (R(x, x), µX (x)) = T (1, µX (x)) =
y∈U
µX (x). ✷

Theorem 3.1 can be read as the fuzzy counterpart of the fact that set X
includes its P-lower approximation and is included in its P-upper approxi-
mation.
Let us observe that µ(x, R(X)) = P os(x, X) while µ(x, R(X)) can be
seen as the credibility that “there is at least one y ∈ U such that y belongs
to X and x is similar to y”.

Theorem 3.2 Considering a fuzzy set X in U and a fuzzy binary reflexive

relation R defined on U , if N is involutory and (T ∗ , T, N ) is a De Morgan
triple, then R(X) is the complement of the set of negative objects with
respect to X, i.e., µ(x, R(X)) = N (N eg(x, X)).

Proof. From De Morgan law we have µ(x, R(X)) = Ty∈U ∗ T (R(x, y),
∗ N (T ∗ (N (R(x, y)), N (µ (y)))) = N (T
µX (y)) = Ty∈U ∗
X y∈U T (N (R(x, y)),
N (µX (y)))) = N (N eg(x, X)). ✷

Theorem 3.3 Considering a fuzzy set X in U and a fuzzy binary reflexive

relation R defined on U , if (T ∗ , T, N ) is a De Morgan triple, then R(X) is
the complement of the set of R(x, U − X), i.e., µ(x, (X)) = N (µ(x, R(U −
X)).

Proof. From De Morgan law we have µ(x, R(X)) = Ty∈U T ∗ (N (R(x, y)),
µX (y)) = N (Ty∈U ∗ N (T ∗ (N (R(x, y)), µ (y)))) = N (T ∗ T (R(x, y),
X y∈U
N (µX (y)))) = N (µ(x, R(U − X)). ✷

Theorem 3.3 expresses in fuzzy terms the following well-known comple-

mentarity property of the rough sets theory: the P-lower (P-upper) approx-
imation of set X is the complement of the P-upper (P-lower) approximation
of its complementary set U − X.

3.2 Modelling Fuzzy Similarity Relations

In this section, for each attribute q ∈ Q we consider a valued binary relation
Rq , i.e., a function Rq : U × U −→ [0, 1], where for all x, y ∈ U, Rq (x, y)
represents the intensity or degree of similarity of x to y with respect to
attribute q. More precisely, for each q ∈ Q and for all x, y ∈ U we have:
160 S. Greco, B. Matarazzo, R. Slowinski

Rq (x, y) = 0 means that there is no similarity of x to y with respect

to attribute q, for all x, y ∈ U ,
Rq (x, y) = 1 means that x is definitely similar to y with respect to
attribute q, for all x, y ∈ U,
Rq (x, y) ≥ Rq (w, z) means that with respect to attribute q the simi-
larity of x to y is at least as credible as the similarity of w to z, for
all x, y, w, z ∈ U .
The similarity classes defined by the valued binary relation Rq cor-
respond to fuzzy granules of knowledge; their credibility varies gradually
between a finite degree equal to one and an infinitely small degree, thus
giving different credibility of rough approximations.
The similarity Rq should satisfy the following mimimal condition:
[f (x, q) = f (y, q)] =⇒ Rq (x, y) = 1,
for all x, y ∈ U and for each q ∈ C.
If Vq = (vq∗ , vq∗ ) ⊆ R, then the similarity Rq should satisfy also the
condition
[f (x, q) ≤ f (y, q) ≤ f (w, q) ≤ f (z, q)] ⇔ Rq (y, w) ≤ Rq (x, z),
for all x, y, w, z ∈ U and for each q ∈ C.
We propose to model the similarity binary relation Rq in the following
way for each q ∈ Q. Let εq1 and εq2 be two functions such that εq1 : Vq −→
R+ , εq2 : Vq −→ R+ and εq1 (f (x, q)) ≤ εq2 (f (x, q)) for each x ∈ U . Given
x, y ∈ U

 1, if |f (x, q) − f (y, q)| ≤ εq1 (f (y, q)),
Rq (x, y) = 0, if |f (x, q) − f (y, q)| > εq2 (f (y, q)),

linear, between the two bounds.
The following form of the functions εq1 , εq2 can be adopted:
εq1 (f (x, q)) = αq1 f (x, q) + βq1 , (15)
εq2 (f (x, q)) = αq2 f (x, q) + βq2 . (16)
To model the comprehensive similarity of a to b ∈ U with respect to
P = {q1 , q2 , . . . , qp } ⊆ Q, denoted by RP (a, b), we consider the credibility
of the proposition “f (a, q1 ) is similar to f (b, q1 ) with respect to q1 and
f (a, q2 ) is similar to f (b, q2 ) with respect to q2 and . . . f (a, qp ) is similar to
f (b, qp ) with respect to qp ”. Thus, we obtain
RP (x, y) = Tq∈P Rq (x, y).
Rough Set Processing of Vague Information 161

Theorem 3.4 For all x, y ∈ U and for each P ⊆ T ⊆ C, RP (x, y) ≥

RT (x, y).

Proof. Let us suppose that T = P ∪ {q} with q ∈ C − P . Thus,

from property (1) of T-norms, we have, for all x, y ∈ U , RT (x, y) =
T (RP (x, y), Rq (x, y)) ≤ min{RP (x, y), Rq (x, y)} ≤ RP (x, y). Iterating the
reasoning we can prove the result for each P ⊆ T ⊆ C. ✷

Theorem 3.5 For each x ∈ U , for each X ⊆ U and for each P ⊆ T ⊆ C,

we have

µ(x, RP (X)) ≤ µ(x, RT (X)) and µ(x, RP (X)) ≥ µ(x, RT (X)).

Proof. From the monotonicity properties of T-norms, T-conorms and nega-

tion and from Theorem 3.4, we have for each x ∈ U , for each X ⊆ U and
for each P ⊆ T ⊆ C,

µ(x, RP (X)) = Ty∈U T ∗ (N (RP (x, y)), µX (y)))

≤ Ty∈U T ∗ (N (RT (x, y)), µX (y)) = µ(x, RT (X)),
∗
µ(x, RP (X)) = Ty∈U T (R(x, y), µX (y))
∗
≥ Ty∈U T (R(x, y), µX (y)) = µ(x, RT (X)),

which proves the theorem. ✷

Theorem 3.5 expresses in a fuzzy context the following property of the

classical rough sets: using greater sets of attributes, it is possible to obtain
more accurate approximations of X: thus, while in the classical rough
sets theory the lower approximation becomes greater (more precisely, not
smaller) and the upper approximation becomes smaller (more precisely,
not greater), in the fuzzy case the membership of the lower approximation
increases (more precisely, it does not decrease) and the membership of the
upper approximation decreases (more precisely, it does not increase).

4 Rule Induction
Using the fuzzy similarity relation one is able to induce decision rules from
a decision table. In the following we consider certain and possible decision
rules. The syntax of a certain decision rule is the following:
162 S. Greco, B. Matarazzo, R. Slowinski

“if f (x, q1 ) is similar to rq1 and f (x, q2 ) is similar to rq2 and

. . . f (x, qp ) is similar to rqp , then x belongs to X”,
where {q1 , q2 , . . . , qp } ⊆ C, (rq1 , rq2 , . . . , rqp ) ∈ Vq1 × Vq2 × . . . × Vqp and
X ⊆ U.
The syntax of a possible decision rule is the following:
“if f (x, q1 ) is similar to rq1 and f (x, q2 ) is similar to rq2 and
. . . f (x, qp ) is similar to rqp , then x could belong to X”,
where {q1 , q2 , . . . , qp } ⊆ C, (rq1 , rq2 , . . . , rqp ) ∈ Vq1 × Vq2 × . . . × Vqp and
X ⊆ U . To each decision rule r there is associated a credibility ℵ(r).
A statement r: “if f (x, q1 ) is similar to r1 and f (x, q2 ) is similar to
r2 . . . and f (x, qp ) is similar to rp , then x belongs to X”, is accepted as a
certain decision rule with a credibility equal to ℵ(r) if there is at least one
w ∈ U such that f (w, q1 ) = r1 and f (w, q2 ) = r2 , . . . , f (w, qp ) = rp , and

T (µ(w, RP (X)), Ty∈U T ∗ (N (RP (y, w)), µX (y))) = ℵ(r) > 0,

where P = {q1 , q2 , . . . , qP }.
Let us observe that ℵ(r) can be interpreted as the credibility of the
proposition “w belongs to the P-lower approximation of X and all the
objects similar to it also belong to X”.
A statement r: “if f (x, q1 ) is similar to rq1 and f (x, q2 ) is similar to rq2
and . . . f (x, qp ) is similar to rqp , then x could belong to X”, is accepted as
a possible decision rule with a credibility equal to ℵ(r) if there is at least
one w ∈ U such that f (w, q1 ) = r1 and f (w, q2 ) = r2 , . . . , f (w, qp ) = rp ,
and
∗
T (µ(w, RP (X)), Ty∈U T (RP (y, w), µX (y)))
∗
= T (µ(w, RC (X)), Ty∈U T (RC (y, w), µX (y))) = ℵ(r) > 0,

where P = {q1 , q2 , . . . , qp }.
Let us observe that, in this case, ℵ(r) can be interpreted as the credi-
bility of the proposition “w belongs to the upper approximation of X and
there is at least one object similar to it that also belongs to X”.
The different formulation of accepted certain and possible decision rules
depends on Theorem 3.5. In simple words, it says that the credibility that
an object x ∈ U belongs to the upper approximation of a set X ⊆ U
does not increase, and very often decreases, when considering a larger set
of attributes. Thus, considering a smaller set of attributes there is the
possibility of overestimating the credibility that an object x ∈ U belongs to
Rough Set Processing of Vague Information 163

the upper approximation of a set X ⊆ U . Therefore, an accepted possible

decision rule may be built on object w ∈ U from an upper approximation
with respect to set P ⊆ C, however, its credibility cannot be greater than
the credibility of a possible decision rule built on the same object w ∈ U
from an upper approximation with respect to set C.
A certain decision rule r: “if f (x, q1 ) is similar to rq1 and f (x, q2 ) is
similar to rq2 . . . and f (x, qp ) is similar to rqp then x ∈ X”, whose credibility
is equal to ℵ(r), is called minimal if there is no other certain decision
rule s: “if f (x, h1 ) is similar to s1 and f (x, h2 ) is similar to s2 . . . and
f (x, hk ) is similar to sk then x ∈ X”, whose credibility is ℵ(s), such that
{h1 , . . . , hk } ⊆ {q1 , . . . , qp }, rh1 = sh−1 , rh2 = sh2 ,. . . , rhk = shk and ℵ(s) ≥
ℵ(r).
A possible decision rule r: “if f (x, q1 ) is similar to r1 and f (x, q2 ) is
similar to r2 . . . and f (x, qp ) is similar to rp then x ∈ X”, whose credibility
is ℵ(r) is called minimal if there is no other possible or certain decision
rule s: “if f (x, h1 ) is similar to s1 and f (x, h2 ) is similar to s2 . . . and
f (x, hk ) is similar to sk then x ∈ X”, whose credibility is ℵ(s), such that
{h1 , . . . , hk } ⊆ {q1 , . . . , qp }, rh1 = sh1 , rh2 = sh2 ,. . . , rhk = shk and ℵ(s) ≥
ℵ(r).
Let us observe that, since each decision rule is an implication, the min-
imal decision rules represent the implications such that there is no other
implication with an antecedent at least of the same weakness and a conse-
quent of at least the same strength.

5 An Illustrative Example
In order to illustrate the methodology proposed in this paper, let us consider
a simple example. First, we will solve this example using the classical rough
set approach based on indiscernibility and crisp granules of knowledge.
Then, we will solve it using the extended rough set approach based on fuzzy
similarity relations and fuzzy granules of knowledge. The example is based
on a decision table describing six firms which have got an approximately
equal credit in a bank. The firms are characterized by three condition
attributes: A1 = value of fixed capital, A2 = value of sales in the year
preceding the application, A3 = kind of activity. Attributes A1 and A2 are
numerical ones, while attribute A3 is a qualitative one with three possible
values.
Decision attribute d makes a partition of the firms as follows: d = “yes”
if the firm paid back its debt, and d = “no” otherwise.
164 S. Greco, B. Matarazzo, R. Slowinski

The decision table is shown in Table 1.

Table 1: Original decision table

Firm A1 A2 A3 d
F1 120 60 A yes
F2 140 72 A no
F3 170 80 A yes
F4 99 49.5 B no
F5 101 50.5 B yes
F6 90 45 B no

In order to apply the classical rough set approach to the above decision
table the values of numerical attributes A1 and A2 must be translated
into some nominal terms, e.g., low, medium or high. This translation,
called coding, involves a division of the original domain of the numerical
attributes into subintervals to which nominal terms are assigned. In our
example the following boundary values were adopted for the subintervals:

fixed capital < 100 =⇒ fixed capital is low,

100 < fixed capital < 150 =⇒ fixed capital is medium,
fixed capital ≥ 150 =⇒ fixed capital is high,
values of sales < 50 =⇒ value of sales is low,
50 < values of sales < 75 =⇒ value of sales is medium,
value of sales ≥ 100 =⇒ value of sales is high.

Therefore, the coded decision table shown in Table 2 was obtained.

Table 2: Coded decision table

Firm A1 A2 A3 d
F1 medium medium A yes
F2 medium medium A no
F3 high high A yes
F4 low low B no
F5 medium medium B yes
F6 low low B no

Denoting by XY the set of firms which paid back the debt and by XN
the set of firms which did not pay back the debt, the lower and the upper
approximations of the set of firms paying back the debt are C(XY ) =
{F 3, F 5} and C(XY ) = {F 1, F 2, F 3, F 5}, while the lower and the upper
Rough Set Processing of Vague Information 165

approximations of the set of firms not paying back the debt are C(XN ) =
{F 4, F 6} and C(XN ) = {F 1, F 2, F 4, F 6}.
Let us observe that F1 and F2 belong to the upper approximations of
XY and XN , i.e., to the boundary between XY and XN because, in the
coded decision table, they have the same description with respect to the
condition attributes from C but they belong to different classes according to
the decision attribute d, i.e., F4 paid back and F6 did not pay back the debt.
Let us observe, moreover, that even if in the coded decision table, F4 and
F5 have quite different description, the original decision table shows that
F4 and F5 have quite similar values on the two first condition attributes.
This is the result of the “boundary effect” of discretization due to the fact
that the thresholds between the intervals of low fixed capitals and medium
fixed capitals is between the corresponding evaluations of F4 and F5 and
the same happens with respect to the intervals of low values of sales and
medium values of sales.
This boundary effect can be avoided when considering the similarity
relation instead of the indiscerniblity relation. For instance, let us assume
that: x is similar to y with respect to A1 (we write xR1 y) iff

|fixed capital of x − fixed capital of y|

≤ 10%,
fixed capital of y

x is similar to y with respect to A2 (we writw xR2 y) iff

|value of sales of x − value of sales of y|

≤ 10%,
value of sales of y

and x is similar to y with respect to A3 (we writw xR3 y) iff x and y have

the same kind of activity.
Furthermore, we state that x is similar to y with respect to C, and
denote this relation by xRC y iff xR1 y, xR2 y, and xR3 y.
Therefore, we have RC = {(F 1, F 1), (F 2, F 2), (F 3, F 3), (F 4, F 4),
(F 4, F 5), (F 4, F 6), (F 5, F 4), (F 5, F 5), (F 6, F 4), (F 6, F 6)} and
−1 −1 −1 −1
RC (F 1) = {F 1}, RC (F 2) = {F 2}, RC (F 3) = {F 3}, RC (F 4) =
−1 −1
{F 4, F 5, F 6}, RC (F 5) = {F 4, F 5}, RC (F 6) = {F 4, F 6}. The lower
and upper approximations of the set of firms paying back the debt are
RC (XY ) = {F 1, F 3} and RC (XY ) = {F 1, F 3, F 4, F 5}, while the lower
and the upper approximations of the set of firms not paying back the debt
are RC (XN ) = {F 2, F 6} and RC (XN ) = {F 2, F 4, F 5, F 6}. Let us observe
that in this case F4 is similar to F5 and F5 is similar to F4, with the conse-
quence that F4 and F5 belong to the boundary between XY and XN rather
166 S. Greco, B. Matarazzo, R. Slowinski

than to the lower approximation of XN and XY , respectively. This result

was obtained because the use of a similarity relation permitted to avoid the
boundary effect typical to discretization of numerical attributes and to the
use of the indiscerniblity relation.
However, let us observe that a more subtle boundary effect is hidden in
the proposed example. In fact F4 is similar to F6, but F5, which in turn is
very similar to F4, is not similar to F6. It is enough a small change in the
definition of similarity with respect to A1 and A2 to obtain rather different
results. For instance, let us suppose that xR1 y iff

|fixed capital of x − fixed capital of y|

≤ 12%,
fixed capital of y

and xR2 y iff

|value of sales of x − value of sales of y|

≤ 12%.
value of sales of y

We obtain RC = {(F 1, F 1), (F 2, F 2), (F 3, F 3), (F 4, F 4), (F 4, F 5),

−1
(F 4, F 6), (F 5, F 4), (F 5, F 5), (F 6, F 4), (F 6, F 5), (F 6, F 6)}, RC (F 1) =
−1 −1 −1
{F 1}, RC (F 2) = {F 2}, RC (F 3) = {F 3}, RC (F 4) = {F 4, F 5, F 6},
−1 −1
RC (F 5) = {F 4, F 5}, RC (F 6) = {F 4, F 5, F 6}. The lower and upper
approximations of the set of firms paying back the debt are RC (XY ) =
{F 1, F 3} and RC (XY ) = {F 1, F 3, F 4, F 5, F 6}, while the lower and the
upper approximations of the set of firms not paying back the debt are
RC (XN ) = {F 2} and RC (XN ) = {F 2, F 4, F 5, F 6}. Let us observe that in
this case F5, which belongs to XY , is similar to F6, which belongs to XN ,
and, therefore, F6 does not belong to the lower approximation of XY as in
the previous case but, instead, it belongs to the boundary of XY and XN .
The rather large sensitivity of the rough approximations to small changes
in the definition of the similarity relation suggests to introduce a graduality
in the concept of similarity and thus justifies a rough approximation based
on a fuzzy similarity.
Therefore, let us consider the following definition of fuzzy similarity: x
is similar to y with respect to A1 with credibility R1 (x, y), where


 capital of x − fixed capital of y|
|fixed

 1, if fixed capital of y ≤ 10%,
R1 (x, y) = |fixed capital of x − fixed capital of y|

 0, if > 15%,

 fixed capital ofy
linear, between the two bounds
Rough Set Processing of Vague Information 167

x is similar to y with respect to A2 with credibility R2 (x, y), where



 of sales of x − value of sales of y|
|value

 1, if value of sales of y ≤ 10%,
R2 (x, y) = |value of sales of x − value of sales of x| .

 0, if > 15%,

 value of sales ofy
linear, between the two bounds

Furthermore, we consider the following fuzzy logical operators:

T (a, b) = min(a, b), for all a, b ∈ [0, 1],

T ∗ (a, b) = max(a, b), for all a, b ∈ [0, 1],
N (a) = 1 − a, for each a ∈ [0, 1].

The following results were obtained. The fuzzy similarity RC (x, y) is

presented in Table 3. We calculated the lower and upper approximation of
XY and XN obtaining the results presented in Table 4.

Table 3: Fuzzy binary relation RC (x, y)

F1 F2 F3 F4 F5 F6
F1 1 0 0 0 0 0
F2 0 1 0 0 0 0
F3 0 0 1 0 0 0
F4 0 0 0 1 1 1
F5 0 0 0 1 1 .56
F6 0 0 0 1 .82 1

Table 4: Lower and upper approximations with respect to {A1, A2, A3}
Firm µ(x, R(XY )) µ(x, R(XN )) µ(x, R(XY )) µ(x, R(XN ))
F1 1 0 1 0
F2 0 1 0 1
F3 1 0 1 0
F4 0 0 1 1
F5 0 0 1 1
F6 0 0.18 0.82 1

From the decision table in Table 1 the following exact decision rules
were induced (between parentheses there is the credibility of the decision
rule):
168 S. Greco, B. Matarazzo, R. Slowinski

1. if fixed capital is similar to 140, then the firm does not pay back its
debt (credibility .86);

2. if fixed capital is similar to 90, then the firm does not pay back its
debt (credibility .18);

3. if fixed capital is similar to 120, then the firm pays back its debt
(credibility .86);

4. if fixed capital is similar to 170, then the firm pays back its debt
(credibility 1);

5. if value of sales is similar to 45, then the firm does not pay back its
debt (credibility .18);

6. if value of sales is similar to 60, then the firm does not pay back its
debt (credibility 1).

Furthermore, from the decision table in Table 1 the following possible

decision rules were induced:

7. if fixed capital is similar to 140, then the firm could not pay back its
debt (credibility 1);

8. if fixed capital is similar to 99, then the firm could not pay back its
debt (credibility 1);

9. if fixed capital is similar to 99, then the firm could pay back its debt
(credibility 1);

10. if fixed capital is similar to 90, then the firm could not pay back its
debt (credibility 1);

11. if fixed capital is similar to 90, then the firm could pay back its debt
(credibility .18);

12. if fixed capital is similar to 120, then the firm could pay back its debt
(credibility 1);

13. if fixed capital is similar to 101, then the firm could pay back its debt
(credibility 1);

14. if fixed capital is similar to 101, then the firm could not pay back its
debt (credibility 1);
Rough Set Processing of Vague Information 169

15. if value of sales is similar to 72, then the firm could not pay back its
debt (credibility 1);

16. if value of sales is similar to 49.5, then the firm could not pay back
(credibility 1);

17. if value of sales is similar to 49.5, then the firm could pay back its
debt (credibility 1);

18. if value of sales is similar to 45, then the firm could not pay back its
debt (credibility 1);

19. if value of sales is similar to 45, then the firm could pay back its debt
(credibility .18);

20. if value of sales is similar to 80, then the firm could not pay back its
debt (credibility 1);

21. if value of sales is similar to 50.5, then the firm could pay back its
debt (credibility 1);

22. if value of sales is similar to 50.5, then the firm could not pay back
its debt (credibility 1);

23. if kind of activity is A, then the firm could not pay back its debt
(credibility 1);

24. if kind of activity is A, then the firm could pay back its debt (credi-
bility 1);

25. if kind of activity is B, then the firm could not pay back its debt
(credibility 1);

26. if kind of activity is B, then the firm could pay back its debt (credi-
bility 1).

Let us remark that the credibility of each decision rule must be inter-
preted in relation with the different syntax of the exact and approximate
decision rule. For example, decision rule 1 and 7 have the same condition
part (“if fixed capital is similar to 140”) and a very similar decision part
(rule 1: “then the firm does not pay back its debt”, and rule 7: “then the
firm could not pay back its debt”). However, the credibility of the two rules
is different: 0.86 for rule 1 and 1.0 for rule 7. This means that the implica-
tion of the rule 1, being an exact decision rule, is not fully credible. This is
170 S. Greco, B. Matarazzo, R. Slowinski

because there is some firm which at least partly satisfies the condition part
of the rule without belonging to the decision class suggested by the rule:
more precisely, with respect to rule 1, firm F1 has a fixed capital partly
similar to 140, but it paid back its debt. On the contrary, the implication
of the decision rule 7, being a possible decision rule, is fully credible. This
is because there is at least one firm which satisfies the condition part of the
rule and belongs to the decision class suggested by the rule: more precisely,
with respect to rule 7, firm F2 has a fixed capital similar (equal) to 140
and it does not pay back its debt.

6 Conclusions
We introduced rough approximations of fuzzy sets by means of similarity
relations defined as reflexive fuzzy binary relations. The general framework
proposed represents a theoretical extension of the rough set approach into
a fuzzy context and also a generalization of the indiscernibility relations to
more general binary relations. The similarity classes defined by the fuzzy
similarity relation correspond to fuzzy granules of knowledge; their credi-
bility varies gradually between a finite degree equal to one and an infinitely
small degree, thus giving different credibility of the rough approximations.
Decision rules induced from these rough approximations have a more gen-
eral syntax and deal more naturally with numerical attributes. Due to
flexibility introduced by fuzzy sets and similarity relations, this new ap-
proach to rough set analysis gives more credible results, as shown by an
exemplary problem.

Acknowledgement: The research of the first two authors has been

supported by the Italian Ministry of University and Scientific Research
(MURST). The third author wishes to acknowledge financial support from
State Committee for Scientific Research, KBN research grant no. 8 T11C
013 13, and from CRIT 2 Esprit Project no. 20288.

References
[1] D. Dubois, H. Prade, Fuzzy Sets and Systems: Theory and Applica-
tions, Academic Press, New York, 1980.

[2] D. Dubois, H. Prade, Rough fuzzy sets and fuzzy rough sets, Int. J.
of General Systems, 17 (1990), 191–200.
Rough Set Processing of Vague Information 171

[3] D. Dubois, H. Prade, Putting rough sets and fuzzy sets together, in
Intelligent Decision Support, Handbook of Applications and Advances
of the Rough Sets Theory (R. Slowinski, ed.), Kluwer, Dordrecht, 1992,
203–233.

[4] J. Fodor, M. Roubens, Fuzzy Preference Modelling and Multicriteria

Decision Support, Kluwer, Dordrecht, 1994.

[5] S. Greco, B. Matarazzo, R. Slowinski, The use of rough sets and fuzzy
sets in MCDM, Chapter 14 in Advances in Multiple Criteria Decision
Making (T. Gal, T. Hanne, T. Stewart, eds.), Kluwer, Dordrecht, 1999,
14.1–14.59.

[6] J. W. Grzymala-Busse, LERS – a system for learning from examples

based on rough sets, in Intelligent Decision Support. Handbook of Ap-
plications and Advances of the Rough Sets Theory (R. Slowinski, ed.),
Kluwer, Dordrecht, 1992, -3-18.

[7] T. Lin, Neighborhood systems and approximation in database and

knowledge base systems, in Proc. 4th International Symposium on
Methodologies for Intelligent Systems, 1989.

[8] S. Marcus, Tolerance rough sets, Cech topologies, learning processes,

Bull. of the Polish Academy of Sciences, Technical Sciences, 42 (1994),
471–487.

[9] K. Menger, Probabilistic theories of relations, Proc. Nat. Acad. Sci.

(Math.), 37 (1951), 178–180.

[10] R. Mienko, J. Stefanowski, K. Toumi, D. Vanderpooten, Discovery-

oriented induction of decision rules, Cahier du LAMSADE 141, Uni-
versitë de Paris Dauphine, Paris, 1996.

[11] Z. Pawlak, Rough sets, Intern. J. of Information and Computer Sci-

ences, 11 (1982), 341–356.

[12] Z. Pawlak, Rough Sets. Theoretical Aspects of Reasoning about Data,

Kluwer, Dordrecht, 1991.

[13] Z. Pawlak, Rough sets and fuzzy sets, Fuzzy Sets and Systems, 17
(1985), 99–102.
172 S. Greco, B. Matarazzo, R. Slowinski

[14] L. Polkowski, A. Skowron, J. Zytkow, Tolerance based rough sets, in

Soft Computing: Rough Sets, Fuzzy Logic, Neural Networks, Uncer-
tainty Management, Knowledge Discovery (T. Y. Lin, A. Wildberger,
eds.), Simulation Councils, Inc., San Diego, CA, 1995, 55–58.

[15] B. Schweizer, A. Sklar, Probabilistic Metric Spaces, North-Holland,

Amsterdam, 1983.

[16] A. Skowron, J. Stepaniuk, Generalized approximation spaces, in Soft

Computing: Rough Sets, Fuzzy Logic, Neural Networks, Uncertainty
Management, Knowledge Discovery (T. Y. Lin, A. Wildberger, eds.),
Simulation Councils, Inc., San Diego, CA, 1995, 18–21.

[17] R. Slowinski, A generalization of the indiscernibility relation for rough

set analysis of quantitative information, Rivista di matematica per le
scienze economiche e sociali, 15 (1993), 65–78.

[18] R. Slowinski, Rough set processing of fuzzy information, in Soft Com-

puting: Rough Sets, Fuzzy Logic, Neural Networks, Uncertainty Man-
agement, Knowledge Discovery (T. Y. Lin, A. Wildberger, eds.), Sim-
ulation Councils, Inc., San Diego, CA, 1995, 142–145.

[19] R. Slowinski, J. Stefanowski, Rough classification in incomplete in-

formation systems, Math. Comput. Modelling, 2, 10/11 (1989), 1347–
1357.

[20] R. Slowinski, J. Stefanowski, Handling various types of uncertainty

in the rough set approach, in Rough Sets, Fuzzy Sets and Knowledge
Discovery (W. P. Ziarko, ed.), Springer-Verlag, London, 1993.

[21] R. Slowinski, J. Stefanowski, Rough set reasoning about uncertain

data, Fundamenta Informaticae, 27 (1996), 229–243.

[22] R. Slowinski, J. Stefanowski, Rough family – software implementation

of the rough set theory, in Rough Sets in Knowledge Discovery (L.
Polkowski, A. Skowron, eds.), vol. 2, Physica-Verlag, Heidelberg, New
York, 1998, 581–586.

[23] R. Slowinski, D. Vanderpooten, Similarity relation as a basis for

rough approximations, in Advances in Machine Intelligence and Soft-
Computing (P. P. Wang, ed.), vol.IV, Duke University Press, Durham,
NC, 1997, 17–33.
Rough Set Processing of Vague Information 173

[24] R. Slowinski, D. Vanderpooten, A generalized definition of rough

approximation based on similarity, IEEE Transactions on Data and
Knowledge Engineering, 1999 (to appear).

[25] Y. Yao, S. Wong, Generalization of rough sets using relationships be-

tween attribute values, Proc. 2nd Annual Joint Conference on Infor-
mation Sciences, Wrightsville Beach, NC, 1995, 30–33.

[26] S. Weber, A general concept of fuzzy connectives, negations and im-

plications based on t-norms and t-conorms, Fuzzy Sets and Systems,
11 (1983), 115–134.

[27] W. Ziarko, D. Golan, D. Edwards, An application of DATALOGIC/R

knowledge discovery tool to identify strong predictive rules in stock
market data, Proc. AAAI Workshop on Knowledge Discovery in
Databases, Washington D.C., 1993, 89–101.
 Relativised Code Concepts
and Multi-Tube DNA Dictionaries
Tom Head
Mathematics Department, Binghamton University
Binghamton, New York 13902-6000 USA
E-mail: tom@math.binghamton.edu

Abstract. Concepts from the theory of variable length codes

are relativised to a language L so that demands are made only
for those strings that belong to L. Decision procedures and reg-
ularity results appropriate for the relativised forms are given.
This suggests the definition of a sequence of code types: the join
codes of level k (1 ≤ k). At level 1 we have precisely the familiar
comma free codes. For a join code of level k, a procedure for
segmenting messages into code words is observed that consists
of k steps, with each step being carried out in the elementary
comma free manner. An illustration of this segmentation pro-
cess is given using an idealization of the Watson-Crick base pair-
ing of complementary strands of DNA. In this process the code
words appear as ssDNA molecules in k distinct test tubes with
each test tube content being washed over each incoming ssDNA
complemented message strand. After k washings the sequence
of attached code words provides the required segmentation.

Preamble
The interplay between finiteness and infinity is particularly clear in studies
of the forms and the essence of computational processes. The specification
and manipulation of infinite sets of formal object is desired. In order to
maintain confidence in the precision of thought finite displays of symbols
are created that can be interpreted to generate or specify the desired infinite
sets. Infinite sets can often be compared by comparing the finite schemes

175
176 T. Head

that specify them. It is tempting to say that finite creatures are able in
this way to interact with the infinite. Perhaps this is the essential spirit of
mathematics. In the present article we use finite automata and finite sets of
code words to manipulate members from an infinite sets of strings – to each
of which we may ascribe a specific meaning. In this way an infinite number
of choices of distinct meanings is provided by the use of finite schemes. At
this point the goals of mathematics and linguistics fuse.

1 Introduction
We allow any subset of the set A∗ of all strings over a finite set A to be
called a code. In this context we refer to A as an alphabet. Any subset
of A∗ may also be called a language. Whether we choose to call a subset
of A∗ a code or a language is determined by the role we intend it to play
in our investigations. Elements of codes will be called words and elements
of languages will be called strings. For C a code, elements of C ∗ will be
called messages. In some expositions of the theory of codes, a subset C of
A∗ is not called a code unless it is uniquely decipherable as defined here in
Section 2. Among the references for this article, only [10] and [1] require
unique decipherability before a set of strings is allowed to be called a code.
By linking the relativisation of code properties [6] with the behaviors of
DNA molecules [7], a sequence of generalizations of the concept of a comma
free code [1] is developed and illustrated. The codes of this sequence are
the join codes of various levels. The comma free codes are precisely the
join codes of the first level. The join codes of level k allow the segmenta-
tion of messages to be made in a sequence of k steps for which each step
has the simplicity of a comma free segmentation. This segmentation pro-
cess is illustrated in Section 6 by using an idealization of the annealing of
complementary single stranded DNA molecules.
As background for the present article, only familiarity with the ele-
mentary aspects of the theory of regular languages and finite automata is
required. For this material see [15] or selected chapters in either [10] or
[1]. If references for the theory of codes are desired, see [10], [1], and [9].
Regarding DNA behavior, see [12]. In the two closing paragraphs of this
section our choice of automata notation is established and used to confirm
a regularity result concerning the set of factors of a language.
Let M = (A, Q, q0 , F, E) be a deterministic automaton where: Q is the
set of states of M ; q0 is the initial state of M ; F is the set of final states;
and E is the set of labeled edges of M , i.e., E is a subset of Q × A ×
Relativised Code Concepts and Multi-Tube DNA Dictionaries 177

Q. Such an automaton recognizes the language consisting of all strings

in A∗ that are the labels of paths that initiate at q0 and terminate at a
state in F . The language recognized by the automaton (A, Q, q0 , F, E)
will be denoted L(A, Q, q0 , F, E). Each such language is regular and each
regular language is recognized by such an automaton. Decision procedures
concerning regular languages commonly use an automaton recognizing the
language.
A string w in A∗ is a factor of a language L if there are strings x, y in
∗
A for which xwy is in L. Fac(L) will denote the set of all factors of L.
When L is regular, Fac(L) is also regular since, for L = L(A, Q, q0 , F, E),
Fac(L) is the union of the set of all languages L(A, Q
, p, {q}, E), where Q

is the set of all states in Q from which F is accessible and p, q are in Q

.

2 Relative Unique Decipherability

Definition 2.1 A code C is uniquely decipherable if, for each message
w in C ∗ , there is only one non-negative integer n and one finite sequence
c1 , c2 , . . . cn−1 , cn of words in C for which w = c1 c2 . . . cn−1 cn .

Definition 2.2 A code C is uniquely decipherable relative to a language

L if, for each string w in L, there is at most one non-negative integer
n and one finite sequence c1 , c2 , . . . cn−1 , cn of words in C for which w =
c1 c2 . . . cn−1 cn .

Note that a code C over an alphabet A is uniquely decipherable if it is

uniquely decipherable relative to either A∗ or C ∗ and that a uniquely deci-
pherable code is uniquely decipherable relative to every language contained
in A∗ . More generally, if C is uniquely decipherable relative to a language
L and S is a subset of L then C is also uniquely decipherable relative to S.
Observe that the code C = {ab, a, ba} is uniquely decipherable relative
to each of the first three of the following four languages, but not to the
fourth: (ab)∗ , a∗ + (ab + ba)∗ , (a + ab)∗ + (ab + ba)∗ , (a + ab)∗ + (a + ba)∗ .
The following proposition was demonstrated in [6], where an algorithm
based on deciding single-valuedness of a-transducers was given.

Proposition 2.1 For a given regular code C and regular language L, a

procedure exists for deciding whether C is uniquely decipherable relative to
L.

Algorithm. Let C be a regular code and L a regular language. Let

M = (A, Q, q0 , F, E) be the minimal deterministic automaton recognizing
178 T. Head

L. For each p in Q let I(p) = L(A, Q, q0 , {p}, E). For each p, q in Q let
L(p, q) = L(A, Q, p, {q}, E). For each q in Q let F (q) = L(A, Q, q, F, E).
Observe that C fails to be uniquely decipherable relative to L if and only if
there are p, q, r in Q for which each of the following seven regular languages
is not empty:

1. C ∗ ∩ I(p),

2. C ∩ F (r), which we call T (r) below,

3. C ∩ L(p, q), which we call C(p, q) below,

4. C ∩ L(p, r), which we call C(p, r) below,

5. L(q, r) ∩ A+ , which we call N (q, r) below,

6. C ∗ ∩ N (q, r)T (r), which we call T (q, r) below,

7. C(p, q)T (q, r) ∩ C(p, r)T (r).

Relative unique decipherability can therefore be decided by deciding the

emptiness of these intersections. ✷
The following known result is a special case of Proposition 2.1.

Corollary 2.1 For a regular code C, a procedure exists for deciding

whether C is uniquely decipherable.

3 Relative Solidity
Definition 3.1 A code C is a solid code if:

1. u and puq in C can hold only if pq is null; and

2. pu and uq in C and u non-null, can hold only if pq is null.

Definition 3.2 A code C is solid relative to a language L if:

1. w = ypuqz in L with both u and puq in C can hold only if pq is null;

and

2. w = ypuqz in L with pu and uq in C and u non-null, can hold only

if pq is null.
Relativised Code Concepts and Multi-Tube DNA Dictionaries 179

Note that a code C is solid if and only if C is solid relative to A∗ and

that a solid code is solid relative to every language contained in A∗ . More
generally, if C is solid relative to a language L and S is a subset of L then
C is also solid relative to S.
Observe that the code C = {ab, c, ba} is not solid, but it is solid with
respect to the language (abc + cba)∗ . It is not solid with respect to C ∗ =
(ab + c + ba)∗ .

Proposition 3.1 For a given regular code C and a regular language L, a

procedure exists for deciding whether C is solid relative to L.

Algorithm. Let X be the union of A+ CA∗ and A∗ CA+ . (1) of the

definition of relative solidity holds if and only if the intersection of the three
regular languages Fac(L), C, and X is empty. Since this intersection is reg-
ular, (1) can be decided by deciding the emptiness of this intersection. We
decide (2) of the definition only in the case that (1) holds. For the purpose
of deciding (2) we use the minimal automaton M = (A, Q, q0 , F, E) recog-
nizing C. Consider each ordered pair (s, t) of states in Q in turn. Associate
with each (s, t) in Q the four languages I(s) = L(A, Q, q0 , {s}, E), T (s) =
L(A, Q, s, F, E), I(t) = L(A, Q, q0 , {t}, E), and T (t) = L(A, Q, t, F, E). For
each of the ordered pairs (s, t), proceed as follows.
Decide whether the intersection of I(s) and T (t) contains a non-null
string. If so, call this intersection U (s, t); if not, do not consider (s, t) fur-
ther and proceed to an ordered pair not yet treated, if one remains. If none
remains, (2) holds and we terminate. If U (s, t) has been defined, decide
whether the intersection of the regular languages Fac(L) and I(t)U (s, t)T (s)
contains a string not in C. If so, (2) fails and we terminate; if not, pro-
ceed to an ordered pair not yet treated. If none remains, (2) holds and we
terminate. ✷
Observe that if C is a code that is solid relative to C ∗ then C is neces-
sarily uniquely decipherable. If C is not uniquely decipherable then there
is a violation of (1), of the definition of relative solidity, with p null and q
non-null. The concept of a comma free code is much older than the concept
of a solid code. The following definition, which is convenient in the present
context, is equivalent to the traditional definition:

Definition 3.3 A code C is comma free if it is solid relative to C ∗ .

Thus the new concept of relative solidity covers two classes of codes,
each of which has a considerable literature: solid codes [9], Section 11, and
180 T. Head

comma free codes [1]. The following known results are special cases of
Proposition 3.1.

Corollary 3.1 For a given regular code C, a procedure exists for deciding
whether C is a solid code.

Corollary 3.2 For a given regular code C, a procedure exists for deciding
whether C is a comma free code.

4 Joins Relative to a Language

Definition 4.1 A string w in A∗ is a join relative to a language L if w is
a factor of L and for every u, v in A∗ for which uwv is in L, both u and v
are also in L.

Proposition 4.1 The set J of joins relative to a regular language L is a

regular language.

Proof. Let L be a regular language over the alphabet A. Let M =

(A, Q, q0 , F, E) be the minimal deterministic automaton recognizing L. J
is regular since it is the complement in Fac(L) of the union of the set of lan-
guages recognized by the automata (A, Q, p, {q}, E) where p, q are in Q and
the following two conditions hold: (1) F is accessible from q; and (2) either
p is not in F or the language recognized by the automaton (A, Q, q, F, E)
is not contained in L. ✷

Definition 4.2 A word w in a code C is a join in C if w is a join relative

to C ∗ . J(C) will denote the set of all joins in the code C.

Corollary 4.1 The subset J(C) of all join words in a regular code C is a
regular language.

Observe that if C is uniquely decipherable then no word in J(C) can be

a factor of any other word in C: If v = puq were in C with u in J(C) and
pq non-null then, since both p and q would necessarily lie in C ∗ , v = puq
would contradict the unique decipherability of C.

Proposition 4.2 If a code C is uniquely decipherable, then J(C) is solid

relative to C ∗ .
Relativised Code Concepts and Multi-Tube DNA Dictionaries 181

Proof. Let C be a uniquely decipherable code. That condition (1), of

the definition of relative solidity, holds is a consequence of the stronger
observation immediately above. We consider condition (2). Suppose that
w = ypuqz is in C ∗ with both pu and uq in J(C) and u non-null. Then
y, yp, qz, z must all be in C ∗ . Consequently, in C ∗ we have the factorizations
w = (y)(pu)(qz) = (yp)(uq)(z). From pu & uq in C and u non-null, it
follows from the unique decipherability of C that pq is null. ✷

Corollary 4.2 For each uniquely decipherable code C, J(C) is comma free.

5 Reading Messages over a Join Code

Let C be a uniquely decipherable code over the alphabet A. Let C0 = C.
We define recursively a descending chain of subsets Ci (0 ≤ i) of C0 . With
Ci defined and J(Ci ) not empty, we define Ci+1 = Ci \J(Ci ). If a non-
negative integer k arises for which J(Ck ) is empty, then C = C0 is the
union of the pairwise disjoint subsets: J(C0 ), J(C1 ), . . . , J(Ck−1 ), Ck . If
no such integer k arises, then C = C0 is the union of an infinite sequence
of pairwise disjoint subsets: J(C0 ), J(C1 ), . . . , J(Ci ), . . . (0 ≤ i), and a
possibly nonempty subset D which would be the complement in C of the
union of all the subsets J(Ci ), (0 ≤ i). If C is regular then, by Proposition
4.1, all the sets arising in the decomposition process are regular, with the
possible exception of D.

Definition 5.1 A uniquely decipherable code C for which the decomposi-

tion procedure carried out above yields either a non-negative integer k with
Ck defined and empty, or an infinite descending chain Ci (0 ≤ i) with D
empty, will be called a join code of level k, in the first case, and of infinite
level, in the second case.

Note that the join codes of level one are precisely the comma free codes.
For these codes we have C1 empty, which is equivalent to C = C0 = J(C).
The join code concept enlarges the comma free concept into a hierarchy
of code classes each of which allows a parallel distributed segmentation
process that extends the natural segmentation process applicable to the
comma free codes.
We continue with the notations set in the first paragraph of this section
and explain this segmentation process. For maximum clarity we consider
first the case in which C = C0 is a finite uniquely decipherable code. In this
182 T. Head

case a non-negative integer k is certain to arise that provides a partition

of C into finite subsets J(C0 ), J(C1 ), . . . , J(Ck−1 ), Ck . Note that it follows
from Proposition 4.2 that, for each i (0 ≤ i ≤ k−1), J(Ci ) is solid relative to
Ci∗ . We may imagine the words of the code C as being listed in a dictionary
consisting of k + 1 successive volumes: J(C0 ), . . . , J(Ck−1 ), Ck . A message
w can be segmented into words as follows. Distributed agents, perhaps
one at each letter occurrence of the message, examine w independently
and in parallel using only volume 0 of the dictionary until all occurrences
in w of words in volume 0 have been enclosed in parentheses. There will
be no overlapping of these occurrences by Proposition 4.2. Agents now
examine independently the remaining unparenthesized subsegments of w
using only volume 1 of the dictionary – inserting parentheses bounding
occurrences of words that appear in volume 1. Similar processes takes
place in succession using volumes 2 through k − 1. If C is a join code of
level k then Ck is empty and w has been completely segmented into its code
words in a parallel distributed manner that involves only local features of
the message. On the other hand, if Ck is not empty then, at this point, we
have performed all of the segmentation that can be done purely locally by
non-communicating agents. The remainder of the segmentation must be
done by standard methods – the most primitive of which is to work from,
say, the left end of each remaining subsegment attempting matches with
words in Ck – backtracking when necessary.
Consider now the case of an infinite uniquely decipherable code C = C0 .
We will consider only the case in which C is a regular join code of infinite
level. The case in which C is a regular join code of finite level will be clear
after a discussion of the case in which the level is infinite. The purpose of
the regularity assumption is to allow the construction of the subsets Ci and
J(Ci ) (0 ≤ i). The restriction to join codes demands that D be empty and
assures that the method proposed here for segmentation of messages will
terminate. In these circumstances a segmentation process is possible that,
conceptually at least, differs very little from the case in which C is finite.
Let C = C0 be a uniquely decipherable regular join code of infinite level.
Each of subsets Ci , J(Ci ) (0 ≤ i) is regular and any finite number of these
can be constructed (in the form of recognizing automata) whenever needed.
Suppose now that we have a message w that we wish to segment into words
in C. We first note the length n of the message w. No word in C of length
greater than n can occur in w. For each i (0 ≤ i), let J
(Ci ) be a listing of
the words in J(Ci ) having length at most n. Each of these finite lists J
(Ci )
is constructible from an automaton recognizing J(Ci ). We regard each of
Relativised Code Concepts and Multi-Tube DNA Dictionaries 183

the lists J
(Ci ) as volume i of a multiple volume dictionary. To segment w
we construct, as need arises, a finite number of the finite volumes J
(C0 ),
J
(C1 ), . . . , J
(Ci ), . . . (0 ≤ i) required for segmenting w exactly as in the
previously treated case when C is finite; except that the use of J(Ci ) as
volume i is now replaced by the use of J
(Ci ) as volume i. That this process
terminates in the complete segmentation of w is guaranteed by the fact that
there are only finitely many code words of length at most n, and each of
these must lie in one of the subsets J
(Ci ), since D is empty.
Examples of join codes of each level are easily given for each alphabet
having at least two symbols. We begin with an example of a finite join code
of level three. Let the alphabet be A = {a, b, c} and the code be C = C0 =
{a, bab, cbabc}. For this code C the following are easily confirmed: J(C0 ) =
{cbabc}; C1 = {a, bab}; J(C1 ) = {bab}; C2 = {a}; J(C2 ) = C2 = {a}; and
C3 is empty. If a two symbol alphabet A = {0, 1} is desired, then the a, b,
and c may be replaced by 101, 1001, and 10001, respectively. An analogous
example of an infinite join code D of infinite level over alphabet A = {0, 1}
is D = {101, 10011011001, 100011001101100110001, . . .}. A finite join code
E of any desired level k is obtained by choosing the first k code words listed
in D.

6 A Conceptual Illustration Suggested by DNA

Complementary Base Pairing
As alphabet we will use the four symbols {A, C, G, T } that are commonly
used in biochemistry for denoting the four deoxyribonucleotides based on
Adenine, Cytosine, Guanine, and Thymine. We consider here how we
might communicate by transferring single stranded (ss) DNA molecules.
We would like to derive the meaning represented by an incoming ssDNA
molecule, m, by allowing the attachment to the message of (usually much
shorter) complementary ssDNA code words. These attachments will be ex-
pected to form spontaneously through the classical Watson-Crick comple-
mentary base pairings which are held together by hydrogen bonds. (Recall
that the pairings are A/T , C/G, G/C, and T /A.) The code molecules may
have easily recognizable (possibly phosphorescent) labels attached. The
meaning conveyed by the message molecule m is expressed by the sequence
of (labels of) the code word molecules that attach through hydrogen bonds
to the message molecule. How can we choose our code words to maximize
the coherence of this proposed decoding scheme? One method is to choose
the code word molecules in such a way that they form a join code.
184 T. Head

Suppose that C0 is a join code of level three over the alphabet

{A, C, G, T }. Then C0 is the union of J(C0 ), J(C1 ), J(C2), with C3 empty.
We regard the words in each of the J(Ci ) as labeled ssDNA molecules. Let
the molecules in each J(Ci ) be stored in aqueous solution in a test tube Ti .
When a message (a ssDNA molecule) arrives that is to be read (decoded),
we assume that it is the Watson-Crick complement of the concatenation
of a sequence of code word molecules lying in C0 . Thanks to the join
code feature of C0 the decoding process can be carried out in the following
sequence of three ultra-parallel steps: Hold the message molecule(s) per-
manently. Wash message with solution from T0 . Wash with solution from
T1 . Wash with solution from T2 . What is the result expected after these
three washes?
If we make the assumption that the code words attach where and only
where the Watson-Crick matching is perfect. During the wash with T0 ,
code words from T0 attach (globally in parallel) in each of a fully unique
set of locations, since J(C0 ) is solid relative to C0 . (These code words are
assumed to remain attached during the two remaining washes.) During the
wash with T1 , code words from T1 attach in each of a fully unique set of
locations, since J(C1 ) is solid relative to C1 . (These code words are assumed
to remain attached during the remaining wash.) During the wash with T2 ,
code words from T2 attach in each of a fully unique set of locations, since
J(C2 ) is comma free. Thus, after the three washes, all bases of the message
molecule are paired with bases of the attached code word molecules. The
meaning of the original message is expressed by the sequence of (labels of)
these attached code word molecules.
From the considerations above it is clear that (under hypotheses that
idealize the biochemical conditions) the decoding of ssDNA messages based
on a finite join code of level k can be carried out with k washes done in the
proper sequence. The multi-volume dictionaries of Section 5 have become
multi-tube DNA dictionaries of this section. The parallel dictionary look
up using volume i in Section 5 has become the i-th wash in this section.

7 Questions
In Sections 2 and 3 the concepts of unique decipherability and solidity were
relativised to regular languages. Which of the many additional code con-
cepts relativise in an interesting way? Is multiset decipherability [11], [13],
[8], [4] relative to a regular language decidable? Can the theory of deci-
pherability with respect to a variety of monoids, which has been developed
Relativised Code Concepts and Multi-Tube DNA Dictionaries 185

by F. Guzman [5] and by E. Gumustop [3], be relativised in an interesting

manner? Does the homophonic extension [14] of the code concept relativise
meaningfully?
Can an algorithm be given that will decide whether any given uniquely
decipherable infinite regular code is: (1) A join code? (2) A join code of
finite level?
In the first paragraph of Section 5 a subset D of a uniquely decipherable
infinite regular code C arose. When is D regular? Which (regular) subsets
can arise in the role of D?
One may wish to extend results of Sections 4 and 5 in either of two
ways: One way would be to allow context to be considered and the other
would be to relativise the join code concept with respect to a language L.
Formal results in Sections 4 and 5 can probably be generalized in either of
these ways, but their biomolecular interpretation, in the manner of Section
6, may be less satisfying.
For a fascinating new example of the use of DNA molecules as message
carriers, see [2].

Acknowledgments.
The results in this article arose in part from an effort to analyze theoret-
ically the potential of biomolecular operations used in various sticker ap-
proaches to DNA computing [12]. Partial support for this research through
NSF CCR- 9509831 and DARPA/NSF CCR-9725021 is gratefully acknowl-
edged.

References
[1] J. Berstel, D. Perrin, Theory of Codes, Academic Press, Orlando, FL,
1985.

[2] C. T. Clelland, V. Risca, C. Bancroft, Hiding messages in DNA mi-

crodots, Nature, 399 (1999), 533–534.

[3] E. R. Gumustop, Varieties of Codes and Codifiable Varieties of

Monoids, Doctoral Dissertation in Mathematics, Binghamton Univer-
sity – State University of New York, Binghamton, NY, 1997.

[4] F. Guzman, A complete list of small proper MSD and SD codes

(manuscript).
186 T. Head

[5] F. Guzman, Decipherability of codes, J. Pure and Applied Algebra (to

appear).

[6] T. Head, Unique decipherability relative to a language, Tamkang J.

Math., 11 (1980) 59–66.

[7] T. Head, Gh. Păun, D. Pixton, Language theory and molecular ge-
netics: generative mechanisms suggested by DNA recombination, in
Vol. II of Handbook of Formal Languages (G. Rozenberg, A. Salomaa,
eds.), Springer-Verlag, Berlin, 1997.

[8] T. Head, A. Weber, Deciding multiset decipherability, IEEE Trans.

Information Theory, 41 (1995), 291–297.

[9] H. Jurgensen, S. Konstantinidis, Codes, in Vol. I of Handbook of For-

mal Languages (G. Rozenberg, A. Salomaa, eds.), Springer-Verlag,
Berlin, 1997.

[10] G. Lallement, Semigroups and Combinatorial Theory, Wiley, New

York, NY, 1979.

[11] A. Lempel, On multiset decipherable codes, IEEE Trans. Information

Theory, 32 (1986), 714–716.

[12] Gh. Păun, G. Rozenberg, & A. Salomaa, DNA Computing. New Com-
puting Paradigms, Springer Verlag, Berlin, 1998.

[13] A. Restivo, A note on multiset decipherable codes, IEEE Trans. In-

formation Theory, 35 (1989), 662–663.

[14] A. Weber, T. Head, The finest homophonic partition and related code
concepts, IEEE Trans. Information Theory, 42 (1996), 1569–1575.

[15] S. Yu, Regular languages, in Vol. I of Handbook of Formal Languages

(G. Rozenberg, A. Salomaa, eds.), Springer-Verlag, Berlin, 1997.
 Uniformly Scattered Factors1
Lucian Ilie2
Turku Centre for Computer Science (TUCS)
Lemminkaisenkatu 14A, Data City 4th floor
FIN-20520, Turku, Finland
E-mail: lucili@utu.fi
Ion Petre
Turku Centre for Computer Science (TUCS) and
Department of Mathematics, University of Turku
FIN-20520, Turku, Finland
E-mail: ipetre@cs.utu.fi
Grzegorz Rozenberg
Department of Computer Science, Leiden University
P.O.Box 9512, 2300 RA Leiden, The Netherlands, and
Department of Computer Science
University of Colorado at Boulder
Boulder, CO 80309, U.S.A.
E-mail: rozenber@wi.leidenuniv.nl

Abstract. A word u appears as a factor of another word v as it

is: in one piece. When u is a subword of v, u may be scattered
as several factors. We consider the case in between and put
some restrictions on the number of factors as which u is allowed
to be scattered. A large class of partial orders which are gen-
eralizations of factors and subwords is obtained. Investigating
the borderline between their finite and infinite antichains, we
are able to fully characterize the property of being well partial
order. The result generalizes Higman’s theorem.
1
This research was partially done during the first author’s visit to Leiden University.
2
Research supported by the Academy of Finland, Project 137358. On leave of absence
from Faculty of Mathematics, University of Bucharest, Str. Academiei 14, R-70109,
Bucharest, Romania.

187
188 L. Ilie, I. Petre, G. Rozenberg

1 Introduction
Two of the most important partial orders on words are the factor and the
subword partial orders. If a word u is a factor of a word v, then u appears
inside v exactly as it is. If u is a subword of v, then u may appear inside
v scattered as several factors. It naturally arises the idea of considering
restrictions on the number of factors as which u is allowed to be scattered.
This idea is not new. It appeared first in a paper of Haines, cf. [2], who just
mentioned the possibility of putting a fixed upper bound on the number
of factors and notices that, in this way, one obtains a relation admitting
infinite antichains (sets containing pairwise incomparable elements only).
A deeper investigation of the problem is performed for the first time in
[4], where some general restrictions are imposed and some necessary and
sufficient conditions for the existence of the infinite antichains. In this
paper, we consider a different approach to the problem. Our restrictions
on the number of factors are similar with the ones of [4] but the way we
consider the factors is different: all factors but the last one have the same
length. In this way, all relations obtained are partial orders and, moreover,
we can iterate the generalization process, obtaining a very large class of
partial orders which are generalizations of factors and subwords. We then
fully characterize those which are well partial orders. As we deal with
well founded partial orders only, the problem consists of investigating the
borderline between finite and infinite antichains. The result generalizes
Higman’s theorem.

2 Basic Partial Orders

In this section, we perform the first step of our generalization. We fix first
some notations and concepts.
For an alphabet Σ, the set of all finite words over Σ is denoted Σ∗ ; the
empty word is ε. For a word w ∈ Σ∗ , |w| denotes the length of w. For two
words u, v ∈ Σ∗ , we say that u is a factor of v, denoted u ≤f v, if there
are x, y ∈ Σ∗ such that v = xuy; u is a subword of v, denoted u ≤s v, if
u = a1 a2 · · · an for some ai ∈ Σ, 1 ≤ i ≤ n, and v = v1 a1 v2 a2 · · · vn an vn+1 ,
for some vi ∈ Σ∗ , 1 ≤ i ≤ n + 1.
Consider a set A and a partial order ≤ on A. A chain of ≤ is a set B ⊆ A
such that, for any a, b ∈ B, either a ≤ b or b ≤ a. An antichain of ≤ is a set
B ⊆ A such that, for any a, b ∈ B, a ≤ b. The partial order ≤ is called well
founded if it has no infinite descending sequence a1 ≤−1 a2 ≤−1 a3 ≤−1 · · ·
Uniformly Scattered Factors 189

such that, for no i ≥ 1, ai ≤ ai+1 , where ≤−1 is the inverse of ≤; ≤ is called

well partial order if it is well founded and has no infinite antichain.
It is easy to see that both the factor partial order ≤f and the subword
partial order ≤s are well founded. Also, the factor partial order is not well
partial order as soon as card(Σ) ≥ 2, since it has infinite antichains, e.g.,
{abn a | n ≥ 1}. The famous theorem of Higman, cf. [3], states that this is
not the case for ≤s (in fact Higman [3] proved a much more general result;
Theorem 2.1 is a consequence of Higman’s result in the case of words).

Theorem 2.1 (Higman [3]) The subword partial order ≤s is well partial
order.

All the bounds we are to put on the number of factors as which a certain
word is allowed to be scattered are given by functions having as argument
the length of the word. We now construct the family G of all these functions.
It contains all g : N −→ R with the following properties ([r] denotes the
integer part or r):
(i) g(0) = 0, 1 ≤ g(n) ≤ n, for any n ≥ 1;
(ii) the function [g(n)] is increasing;
n
(iii) the function [g(n)] is increasing, where g(0) = 0 and g(n) = g(n) , for
any n ≥ 1.
For a non-empty word w ∈ Σ+ and a function g ∈ G, we define the
g-factorization of w as the factorization

w = w1 .w2 . · · · .w[g(|w|)] .w[g(|w|)]+1

where
|w1 | = |w2 | = · · · = |w[g(|w|)] | = [g(|w|)].
From the definition of the family G, the g-factorization of a non-empty
word is well defined. We may assume, for the sake of completeness, that
the g-factorization of the empty word has just one factor.
Next, for a function g ∈ G, we define the binary relation ≤f,g on Σ∗ as
follows.
(a) ε ≤f,g ε; ε ≤f,g w, for any w ∈ Σ∗ ; w ≤f,g ε implies w = ε;
(b) for two non-empty words u, v ∈ Σ+ , if their g-factorizations are, respec-
tively,
u = u1 .u2 . · · · .u[g(|u|)] .u[g(|u|)]+1 ,
v = v1 .v2 . · · · .v[g(|v|)] .v[g(|v|)]+1 ,
then u ≤f,g v if and only if (i)–(iii) below are fulfilled:
190 L. Ilie, I. Petre, G. Rozenberg

(i) |u| ≤ |v|;

(ii) there is a subsequence 1 ≤ k1 < k2 < · · · < k[g(|u|)] ≤ [g(|v|)] such that
ui ≤f vki , for any 1 ≤ i ≤ [g(|u|)];
(iii) if |u| = |v|, then u[g(|u|)]+1 = v[g(|v|)]+1 .

Remark 2.1 For any u, v ∈ Σ∗ , if u ≤f,g v and |u| = |v|, then u = v.

Also, for any g ∈ G, ≤f,g is well founded.
Consider next a few examples.
Example 2.1 Our first example shows that indeed our relations intro-
duced above (and which are partial orders as proved below) are generaliza-
tions of the factor and subword partial orders. Consider the two functions
1, 1N : N −→ N,
given by
1(0) = 0, 1(n) = 1, 1N (n) = n,
for any n ≥ 0. It is clear that we have
≤f,1 =≤f , ≤f,1N =≤s .
Example 2.2 Consider the function g1 (n) = min(n, 2), for any n ≥ 0.
Then, for the words u = aabab and v = bbaaabba we have the g1 -
factorizations respectively aa.ba.b and bb.aa.ab.ba and hence u ≤f,g1 v.
Also aa.ba ≤f,g1 ba.ab.ab.
Example 2.3 For the function
#
1, if 1 ≤ n ≤ 3,
g2 = n
3, if n ≥ 4,
we have that aab.abb.aba.ab ≤f,g2 baab.abbb.abaa but ab.aa.bb.a ≤f,g2
ab.aa.bb.b.
√
Example 2.4 Consider next a little bit stranger function, g3 (n) = n.
Clearly, g3 ∈ G. We have then, for instance, aba.aba.bba.bbbaba ≤f,g3
abba.abaa.baba.abba.
Lemma 2.1 For any g ∈ G, the relation ≤f,g is a partial order.
Proof. The reflexivity of the relation is obvious and the antisymmetry
follows from Remark 2.1. For the transitivity, we just have to use the
transitivity of the factor partial order ≤f . ✷
Uniformly Scattered Factors 191

3 Characterization of Well Partial Orders

We give in this section a characterization of those partial orders introduced
above which are well partial orders. As they are all well founded, we shall
be concerned only with the finiteness of antichains.
We first notice that, for any g ∈ G, since g is increasing, the limit
lim g(n) always exists in R ∪ {∞}.
n→∞

Theorem 3.1 If g ∈ G such that lim g(n) = ∞, then ≤f,g is not well
n→∞
partial order as soon as card(Σ) ≥ 2.

Proof. All we have to do is to find an infinite antichain for ≤f,g . As g(n)

goes to infinity with n, it follows that there is a subsequence (kn )n≥1 of
(n)n≥1 such that

2 ≤ g(k1 ) < g(k2 ) − 1 < g(k3 ) − 2 < g(k4 ) − 3 < · · ·

Consider then the following set of words over the two-letter alphabet {a, b}:

A = {wn = (ab[g(kn )]−2 a)[g(kn )] akn −[g(kn )][g(kn )] | n ≥ 1}.

We claim that A is an antichain of ≤f,g . Assume that there are 1 ≤ n < m

such that wn ≤f,g wm . For any n ≥ 1, the g-factorizations of wn is

wn = ab[g(kn )]−2 a.ab[g(kn )]−2 a. · · · .ab[g(kn )]−2 a.akn −[g(kn )][g(kn )] ,

and so, from the fact that wn ≤f,g wm , we get that ab[g(kn )]−2 a ≤f
ab[g(km )]−2 a, which is impossible since g(kn ) < g(km ) − 1 implies that
[g(kn )] < [g(km )]. The result is proved. ✷

We now prove that the converse of Theorem 3.1 hold true as well, thus
obtaining a characterization of the relations in the family (≤f,g )g∈G which
are well partial orders.

Theorem 3.2 If g ∈ G such that lim g(n) < ∞, then ≤f,g is well partial
n→∞
order.

Proof. Assume that lim g(n) = d, for some d ∈ R. Therefore, as g is

n→∞
increasing, there are c, n0 ∈ N, c ≥ 1, such that, for any n ≥ n0 , [g(n)] = c
(in fact, c ∈ {[d], [d] − 1}).
192 L. Ilie, I. Petre, G. Rozenberg

We argue by contradiction. Assume that ≤f,g is not a well quasi order

and consider an antichain of it, say {w1 , w2 , w3 , . . .}. We may assume that
n0 ≤ |w1 | < |w2 | < |w3 | < · · ·
Consider a new alphabet ∆ which has as letters the words of length
c over Σ, that is, ∆ = {w ∈ Σ∗ | |w| = c}. For any i ≥ 1, consider the
g-factorizations of wi wi = wi,1 .wi,2 . · · · .wi,[g(|wi |)] .wi,[g(|wi |)]+1 and denote
vi = wi,1 wi,2 · · · wi,[g(|wi |)] . Then the set {v1 , v2 , v3 , . . .} ⊆ ∆∗ , where each
word vi is viewed as a word over ∆ (its letters being the factors in the cor-
responding g-factorization), is an infinite antichain of the subword partial
order ≤s on ∆∗ . But this contradicts Higman’s theorem. ✷

From Theorems 3.1 and 3.2 we obtain the characterization of the rela-
tions in the family (≤f,g )g∈G which are well partial orders.

Theorem 3.3 For any g ∈ G, ≤f,g is well partial order if and only if
lim g(n) < ∞.
n→∞

4 Further Generalization
It is not difficult to see that in defining our general relations we may start
with any partial order instead of the factor partial order ≤f and still the
obtained relations are partial orders. In this way, a natural question arises:
what are the well partial orders if we start with the subword partial order
≤s instead of the factor one ≤f ? More generally, we can iterate the process
of defining our partial orders above. So, as ≤f,g1 is a partial order, for any
g1 ∈ G, we can use it as a start instead of ≤f . If the function of G we use
is g2 , then denote the obtained partial order by ≤f,g1 ,g2 . We then iterate
this process and get the most general partial orders which are of the form

≤f,g1 ,g2 ,...,gn , (1)

where n ≥ 1, gi ∈ G, for any 1 ≤ i ≤ n. We notice that starting with ≤s is

just a particular case here, namely, when g1 (n) = 1N .
In this section we make all preliminaries for the main result of the paper,
which characterizes those partial orders in (1) which are well partial orders.
As all partial orders in (1) are well founded, we shall investigate only the
existence of infinite antichains.
The following result on arbitrary partial orders is well known. We give
here a short proof for the sake of completeness.
Uniformly Scattered Factors 193

Lemma 4.1 Let A be an infinite set and ≤ a partial order on A. Then A

contains either an infinite antichain or an infinite chain of ≤.

Proof. Consider the set M of all maximal elements of A with respect to ≤.

M is antichain of ≤. It it is infinite, then we are done. It it is empty, then,
clearly, A contains an infinite chain of ≤. Assume M is finite non-empty.
Then there is x ∈ M such that the set N = {y ∈ A | y ≤ x} is infinite.
We now repeat the above reasoning with N − {x} instead of A, namely,
we consider the cardinality of the set of maximal elements of N − {x}. If
the above reasoning can be repeated indefinitely, that is, at each step, the
respective set of maximal elements is finite non-empty, then, again, we can
construct an infinite chain of ≤ in A. ✷

The next corollary follows immediately from Lemma 4.1.

Corollary 4.1 Let A be an infinite set and ≤ a partial order on A. If ≤

is well partial order, then A contains an infinite chain of ≤.

For any partial order ≤ on Σ∗ and any g ∈ G, we denote by ≤g the

relation obtained as in Section 2, using ≤ instead of ≤f .
For any positive integer c ≥ 1, consider the function c : N −→ N,
defined by c(n) = min(n, c), for any n ≥ 0. Obviously, c ∈ G.

Lemma 4.2 For any c ≥ 1 and any partial order ≤ on Σ∗ , the relation
≤c is well partial order.

Proof. Our idea is similar with the one used by Conway [1] for proving
Higman’s theorem and which is originally due to Nash-Williams, cf. [5].
Clearly, it is enough to prove that the relation

≤
c =≤c −{(u, v) | u ≤c v, |u| ≤ c}

has no infinite antichain. We argue by contradiction. Assume that there is

an infinite antichain of ≤
c , say

w1 , w2 , w3 , . . . (2)

It follows that

for any i, j ≥ 1, if i < j, then wi ≤

c wj . (3)

We may assume that (2) is an “earliest” sequence of words such that (3) is
fulfilled, i.e., w1 is a shortest word beginning a sequence satisfying (3), w2
194 L. Ilie, I. Petre, G. Rozenberg

is a shortest word such that w1 , w2 begin a sequence satisfying (3), and so

on and so forth.
Now, there exists a word w ∈ Σ∗ , |w| = c, such that infinitely many of
the words of (2) start with w as a prefix. Assume that the indices of all
such words are (ij )j≥1 such that 1 ≤ i1 < i2 < i3 < · · · and put, for any
j ≥ 1, wij = wuj . Construct then the sequence

w1 , w2 , . . . , wi1 −1 , u1 , u2 , . . . (4)

It is not difficult to check that the sequence (4) satisfies (3). Also, as c ≥ 1,
it is “earlier” than (2), a contradiction. The lemma is proved. ✷

The next result generalizes Theorem 3.2.

Lemma 4.3 For any partial order ≤ and any function g ∈ G such that
lim g(n) < ∞, the relation ≤g is well partial order.
n→∞

Proof. Assume lim g(n) = d, for some d ∈ R. Since g is increasing,

n→∞
there are c, n0 ∈ N, c ≥ 1, such that, for any n ≥ n0 , [g(n)] = c (in fact,
c ∈ {[d], [d] − 1}).
Assume that ≤g is not a well partial order. Then, there exists an infinite
antichain for it, say A = {w1 , w2 , w3 , . . .}. By possibly eliminating some of
the words of A, we may assume that n0 ≤ |w1 | < |w2 | < · · · .
Consider, for each wi , i ≥ 1, its g-factorization, say

wi = wi,1 .wi,2 . · · · .wi,[g(|wi |)] .wi,[g(|wi |)]+1

and denote, for any i ≥ 1,

vi = wi,1 wi,2 · · · wi,[g(|wi |)] .

Notice that |wi,j | = c, for any 1 ≤ j ≤ [g(|wi |)].

Fix a letter a ∈ Σ. We then claim that

B = {v1 a, v2 a, v3 a, . . .}

is an antichain of ≤c . To prove this, consider vi a, vj a ∈ B and assume that

i < j. Then, the c-factorizations of these two words are clearly

vi a = wi,1 .wi,2 . · · · .wi,[g(|wi |)] .a,

vj a = wj,1 .wj,2 . · · · .wj,[g(|wj |)] .a
Uniformly Scattered Factors 195

and hence, using the fact that |wi | < |wj |, vi a ≤c vj a would imply wi ≤g
wj , a contradiction. Therefore, B is an infinite antichain of ≤c , which
contradicts Lemma 4.2. ✷

For the rest of the paper, we shall make the following assumption: any
partial order ≤ on words we shall consider has the property that u ≤ v
implies |u| ≤ |v|. Remark that all partial orders ≤f,g1 ,...,gn do have this
property.
Consider an alphabet Σ and a partial order ≤ on Σ∗ . Denote by Σ∗ the
set of all tuples of words with equal length over Σ, that is,

Σ∗ = {(w1 , w2 , . . . , wn ) | n ≥ 1, wi ∈ Σ∗ , |w1 | = |w2 | = · · · = |wn |} ∪ {⊥},

where ⊥ stands for the empty tuple. We define the partial order ≤ on Σ∗
as follows: for any x = (x1 , x2 , . . . , xn ), y = (y1 , y2 , . . . , ym ) ∈ Σ∗ , we have
x≤y if and only if (i)–(iii) below are fulfilled (we assume that the length of
the components of ⊥ is zero):
(i) n ≤ m;
(ii) |x1 | ≤ |y1 |;
(iii) there is a subsequence 1 ≤ k1 < k2 < · · · < kn ≤ m such that xi ≤ yki ,
for any 1 ≤ i ≤ n.
It is easy to see that ≤ is a partial order on Σ∗ . The next lemma shows
that the property of being well is inherited from ≤ to ≤.

Lemma 4.4 If ≤ is a well partial order on Σ∗ , then ≤ is a well partial

order on Σ∗ .

Proof. By contradiction. Assume ≤ is not well partial order. We shall con-

struct an infinite antichain of ≤, thus contradicting the hypothesis. This
infinite antichain is constructed inductively, the main tool for the construc-
tion being the next claim.
Claim. Assume we have B ⊆ Σ∗ and C ⊆ Σ∗ which verify (∗)–(∗ ∗ ∗)
below:
(∗) B is an antichain of ≤;
(∗∗) C is an infinite antichain of ≤;
(∗ ∗ ∗) for any u ∈ B and v = (v1 , v2 , . . . , vn ) ∈ C, we have u ≤ vi , for any
1 ≤ i ≤ n.
Then, there is u ∈ Σ∗ and C
⊆ Σ∗ such that (∗)–(∗ ∗ ∗) above are
fulfilled with B
= B ∪ {u} and C
instead of B and C, respectively.
196 L. Ilie, I. Petre, G. Rozenberg

Assuming that the Claim holds, we construct an infinite antichain B of

≤ starting with B = ∅ and C an infinite antichain of ≤ (there exists such
a C by our assumption) and applying indefinitely the Claim.
Proof of Claim. Assume C ⊆ Σ∗ is an infinite antichain of ≤ and denote

C = {wn = (wn,1 , wn,2 , . . . , wn,l(n) ) | n ≥ 1}.

In virtue of Lemma 4.2, we may assume that, for any n < m, |wn,i | < |wm,j |.
For any n ≥ 1, denote by i(n) the smallest i, 1 ≤ i ≤ l(n), such that
w1,1 ≤ wn,l(n) ; if there is no such i, then set i(n) = 0 (of course, i(1) = 1).
If there are infinitely many n such that i(n) = 0, then B
= B ∪ {w1,1 }
and C
= {wn ∈ C | i(n) = 0} are good choices. It is important to notice
that when l(1) = 1, i.e., w1 = w1,1 , then i(n) = 0, for all n ≥ 2, and we can
construct B
and C
as required.
If there are finitely many n with i(n) = 0, then consider the set

C1 = {(wn,1 , wn,2 , . . . , wn,i(n)−1 ) | i(n) = 0}.

By Lemma 4.1, C1 contains either an infinite antichain or an infinite chain

of ≤; denote it C2 in either case. In the former case, take B
= B ∪ {w1,1 }
and C
= C2 ; clearly, B
and C
satisfy (∗)–(∗ ∗ ∗).
Consider the latter case and put

C3 = {wn,i(n) | (wn,1 , wn,2 , . . . , wn,i(n)−1 ) ∈ C2 }.

Applying Corollary 4.1 to C3 and ≤ we get an infinite chain of ≤ in C3 ,

say C4 . Therefore, the set

C5 = {(w1,2 , w1,3 , . . . , w1,l(1) )}

∪ {(wn,i(n)+1 , wn,i(n)+2 , . . . , wn,l(n) ) | wn,i(n) ∈ C4 }

is an infinite antichain of ≤ and also (∗)–(∗ ∗ ∗) are fulfilled with C5 in-

stead of C. Notice the essential difference between C and C5 : w1,1 has
been eliminated. We next repeat the same reasoning with C5 instead of C
and continue in this way until either some good B
and C
are found or
w1,2 , w1,3 , . . . , w1,l(1)−1 are all eliminated, one by one; in the latter case, B

and C
as required can be again constructed, as we already noticed. The
proof is concluded. ✷

Corollary 4.2 For any well partial order ≤ on Σ∗ and any g ∈ G, the
relation ≤g is well partial order.
Uniformly Scattered Factors 197

Proof. Assume that ≤g is not well partial order and consider an infinite
antichain {wn | n ≥ 1} of it. Without loss of generality we can assume that
|wn | < |wn+1 | for any n ≥ 1. If the g-factorization of wn is

wn = wn,1 .wn,2 . · · · .wn,[g(|wn |)] .wn,[g(|wn |)]+1 ,

then the set

{(wn,1 , wn,2 , . . . , wn,[g(|wn |)] ) | n ≥ 1}
is an infinite antichain of ≤. This contradicts Lemma 4.4. ✷

5 The General Characterization

We are now ready to prove the main result of the paper: the characteriza-
tion of the partial orders in (1) which are well partial orders. The part (i)
in Theorem 5.1 is a generalization of Higman’s theorem.

Theorem 5.1 Let m ≥ 1 and g1 , . . . , gm ∈ G.

(i) If lim gk (n) < ∞ for some 1 ≤ k ≤ m, then ≤f,g1 ,...,gm is well partial
n→∞
order.

(ii) If lim gk (n) = ∞ for all 1 ≤ k ≤ m, then ≤f,g1 ,...,gm is not well partial
n→∞
order as soon as card(Σ) ≥ 2.

Proof. (i) By Lemma 4.3 we obtain that ≤f,g1 ,...,gk is well partial order.
Then, by Corrolary 4.2, all relations ≤f,g1 ,...,gl with k ≤ l ≤ m, are well
partial orders.
(ii) For any g ∈ G, let [g] denote the function given, for any n ≥ 0, by
[g](n) = [g(n)]. It is clear from the hypothesis that

lim ([g1 ] ◦ · · · ◦ [gm ])(n) = ∞.

n→∞

Then, there is a subsequence (kn )n≥1 of (n)n≥1 such that 2 ≤ ([g1 ] ◦ · · · ◦

[gm ])(k1 ) and ([g1 ] ◦ . . . ◦ [gm ])(kn ) < ([g1 ] ◦ . . . ◦ [gm ])(kn+1 ) − 1 for all n ≥ 1.
(i)
Denote kn = ([gi ] ◦ . . . ◦ [gm ])(kn ) for any n ≥ 1 and i, 1 ≤ i ≤ m;
(m+1) (2) (2)
put also kn = kn . We have then 2 ≤ [g1 ](k1 ) < [g1 ](k2 ) − 1 <
(2)
[g1 ](k3 ) − 2 < · · · This, by the proof of Theorem 3.1, implies that the set
(2) (2) (2) (2) (2)
)] kn −[g1 (kn )][g1 (kn )]
A1 = {(ab[g1 (kn )]−2
a)[g1 (kn a | n ≥ 1}
198 L. Ilie, I. Petre, G. Rozenberg

is an antichain of ≤f,g1 .
Assume that we have constructed an infinite antichain of ≤f,g1 ,...,gi−1 ,
(i)
say Ai−1 = {un | n ≥ 1}, such that |un | = kn , for all n ≥ 1. We
claim that we can construct an infinite antichain Ai = {vn | n ≥ 1} of
(i+1) (i)
≤f,g1 ,...,gi such that |vn | = kn , for all n ≥ 1. Notice that 1 ≤ kn =
(i+1) (i+1)
([gi ] ◦ [gi+1 ] ◦ . . . ◦ [gm ])(kn ) = [gi ](kn ) ≤ kn .
(i+1) (i)
We choose, for every n ≥ 1, the word vn = (un )[kn /kn ] wn , where wn
(i+1)
is a prefix of un such that |vn | = kn (actually, wn can be any other word
of the same length).
Suppose that Ai is not an antichain of ≤f,g1 ,...,gi , i.e., for some r < s
(i)
we have vr ≤f,g1 ,...,gi vs . Since [gi ](|vn |) = [gi ](kni+1 ) = kn = |un |, for all
n ≥ 1, the gi -factorizations of vr and vs are

vr = ur . · · · .ur .wr , vs = us . · · · .us .ws .

Thus, vr ≤f,g1 ,...,gi vs implies that ur ≤f,g1 ,...,gi−1 us , which contradicts the
fact that Ai−1 is an antichain of ≤f,g1 ,...,gi−1 .
For i = m we obtain an infinite antichain Am of ≤f,g1 ,...,gm , which
concludes the proof of the theorem. ✷

References
[1] J. H. Conway, Regular Algebra and Finite Machines, Chapman and
Hall, 1971.

[2] L. Haines, On free monoids partially ordered by embedding, J. Combin.

Theory, 6 (1969), 94–98.

[3] G. Higman, Ordering by divisibility in abstract algebras, Proc. London

Math. Soc., 2 (3) (1952), 326–336.

[4] L. Ilie, Generalized factors of words, Fundamenta Inform., 33 (1998),

239–247.

[5] C. Nash-Williams, On well quasi-ordering finite trees, Proc. Cambridge

Philos. Soc., 59 (1963), 833–835.
 Splicing Normalization and Regularity
Vincenzo Manca
Università di Pisa, Dipartimento di Informatica
Corso Italia, 40 - 56125 Pisa, Italy
E-mail: mancav@di.unipi.it

Abstract. Some canonic forms of splicing derivations are in-

troduced and the notion of ω-splicing is used for proving that
H systems with a finite number of splicing rules and a regular
set of axioms generate regular languages.

1 Introduction
Splicing is one of the basic operations of DNA Computing. It was intro-
duced in [5], [6] as a formal representation of DNA recombinant behavior:
nucleic strands are cut in specific sites, by enzymes, so that pieces of differ-
ent strands whose sticky ends match can be concatenated producing new
DNA molecules.
This mechanism suggested new generative systems in formal language
theory, H systems, and introduced new perspectives in the combinatorial
analysis of strings, languages, grammars, and automata.
More recently, going in the opposite direction, towards a Biological
Mathematics, rather than a Mathematical Biology, it was discovered that
a DNA soup can encode a combinatorial problem and can be transformed,
by means of test tube genetic engineering techniques, into a final soup that
encodes the solution of the problem [1], [11]. This approach disclosed new
horizons in the search for new ideas, and applications. Thus, the biologi-
cal trend, initiated by Kleene’s finite state automata [9] and Lindenmayer
systems [10], was continued, in a new direction, with H-systems. Thereby,
new computational models, inspired by biological metaphors were intro-
duced, and biochemical interpretations were found for concepts and results
in formal language theory [19], [15], [16]. The field covering these subjects
is now referred as DNA Computing, or in wider perspectives, Molecular

199
200 V. Manca

Computing, or Natural Computing. The benefits of this synergy (we could

say metasplicing) between Mathematics, Computer Science and Biology are
apparent and constitute an exceptional stimulus for a deep understanding
of the combinatorial mechanism underlying splicing, the catalyst of all this
process.
One of the most important mathematical property of splicing, in its
original formulation, was that the class of languages generated by splicing
rules is a proper subclass of regular languages. Nevertheless, the proof of
this result has a long story. It originates in [2], [3] and was developed in
[17], in terms of a complex inductive construction of a finite automaton
(in [16] is presented this proof). More general proofs, in terms of closure
properties of abstract families of languages, are given in [18], [7], [15].
In this paper we show a different direct proof, as a natural consequence
of some properties of normalization of splicing processes. We present a
sort of geometrical representation of splicing derivations and introduce the
notion of ω-splicing that allows us to clarify a crucial phenomenon on which
regularity depends.
Consider an alphabet V and two symbols #, $ not in V . A splicing rule
over V is a string r = u1 #u2 $u3 #u4 , where u1 , u2 , u3 , u4 ∈ V ∗ . For such a
rule r and for x, y, w ∈ V ∗ we define the (ternary) splicing relation =⇒r

(x, y) =⇒r w iff x = x1 u1 u2 x2 ,
y = y 1 u 3 u 4 y2 ,
w = x1 u1 u4 y2 , for some x1 , x2 , y1 , y2 ∈ V ∗ .

The string x is the up premise and the string y is the down premise of the
r-splicing step; the string w is the conclusion. Strings u1 , u2 , u3 , u4 are the
left up, right up, left down, and right down components of the rule r; the
pairs of strings (u1 , u2 ), and (u3 , u4 ) are the up site and the down site of
the rule r. The string x1 u1 is the r-head of x and w; the string y1 u3 is the
r-head of y. The string u4 y2 is the r-tail of y and w; while u2 x2 is the r-tail
of x.
Therefore, when an r-splicing step is applied to two strings, they are cut
in between the left and right components of sites of r, and then the r-head
of the up premise is concatenated with the r-tail of the down premise. The
resulting string is the conclusion of the step.
An H system system [16] Γ is given by: an alphabet V , a set A of strings
over this alphabet, called axioms of the system, and a set R of splicing
rules over this alphabet. The language L(Γ) generated by Γ consists of the
Splicing Normalization and Regularity 201

axioms and the strings that we can obtain starting from the axioms, by
applying to them iteratively the splicing rules of Γ. If a terminal alphabet
is considered, then we obtain an extended H system. We say finitary an H
system with a finite number of splicing rules. It is known that:
i) finitary H systems with finite axioms characterize a proper subfamily
of regular languages;
ii) finitary H systems with regular sets of axioms characterize the regular
languages;
iii) extended finitary H systems with finite axioms characterize the reg-
ular languages;
iv) extended H systems with regular sets of rules and finite axioms char-
acterize the recursively enumerable languages;
v) and that extended finitary H systems with finite axioms plus certain
controls on the use of rules or with certain distributed architectures
characterize the recursively enumerable languages.
Comprehensive details can be found in [16].

2 Splicing Derivations
In [13] we introduced the notion of a Derivation System that allows us
to analyze in an uniform way a great variety of symbolic systems, and to
determine their common structure based on two main aspects: the com-
binatorial mechanism of the rules (e.g., replacement, parallel replacement,
insertion/deletion, splicing, . . . ), and the regulation strategy that specifies
the ways rules can be applied.
In this section, we consider derivations in the specific case of splicing
systems, and show that we can express, in terms of splicing derivations,
several interesting concepts useful in the analysis of splicing.
Definition 2.1 Given an H system Γ = (V, A, R), a splicing derivation δ
of Γ, of length n, is a sequence of n strings and n labels, where each string
associated with a corresponding label (written before an arrow that points
to the string). A label is, either a triple (rule, string, string), or a special
label indicated by λ:
(λ → δ(1), l(1) → δ(2), . . . , l(n − 1) → δ(n))
where for 1 ≤ i < n:
202 V. Manca

• if l(i) = λ then l(i) ∈ R × {δ(1), . . . , δ(i − 1)} × {δ(1), . . . , δ(i − 1)};

• if l(i) = λ, then δ(i + 1) ∈ A;
• if l(i) = (ri , β(i), γ(i)), then (β(i), γ(i)) =⇒ri δ(i + 1);
• ∀i ∃j, i < j ≤ n such that δ(i) = β(j);
• if n = 1, then δ(n) ∈
/ A.

At each step i δ(i) is called the current string of that step. According
to the definition, any current string, apart the final one, has to be the
conclusion of some splicing step with premises which are current strings of
steps preceding i (no useless strings can occur in a derivation δ).
We indicate by ∆(Γ) the set of (splicing) derivations of a H system Γ.
If the last element δ(n) of a derivation δ is the string α, then we say
that δ derives α and we write
δ -Γ α.
Two derivations of ∆(Γ) are said to be equivalent if they derive the
same string.
A derivation δ ∈ ∆(Γ) of length n is said to be linear when, for any
1 ≤ i < n, if the element δ(i + 1) is not an axiom, then it is obtained by
applying a rule of R with δ(i) as a premise.

Lemma 2.1 (Linearity Lemma) Given a derivation δ ∈ ∆(Γ) there is al-

ways a linear derivation equivalent to it.

Proof. By induction on the length of derivations. For derivations where

no rules are applied the linearity is trivial. Assume that for derivations of
length smaller or equal to n we have linear derivations equivalent to them.
Let δ be a derivation of length n + 1. The string δ(n + 1) has two premises
δ(i), δ(j) with i, j ≤ n.
By the induction hypothesis there are two linear derivations δ
, δ

of
length at most n that derive δ(i), δ(j), respectively. Consider the concate-
nation of δ
, δ

where the steps of δ

that are already in δ

are removed, then
add to it the last splicing step of δ. This derivation is a linear derivation
equivalent to δ. ✷

A linear derivation can be represented in the following way:

(λ → δ(1), l(1) → δ(2), . . . , l(n − 1) → δ(n)),
Splicing Normalization and Regularity 203

where if at step i the label is different from λ, then it is (ri , β(i), p) with
β(i) ∈ {δ(1), . . . , δ(i − 1)}, and p ∈ {0, 1}, in such a way that:

• if l(i) = (ri , β(i), 0), then (β(i), δ(i)) =⇒ri δ(i + 1);

• if l(i) = (ri , β(i), 1), then (δ(i), β(i)) =⇒ri δ(i + 1).

All the derivations we consider in the following are (tacitly) assumed to

be linear derivations.
We introduce the notion of ω-splicing in order to consider infinite pro-
cesses of splicing (ω stands for the set of natural numbers).
Given two (linear) derivations δ, δ
we say that δ is an expansion of δ
,
and we write δ < δ
, if δ
is obtained by interrupting δ at some step i,
performing some further steps that derive a string, say β, and then, by
continuing from β by applying the same labels of the steps that in δ are
after the step i.
An ω-splicing is an infinite sequence of derivations:

δ = (δ i | i ∈ ω),

where, for every i ∈ ω, δ i is called a component of δ, and for every i > 1

there is a j ≤ i such that δ j < δ i .
The set ∆ω (Γ) is the set of ω-splicings δ such that all their components
belong to ∆(Γ). Any ω-splicing δ determines a language L(δ) constituted
by the strings derived by its components.
The notion of ω-splicing allows us to give a necessary condition for the
infinity of finitary H systems with finite axioms.

Lemma 2.2 (Infinity Lemma) Let Γ be a finitary H system with finite

axioms A. If the language L(Γ) is infinite, then the set ∆ω (Γ) is not empty.

Proof. Consider all the possible ways to apply the splicing rules starting
from the axioms. We can arrange all these possibilities in a rooted tree
where at first level we have the axioms and for each node its sons are
obtained by applying a splicing step between it and some other node of a
preceding level. Of course if L(Γ) is infinite, then this tree is infinite. But,
this tree is a finitary tree, therefore, by König’s lemma it is infinite if it has
an infinite path. This path is essentially an ω-splicing. ✷

An ω-splicing is linear iff all its components are linear. Now we intro-

duce important classes of linear ω-splicings.
204 V. Manca

Given a step i of a linear derivation δ, it is k bounded if δ(i) is an axiom,

or it is the conclusion of a splicing step where one of the two premises is
the conclusion of the previous step, while the other one is at most k-old,
that is, it is a string obtained at most k steps before that step. Splicing
processes with an arithmetical control, labeling the age of strings in the
derivation, were studied in a different context, and with different aims in
[12].
Let Γ = (V, A, R) be an H system. The set Bk (Γ) of k bounded deriva-
tions is the subset of derivations δ such that, every splicing step of the
derivations is k bounded. Two subclasses of Bk (Γ) are the class RDk (Γ)
of right down k bounded derivations and the class LUk (Γ) of k bounded
left up derivations. In a derivation of RDk (Γ) the requirement of being at
most k-old has to be satisfied for the down premise of every step, while in
derivations of LUk (Γ) the same requirement has to hold for the up premise
of every step.
The terms right-down and left-up are due to a graphical representation
of splicing where the up premise is a (rectangular) frame, and the down
premise another (rectangular) frame put under it.
In this manner the r-head of the superior frame (up premise) and the
r-tail of the inferior frame (down premise), connected by an arrow provide
a representation of the result of the r-splicing. If the result of this step is
the up premise of another step, we go in the right down direction (this kind
of staired representation was adopted in [15]); otherwise if it is the down
premise of a further step we go in the left up direction.
The following is the graphical representation of a (0 bounded) RD
derivation with three steps.

The following is the graphical representation of a (0 bounded) LU

derivation with three steps.
Splicing Normalization and Regularity 205

ω ω ω
The classes Bk (Γ), RDk (Γ), LUk (Γ) are constituted by ω-splicings
whose components belong to Bk (Γ), RDk (Γ), and LUk (Γ) respectively.
We can define the notion of RD and LU ω-splicing in terms of rewriting
relations.
Given a splicing rule r : u1 #u2 $u3 #u4 over the alphabet V and z ∈ V ∗ ,
we write, for x, y ∈ V ∗

x =⇒RD
(r,z) y iff (x, z) =⇒r y;

we say that =⇒RD(r,z) is the RD (rewriting) rule associated to the splicing r

with the index string z.
Analogously, =⇒LU (r,z) is the LU (rewriting) rule associated to the splicing
r with the index string z:

x =⇒LU
(r,z) y iff (z, x) =⇒r y.

Therefore, a 0 bounded RD derivation is obtained by applying RD

rewriting rules indexed by the axioms, while a 0 bounded LU derivation is
obtained by applying LU rewriting rules indexed by the axioms. Deriva-
tions of RD0 or LU0 are also called axiomatic (RD or LU ) derivations.
Note that, in general, an axiomatic LU derivation cannot be trans-
formed into an axiomatic RD derivation by changing the verse of the ar-
rows in the staired representation. For example, given the splicing rule
γδ#β$γ#δ with γ = δ, you can go in the left up direction any number of
times, but only one step in the right down direction (analogously, a rule
α#β$γ#αβ with γ = α shows that, in general, axiomatic RD derivations
cannot be transformed into LU axiomatic derivations by reversing the ar-
rows).
Any derivation can be factorized in terms of different nested levels of
axiomatic derivations. We can illustrate this factorization by using the
206 V. Manca

following diagram, where horizontal lines indicate sequence of steps and

bullets represent start or end strings of these sequences. Strings α2 , α3 are
premises of a splicing step of conclusion α4 ; and α1 , α5 are premises of a
splicing step of conclusion α6 .

α1 α6
✉ ✉

α2 α4 α5
✉ ✉ ✉

α3
✉

Given a derivation δ of length n, a subderivation of δ is a subsequence

δ
= (l(j1 ) → δ(j1 ), ..., l(jm ) → δ(jm )), with 1 ≤ j1 ≤ jm ≤ n, that is also
a derivation. A component of δ is a proper subderivation if it is different
from δ.
A derivation is nonredundant if it has no proper subderivation equivalent
to it. For example, an H system with a splicing rule u#v$u
#v can generate
derivations of any length that derive the same string. A less trivial situation
of redundancy is illustrated by the following diagram, where frames are
replaced by horizontal lines and only the parts remaining after splicing are
indicated.
r1

α ❄ u v
r2
❄
r3
❄
r4
❄

Consider the sequence of splicing steps r1 , r2 , r3 , r4 . Let α be the r1 -tail

of the conclusion of the first splicing step. If the up site (u, v) of the fourth
step is included in α, then the sequence of rules r1 , r2 , r3 , r4 is equivalent
to the sequence r1 , r4 , where the two steps r2 , r3 are avoided.
Splicing Normalization and Regularity 207

An ω-splicing is nonredundant iff all its components are nonredundant.

In a linear splicing derivation, each splicing step is constituted by two
substeps: the current string is cut in at a splicing point: a part of this string
is kept (the left part in the case of an RD step, or the right part in the case
of an LU step) and a string is appended to this point (on the right in an
RD step, and on the left in an LU step). If at a step i, the splicing point
is internal to a part of the current string that was appended in a previous
step j, then we say that the step i is counterdirectional with respect to the
step j (a splicing step is counterdirectional if it is counterdirectional with
respect to some splicing step). For example, in the derivation (r1 , r2 , r3 , r4 )
represented by the diagram above the fourth splicing step is counterdirec-
tional with respect to the second step and to the third step. In other words,
a counterdirectional step removes, in the current string, the splicing point
of a previous step.
In an RD axiomatic derivation, a counterdirectional step occurs when
at some step the left up component of a rule is in the head of the conclu-
sion of some previous splicing step. In an LU axiomatic derivation, in a
counterdirectional step, the right down component of the rule is in the tail
of the conclusion of some previous splicing step.
A step is monotone if it is not counterdirectional; a derivation is mono-
tone, if all its splicing steps are monotone. An ω-splicing is monotone if its
components are monotone.
In any string we have a finite number of splicing points, moreover, by
the definition above, all the splicing points of the counterdirectional steps of
a given step are located in the same string. Therefore the following lemma
holds.
Lemma 2.3 (Finite Counterdirectionality Lemma) In a nonredundant ω-
splicing the steps that are counterdirectional with respect to a given step i,
if any, are a finite number.
Monotone axiomatic RD or LU derivations can be represented in a
schematic way with an oriented line going from the left up corner to the
right down corner, or viceversa. In this representation, any splicing step
increases the line with a further piece.
There are nonredundant derivations that are not monotone. Consider
the derivation (r1 , r2 , r3 , r4 ) represented before in the diagram. If the right
up component of the last splicing step r4 is so long that it reach the right-
most horizontal line, then we cannot reduce the derivation (r1 , r2 , r3 , r4 ) to
any proper subderivation of it.
208 V. Manca

In conclusion, modulo elimination of redundant steps, we can factorize

a derivation by composing RD and LU axiomatic derivations, by a diagram
of the following type.

t
❥
✲ t ✲t
t ~
z
❦
t ✐
t y
❦ qt
t
t
s

✎   ✏

In this picture arrows represent monotone axiomatic RD (descending

arrows) or LU (ascending arrows) subderivations. These subderivations
are structured at different levels. At each level there is a current string
of that level. The initial level is the most external level (the highest ar-
rows). Vertical lines indicate the initial and the final splicing points of the
subderivations.
When no counterdirectional step arises, a monotone axiomatic sub-
derivation appends a string to (after in the RD case, or before in the LU
case) the final splicing point of the current string of the superior level.
When a counterdirectional step occurs (in the picture, the arrows with
two bullets), then the result of a monotone axiomatic subderivation is ap-
pended to the current string of the superior level, but at a splicing point
that is not the last of that level. This produces a gap (in the picture, in-
dicated by a bowed line) in the process of concatenating the substrings of
the different levels.
The projections of arrows along the horizontal line indicate the ap-
pended strings. The horizontal line at the bottom represents the string
globally derived.

Consider a derivation. There are two possibilities: at any step, we can

go on by expanding it with further splicing steps, or we cannot continue
anymore. This means that we can distinguish between derivations that
Splicing Normalization and Regularity 209

cannot be extended after a certain number of steps and derivations that

are components of ω-splicings. The next lemma explains the role of the
derivations of the first type.

Lemma 2.4 (Finite Splicing Lemma) Given a finitary H system Γ with

a regular set of axioms, and a natural number k, the language of strings
derived by derivations of Γ of length at most k is a regular language.

Proof. It is a simple consequence of closure properties of noniterated splic-

ing (cf. Lemmas 7.1, 7.2, and 7.3 in [16]). In fact, by using a more direct
argument, we know that regular languages are closed with respect to pre-
fix, subfix and concatenation. But any splicing step can be expressed by
means of these operations applied to the premises. Therefore if we apply a
splicing step to the axioms of Γ we get a regular language; if we continue
this process k times, at end we reach a regular language. ✷

This lemma can be easily extended by saying that, given a finitary H

system Γ whose axioms belong to a class F L of formal languages, if ∆ω (Γ) =
∅, then also L(Γ) belongs to F L. Therefore, when Γ has finite axioms, if
L(Γ) is infinite, then ∆ω (Γ) = ∅, as we already proved via König lemma.
If we consider only nonredundant ω-splicings, then we can reverse this
implication, that is, L(Γ) results to be infinite iff the set of nonreduntant
ω-splicings of ∆ω (Γ) is not empty.

3 Regularity
In this section we prove that finitary H systems with regular sets of axioms
generate regular languages. This result referred in [16] as Regularity Pre-
serving Lemma is a very important aspect of (nonextended) splicing. In
literature there are known indirect proofs [18], [16], [7], via algebraic clo-
sure properties of AFL (abstract families of languages), and direct proofs
[2], [3], [17], [16] via a finite automaton that recognizes the language gen-
erated by a given H system. The proof we present here is another direct
proof using the notion of ω-splicing, that makes evident a general aspect of
finitary splicing that implies its regularity.
A crucial aspect of finitary H systems is that we can build arbitrarily
long derivations, where an unbounded number of steps depend completely
on a substring that occurred before in the derivation. We consider this
aspect by using the notion of locality of a step.
210 V. Manca

Given a H system Γ we call the left radius of Γ the maximum length

of left components of rules of Γ, and the right radius of Γ the maximum
length of its right components (the maximum of these two lengths is the
radius of Γ).
Let us call locality of a splicing step i, in a derivation of ∆(Γ), the
substring θ = u
uvv
where uv is the site of the rule applied to δ(i) and
u
uvv
is the substring of δ(i), in a such a way that the length of u
u equates
the left radius of Γ, and the length of vv
equates the right radius of Γ.
At any step, if its splicing point will not be removed in a following
step, what can happen after that step depends completely on its locality.
In a finitary H system the localities are a finite number, therefore their
lengths are bounded by some natural number; that is, in finitary H systems
we have a sort of local generation. This phenomenon, illustrated by the
following proof, resembles the notion of locally testable language [14], whose
connection with splicing played a crucial role in the definition and analysis
of this combinatorial mechanism [20], [4], [8].
Given two ω-splicings δ1 , δ2 , we can define a generation pre-ordering
relation ≤L such that δ1 ≤L δ2 if L(δ1 ) ⊆ L(δ2 ). We say that an ω-splicing
of ∆(Γ) is maximal if it is maximal with respect to this pre-ordering relation.
Moreover, we say that two ω-splicings δ1 , δ2 are generation equivalent if
L(δ1 ) = L(δ2 ). Let us indicate by [δ] the class of generation equivalence of
the ω-splicing δ. The existence of maximal ω-splicings is a consequence of
Zorn’s Lemma (in an ordered set where each nonempty chain has the least
upper bound, any chain has a maximal element too).
The essential tools of the following argumentations are the Pigeonhole
principle and the induction principle.

Lemma 3.1 (Infinity Gate Lemma) Given a finitary H system Γ, there

exists a number gΓ such that in Γ any derivation of length greater than gΓ
is a component of an ω-splicing.

Proof. Having in Γ a finite number of localities, we are sure that there exists
a number g such that in g steps two occurrences of the same locality have
to occur. If the second occurrence is not counterdirectional with respect
to the first one, then gΓ = g, because we can concatenate any number of
times the string between the splicing points relative to the two occurrences
of that locality. Otherwise, if we consider 2g, 3g, . . . steps we get further
occurrences of the same locality. We know from Lemma 2.3 that all these
Splicing Normalization and Regularity 211

occurrences cannot be counterdirectional to the first one. Therefore, for

some positive natural number k, gΓ = kg. ✷

Lemma 3.2 (Finite and Infinite Splicing Representation Lemma) Given a

finitary H system Γ with a regular set of axioms,

L(Γ) = L0 ∪ L(δ)
[δ]∈M

where M is the set of generation equivalence classes of maximal ω-splicings

in ∆ω (Γ), and L0 is a regular language.

Proof. Consider a derivation in ∆(Γ). If it can be carried on with an

unbounded number of steps, then it is a component of an ω-splicing δ, thus
what we can generate by means of it is generated by δ; otherwise, if it stops
after some number of steps, then it can generate a string that we could not
generate with an ω-splicing.
Let gΓ be the number given by the previous lemma, and define L0 to be
the language of strings derived with at most gΓ steps. As we have pointed
out (Lemma 2.4), L0 is regular; therefore, L(Γ) = L0 ∪ L1 , where L1 is
the set of strings derived with more than gΓ steps. But, L1 consists of the
strings generated by ω-splicings, and therefore, the statement follows by the
definitions of maximal ω-splicings and of generation equivalent ω-splicings.
✷

Theorem 3.1 (Regularity Preserving Lemma) If Γ is a finitary H system

with a regular set of axioms, then L(Γ) is a regular language.

Proof. In virtue of the previous lemmas, it is sufficient to prove that:

• The number of the generation equivalence classes of nonredundant

and maximal ω-splicings in ∆ω (Γ) is finite.

• If δ is a nonredundant and maximal ω-splicing δ ∈ ∆ω (Γ), then L(δ)

is a regular language.

Consider a nonredundant and maximal (n.m.) ω-splicing δ. It has

at least a rule that applies infinitely many times with the same locality,
because rules and localities of Γ are finite, while δ is an ω-splicing and
thus has infinitely many steps. Let (r, θ) be a pair (rule, locality) of this
212 V. Manca

kind, that we call a knot. Consider the occurrence of a knot (r, θ) that
we call initial because it is the first occurrence of (r, θ) in δ such that no
splicing step of a further occurrence of (r, θ) can be counterdirectional with
respect to (the step of) this occurrence. Such an initial occurrence must
exist because otherwise, the occurrences of (r, θ) being infinitely many, we
could have an infinite sequence of steps that are counterdirectional with
respect to the same splicing step, contradicting Lemma 2.3. For the same
reason, after the (first) initial occurrence of (r, θ) we have the second initial
occurrence of (r, θ) such that no other further occurrences of (r, θ) could
be counterdirectional with respect to it. Analogously we can define all the
other initial occurrences (the third, the fourth, and so on).
The following picture represents the factorization of a n.m. ω-splicing
δ with respect to the initial occurrences of (r, θ).

σ (r, θ) (r, θ) τ
✈ q ✈
✻ q
q

π1 , π2 , ...

In this factorization σ, {π1 , π2 , . . .}, τ are called the seed, the power, and
the termination of δ. The power {π1 , π2 , . . .} contains all the sequences of
steps (possibly infinitely many) that starting from an initial occurrence of
the knot (r, θ) can reach again another initial occurrence of the same knot.
The seed and the termination could be empty.
By the assumption on the initial occurrences of (r, θ), we know that if α
is the string that we keep at the step of an initial occurrence of (r, θ), then
when the next initial occurrence of (r, θ) occurs, some string β is added to
α; beside, after appending α the same locality occurs, therefore, since δ is
maximal, for every n ∈ ω in δ also β n is appended to α.
Moreover, if some step in the termination would be counterdirectional,
with respect to all the initial occurrences of the knot that can occur after
its first initial occurrence, then the knot could be avoided, because only
a finite number of occurrences of (r, θ) could be necessary in performing
the termination. This would imply that δ should be redundant, but we
assumed δ to be nonredundant.
Now, we proceed by induction on the number of knots in n.m. ω-
splicings.
Splicing Normalization and Regularity 213

Consider a n.m. ω-splicing with only the knot (r, θ). In its seed, termi-
nation, and power no other knot can occur. In fact, any locality can occur
a finite number of times: m1 the first one, m2 the second one, and so on.
But the derivations that verify this condition are in a finite number.
Therefore, we have only a finite number of possibilities for
{σ, π1 , π2 , . . . , τ }, and even if we can arrange the elements of the power in
different ways, all maximal nonredundant ω-splicings are equivalent inas-
much as they generate the same language:

L(δ) = α{β1 , β2 , . . . , βn }∗ γ,

where α is the string that is kept at the first initial occurrence of (r, θ),
{β1 , β2 , . . . , βn } is the set of strings that can be appended to α by the
(finite) elements of the power, and γ is the string that can be appended
to α without producing an initial occurrence of the knot (r, θ). Of course,
L(δ) is regular. Therefore, since the number of possible knots is finite, the
ω-splicings with 1 knot generate a finite number of regular languages.
Now consider the inductive case. Assume that n.m. ω-splicings with at
most k knots generate a finite number of regular languages. Let δ be a n.m.
ω-splicing with k +1 knots. We could factorize δ in the same manner we did
before: seed, power and termination; but in this case it can happen that
among the seed, power or the termination of δ there are infinitely many
steps, because some knots can occur repetitively. Therefore,

L(δ) = L0 (L1 + L2 + . . . + Lm )∗ L
,

where: L0 are the strings kept when the steps of σ finish; the languages
L1 , L2 , . . . , Lm consist of strings appended to L0 by {π1 , π2 , . . .}, and L

is the language of strings appended by τ . By the induction hypothe-

sis, L0 , L1 , L2 , . . . , Lm , L
are regular languages, because σ, π1 , π2 , . . . , τ are
n.m. ω-splicings with at most k knots. Therefore, L(δ) is regular.
Moreover, since the number of knots is finite, we have a finite number
of possibilities for the first of k + 1 knots, thus the n.m. ω-splicings with
k + 1 knots can generate a finite number of regular languages.
In conclusion, n.m. ω-splicings determine a finite number of generation
equivalence classes, and every n.m. ω-splicing generates a regular language.
This concludes the proof. ✷

If Γ is an H system with an infinite set of rules, then the argument of

finite set of localities does not apply.
214 V. Manca

The proof of this theorem is based on a sort of a pumping mechanism.

It would be interesting to define a notion of locally generated languages
where this pumping phenomenon is a direct consequence of general locality
conditions that do not depend on the particular combinatorial mechanism
underlying the derivations.

References
[1] L. Adleman, Molecular computation of solutions to combinatorial
problems, Science, 266 (Nov. 1994), 1021–1024.

[2] K. Culik II, T. Harju, The regularity of splicing systems and DNA, in
Proc. ICALP 1989, Lecture Notes in Computer Science, 372, Springer-
Verlag, 1989, 222–233.

[3] K. Culik II, T. Harju, Splicing semigroups of dominoes and DNA,

Discrete Applied Mathematics, 31 (1991), 261–277.

[4] A. De Luca, A. Restivo, A characterization of strictly locally testable

languages and its application to subsemigroups of a free semigroup,
Information and Control, 44 (1980), 300–319.

[5] T. Head, Formal language theory and DNA: an analysis of the gen-
erative capacity of recombinant behaviors, Bulletin of Mathematical
Biology, 49 (1987), 737–759.

[6] T. Head, Splicing schemes and DNA, in: Lindenmayer Systems: Im-
pacts on Theoretical Computer Science, Computer Graphics, and De-
velopmental Biology, (G. Rozenberg, A. Salomaa, eds.), Springer-
Verlag, Berlin, Heidelberg, 1992, 371–383.

[7] T. Head, Gh. Păun, D. Pixton, Language theory and molecular ge-
netics, Chapter 7 in Handbook of Language Theory (G. Rozenberg,
A. Salomaa, eds.), vol 2., Springer-Verlag, Berlin, Heidelberg, 1997,
295–360.

[8] T. Head, Splicing representations of strictly locally testable languages,

Discrete Applied Mathematics, 87, 1-3 (1998), 139–147.

[9] S. C. Kleene, Representation of events in nerve nets and finite au-

tomata, in Automata Studies, Princeton Univ. Press, Princeton, NJ,
1956.
Splicing Normalization and Regularity 215

[10] A. Lindenmayer, Mathematical models for cellular interaction in de-

velopment, I and II, J. Theoret. Biol., 18 (1968), 280–315.

[11] R. Lipton, DNA solution of hard computational problems, Science,

268 (April 1995), 542–545.

[12] V. Manca, Gh. Păun, Arithmetically controlled H systems, Computer

Science Journal of Moldova, 6, 2(17) (1998), 103–118.

[13] V. Manca, Logical string rewriting, Theoretical Computer Science, to

appear.

[14] R. McNaughton, S. Papert, Counter-Free Automata, MIT Press, Cam-

bridge, MA, 1971.

[15] Gh. Păun, On the splicing operation, Discrete Applied Mathematics,

70, 1 (1996), 57–79.

[16] Gh. Păun, G. Rozenberg, A. Salomaa, DNA Computing. New Com-

puting Paradigms, Springer-Verlag, Berlin, 1998.

[17] D. Pixton, Regularity of splicing languages, Discrete Applied Mathe-

matics, 69, 1-2 (1996), 101–124.

[18] D. Pixton, Splicing in abstract families of languages, Technical Report

26/97 of SUNY, Univ. of Binghamton, New York, 1997.

[19] G. Rozenberg, A. Salomaa, eds., Handbook of Language Theory, 3

Volumes., Springer-Verlag, Berlin, Heidelberg, 1997.

[20] M. P. Schützenberger, Sur certaines operations de fermeture dans les

languages rationels, Sympos. Math., 15 (1975), 245–253.
 Infinitely Many Infinities
Carlos Martı́n-Vide
Research Group on Mathematical Linguistics and Language Engineering
Rovira i Virgili University
Pl. Imperial Tarraco, 1
43005 Tarragona, Spain
E-mails: cmv@astor.urv.es, cmv@tinet.fut.es

[The Country, the Olympic year 1992, sunny though a bit cloudy – fine
weather, as ever; P and Q seated on the ground amongst the pines; tulips,
cypresses and rosemary nearby.]
P: Always, we come here to spend the day. Always, the weather is the
same, hot and dry. Always, we become thirsty by mid-morning. Meteorolo-
gists call this a fine weather! It must be fine for them. To me, it’s a torture.
Above all, because of how monotonous it is. When it’s raining, there exists
atmospheric activity. When the sun is shining as today, it doesn’t...
Q: And now it happens that you say that meteorologists are conserva-
tive!
P: Of course! Their television says, for example, that the weather is, will
be or was fine. I don’t have to remain impassive when listening a statement
of that kind. I like atmospheric uprising, simply that something is occurring
in the sky. And I’m about to take accelerated classes for learning to pray for
rain. Afterwards, I’ll set up a company together with some old colleagues
from the seminary...
Q: Your tendency to excesses, to go too far, to say the first thing that
crosses your mind appears again. If you had some more sense, your affairs
in the life would have gone better.
P: They have not gone so bad to me. After all, I can waste time as I’ve
been doing the last five years, since I left that old building which welcomed
us every morning reluctantly, and where we came in with the same spirit
and the same little passion. Let me get a bit excited now, when my passions

217
218 C. Martı́n-Vide

can no longer bother anybody, as there is nobody. Well, yes, you. But my
moods don’t affect you, do they?
Q: You say everything. Of course, they affect me, the same as it must
happen to the same chickens... I mean to the same birds... that come to
meet us every morning. But I cannot allow myself a passion for this.
P: Why cannot you allow yourself a passion?
Q: Because it’s not a part of my programme and gets me out of my
rails.
P: Well, be derailed, my dear fellow. At your age, have you not already
realized that passions move the world?
Q: You are a disbeliever.
P: Of course. And it has costed me a great deal. I prefer to be that
than conceited, or than a believer.
Q: Well, bah! Don’t start again with the bore of last Sunday. It is
obvious that you like war, conflict, risk, struggle. I prefer placidity and
contemplation. Since nearly two hundred consecutive days ago, we take
the train exactly at the same time and come exactly to the same place. On
top of that, you cannot want us to talk about the same topics always.
P: There exists a rather small variety of topics to talk about. Notice
that I always bring the same book.
Q: Yes, I noticed it. Have you not yet read it?
P: I never read it. I think that even I didn’t open it since the pretty
pupil girl in the school gave it to me last year.
Q: Not even you’ll open it.
P: Not even I’ll open it.
Q: And what is it about?
P: About infinity. It’s a book that Cantor wrote about infinite sets last
century. It looks very interesting.
Q: And if it looks so to you, why not glancing through it at least?
P: Well, look, because I think that, if I open it, I’ll finally find out
everything. And I prefer to remain in my ignorance, since it means to me
a kind of secret.
Q: How very strange things you mind!
Infinitely Many Infinities 219

P: I read only encyclopedias. Alphabetically.

Q: Alphabetically?
P: Yes, because nothing is of more interest to me than any other thing.
Q: And don’t you believe that, at your age, you ought to have already
chosen?
P: Yes, look at what happened to me with Marilin...
Q: Well... keep apart your conjugal problems, as I’d have much to say
about. What is surprising to me is that for a mathematician like you...
P: ...a wretched teacher of mathematics until his retirement...
Q: ...well, so for a wretched teacher of mathematics like you everything
is interesting. What is common is that just the opposite happens: that one
is only interested in mathematics.
P: No, look. Shortly after I began to teach, I realized that my colleague
teaching French and I were doing not very different tasks: both were teach-
ing languages. Our pupils did not perceive so, but I was seeing it more and
more in this way.
Q: As much as doing the same...!
P: Definitely yes. That is why I began to be interested in languages. In
the plenty of time I had after lunch and before coming back to the school,
I took the habit to sit down in the yard of the opposite faculty and to start
to read some stuff that I was lent at the library of the university. I was
reading without any order things by Martinet, Vossler, Harris and a lot of
more people. I felt them very close to the way of arguing I was used to.
After a time, some day I decided to stand up from my seat and to look at
the offices for some professor who could guide me. I was successful: he was
an old, bearded but kind professor, which was with me for a good while,
until the necessity made me to run to my loved school.
Q: And did he take you seriously?
P: Yes, in a discoloured filing card that he gave to me, he noted down
in his own way a brief guide of readings that I should do. Look, I got a
very precious information by chance. Good things of life occur by chance:
loves, lottery and everything else!
Q: God doesn’t play dice, my dear.
P: And how! I’ve just read a quotation from Hawking in the newspaper:
220 C. Martı́n-Vide

“God not only does play dice, but sometimes throws them where they
cannot be seen”.
Q: Very nice! But that’s a lie.
P: Beauty and truth must be put side by side. Okay? Well, it doesn’t
matter, the fact is that I started to read what that kindly man had wanted
to note down in my list. I was amazed by a variety of authors’ professional
origins I was not accustomed to: some of them were grammarians, some
other ones philosophers, there were a few psychologists and mathematicians
and many other people. I had thought that the study of languages was a
matter of grammarians’ concern, and, however, was seeing that it was not
the case, that it was not a matter only of their concern.
Q: Perhaps it was that man’s modern own oddity.
P: No, time and readings were convincing me that he was not on the
wrong road. Because, moreover, my anonymous guide stressed very much
that I should read the classics.
Q: Plato, for sure!
P: Among others.
Q: Plato said everything – don’t forget it.
P: Well, I don’t know. I read a good handful of works from previous
centuries and noticed that the study of language is largely accumulative,
unlike what we normally are used to see in the sciences.
Q: In other words, the study of language is not a science like others.
P: No, I’m not saying that, wait. I’m just telling you about my per-
plexity. Little by little, I was getting excited and learnt to put my concept
of progress at issue. This people didn’t seem to make progress, but to give
different explanations to the same problem each time: how is it that we
speak so good, so much and so early. They didn’t raise it in such a way,
but I was intrigued by this, because my experience as a teacher had showed
to me that it didn’t happen the same with the learning of mathematics.
Q: You should take note of that as an additional piece of information
of a difference that you cannot deny.
P: I don’t know whether or not I can deny it. But while reading some
treatises written in the nineteenth century about the evolution of languages
along the history, I tended to pay especial attention to resemblances rather
than to differences. Maybe it was the deformed view of a person with
Infinitely Many Infinities 221

a mathematical training, but the fact is that it was so and that I didn’t
believe to be the only one in the world to think so. Well, but some further
detail awaked my curiosity: people who had worked on languages had been
intrigued by the relationship that words have with things. Since my years
in the faculty, I had retained my interest in discovering what what I was
learning had to do with the real life, not only with machines and the world
of development, but also with human beings and their world of passions and
feelings. It turned out to be very comfortable to me that other people were
concerned before with the relationship that languages have with reality.
Q: A merely instrumental relationship?
P: Not at all. Both the mathematical language and any language, which
are of course structures composed of elements that perform certain func-
tions, the same as cars which you like so much, just the same as cars, have
their history, like rivers, which look for their bed in the land.
Q: Wait, I’m wondering whether this whole trip was really of some use
to you.
P: It gave pleasure to me.
Q: That’s all?
P: That’s all.
Q: Just?
P: Just. Is it not so much?
Q: Frankly speaking, no.
P: Please let me continue, since that’s not all. I tried to take a historical
order, as was that man’s first tip should I wished to acquire a certain
linguistic culture. And in this way I reached a book that I was about not
to read, because it seemed to me that it was written in Latin: by a German
– I believe – with a bit hectic life. His small book made me see that language
was the key for clearing up something that had ever appeared to me as a
huge bundle: what still continues to be called philosophy. Moreover, its
style impressed me.
Q: Why have you never written a book, instead of being so much reading
others’ ones?
P: Well, look, because I’ve seen much redundancy circulating. That
teutonic sir... well, on second thoughts, I don’t know whether he was born
in Vienna... that sir taught to me that there are just very few things to say
222 C. Martı́n-Vide

that have not been already said.

Q: Okay, but perhaps it is worth saying them in another way.
P: Perhaps it is. I don’t know. Many styles are possible, but the best
is just one.
Q: The best! The best! What do you mean?
P: No, please, forget it, it doesn’t matter, I’m in no mood now for
profound philosophies with this so intense heat. I just know that such a
book, which looks to have been written unwillingly and is the only by him
I know, impressed me.
Q: It is evident that you need strong excitements.
P: Of course, otherwise how do you stir your imagination up? Those
thrills are driving you, until one ends up living a life that is spinned with
the remnants which often one inadvertently runs into.
Q: And which was, if I may ask, your next indescribable experience?
P: I’ll tell it to you provided that you allow me to take a bunch of grapes
out of the bag that is between your legs.
Q: All of them for you. You know that I’ve never been alcoholic.
P: I am, and many ancestors that precede me in that book of history
where I’m sure I’ll appear accompany me in this hobby. I like to be out of
place, to make me conspicuous...
Q: You don’t need to remind me of it, I know it. You discuss the price
of the ticket with the same clerk every day.
P: It’s never with the same ticket clerk. We are different every day.
Q: Bah, let you continue!
P: Don’t get excited. So be it. I had heard that a certain Chomsky had
been a specialist in politely not fitting in with the ideas that his colleagues
supported in his time.
Q: What century are you speaking now about?
P: About this, the century when you and I will surely die. I had heard
to qualify his work as revolutionary. Not only in grammar, but also in phi-
losophy, psychology, logic, politics and more fields. That is why I was quite
favourably predisposed when I started to read him. The most attractive
feature of his setting out was that he was opening a specialized territory to
everybody, no matter their origin, which had something to say about lan-
Infinitely Many Infinities 223

guage. That is, he invested even me with a very small authority. I shared
with him his insistence on the necessity both to devote oneself to theory
and to fearless go through the borders that academic institutions had set
up, in their vain aspiration for drawing some limits on a field that is much
larger than this one which you and I are kindly sharing.
Q: Pretty nice!
P: That’s how it is. After Chomsky, one talks about language sciences,
in plural. Among them, grammar, or linguistics as now it is usually called
– I like grammar much more –, occupies a relatively privileged position,
because of the tradition that it has accumulated along the time. Moreover,
that sir was farsighted for summarizing which are the great areas in the
study of languages in a quite simple way: their structure, acquisition and
use. After him, there have been contributions to those areas by linguists as
well as mathematicians, psychologists, philosophers and many others, since
he made categorically clear that a language is an immensely polyhedric
object, and that any well aimed path we take for approaching it is perfectly
legitimate at the same time as it will enrich the other ones without any
doubt. That is why you’ll see today that one talks about language sciences
and no sensible person is wondering who is the boss there. Everyone uses
the methods he learnt, and learns to listen the others and to understand
what they are doing.
Q: Don’t you think that all this is celestial music, in view of what is
happening with the overspecialization of sciences?
P: They are not incompatible trends. Maybe this people are at a certain
stage when they need to sit down and to interchange their results and tools,
in order to get up later equipped with a greater strength and a more detailed
map of the land that they have to continue to explore. It has occurred along
the history of science several times and is not bad. It is a kind of spiritual
retreat...
Q: I observe once again that you don’t take seriously even what is of
your interest. Poor Marilin!
P: I do. However, just a little bit of irony doesn’t harm at all. Not even
wise people have always learnt to take what they are doing with a bit of
distance and a sense of humour.
Q: Nor politicians.
P: Those even less, engaged as they are to organize their fellow citizens’
lives. Let them get on with it!
224 C. Martı́n-Vide

Q: And have your readings arrived up to there?

P: Yes, not much further. Those problems of language structure, ac-
quisition and use seem big enough for me to daringly go forward: how are
languages built, how is it that we are able to learn them in such a short
time, and which abilities do we develop that allow us to use them so appro-
priately as you and I are doing for years. Moreover, everything I’m reading
on language has brought before my eyes some problems that interest me for
the first time in my life. They interest me, they excite me, they intrigue me.
Probably because they are the modern version of some ancient problems of
the damned philosophy that I never succeeded in understanding and that
I now finally grasp. While I was never a humanist, I’ll end up being con-
verted to the cause. I’m not interested in mathematical problems: they are
like riddles or crosswords. I’m interested in bigger and pastier problems.
But at the same time they overwhelm me, because they bring me from one
matter to other even more obscure pith, and I’ll finish by asphyxiating. Do
you see why I refuse to open the book that my ex-pupil girl was so kind to
give away to me that day when she knew that she will not meet me again?
Q: Don’t get sentimental.
P: No, you should understand that the sentimental state is the natural
state, and any other one is a degeneration of it. That, thus, the feeling
is the primary fact, the organ precedes the function, and the earth goes
before the sky.
Q: I already live in my heaven well.
P: But you will fall down, and you have not become accustomed to
crashes.
Q: Well, as you like. After all, you are the boss. And then you hit me.
P: It’s not that, disagreements concerning principles can only be solved
by force. And you will agree with me that it’s not worth it.
Q: Certainly, and I’d like you to remember it every time you get angry.
P: I’ll remember it. I promise it to you.
Q: Look, you’ll die soon and we’ll give up coming here as every day.
What will it happen afterwards?
P: I don’t know, that I’ll finally go to your heaven – I guess. I don’t
know, I don’t have qualities of a fortune-teller.
Q: All things you thought while reading, should you not put them in
Infinitely Many Infinities 225

writing?
P: What for? No.
Q: Even if for others to spare your path.
P: No, it’s better for them to go through it step by step. The only thing
that it would not matter to me to write down in a few pages, very few pages,
is my feeling that the reflection on language, which is one of the few useless
jobs that I find interesting in the life, is becoming more and more immersed
in another more general interdisciplinary environments: the worlds of com-
putation, cognition and communication. I believe that these three words
can well inherit in the future the attention paid to structure, acquisition
and use in recent times. The attention and the money, as wherever there
are dollars there is really some seed of utility. I’ve noticed recently that
what are all these concerns on language for is starting to interest. And
it seems that they are useful for something: for contributing to develop
the computer industry, for preventing and correcting the problems some
people encounter with their mother tongue, for making the social conflicts
arisen from the cohabitation of peoples with different languages and cul-
tures more bearable, or for improving – and about time too! – the methods
for teaching foreign languages.
Q: And is that bad?
P: No, I don’t think so. If there is no money, this people will not be able
to carry on. I’d like to be still here in order to see how they accommodate
to this new situation. It has always been said that the money rots good
intentions and deviates fair purposes. It seems to me that it remains to be
proved. Anyway, there don’t exist some intentions or purposes that can be
so, abstractly, qualified as good.
Q: However, on that track, everything is right in the end.
P: Yes, but everything might be right. We tend to constantly value ev-
erything. One becomes really old when says that things are unfortunately
no longer like before. Of course, today is different from before. And it’s
all right to be so. Do you see?, it’s all right. People working on artifi-
cial intelligence, which are also concerned with language – after all, they
are intelligent and know what is worth spending the time –, say that the
very characteristic of human beings is their capacity to learn and to adapt
themselves to new situations. I don’t know. Anyway, one must always be
attentive to what is happening around. And what is happening is what is
happening: there is no alternative.
226 C. Martı́n-Vide

Q: Yes, master. So be it.

P: What a shame I have only little time!
Q: No!
P: I’m hungry.
Q: What a change, now when I was already got caught on the tone of
your dissertation!
P: Let’s go, my colleague, as Sister Trinidad must be waiting for us.
And let me leash you, since the ticket inspector clearly doesn’t understand
you.
[After a few weeks, one morning P was discovered died on the floor at
his room. Within the drawer of his bedside table, the police could find,
among many small, yellowish and muddled cigarette papers, three sheets
with this text, which we literally transcribe for philological and psychiatric
reasons. Apparently, he didn’t sleep for some time and spent the nights,
among the general silence, by giving up his mind through the recesses of
so thick problems that he was unable even to assign them to some con-
crete professionals. It should happen that his linguistic interests roused his
concern with the subject you’ll read about, that he then revived what his
mathematical training had provided him with, and that, finally, he was car-
ried away with his almost senile liking for philosophical speculation. This
attitude was surely the last precipitate of a life that was zigzagging and
eluding answers. Or maybe it was all just about a matter of style.]

THE INFINITY, THE PARADOX AND THE HORROR:

SOME QUESTIONS FOR TWO NIGHTS REFLECTING

(1) Are the infinite, the nonfinite, the indefinite and the unlimited all
different things?
(2) Is the infinite something negative?
(3) According to the Apollonian view, Greeks show horror to the infi-
nite, since they consider the reason as unable to understand it; opposite to
it, the Faustic view emphasizes the passionate tendency towards the infi-
nite: is the reason moving in the emptiness when dealing with the infinite?
(4) How to reconcile that the Greek art rejects the infinite whereas the
Greek philosophy admits it as a problem?
Infinitely Many Infinities 227

(5) Are the notion of the infinite, the belief that there exists some
infinite reality, the feeling of the infinite, the expression of the infinite, the
imagination of the infinite, etc. all quite different things?
(6) Sometimes one of the features showed by the infinity is the cyclicity
or the eternal return. If there is repetition, could one claim the infinity not
to exist?
(7) Does the infinity always show the divisibility of the continuum as a
feature?
(8) Does the infinity always show the eternity as a feature?
(9) Are there things that are unlimited in themselves, for example plea-
sure?
(10) Does the infinity emerge – as it looks to be – from the concept of
continuum and its associated one of limit?
(11) The infinity can be: by division (nondenumerable) or by addition
(denumerable). Does our belief in the infinity result from the infinity of
time and the divisibility of magnitudes?
(12) Is the infinity either a substance, an attribute or an exten-
sion/quantity?
(13) Is the infinity such that either it can not be gone through or, even
being able to be, it is not actually gone through?
(14) Is it true that the infinity is something that there is something
further on (not that there is nothing further on)?
(15) Does the (actual) infinity exist while not being enumerated?
(16) What has the infinity to do with the infinitive?
(17) Are infinity and perfection, as definability, attached each other
always?
(18) The most resolute infinity is God. Is the divine infinity different
from either the numerical infinity or the infinite divisibility of a line?
(19) Could a point be regarded as infinite insofar as it has no divisible
extension?
(20) Does the modern (post-medieval) age involve a certain movement
from finitism towards infinitism, or from a closed world to an open one?
(21) To what extent the scientific revolution and the progresses of math-
ematical thinking (especially, the concept of function and the infinitesimal
228 C. Martı́n-Vide

calculus) have contributed to the infinitization of our idea of the world, and
to what other extent should one attribute it to philosophy, literature and
art?
(22) Is it legitimate to interpret infinitist trends in the Baroque art?
(23) Does Descartes’ ontological argument get right when stating the
idea that one must move from the infinite (as the most objective thing) to
the finite, and not backwards?
(24) Is infinitism necessarily linked to pluralism?
(25) Is it true that the infinity can not be the mere feeling of something
incommensurable, but is measurable and can be calculated?
(26) Is an ascending hierarchy of infinities like the following acceptable:
numerical infinity first, then infinity of the world, and lastly infinity of God?
(27) Is the notion of the infinity inseparable from the feeling of the
infinity, and in particular from the paradoxes and sublime truths it gives rise
to and from the admiration it inspires, insofar as it makes the imagination
burn and asks for both this latter as well as the thought?
(28) Is it true that the contemplation of the eternal silence of the infinite
universe frightens and terrifies humans and makes them to feel lost in the
middle of that amazing marvel?
(29) How do we acquire the idea of the infinity and how to justify that
we accept the infinity in duration better than in extension?
(30) Does the idea of the infinity arise always from the additive con-
catenation of finite segments?
(31) Might one say that the infinity should only be predicated about
space, time (both quantitative notions, as they are related to numbers) and
God (a qualitative idea, which is connected with perfection)?
(32) Why the feeling of the infinite occupies a central position just in
the Romanticism?
(33) There are those who think that every passing from the finite to
the infinite is rationally illegitimate. Some others believe that the infinite
is just a way of speaking. Is it true that accepting the infinite generates
unsolvable paradoxes, and that from the finite to the infinite there is a
jump rather than a step?
(34) Is it of some use here to refer to Jorge Luis Borges’ total library?
Infinitely Many Infinities 229

(35) “God made the integers; everything else is a human work” (L.
Kronecker). Are there infinite sets?
(36) As the idea of the infinity i) can not come from the experience,
because every object from the experience is finite, and ii) can not come
from the imagination, because this latter is able only to reproduce the data
from the senses, and this would in any case give rise to the indefinite, might
it be an ‘a priori’ idea?
(37) From the viewpoint of generalized relativity theory, the universe is
finite but unlimited, like a sphere, which is finite but can be gone through
in all directions without reaching any limit. Is being infinite clearly distinct
from being unlimited?
(38) What would become of mathematical analysis without both the
infinitely large and the infinitely small magnitudes?
(39) It was believed some time that finite magnitudes consisted of in-
finitely many infinitely small indivisible parts, which were conceived as
constants smaller than any finite magnitude. Should one rather think to-
day that finite magnitudes are composed of an unlimited increasing number
of components that decrease unlimited?
(40) Both infinitely small and infinitely large magnitudes are the basis
for the definition of derivative (as the ratio of infinitely small magnitudes)
and integral (as the addition of infinitely many infinitely small magnitudes).
Does that coincide with some primary intuition?
(41) Can we calmly accept the numbers +∞ and −∞ in the theory of
functions of one real variable?
(42) To say that a variable is, for instance, infinitely small makes sense
only if the nature of the variation is described as a function of some other
variable. What is the nature of the infinite sets of mathematical objects?
(43) Is enumeratively characterizing infinite sets of mathematical ob-
jects unimaginable?
(44) An infinite game is a noncooperative game, in particular a two-
player game amounting zero and with infinite sets of strategies at the play-
ers’ disposal. Is it all an entelechy?
(45) The existence of infinite objects in a formal theory is guaranteed
by an infinity axiom. Dedekind’s infinity axiom, specifically, assures an
one-to-one correspondence to exist between the set and some of its proper
subsets. Is it an entelechy again?
230 C. Martı́n-Vide

(46) For intuitionism, a philosophy close to nominalism and idealism,

there are no other infinite sets than the denumerable ones: abstract objects
are creations of the human mind. Opposite to it, realism or platonism
stands. Should we stop being platonists?
(47) May one talk about God’s infinite love for human beings or your
infinite look?
(48) Is the infinity a fiction? Why (what) the horror (of)?
 Finite and Infinite in Geography
Georges Nicolas
15 Rue Alfred de Musset
25300 Pontarlier, France
E-mail: nicorad@fc-net.fr

Abstract. Geography regards macroscopic entities perceived

and represented at different scales. Microscopic or macrocos-
mic entities are not in the field of geography. Two conceptions
of space coexist in geography. As empirical “reality”, relations
between the geographical locuses-objects observed by a geog-
rapher generate a geographical space. This space is “full” and
constituted by a finite set of locuses-objects. To represent these
geographical spaces with “geomaps” or “maps”, geographers
use a “void” cartographic space formed by an infinity of points.
The transition between these two types of spaces needs a defini-
tion of the geographical entity as a Cartesian product between
a locus and an object, which are differentiated on the surface of
the Earth.

1 Introduction
The oldest reflections of the Greek thought concerned nature and origin of
Cosmos. They included an enquiry on the shape of the Earth considered
as a Whole and on the pattern of the inhabited world to determine the
situation of it Parts, the ones in comparison with the others.
Eratosthene (-275, -193) is the inventor of the word geographiká (from
gê = the Earth, and gráphein = to make an incision in order to write or
to draw). His Geography included two books. The first one was a polemic
about the geographical value of the descriptive poetry. It also contained a
project for a scientific description of the Earth. The second book included a
calculation of the dimensions of the Earth (considered as a sphere) built up

231
232 G. Nicolas

with angular astronomical measures and evaluations of earthly distances.

Erastosthene proposed the first geo-metrical picture of the world that will
be later called chart or map (carte) from the 16th century in the Occident.
This use will definitely triumph only at the end of the 19th century: in
French, the word cartographe (cartographer) appears only in 1877.
History of geography is inseparable from history of cartography. Recip-
rocal assimilation of geographical space and cartographic space is a fact in
the middle of the 20th century. In this conception, geography begins with
the location of objects with coordinates projected on axes drew on a plane.
This know-how offers the advantage of allowing the use of Euclidean defi-
nition of the Earth in two or three dimensions. But the Whole/Parts logic
applied to the surface of the Earth, with or without the help of cartogra-
phy, is also present in the non-linear continuity of the history of occidental
geography.
This very long story unfolded in a context where relation between the
words finite and infinite inverted, [4]. Until the Renaissance, cartography
evolves chaotically and not always so rigorously as his founders wished. It
used implicitly infinite sets to represent points, lines and areas. Geography
for its part remained dependant of an Aristotelian world made of spherical
universes fitted together. The Earth was the most central of these uni-
verses and the most external sphere the finite limit. From the Renaissance
onward, cartography assimilates the revolution of Copernic and Newton,
which inverts the relation between finite and infinite. Closely submitted to
political powers, geography puts a central humanist vision in place of its
central cosmic vision of the Earth. It has not yet completely got free of
this vision nowadays.
At the middle of the 20th century, classical cartography (principally
topographical) has become completely mathematical. Geographers use in-
tensively these maps in the field and in their laboratory. They make also
graphical representations, paradoxically called “geographic maps”. But
this activity is not accompanied by a mathematical elaboration of their ob-
jects and of their methods. The making of their maps borrows techniques
of classical cartography. However scientific support of geographic cartog-
raphy disintegrates rapidly in the years 50–60 when geographers try to
quantify invisible facts on topographical maps. Quantitative geographers
borrow then their procedures to statistics without elaborating a properly
geographical object on which measures can be made.
The absence of properly geographical formalisation generates today dif-
ficulties in geographical computer science (geomatics). Methods and tech-
Finite and Infinite in Geography 233

niques being the same, data processing images of the surface of the Earth
made by statisticians, infographists and geographers are at the present time
similar. Assuming that, progresses have been realised by mathematical re-
search since the beginning of the 20th century concerning finite and infinite.
To what degree do they allow to devise an object for properly geographical
measuring?

2 Spatial Geographical Differentiation

Geography concerns macroscopic entities perceived at different scales. Mi-
croscopic or macrocosmic entities are not in the field of geography.

Definition 2.1 An entity is spatial if it is formed by the association of a

locus and an object.

Definition 2.2 An information is geographical if it differentiates either

the locus or the object, or the locus and the object altogether, of a spatial
entity situated or localised on the surface of the Earth.

If Λ is a finite set of locuses and O is a finite set of objects, then the

Cartesian product: P = Λ×O is the set of ordered pairs p = (λ, o), where λ
belongs to Λ and o belongs to O. Two pairs p1 = (λ1 , o1 ) and p2 = (λ2 , o2 )
are distinct, and we write p1 = p2 , if there is a differentiation (written with
a t) of at least one of their components, the locus or the object. There are
then four possibilities.
a) Locus and object differentiation: λ1 = λ2 and o1 = 02 .
Assuming that the relation = (negation of =) is anti-reflexive, sym-
metrical and not transitive, strong differentiation would be anti-
reflexive, symmetrical and not transitive.

b) Locus differentiation: λ1 = λ2 with o1 = o2 .

Assuming the properties of the relations = and =, weak locus differ-
entiation would also be anti-reflexive, symmetrical and not transitive.

c) Object differentiation: λ1 = λ2 with o1 = o2 .

For the same reasons as at point b), weak object differentiation is
anti-reflexive, symmetrical and not transitive.

d) Indifferentiation: λ1 = λ2 and o1 = o2 .
Assuming the properties of equality =, indifferentiation (or equiva-
lence) is reflexive, symmetrical and transitive.
234 G. Nicolas

Placing from top to bottom the four possibilities, that is from the
strongest to the weakest, we have:

Locus and object differentiation

λ1 = λ2 and o1 = o2
✟ ❍
✟ ❍
✟ ✟ ❍❍
✟ ❍
✙
✟✟ ❍
❥
❍
Weak locus differentiation Weak object differentiation
λ1 = λ2 with o1 = o2 λ1 = λ2 with o1 = o2
❍ ✟
❍❍ ✟✟
❍ ✟
❍❍ ✟✟
❥
❍ ✙
✟
Indifferentiation
λ1 = λ2 with o1 = o2

3 Geomaps and Maps

Given the situation S, which is the relative position of certain geographical
objects, expressed with order or non-metrical structures, a geomap M could
be worked out and processed with S.
Given the localisation L of geographical objects with numeric coordi-
nates and the graphic representation C of these objects, a map C could be
worked out and processed with L.
There is a duality between the pair (M, S) and the pair (C, L). Each
relation between M and S involves automatically a similar relation where
M can be replaced by C and S by L. This situation is analogous to the
one of the De Morgan laws. On the other hand, analogy between (M, S)
and (C, L) involves that M is to C what L is to S, that allows writing:
S | L ↔ M | C. In other words, geomappography is to cartography what
situation is to localisation.
Scale expresses the relation between objects and their representations
on a map or a geomap. The scale of a map is a connection between the
dimensions of the object and the dimensions of its representation. This
connection is opposite: the bigger the represented object the smaller its
cartographic scale. The scale of a geomap is given by the ratio of the
dimensions between represented geographical objects. This ratio is direct:
the dimension and the scale of a geomap vary in the same way. The geomap
scale of big geographical objects is a big scale. Inversely, the geomap scale
Finite and Infinite in Geography 235

of little geographical objects is a little scale.

Geomap represents relations between locuces-objects. It does not nec-
essarily use localisation. It can, for commodity reasons, be established on
a map to evoke in the observer’s mind the cartographic representation of
the geographical space. But, since a geomap is not supposed to be used
in the field, its design does not need to be as precise as a map. Geomap
allows to represent locus and object differentiation altogether (λ1 = λ2 and
o1 = o2 ), that is, the identity of the geographical locus-object.
A map, built on localisation, can express only one element of the entity
locus-object. If the differentiation is generated by the locus (weak differ-
entiation: λ1 = λ2 ), then only one sort of object (o1 = o2 ) corresponds
to all localisations and for every object a map must be worked out and
processed (analytical map). If, on the contrary, differentiation is generated
by the object (weak differentiation: o1 = o2 ), then all locuses are equiva-
lent (λ1 = λ2 ) and it is possible to represent on a same map many objects
in every localisation (synthetic map). Finally, if neither the locus nor the
object are differentiated, it is impossible to make a geomap or a map.

GEOMAP
Locus and object differentiation
✟ ❍
✟✟ ❍❍
✟ ✟ ❍❍
✙
✟✟ ❍❥
❍
ANALYTICAL MAP SYNTHETICAL MAP
Weak locus differentiation Weak object differentiation
❍❍ ✟
❍ ✟✟
❍❍ ✟✟
❍
❥
❍ ✙
✟✟
NEITER GEOMAP NOR MAP
Indifferentiation

4 Finite and Infinite in Geography

The differentiate character of geographical entities is not a mathematical
problem. It is however necessary to understand of what it consists its
mathematical properties. Difference indicates subject’s exteriority in com-
parison with reality. It also characterizes subject’s capacity to enter in
relations with reality and to recognize, to internalize, to incorporate and to
transform it into an object. Difference generates identity because it allows
236 G. Nicolas

to distinguish what belongs to an object, to an individual, to a sensation

or to an idea.
Historically, geographers use difference to know the Earth as a Whole.
Earth differentiation is therefore the characteristic of geographical reality
discovered by the geographer as a subject thanks to the cognitive mecha-
nism of difference. But geographers cannot go beyond differentiation of the
Earth as a Whole. The earthly Whole can be divided in Parts and every
Part can be equaled to a new Whole. But it is impossible to equal a Part
to the Earth considered as a Whole. The earthly Whole is therefore at the
start of comparison with other geographical Wholes.
Since there is no object without locus on the surface of the Earth, geo-
graphical differentiation regards all pairs formed by a locus and an object.
Geographical differentiation as a process shows the two historical and logi-
cal ways followed to do geography. The first way is thinking on the entities
formed by locuses and objects, which are simultaneously geographically
differentiated on the surface of the Earth (strong differentiation). The sec-
ond way is thinking on entities situated on the surface of the Earth whose
locuses (but not objects) or objects (but not locuses) are geographically dif-
ferentiated (weak differentiations). Finally, the last way, distinct from the
two others, the way of undifferentiated entities on the surface of the Earth,
concerns non-geographical spatial sciences (economy, geometry, psychology,
topology, etc.).
Researches on number theory by Georg Cantor (1845-1918) proved that
there are many kinds of finites and infinites, [1]. Cartographic space is
evidently constituted by an infinity of points (points, lines and areas). This
infinity allows two things. First, it makes possible the localisation of locuses
or objects with the help of coordinates projected on a system of axes on
a plane, which is a blank space (without physical objects). Secondly, this
non-enumerable infinite allows calculating localisations with precision. On
the other hand, geographical space generated by locuses-objects relations
has two limits: the empty set and the Earth considered as a Whole. The
relation between the geographer subject and the locus-object “Earth” sets
up the geographical objects. The finite number of geographical locuses-
objects is an ordered set generated by this relation.
Consequently, there is not one single geographical space but as many
spaces as possible ways of connecting locuses-objects. These geographical
spaces are not “void” but “full”. Their cardinal is finite and their limits
are the empty set (∅) and the Earth considered as a Whole.
Finite and Infinite in Geography 237

GEOMAP
Locus and object differentiation
Space resulting of the relation between a finite number
of locuses-objects ordered by the geographers.
This space has two limits, ∅ and the Earth as a Whole.
The cardinality of this “full” space of locuses-objects is a finite set.
✏✏ PP
✏ ✏✏ PP
PP
✮
✏ ✏ q
P
ANALYTICAL MAP SYNTHETICAL MAP
Weak locus differentiation Weak object differentiation
Representation of locuses or objects can be enumerated
and ordered in a “void” cartographic space.
This cartographic space is constituted of non-denumerable
sets of points (points, lines or areas)
❍ ✟
❍❍ ✟✟
❍ ✟
❍❍ ✟✟
❥
❍ ✙
✟✟
NEITHER GEOMAP NOR MAP
Indifferentiation
“Void” space constituted by a non-denumerable set of points

References
[1] S. Baruk, Dictionnaire de mathématiques élémentaires, Paris, Seuil,
1992-1995.

[2] A. Bouvier, M. George, F. Le Lionnais, Dictionnaire des mathématiques,

Paris, PUF, 1993.

[3] G. Nicolas, S. Marcus, Logique Tout/Partie, Actes du Colloque IRI,

Sion, IKB-Eratosthéne, to be released in September 1999; diffusion:
http://www.ikb.vsnet.ch/era3.html.

[4] J. P. Osier, Infini, Encyclopédie philosophique universelle, Les notions

philosophiques, Paris, PUF, 1990, vol. 1, 1289–1291.
 Ultrafilters, Dictators, and Gods1
Piergiorgio Odifreddi
Dipartimento di Informatica, Universita di Torino
Corso Svizzera 185, 10149 Torino, Italy
E-mail: piergior@di.unito.it

Abstract. We will show how a well-known and elementary

fact about ultrafilters can be reinterpreted both politically and
theologically, thus providing simple proofs of two central results
in these areas. Mathematicians can consider such formulations
as the true essence of such results. Outsiders can find comfort in
a saying by Goethe, who once said (Maxime und Reflectionen,
1729): “Mathematicians are like Frenchmen: as soon as you tell
them something, they translate it into their own language, and
it immediately appears different”.

1 Ultrafilters
In 1937 Henri Cartan introduced the following notions. Given a set A of
elements, let A be a subset of the power set of A, i.e., a set of subsets of
A. We use small letters such as x and y to refer to elements of A, and
capital letters such as X and Y to refer to subsets of A. Let us consider
the following possible properties of A:
1. A is upward closed with respect to to inclusion, i.e., if X ∈ A and
X ⊆ Y , then Y ∈ A;
1
Solomon Marcus has often stressed the broad cultural interest of mathematics, writ-
ing extensively on its connections with apparently distant areas such as poetics and the-
ology. A few years ago he published, together with Cristian Calude and Doru Ştefănescu,
an interesting paper with an unusual title, The Creator versus its creation. From Sco-
tus to Gödel, Collegium Logicum, Annals of the Kurt-Gödel-Society, Vol. 3, Institute of
Computer Science, AS CR Prague, Vienna, 1999, 1-10, in which he was kind enough to
quote a manuscript of ours on the mathematical modelling of God. His Festschrift seems
a perfectly suited occasion not only to publish those observations, but also to dedicate
them to him.

239
240 P. Odifreddi

2. A is closed with respect to intersection, i.e., if X ∈ A and Y ∈ A,

then X ∩ Y ∈ A;

3. A does not contain all subsets of A, i.e., there exists at least an X

which is not in A;

4. given any subset X of A, either it or its complement belongs to A,

i.e., X ∈ A or X ∈ A.
A is called a filter on A if it satisfies Properties 1 and 2. The idea is that
A consists of “large” subsets of A, and Properties 1 and 2 are obviously
necessary: if a set is larger than a large one, it is a fortiori large; and if
two sets are large, their intersection should still be large. The following are
examples of filters:
• The subsets of A containing a given element x. This is called the
principal filter generated by x.

• The cofinite subsets of A, i.e., the subsets containing all elements of

A with at most finitely many exceptions. Such a filter is not principal
because, given any element x, the set A − {x} obtained by taking x
out of A is cofinite, and hence it belongs to the given filter, but it
does not contain x, and hence it does not belong to the principal filter
generated by x.
Properties 1 and 2 are not sufficient to forbid A to contain also “small”
subsets of A. For example, nothing forbids A to contain all subsets of A.
Property 3 is introduced to avoid this extreme possibility, and a filter that
satisfies it is called a proper filter. Because of Property 1, a filter is proper
if and only if it does not contain the empty set ∅, because the latter is
contained in every subset of A.
Property 4 is, in a sense, opposite to 3: while the latter forbids the filter
to be too large, the former forbids it to be too small, and a proper filter
satisfying it is called an ultrafilter. Going back to the previous examples,
we note that:
• A principal filter is an ultrafilter. Indeed, if it is generated by x, then
either x is in X, or it is in X.

• The cofinite subsets of A are not an ultrafilter. Indeed, if A is finite,

then they form a filter which is not proper. And if A is infinite, there
is a subset X of A such that both X and X are infinite, and thus
neither X nor X is cofinite.
Ultrafilters, Dictators, and Gods 241

The fact we alluded at above is the following.

Theorem 1.1 If A is finite, then every ultrafilter on A is a principal filter.

Proof. Let U be an ultrafilter on a finite set A, and I be the intersection

of all elements of U.
Since A has only a finite number of elements, it also has a finite num-
ber of subsets, and U is finite. Thus I is obtained by a finite number of
intersections of elements of U, and it belongs to U by Property 2.
By Property 3, I cannot be empty. Thus, it contains an element x. By
Property 4, either {x} or {x} is in U. But it is impossible that {x} be in
U, since it does not contain x, which is in I (and hence in all elements of
U). Then {x} is in U, and I = {x}.
It follows that U is the principal filter generated by x. Indeed, on the
one hand, every element of U contains x, by definition of I. On the other
hand, if x belongs to X as an element, then X contains {x} as a subset,
and thus it is in U by Property 1. ✷

Corollary 1.1 If an ultrafilter on A contains the intersection of all its

elements, then it is a principal filter.

Proof. The finiteness assumption in the previous theorem is used only to

deduce that the intersection of all elements of the ultrafilter is still in the
ultrafilter. ✷

We note in passing that the results just stated are not trivially true,
since on every infinite set A there are ultrafilters that are not principal (for
example, any ultrafilter containing the cofinite sets).

2 Dictators
In 1785 Jean Antoine Nicolas Marie de Caritat, better known as the Mar-
quise of Condorcet, discovered the following paradox of the majority voting
system. Consider three voters 1, 2 and 3, that have to choose among the
alternatives A, B and C. Suppose they have the following cyclic orders of
preferences:
1: A B C
2: B C A
3: C A B,
242 P. Odifreddi

to be read in the following way: 1 prefers A to B and B to C, 2 prefers B

to C and C to A, and 3 prefers C to A and A to B. When the alternatives
are put to a vote two by two, A wins over B by two votes (those of 1 and
3) to one (that of 2), and similarly B wins over C by two votes (1 and 2) to
one (3). One could thus think that A should win over C, while the opposite
happens, and C wins over A by two votes (2 and 3) to one (1).
The paradoxical aspect is that, while the individual preferences of the
voters are linearly ordered, the social order that majority vote produces
becomes circular.
In 1951 Kenneth Arrow asked himself whether one could find a voting
system that would preserve the linear order of individual preferences. Given
a set of voters, let us call a subset X of it decisive if, given two alternatives,
one of them wins when all the elements of X vote for it, and all the elements
of X vote for the other.2
The notion of a decisive set makes sense under the hypothesis of vote
dependence, i.e., when the choice of the winner between two alternatives de-
pends only on the votes they receive according to the individual preference
orders, but not on other factors dependent on the alternatives.
Arrow’s basic assumption is the following:
Axiom 2.1 The decisive sets form an ultrafilter on the set of voters.
This assumption can be justified as follows. Property 1 says that if a
set of voters is decisive, then so is any set containing it. In other words, if
an alternative wins when it receives a certain number of votes, it continues
to win if it receives more votes (monotonicity).
Property 3 says that not every set is decisive, which is obvious if at
most one alternative has to win.
Property 4 says that either a set or its complement are decisive, which
is obvious if at least one of the two alternatives has to win, i.e., if there has
to be no tie.
The following is thus the essential justification of Arrow’s assumption.
Proposition 2.1 Property 2, i.e., the fact that the intersection of two de-
cisive sets is decisive, is equivalent to the fact that the social order is linear.
Proof. Suppose X and Y are decisive, but X ∩ Y is not. Then X ∩ Y
is decisive by Property 4. Consider the following orders of preferences,
2
In a majority vote, the decisive sets are those containing at least one half of the
voters plus one. They satisfy Properties 1, 3 and 4, the latter under the hypothesis that
the set of voters has an odd number of elements, but not Property 2, by Proposition 2.1
and Condorcet’s Paradox.
Ultrafilters, Dictators, and Gods 243

and note the analogy with the system of preferences used in Condorcet’s
Paradox:
X ∩Y : A B C
Y −X : B C A
X −Y : C A B
X ∪Y : C B A,
to be read as follows: every voter on the left set has the order of preference
indicated on the right. Then A wins over B because every element of X
prefers A to B, every element of X prefers B to A, and X is decisive.
Similarly, B wins over C because Y is decisive, and C wins over A because
X ∩ Y is decisive. Thus the social order is not linear.
Conversely, given any system of individual preference orders, consider
the sets X, Y and Z of voters preferring, respectively, A to B, B to C and
A to C. If A wins over B, then X must be decisive, otherwise X would
be decisive by Property 4, and B would win over A. Similarly, if B wins
over C, then Y must be decisive. If the intersection of two decisive sets
is decisive, then X ∩ Y is decisive. Since every element of it prefers A to
B (being in X) and B to C (being in Y ), and hence A to C (because the
linear orders are linear), it follows that X ∩ Y ⊆ Z. By Property 4, then
Z is decisive and A wins over C. Thus the social order is linear. ✷

Note that the previous proof uses what Arrow calls the hypothesis of
individual freedom: every possible combination of individual preference or-
ders is admissible.
Now, since the set of voters is obviously finite in any application, we
can use Theorem 1.1 and get a version of Arrow’s result.
Theorem 2.1 The decisive sets form a principal filter.
In other words, there exists a voter whose vote alone determines the
result of any election. Arrow calls such a voter a dictator , and deduces
from this that it is impossible to find a democratic voting system satisfying
the minimal conditions used above, in particular vote dependence, mono-
tonicity and individual freedom. Thus every possible system produces a
dictatorship, and democracy is impossible. For this result Arrow obtained
the Nobel Prize for Economics in 1972.

3 Gods
In 1077 Anselm of Aosta discovered the following ontological proof of the
existence of God: if we define God as a being with all perfections, then
244 P. Odifreddi

God exists because existence is a perfection.

The unsatisfactory aspects of this proof are that, on the one hand,
nothing ensures that the definition is not contradictory, and on the other
hand, that existence cannot be considered as a property, and hence as a
perfection.
In 1970 Kurt Gödel asked himself whether one could formalize the on-
tological proof, thus making it logically more acceptable. The idea is to
replace perfections by positive properties, intended as particular subsets of
the set of elements constituting the world.3
Gödel’s assumption is the following:

Axiom 3.1 The positive properties form an ultrafilter on the world.

More explicitly, one assumes that a property that contains a positive

property is positive, the intersection of two positive properties is positive, a
positive property is non empty, and the complement of a positive property
is not positive.4
Such assumptions can be justified by analogy, i.e., by noticing that
a number larger than a positive number is positive, the product of two
positive number is positive, a positive number is not zero, and the opposite
of a non-positive number is positive.
If we make the further assumption of finiteness of the world , we can
apply Theorem 1.1 and get a version of Gödel’s result.

Theorem 3.1 The positive properties form a principal filter.

In other words, there exists an object determined by the positive prop-

erties. Gödel calls such an object God , and deduces from this the existence
of God, on the basis of the minimal conditions used in the justification of
the axiom that the positive properties form an ultrafilter, and on the basis
of the finiteness of the world.
To eliminate the latter and unsatisfying assumption, Gödel replaces
it by an additional axiom which, probably, Dostoevsky would not have
accepted.
3
A property can be identified with the set of elements satisfying it, and a set can be
identified with the property of belonging to it. Then logical implication, conjunction and
negation correspond to set theoretical inclusion, intersection and complement.
4
Notice that the set of positive properties, being a filter, is closed with respect to
finite intersection and arbitrary union: in particular, it forms a divine topology on the
universe. Since the filter is proper, this topology is connected and not separable. More
information on it can be found in the paper by Calude, Marcus and Ştefănescu quoted
at the beginning.
Ultrafilters, Dictators, and Gods 245

Axiom 3.2 Being God is a positive property.

Since the property of being God is, by definition, the intersection of all
positive properties, the effect of this axiom is to make such an intersection
an element of the ultrafilter of all positive properties. By Corollary 1.1
such an ultrafilter is then principal, and the existence of God now follows
as above.

4 Conclusion
Theorem 1.1 has shown that there is a connection between the existence
of dictators and gods. We imagine that it would be difficult to make such
a connection more explicit, if one excludes trivialities such as: God can be
considered as a dictator, or dictators like to be considered as gods. However,
as Umberto Eco as taught us, on what we cannot theoretize, we must tell
a story. We thus end with an anecdote.
In 1948 Gödel had to take the examination to become an American
citizen. He studied the Constitution, and found out that in it there was a
logical-legal possibility of turning the United States into a dictatorship. In
the car trip to the ceremony, he communicated the discovery to Einstein,
who was his witness. The latter tried to persuade Gödel that the ceremony
would not have been the best time to talk about this, but by chance the
judge noticed that Gödel was of Austrian citizenship, and referring to the
Anschluss told him that in the United States he should not fear anymore the
advent of a dictator. Einstein had to jump in, since Gödel had immediately
seized the chance and started a communication of his most recent discovery.
We do not know which proof of the possibility of a dictatorship in the
United States Gödel had found. But the developments reported in this note
allow the reader to formulate an obvious conjecture.
 On the Behaviour of Semiautomata
Sergiu Rudeanu
Faculty of Mathematics, University of Bucharest
Str. Academiei 14, 70109 Bucureşti, Romania
E-mail: rud@funinf.math.unibuc.ro

Abstract. It has been known for a long time that the Moore
automaton constructed by Nerode is minimal, i.e., it has the
smallest number of states among all the automata with the
same behaviour. A theorem due to Goguen has revealed that
the deep background of this property is the adjunction of two
functors between the category of reachable Moore automata and
the category of their behaviours. In this paper we introduce the
category of F -automata and the category of F -behaviours and
prove a Goguen-like theorem which includes as particular cases
both the original Goguen theorem and a similar theorem for
Mealy automata. We also characterize those functions which
occur as behaviours of semiautomata and prove that the cate-
gory of reachable semiautomata is isomorphic to the category
of their behaviours.

1 Introduction
The starting point of this paper is a theorem due to Goguen [4], which
establishes a pair of adjoint functors between the category of reachable
Moore automata and the category of surjective behaviours f : X ∗ → Y .
Since semiautomata can be viewed as particular cases of Moore automata,
it is natural to ask which behaviours can be realized by reachable semiau-
tomata. Another natural question is whether the Goguen theorem can be
extended to Mealy automata.
In Section 2 we prove that a surjection f : X ∗ → Y can be realized by a
reachable semiautomata if and only if kerf is right invariant. In Section 3
we prove a theorem which includes as particular cases the Goguen theorem

247
248 S. Rudeanu

mentioned above and a quite similar result for Mealy automata. In Section
4 we deal with the particular case of semiautomata and prove the sharper
result that the category of reachable semiautomata is isomorphic to the
category of their behaviours.
All necessary automata-theoretical prerequisites will be recalled; cf. [1],
[2], [3], [5].

2 Behaviours Realized by Semiautomata

In the sequel we use the following standard notation: A∗ is the monoid of all
the words over the alphabet A, λ denotes the empty word, A+ = A∗ − {λ},
while α∗ : A∗ → A
∗ stands for the monoid homomorphism which extends
a map α : A → A
.
A semiautomaton is an algebra Σ = (S, I; δ, s0 ), where δ : S × I → S
and s0 ∈ S are known as the transition function or next-state function and
the initial state or central state or source, respectively.
The same letter δ is used for the function δ : S × I ∗ → S which extends
the transition function and satisfies
(1.1) δ(s, λ) = s, ∀s ∈ S
(1.2) δ(s, w1 w2 ) = δ(δ(s, w1 ), w2 ), ∀s ∈ S, ∀w1 , w2 ∈ I ∗ .
The reachability function δ 0 : I ∗ → S is defined by
(2) δ 0 (w) = δ(s0 , w), ∀w ∈ I ∗ ,
and the semiautomaton is said to be reachable provided δ 0 is surjective.
A Moore automaton (also called a (sequential) machine in [1], [3], [4])
is an algebra A = (S, I, O; δ, µ, s0 ) where Σ(A) = (S, I; δ, s0 ) is a semiau-
tomaton and µ : S → O is known as the output function. The automaton
is said to be reachable provided its semiautomaton reduct Σ(A) is so.
The function β : I ∗ → O defined by
(3) β(w) = µ(δ 0 (w)), ∀w ∈ I ∗
is called the behaviour of A ; we also say that the automaton A realizes the
function β.
This algebraic model represents a discrete-time technical device which
has the set S of internal states, receives input signals from the set I and
produces output signals belonging to the set O. The operation of the device
obeys the law
(4) sn+1 = δ(sn , in ), on = µ(sn ),
where xk denotes the value of x ∈ X at time t := k ∈ N . So the internal
On the Behaviour of Semiautomata 249

state at time t := 0 is s0 and if the device receives the sequence of inputs

i0 , i1 , . . ., in−1 , then at time t := n it will assume the state δ 0 (w) and
produce the output signal β(w), where w = i0 i1 . . . in−1 ∈ I ∗ . Note that
δ 0 (λ) = s0 and β(λ) = o0 .
Now remark that a semiautomaton Σ = (S, I; δ, s0 ) can be identified
with the Moore automaton A(Σ) = (S, I, S; δ, 1S , s0 ), where 1S is the iden-
tity map of the set S. The behaviour of A(Σ) is simply δ 0 . Having in view
this remark it is natural to rename the reachability function δ 0 as the be-
haviour of the semiautomaton Σ and to say that Σ realizes the function δ 0 .
Proposition 2.1 below provides the answer to the following natural ques-
tion: what are the functions f : X ∗ → Y that can be realized by reachable
semiautomata?
Recall first that kerf is the equivalence defined on the domain of the
function f by x kerf y ⇔ f (x) = f (y) and that an equivalence ≡ on a
monoid X ∗ is called right invariant provided w1 ≡ w2 ⇒ w1 w ≡ w2 w.

Proposition 2.1 A function f : X ∗ → Y is realized by a reachable semi-

automaton if and only if f is surjective and kerf is right invariant.

Proof. If the function f satisfies the above conditions, then the semiau-
tomaton
(5.1) SN f = (Y, X, σf , f (λ)),
(5.2) σf : Y ×X → Y , σf (f (w), x) = f (wx), ∀w ∈ X ∗ , ∀x ∈ X,

is well defined.
Further we show that the extension σf : Y × X ∗ → Y of σf is given by
(6) σf (f (w), w
) = f (ww
), ∀w, w
∈ X ∗ .
The proof is by induction on w
. For w
:= λ relation (6) holds by (1.1),
while (6) implies, via (1.2) and (5.2), that for every x ∈ X,
σf (f (w), w
x) = σf (σf (f (w), w
), x) = σf (f (ww
), x) = f (ww
x).
Now (6) implies
(σf )0 (w) = σf (f (λ), w) = f (λw) = f (w),
that is
(7) (σf )0 = f .
Conversely, if Σ is a reachable semiautomaton, then δ 0 is surjective and
ker δ 0 is known to be right invariant (the easy proof uses (2) and (1.2)). ✷

The deep background of Proposition 2.1 will be pointed out in Section

4; cf. Section 5.
250 S. Rudeanu

3 Mealy Automata and a Generalization of

Goguen’s Theorem
The definition of a Mealy automaton A = (S, I, O; δ, µ, s0 ) is similar to that
of a Moore automaton, with the difference that we have now µ : S × I → O.
The technical meaning of this concept is similar to that of a Moore
automaton; in particular on = µ(sn , in ). So in this case it is natural to
define the behaviour of the automaton as the function β : I + → O given by
(8) β(wi) = µ(δ 0 (w), i), ∀w ∈ I ∗ , ∀i ∈ I,
whose meaning is the same as for the behaviour of a Moore automaton.
Note that the initial output o0 of a Mealy automaton is undetermined.
In this section we generalize the Goguen theorem mentioned in the
Introduction in such a way as to include a similar result for Mealy automata;
as a matter of fact, the proof itself will be quite similar to the original one.
Let F : Set → Set be a functor which preserves surjections.
By an F -automaton we mean an algebra A = (S, I, O; δ, µ, s0 ), where
ΣA = (S, I; δ, s0 ) is a semiautomaton and µ : S × F I → O is called the
output function. The automaton A is said to be reachable if ΣA is so. The
behaviour of A is the function
(9) β : I ∗ × F I → O, β(w, j) = µ(δ 0 (w), j), ∀w ∈ I ∗ , ∀j ∈ F I.
In the case when F is the constant functor F I = {•}, where {•} is
a fixed singleton, and F f = 1{•} , we identify I ∗ × {•} with I ∗ and the
behaviour (9) with the behaviour (3), so that we obtain the concept of a
Moore automaton. Taking F := IdSet = the identity functor of Set, we
have I ∗ × F I = I + and we identify the behaviour (9) with the behaviour
(8), so that we obtain the concept of a Mealy automaton.
We define the category F Aut of F -automata in the spirit of universal al-
gebra: the objects of F Aut are the F -automata, while the morphisms from
A = (S, I, O; δ, µ, s0 ) to A
= (S
, I
, O
; δ
, µ
, s
0 ) are the triples (a, b, c)
where a : S → S
, b : I → I
and c : O → O
satisfy the identities:
(10.1) a(δ(s, i)) = δ
(a(s), b(i)), ∀s ∈ S, ∀i ∈ I,
(10.2) c(µ(s, j)) = µ
(a(s), F b(j)), ∀s ∈ S, ∀j ∈ F I,
(10.3) a(s0 ) = s
0 ,
the composition of morphisms being defined componentwise.
It is easy to check, using the fact that F is a functor, that F Aut is
actually a category.

Lemma 3.1 Properties (10.1) and (10.3) imply

On the Behaviour of Semiautomata 251

(11) a(δ(s, w)) = δ

(a(s), b∗ (w)), ∀s ∈ S, ∀w ∈ I ∗ ,
(12) a(δ 0 (w)) = δ
0 (b∗ (w)), ∀w ∈ I ∗ ,
and if (10.2) holds also, then
(13) c(β(w, j)) = β
(b∗ (w), F b(j)), ∀w ∈ I ∗ , ∀j ∈ F I.

Proof. Identities (11) and (12) are well known: (11) is established by
induction on w, via (1.1), (10.1) and (1.2), while (12) is obtained from (11)
written for s := s0 and taking into account (10.3). To prove (13) we use
(9), (12) and (10.2):

β
(b∗ (w), F b(j)) = µ
(δ
(b∗ (w)), F b(j))
0

= µ
(a(δ 0 (w)), F b(j))
= c(µ(δ 0 (w), j))
= c(β(w, j)).

Let further F Beh be the category of F -behaviours, whose objects are

the functions of the form f : X ∗ ×F X → Y , while the morphisms from f to
f
: X
∗ × F X
→ Y
are the pairs (b, c) where b : X → X
and c : Y → Y

satisfy the identity

(14) c(f (w, z)) = f
(b∗ (w), F b(z)), ∀w ∈ X ∗ , ∀z ∈ F X,
the composition of morphisms being again defined componentwise.
It is easy to check, using the fact that ∗ and F are functors, that F Beh
is actually a category.
Now let F A be the subcategory of F Aut which consists of those
F -automata that are reachable and of those morphisms (a, b, c) for which
b is surjective. Let also F B be the subcategory of F Beh which consists of
the same objects f : X ∗ × F X → Y and of those morphisms (b, c) between
them for which b is surjective.
The functor external behaviour
(15) E : FA → FB
is defined by EA = β = the behaviour (9) of A, and E(a, b, c) = (b, c). (As
a matter of fact the morphism sets of a category have to be disjoint, so that
the morphisms of F Aut are in fact of the form ((a, b, c), A, A
) and accord-
ingly the latter definition should read E((a, b, c), A, A
) = ((b, c), EA, EA
).
However we conform to common practice and use a less sophisticated no-
tation.)
252 S. Rudeanu

Since the behaviour of an F -automaton is an F -behaviour by Lemma

3.1, (13), it is plain that E is actually a functor (as a matter of fact we
could have defined E : F Aut → F Beh). We are now going to construct a
functor working in the opposite direction, from F B to F A.
We associate with each function f : X ∗ × F X → Y the equivalences
ker1 f and ∼f on X ∗ defined as follows: ∀w1 , w2 ∈ X ∗ ,
(16) w1 ker1 f w2 ⇔ f (w1 , z) = f (w2 , z), ∀z ∈ F X,
(17) w1 ∼f w2 ⇔ f (w1 w, z) = f (w2 w, z), ∀w ∈ X ∗ , ∀z ∈ F X.

Lemma 3.2 ∼f is the greatest right invariant equivalence relation included

in ker1 f .

Proof. It is routine to check that ∼f is a right invariant equivalence relation

included in ker1 f . If ≡ is another relation with these properties then
w1 ≡ w2 ⇒ w1 w ≡ w2 , ∀w ⇒ w1 w ker1 f w2 w, ∀w ⇔ f (w1 w, z) =
f (w2 w, z), ∀z, ∀w ⇔ f (w1 w, z) = f (w2 w, z), ∀w, ∀z ⇔ w1 ∼f w2 . ✷

Corollary 3.1 For every F -automaton, ker δ 0 ⊆∼β ⊆ ker1 β.

Proof. Recall that kerδ 0 is right invariant, note that kerδ 0 ⊆ker1 β by (9)
and apply Lemma 3.2. ✷

For every w ∈ X ∗ , let [w]f denote the coset of w modulo ∼f .

Proposition 3.1 Let f : X ∗ × F X → Y . Then:

(i) The following construction is a reachable F -automaton:
(18) N f = (X ∗ / ∼f , X, Y ; δf , µf , [λ]f ),
(19) δf : (X ∗ / ∼f ) × X → X ∗ / ∼f , δf ([w]f , x) = [wx]f , ∀w ∈ X ∗ ,
∀x ∈ X,
(20) µf : (X ∗ / ∼f ) × F X → Y , µf ([w]f , z) = f (w, z), ∀w ∈ X ∗ ,
∀z ∈ F X;
(ii) δf ([w]f , w
) = [ww
]f , ∀w,w
∈ X ∗ ;
(iii) (δf )0 =nat∼f ;
(iv) EN f = f .

Comment. N f will be called the Nerode F -automaton associated

with f .
Proof. If [w1 ]f = [w2 ]f then w1 x ∼f w2 x and f (w1 , z) = f (w2 , z) by
Lemma 3.2, therefore N f is well defined.
On the Behaviour of Semiautomata 253

Property (ii) is easily proved by induction on w

.
From (ii) we obtain (δf )0 (w) = δf ([λ]f , w) = [w]f , that is (iii), which
completes the proof of (i) as well.
Finally from (9)(iii) and (20) we infer
EN f (w, z) = µf ((δf )0 (w), z) = µf ([w]f , z) = f (w, z). ✷

Lemma 3.3 Let A = (S, I, O; δ, µ, s0 ) be a reachable F -automaton,

f : X ∗ × F X → Y and (b, c) ∈ F B(EA, f ). Then there is a unique
function a : S → X ∗ / ∼f such that (a, b, c) ∈ F A(A, N f ), namely
a(δ 0 (w)) = [b∗ (w)]f , ∀w ∈ I ∗ .

Proof. Since A is reachable, the elements of S are actually of the form

δ 0 (w) and in order to prove that a is well defined it remains to show
that if δ 0 (w1 ) = δ 0 (w2 ) then b∗ (w1 ) ∼f b∗ (w2 ), that is f (b∗ (w1 )w
, z) =
f (b∗ (w2 )w
, z) for every w
∈ X ∗ and every z ∈ F X. Recall that
EA = β : I ∗ × F I → O. Since b : I → X is a surjection, so is b∗ and since F
preserves surjections it follows that F b is a surjection. Therefore w
= b∗ (w)
for some w ∈ I ∗ and z = F b(j) for some j ∈ F I. Taking into account the
homomorphism condition (14) written for f := EA = β and f
:= f , then
(9) and the right invariance of kerδ 0 , we obtain

f (b∗ (w1 )w
, z) = f (b∗ (w1 )b∗ (w), F b(j))
= f (b∗ (w1 w), F b(j)) = c(β(w1 w, j))
= c(µ(δ 0 (w1 w), j)) = c(µ(δ 0 (w2 w), j)) = · · ·
= f (b∗ (w2 )w
, z).

We prove similarly the homomorphism conditions (10) for δ

:= δf ,
µ
:= µf and s := δ 0 (w):

a(δ(δ 0 (w), i)) = a(δ(δ(s0 , w), i)) = a(δ(s0 , wi))

= a(δ 0 (wi)) = [b∗ (wi)]f
= [b∗ (w)b(i)]f = δf ([b∗ (w)]f , b(i))
= δf (a(δ 0 (w)), b(i)),
c(µ(δ 0 (w), j)) = c(β(w, j)) = f (b∗ (w), F b(j))
= µf ([b∗ (w)]f , F b(j))
= µf (a(δ 0 (w)), F b(j)),
a(s0 ) = a(δ(s0 , λ)) = a(δ 0 (λ)) = [b∗ (λ)]f = [λ]f .
254 S. Rudeanu

To prove uniqueness, suppose that (a

, b, c) ∈ F A(A, N f ) for some
a
: S → X ∗ / ∼f . Then Lemma 3.1 and Proposition 3.1 (iii) imply
a
(δ 0 (w)) = (δf )0 (b∗ (w)) = [b∗ (w)]f . ✷

Proposition 3.2 The following construction defines a functor

N : F B → F A. Take an object f of F B to the Nerode F -automaton N f
and if (b, c) ∈ F B(f, f
) define N (b, c) = (a, b, c), where
(21) a : X ∗ / ∼f → X
∗ / ∼f
, a([w]f ) = [b∗ (w)]f
, ∀w ∈ X ∗ .

Comment. N will be called the Nerode functor.

Proof. We have seen in Proposition 3.1 that N f is an object of F A. Further
let (b, c) ∈ F B(f, f
).
Since EN f = f by Proposition 3.1 (iv), we have (b, c) ∈ F B(EN f, f
)
and we can apply Lemma 3.3 with A := N f and f := f
, hence S :=
X ∗ / ∼f and X := X
. So there is a unique function α : X ∗ / ∼f → X
∗ / ∼f

such that (α, b, c) ∈ F A(N f, N f
), namely α((δf )0 (w)) = [b∗ (w)]f
. But
(δf )0 (w) = [w]f by Proposition 3.1 (iii), so that α = a, therefore a is well
defined and (a, b, c) ∈ F A(N f, N f
). Checking the functorial conditions
for N is routine. ✷

Theorem 3.1 (E, N ) is a pair of adjoint functors between the categories

F A and F B.

Proof. It follows from Proposition 3.1 (iv) and EN (b, c) = E(a, b, c) =

(b, c), that EN = IdFB . This enables us to prove the desired conclu-
sion in the following form: for every object f of F B, the pair (EN f, 1f ),
where 1f : EN f → f is the identity morphism, is a final object in the
category E/f . To be specific, this claim means that for every object
u = (b, c) : EA → f of E/f there is a unique morphism u : A → N f such
that u = 1f ◦ Eu = Eu. Setting u = (a, b
, c
), the latter condition becomes
(b, c) = (b
, c
), hence the claim reduces to the existence of a unique a such
that (a, b, c) ∈ F A(A, N f ), which is true by Lemma 3.2. ✷

4 An Isomorphism of Categories
Proposition 2.1 characterizes the behaviours of reachable semiautomata.
In this section we apply Goguen’s technique to obtain an isomorphism
On the Behaviour of Semiautomata 255

between the category of reachable semiautomata and the category of their

behaviours.
Let SA be the category whose objects are the reachable semiautomata
Σ = (S, I; δ, s0 ) and whose morphisms from Σ to Σ
= (S
, I
; δ
, s
0 ) are the
pairs (a, b) where a : S → S
, b : I → I
is a surjection and the identities
(10.1) and (10.3) hold.
Let SB be the category whose objects are the surjections of the form
f : X ∗ → Y for which kerf is right invariant and whose morphisms from
f to f
: X
∗ → Y
are the pairs (b, a) where b : X → X
is surjective,
a : Y → Y
and the following identity holds:
(22) a(f (w)) = f
(b∗ (w)), ∀w ∈ X ∗ .
In both categories the composition of morphisms is defined componen-
twise.
Define a functor SE : SA → SB as follows. Using the same notation
Σ, Σ
as before, set SEΣ = δ 0 : I ∗ → S and for (a, b) ∈ SA(Σ, Σ
) set
SE(a, b) = (b, a). Then δ 0 is an object of SB by Proposition 2.1. On the
other hand, it follows from Lemma 3.1 that identity (12) holds, which is
condition (22) for δ 0 and δ
0 ; so (b, a) ∈ SB(δ 0 , δ
0 ) = SB(SEΣ, SEΣ
).
A functor SN : SB → SA is obtained which takes an object f of SB
to the object SN f of SA defined by (5) in the proof of Proposition 2.1
and which takes a morphism (b, a) ∈ SB(f, f
) to SN (b, a) = (a, b). Then
conditions (10.1) and (10.3) are fulfilled for the semiautomata SN f and
SN f
, because conditions (5) and (22) imply
a(σf (f (w), x)) = a(f (wx)) = f
(b∗ (wx)) = f
(b∗ (w)b(x))
= σf
(f
(b∗ (w)), b(x)) = σf
(a(f (w)), b(x)),
a(f (λ)) = f
(b∗ (λ)) = f
(λ).
Theorem 4.1 The functors SE and SN establish an isomorphism between
the categories SA and SB.
Proof. It follows from Proposition 2.1, (7) that SE(SN f ) = (σf )0 = f ,
while SE(SN (b, a)) = SE(a, b) = (b, a). Therefore SE ◦ SN = IdSB .
On the other hand, using (5) we obtain
(23) SN (SE(Σ)) = SN (δ 0 ) = (S, I; σδ0 , δ 0 (λ)),
σδ0 (δ 0 (w), i) = δ 0 (wi) = δ(s0 , wi) = δ(δ(s0 , w), i) = δ(δ 0 (w), i),
that is, σδ0 = δ. Besides, δ 0 (λ) = δ(s0 , λ) = s0 . Therefore property (21)
becomes SN (SE(Σ)) = Σ. Since we have also SN (SE(a, b)) = SN (b, a) =
(a, b), it follows that SN ◦ SE = IdSA . ✷
256 S. Rudeanu

5 Conclusions
As already noticed, taking the constant functor F X := {•}, F f := 1{•} ,
we obtain Moore automata and the behaviour (3). Moreover, the theory
constructed in Section 3 is simplified by the disappearance of the factors
F I and F X and of the second argument of the functions; in particular
ker1 f reduces to kerf . Briefly, we obtain the original Goguen theorem and
its proof.
We have also noted that for F := IdSet we obtain Mealy automata
and the behaviour (8). So, the corresponding theorem for Mealy automata
and its proof are obtained from Section 2 by replacing F I, F X, I ∗ × F I
and X ∗ × F X by I, X, I + and X + , respectively.
It has been known for a long time that the Nerode automaton is mini-
mal, i.e., it has the smallest number of states among all automata with the
same behaviour. The Goguen theorem has revealed that the deep back-
ground of this property is the adjunction between his functors E and N .
The Nerode F -automaton is also minimal, as shown by the same (easy)
proof as for Moore automata.
Note also that Proposition 2.1 is in fact included in the isomorphism
constructed in Theorem 4.1 between the category of reachable semiau-
tomata and the category of their behaviours.

Acknowledgement. The author wishes to thank Professor Cristian

Calude for his valuable remarks and in particular for the conjecture which
led to Theorem 3.1.

References
[1] M. A. Arbib, E. G. Manes, Arrows, Structures and Functors. The Cat-
egorical Imperative, Academic Press, New York, 1957.
[2] P. Deussen, Halbgruppen und Automaten, Springer-Verlag, Berlin, 1971.
[3] A. Ginzburg, Algebraic Theory of Automata, Academic Press, New
York, 1968.
[4] J. A. Goguen, Realization is universal, Math. Systems Theory, 6, 4
(1973), 359–374.
[5] R. E. Miller, Switching Theory, Vol. 2, John Wiley & Sons, New York,
1965.
 Infinity – An Outline of Conceptions in
Mythology, Cosmology and Natural Philosophy
Jouko Seppänen
Helsinki University of Technology
Department of Computer Science
02150 Espoo, Finland
E-mail: jouko.seppanen@hut.fi

Abstract. The earliest conceptions of infinity and eternity ap-

parently occurred in Oriental cosmological and cosmogonical
myths and fables that have been passed down in popular tradi-
tion and epic texts. Some of these became codified in literary
religions as holy scripts and were subsequently refined in inter-
pretations and commentaries. The pre-Socratic Greek natural
philosophers were the first to detach themselves from myths
and religious authority and began to ponder questions of cos-
mology and cosmogony independently of religious tradition and
offered a variety of original views many of which involved no-
tions of infinity and eternity in various senses and contexts. The
Medieval patriarchs, theologians and scholastic philosophers di-
gested and filtered the Oriental religious and the Greek philo-
sophical thought, transferred the idea of infinity from cosmol-
ogy to theology and carried speculations about the immensity
of God and His virtues to the extreme. In the Renaissance the
legacy of the Greek natural philosophy and Indian arithmetics
were mediated to Europe by the Arabs which allowed the nat-
ural philosophers of the Renaissance to bring the notion back
to physical cosmology and the astronomers and mathematicians
to pave way for the rise of science and new manifestations and
interpretations of the notion of infinity.
In the history of classical and modern astronomy and cosmology
infinity continued to play a variety of roles until it was finally

257
258 J. Seppänen

banished by the relativistic cosmology in the large and by the

quantum gravity in the small and became substituted for by a
conception of a universe of finite age, size and scale but hav-
ing unbounded space-time geometry and cosmology. Yet, the
notion of infinity was to undergo a Renaissance of its own in
mathematics – in geometry, number theory, analysis, set theory,
classical and mathematical logic, discrete mathematics and the
theory of computation. Moreover, the dimension of complexity
of the real world would come to reveal new potential infinities in
the dimensions of the evolution of life, mind, language, culture
and human thought and civilization, including science, technol-
ogy and arts – and the notion of infinity as reduced merely to
a construction of the human mind.

1 Introduction
The notion of ‘infinity’ occupies a unique place in the mental and cultural
history of human civilization. In implying something beyond any conceiv-
able bound, endless, unlimited, boundless, eternal etc. in terms of space,
time, quantity, quality or other dimension of meaning the notion has come
to stand for the negation of anything actual and comprehensible. Its be-
longing to the world of imagination is witnessed by its occurrence typically
in mythology, theology, philosophy and mathematics, all of which are ex-
pressly products of mental construction, rather than in everyday life in the
real world. Yet, the idea has proved to be equally important to the early
conceptions about the world and heavens, theological and philosophical
speculations about the cosmos and gods as to scientific theories about the
universe. The notion itself has been so central that it has become lexicalized
as a variety of words on its different senses and aspects in all languages.
The idea of infinity has a long history of discovery, interpretation and
refutation in different fields of knowledge. In discussing the cultural or
conceptual history of some notion in a multidisciplinary setting many view-
points and approaches can be taken. One can choose to consider aspects
such as the words or terms used to refer to the notion in question in various
languages. Lexical and semantic analyses can reveal important aspects of
the way of thought and world view of different peoples and cultures and how
they have arrived to the idea. Etymological analyses can disclose interest-
ing deep connections between languages, myths and religions and between
cultures. The philosophical and scientific definitions and interpretations of
Infinity – An Outline of Conceptions 259

the notion as used today will then be seen against a broader cultural and
historical context.
In historical perspective the use and meaning of infinity as well as its
status in different fields knowledge has undergone a complex evolution.
In various phases infinity and infinite things have received diverse inter-
pretations and raised a variety of problems and paradoxes many of which
continue to be discussed and remain by far not settled questions. In the
following we will take an overview of the conceptual history of ‘infinity’
only as it has occurred in the mythologies of a few Oriental and Middle
Eastern civilizations, in the ancient Greek and Roman natural philosophy,
in the Mediaval patristic Christian theology and scholastic philosophy and
in the physical cosmologies and natural philosophy of the Renaissance.
The earliest conceptions of infinity and eternity apparently occurred in
Oriental cosmological and cosmogonical myths and fables that have been
passed down in popular tradition and epic texts. Some of these became cod-
ified in literary religions as holy scripts and were subsequently refined in
interpretations and commentaries. The pre-Socratic Greek natural philoso-
phers were the first to detach themselves from myths and religious authority
and began to ponder questions of cosmology and cosmogony independently
of religious tradition and offered a variety of original views many of which
involved notions of infinity and eternity in various senses and contexts.
The Medieval patriarchs, theologians and scholastic philosophers digested
and filtered the Oriental religious and the Greek philosophical thought,
transferred the idea of infinity from cosmology to theology and carried
speculations about the immensity of God and His virtues to the extreme.
In the Renaissance the legacy of the Greek natural philosophy and Indian
arithmetics were mediated to Europe by the Arabs which allowed the nat-
ural philosophers of the Renaissance to bring the notion back to physical
cosmology and the astronomers and mathematicians to pave way for the
rise of science and new manifestations and interpretations of the notion of
infinity.

2 Prehistory – Infinities in Mythology

Looking at the sky or even at open sea or desert or realizing the unend-
ing flow of time and events from the future to the past must have struck
already the early man with the idea of infinity. Observation and wonder-
ing of the immensity of the heavens and the world around us go back to
the dawn of human thought and cognition. The awe of the stars and of
260 J. Seppänen

the celestial sphere must have incited the idea of what lay beyond and of
endlessness in terms of spatial extensions and cyclic time. The questions
of the constitution, origin and purpose of everything began to intrigue the
human curiosity, awoke admiration and inspired imagination, and all these
received their expression in the emergence of religions.
Archaic cultures conceived the universe as having three levels – sky,
earth and underworld – connected by a central axis and themselves as liv-
ing in the centre. The Earth appeared as flat and finite though so large
that only gods and perhaps heros could reach the boundary. The huge
heavenly vault gave shelter to the Earth and its endless turning together
with the heavenly bodies conceived as gods determined the life on Earth
and marked the boundary between the human and the divine realms while
the underworld was the world of dead. The idea of the human habitat being
located in the centre of the Earth and heavenly spheres became archetypal
cosmological symbols which received expression and interpretation in the
mythologies of the prehistoric peoples of the world. Cosmographical no-
tions such as World Axis, World Pillar, Pole Star, Sky Nail, World Tree,
Cosmic Mountain etc. implied as their complement if not infinite then at
least immense dimensions of the world and stirred speculations about what
possibly obtained beyond the visible world.
The idea of infinity may have occurred in the context of questions about
the conditions before the coming into being of the world as reflected in the
cosmogonical myths and stories of creation. The oldest conceptions of infin-
ity are, indeed, found in the mythical cosmogonies as witnessed by notions,
names and words denoting or connotating the idea of infinity in one sense
or another. The abundance of ordinary words like endless, boundless, un-
limited, unterminating, inexhaustible, eternal etc. in all natural languages
also witnesses of a long mental, linguistic and cultural history of the idea
of infinity.
Explicit terms for infinity occur in all ancient Oriental and other lit-
erary cultures, as in the holy scripts and philosophical texts of Chinese
Daoism, Indian Hinduism and Buddhism, Iranian, Babylonian, Semitic,
Egyptian, Greek and others. Many of these notions have been carried over
and survived in the western spiritual, philosophical and literary traditions
and modern language. In the following subsections, infinity in the Indian,
Greek and Jewish mythical traditions are briefly introduced.
Infinity – An Outline of Conceptions 261

Indian – Aditi, the Unlimited

An ancient Hindu myth deriving latest from the time 1500 B.C., when
the Aryan tribes migrated from Iran to Hindustan, tell about the battle
between the Danavs, sanskr. danaw, bounded, and the Adityas (Vilenkin
1982), sanskr. aditya, unbounded, also sun, and growth. In the Veda texts
Aditias are identical at times with Savitri or Surya, the sun god, sanskr.
surya, sun, cf. also Finn. suuri, great, the greatest of the heavenly bodies.
With the help of Indra, the son of Heaven and Earth, the Adityas triumphed
and, by a miracle, grew to frightful size terrifying the Heaven and Earth to
break off into two opposite parts making up the world.
In another version of the legend (Barber 1979) Aditi, sanskr. aditi, un-
limited, eternity, infinite consciousness, mother of Adityas, gave birth to
originally three and in later versions of the text to eight or twelve gods, one
for each month of the year, including Mitra, Aryman, Varuna, Ansa, Indra,
Dhatri and others. She rejected one of them, Marttanda, who became the
sun. Other Adityas were Aryaman, the god of truth and the light of divine
consciousness, Mitra, the ruler of early things and Varuna, his opposite,
the ruler of spiritual things, Bhaga, the god of wealth, cf. Slav. bog, god,
bogatyi, rich, Indra, the god of weather, Varuna, cf. Slav. vorona, mag-
pie, the god of heavens, later of the oceans and rivers, etc. (Schumacher,
Woerner 1994). Cf. also Finn. äiti, mother, Maa-äiti, mother-earth, maa,
earth, Gr. gaia, geo, mother-earth.
Similar myths are found in the ancient Iranian tradition according to
which the God Ormazd created Earth, man and all creatures and things
on Earth, and Heaven in the form of an egg reaching at its highest point
the infinite world. The theme of creation from a primal egg occurs in the
myths of many other peoples, too, among them in Finnish Kalevala.
In the Hindu and Buddhist traditions infinite was manifested also in
the temporal dimension as the idea of an eternal cycle, Skr. rta, cycle,
ratha, carriage, cf. Finn. rata, rail, path, ratas, wheel, rattaat, carriage, an
endless return in incarnations of the human soul.

Greek – Empyreon and Cosmos

The Greek inherited elements from the ancient Egyptian and Oriental cos-
mologies and theologies in which the Sun continued to occupy a central role
and was worshipped as the main God. In the Greek mythology the realm
of the Sun God became known as the Empyreon, Gr. en, in + pur, py-
ros, fire, empurein, set in fire, empurion, empurios, the sphere of fire, Lat.
262 J. Seppänen

empyreus, caelum, heaven, cf. Finn. ympyrä, circle, ympäri, around. Note
also Gr. peri, about, around, roundabout, Lat. per, through, via, around,
period, Gr. peri + hodos, way, course, periphery, perimeter, perihelion,
etc., Finn. peri, perä, end, periaate, principle, perusta, foundation.
According to the Greek cosmogony the world emerged from chaos, Gr.
khaos, a primordial abyss, formless void, utter confusion, also chasm, Gr.
khasm, deep cleft, gorge, and turned into an ordered and harmonic cosmos,
Gr. kosmos, jewel, beauty, order, ornament. In some myths of this type
the chaotic state is never completely overcome but continues as an eternal
struggle of the opposites, a central principle in the Chinese taoist tradition.
While order emerges and may continue to develop in space and time, ves-
tiges of chaos remain and the created order is always in danger of slipping
back into chaos or else, chaos appears as the destiny of the cosmos when it
has exhausted all of its meaning and potential. It is worthwhile to notice
that chaos, as a modern mathematical and complex dynamical notion has
the same attribute of infinity and disorder, something endlessly changing
and never repeating itself or any part thereof in motion, form, order or
scale.

Jewish – Sephirot and Ein Sof

In the ancient Jewish cosmology as transmitted down with the Kabbalah,
Hebr. kabbalah, tradition, the world is conceived as created by God from
nothing, Lat. ex nihilo, and consisted of a system of ten spheres, Hebr.
sephiras, sphere, pl. sephitot, emanating from infinity, Hebr. ein sof, ein,
non + sof, end, endless, the symbol of God, who never revealed himself
directly but only through his virtues and deeds. God transcended all the
spheres the upper most of which was called the crown. Below it came the
spheres of wisdom, intelligence, mercy and lesser realms down to the lowest
kingdom of earthly life. It was only through the sephirot that man could
approach the holy spirit and himself pursue spiritual fulfillment. In the
Old Testament one can find statements about the omnipotence and other
properties of God without bound or limit, Hebr. ein gebul, bound, limit.
Over time the Kabbalah spread via Alexandria to Spain where its phi-
losophy came to culmination in the thinking of the Jewish philosopher and
theologian Moses Maimonides (1135–1204) as documented in his “Moreh
nebouchim”. His “Guide for the perplexed” spread to Europe, gained wide
popularity and exerted considerable influence on Medieval mystics.
In Catholic philosophy the Span. Franciscan theologian and mystic
Raimundus Lullus (1235–1315) wrote, instincted by a vision, “Ars magna
Infinity – An Outline of Conceptions 263

et ultima” (1273) in which he developed a diagrammatical method known as

Lullian art or Lullian circles, with the aim of mechanically proving the infi-
nite and absolute qualities of God. His circles comprised of concentric discs
which quoted the qualities of God, in Latin “Gloria”, “Bonitas”, “Magni-
tudo”, “Duratio”, “Potestas”, “Sapientia”, “Voluntas”, “Virtus”, “Veritas”
as “Bonitas”. By turning the discs he could obtain new combinations of
these and other notions on other discs and thereby supposedly be able to
prove higher order properties and to derive new knowledge. The work was,
however, condemned in 1376 by the church for its attempt to link faith
with reason, but the idea was sympathetically received by some scholastic
philosophers who continued to refine their speculations. The Lullian ideas
influenced also the Germ. mathematician and philosopher Gottfried Leib-
niz (1646–1716) to dream of a universal language and machine that would
allow to mechanically derive all possible knowledge.

3 Antiquity – Infinities in Philosophy

The Greek philosophers were the first to detach themselves from myths
and religious authority and began to ponder questions of cosmology and
cosmogony independently. The pre-Socratic scholars were to found natu-
ral philosophy and to introduce a number of new and original conceptions
about the constitution and origin of the world. Many of the natural cos-
mologies involved also notions of infinity in some senses.

Apeiron and Myriad

The early Greek cosmologies of the pre-Socratic philosophers were, for obvi-
ous reasons, based more on imagination than evidence. The first naturalis-
tic conception about the origin and nature of the world is generally ascribed
to the Greek philosopher Thales of Miletus (c.625–547) often quoted as the
first scientist. He was, indeed, engaged with astronomy and geometry, and
began to suspect the myths and role of gods in the events of nature. To him
the Earth appeared as flat and surrounded by an ocean and an atmosphere
of vapour. The heavenly bodies were floating in the atmosphere and all
things had come about from water and vapour. Thales, yet, is not known
to have given any opinion about infinity. With his attempts to find natural
causes and explanations to the origin and order of the world and physi-
cal phenomena and in basing them on observation and original reasoning
rather than belief he become the founder of naturalism.
264 J. Seppänen

His disciple Anaximander of Miletus (c.610–c.547), the author of the

first philosophical text “On Nature”, was the first to take up the notion
of infinity explicitely. To him the world was cylindrical with the heavenly
bodies moving each on its shell at different distances from the Earth. To
account for the origin of the world he introduced the terms arkhe, Gr.
arkhe, origin, first principle, beginning of all things, and apeiron, Gr. a,
not + peirein, try, apeiron, boundless, infinite, indefinite, eternal, endless
motion. The universe had come into being from apeiron which was an
inexhaustible and eternal chasm. The varieties of things and their qualities
were formed by differentiation and isolation from the primal chasm and
the eternal struggle of its opposites. Moreover, there were infinitely many
other worlds and new worlds could come into being from the apeiron until
they perished back into apeiron.
The last of the great Milesian philosophers Anaximenes (c.585–527),
a disciple of Anaximander, identified air as the primal element which by
condensation and rarefaction gave birth to all things. He reasoned that
since air is the breath of life for man it must also be the main principle of
the universe. The infinite air was the principle from which all things that
are, that are becoming and that shall be, including men and gods.
Latest in the 5th century Greek began to consider counting with num-
bers and the problem of boundlessness in mathematical terms. The first
Gr. mathematician Pythagoras (c.580–500), who became the founder of
the field by introducing the word ‘mathematics’, Gr. mathematike, man-
the, mind, thought that the world was made out of numbers and that all
things could be explained by arithmos, Gr. arithmos, number, natural
numbers and their properties, ratios and arrangements.
The Greek had the word myriad, Gr. murios, countless, innumer-
able, meaning an unconceivably large number. Yet, the Pythagoreans
were shocked to discover the idea of genuine infinity as manifested in the
incommensurate ratio of the side of a square to its diagonal. According
to the legend, which goes in many versions, the Pythagoras and his dis-
ciples became horrified and kept the discovery as a divine secret until it
was revealed to outsiders by a former but expelled member Hippasus of
Metapontum (5th c.), and was duly punished. One account has it that a
tombstone was raised to him while still alive, while another version has it
that he was left to drown in a shipwreck.
According to the Pythagorean cosmology what lay outside the heavens
was an infinite void by which they understood endless extent of air. The
notion of apeiron occurs also in Pythagorean dualism where it is opposed
Infinity – An Outline of Conceptions 265

to the principles of peras, Gr. peras, limit, boundary, and arithmos, num-
ber. According to Pythagoreans, apeiron was the boundless and amorphous
which together with its opposite, the limit and number were the basis of
everything that exists.
Interesting enough, by chance or by true connection, the Greek words
apeiron, and peras bear close morphological and semantic, possibly also
etymological connection either as a cultural loan or even true genetical
relationship, with the Finn. root per, peri, perä, perus, etc. and their
derivatives, cf. Finn.
– peri, destination, range,
– perä, rear, tail, end, back, foundation, perässä, after, following, next,
– perin, thoroughly, extremely, exceedingly, utterly,
– perinne, tradition, perintö, heritage, legacy,
– periaate, principle,
– perus, base, basis, foundation, bottom, primary, perukka, bottom,
end, perusaine, primary matter, perusluku, cardinal number, perussyy, pri-
mary cause, prime mover, perusta, foundation, peruste, base, ground, basis,
reason, cause.
Note also Engl. experience, Lat. experiri, Engl. empirical, empirism,
Lat. empiricus, Gr. empirikos, ancient physician, empeireia, experience,
empeiros, skilled, en + peirein, try, peira, trial, attempt, attack, cf. also
Engl. peril, Lat. periculum, experiment, risk, danger.

Eternity and Infinity of Particles

The materialist philosopher Heraclitus of Ephesus (c.544–c.483), the
philosopher of becoming, held in his “On Nature” that the primal material
of the world was fire for it was the most capable of change. According to
him, the world was eternal, not created by any god nor by man. It had
always been and would continue to exist for ever and everything was in
nonending flux of change. One could not step into the same river twice for
he who did would be “washed by other waters”.
The head of the Eleatic school Parmenides (c.540–480), the philosopher
of ‘Being’, vigorously opposed the idea of apeiron and conceived the world
as eternal but finite, immobile and having the form of a sphere filled with
things and free from void. He developed the principle of being in opposition
to the principle of becoming of Heraclitus.
Another materialist philosopher Anaxagoras of Clazomenae (c.500–
c.428) conceived in his book “On Nature” the world as consisting of an
infinite variety of qualitatively dissimilar primary elements of imperishable
266 J. Seppänen

and infinitely divisible particles of matter which he called homoeomeries,

Gr. homo, same + eomer, particle. All things were made of these ‘seeds’
which occurred throughout the universe. Their coming together by chance
or necessity gave rise to individual things, and their separation entailed
passing away of individual things. According to Anaxagoras “there is no
smallest among the small and no largest of the large, but always some-
thing still smaller and something still larger”. The particles retained their
character on division and there was a portion of everything in everything,
except Mind. The motive force that conditioned the union and division of
particles was nous, Gr. nous, mind, cf. Finn. nousta, arise, nousu, ascent,
which he understood to be the substance of the lightest and finest kind,
eternal and imperishable and capable of motion and communication of its
properties to things.

Infinite Cosmos and Infinity of Worlds

The idea of infinite divisibility of matter, in its turn, was rejected by a third
breed of materialist philosopher Leucippus (c.500–440), the founder of the
School of Abdera, who introduced the notions of absolute vacuum, causality
and atom, Gr. a, not + tomos, part. He regarded space without matter as
the essential ingredient of cosmos, an infinite void in which atoms moved
and without which no change would be possible. “Nothing arises without
cause but everything arises on some grounds and by the force of necessity”.
He also proposed the principle that all qualitative differences in nature may
be reduced to quantitative ones, opposing the principle of homogeneous
being advocated by Parmenides as an infinite variety of dissimilar kinds
but homogeneous particles.
To Leucippus atoms were also of infinite variety but indivisible and
occurred throughout the infinite space and were separated from one another
by not-being, that is empty space. The coming into being of things, was
caused by atoms coming together by motion in the otherwise empty space.
The cosmos appeared to him not only as infinitely large but with infinitely
many other worlds in it. The ideas of Leucippus were developed further
and applied to explain human perception, memory and mind by his disciple
Democritus (c.460–360) who became the founder of atomism as a general
doctrine.
In the Hellenic period the materialist and atomist philosophies and cos-
mologies were adopted and cultivated notably by the Greek atheist philoso-
pher Epicurus (341–270) and in the Roman Time by the Roman materialist
philosopher and poet Carus Lucretius (c.99–55). They recognized eternity
Infinity – An Outline of Conceptions 267

of matter and motion of atoms as central principles, developed atomism

into a materialist philosophy of life and became founders of sensationalism.

Infinitely Small and Infinitely Large

The question of infinity arose also in attempts to understand the nature
of motion which was taken up by Zeno of Elea (490–430), a disciple of
Parmenides. Defending the doctrine of being of his master as opposed to the
doctrines of becoming, change and motion, he presented logical paradoxes,
Gr. para, side, alongside + doxa, opinion, about the nature of motion and
developed the method of logical argument and proof. With the paradoxes
known as “Dichotomy”, “Achilles and the Tortoise”, “The Flying Arrow”
and “The Stade” Zeno argued against the possibility of motion. Under the
assumption of the infinite subdivisibility of space and time, he showed that
the idea of motion was self-contradictory, and hence, not a valid principle
for explanation of natural phenomena as the atomists had taught. His
reasoning made also Anaxagoras (c.500–c.428) to reconsider his thought
and to convince himself that there was no smallest quantity of anything.
If Zeno raised the question about infinity in the small, the mathemati-
cian Archytas of Tarentum (c.428–347) from the Pythagorean school, who
was told to have solved the problem of doubling the cube, became the first
to reason logically about infinite cosmos in asking: “If I am at the extremity
of the heaven of the fixed stars, can I stretch outwards my hand or staff?”
and answering that “It is absurd to suppose that I could not. And if I can,
what is outside must be either body or space. We may then in the same
way get to the outside of that again, and so on. And if there is always a
new place to which the staff may be held out, this clearly involves extension
without limit”.
The first of the great philosophers of Athens, Socrates (469–399), ini-
tiated the turn from naturalism and materialism to idealism. To him the
structure of the world and the physical nature of things were unknowable.
One can know only oneself as he expressed it “know thyself”. With this
turn Socrates opened a new dimension for philosophy to explore, that of
mind and knowledge, which were to reveal a multitude of new kinds of
infinity, although Socrates himself did not specifically discuss the question
of infinity himself.
His disciple Plato (428–348) adopted from his teacher the notion of idea,
Gr. eidos, and developed in his more than 30 dialogues the doctrine of
objective idealism to fight the materialist teachings of his time. To Plato
ideas were unworldly, eternal and independent of time and space. Ideas
268 J. Seppänen

could neither come into being nor perish because they already consisted
of a perfect and absolute totality. The source of such knowledge was the
immortal human soul, which was, however, defective and only reminescent
of the perfection of the divine world of ideas.
But, strange enough, to Plato there was no infinity, not in this world
nor in the world beyond, which makes his idealism defective if not self-
contradictory. He believed that even his ultimate idea, form, the Good,
must be finite and definite. This was in sharp contradistinction with al-
most all later metaphysicians, who assumed the absolute necessarily to be
infinite.
The Gr. mathematician and astronomer Eudoxus of Cnidus (c.400-
c.347) conceived the first model to account for the motions of celestial
bodies which he thought to be carried around the Earth each on their own
sphere. In mathematics he removed the limitation imposed by Pythagoras
of allowing only rational proportions accepting also incommensurable ratios
and magnitudes. Based on the idea of continued divisions and proportions
he developed in geometry the method of exhaustion which made it possible,
for the first time, to calculate areas under curves and volumes under curved
surfaces by dividing them into ever smaller sections. With these ideas,
which were to be adopted by Euclid (c.330–c.260) and were included in his
Book V, Eudoxus anticipated by nearly two thousand years the calculi of
indivisibles and infinitesimals.

Actual and Potential Infinity

The cosmology of multiple spheres was refined by the great philosophers
of Athens and subsequently by astronomers. The greatest of all philoso-
phers, logician and scientist, Aristotle (384–322), perfected the system of
spheres with the Sun, Moon, planets and stars circling the Earth each on
their spheres. As described in his “Physike”, the Aristotelian cosmology
consisted of a series of nine spheres and later up to 56 all set in motion
by the outermost permanent and divine crystal sphere, the one which he
called the prime mover that lay beyond the sphere of the fixed stars.
Yet, Aristotle denied the idea of an infinite void space on the grounds
of his conception of mechanics. In his “De caelo” he declared categorically
that there was neither time nor space, nor void outside of the heaven.
There was nothing infinite nor void, since things, possessed of dimensions,
such as space, would have been equivalent to body and therefore incapable
of receiving other material bodies. But even if void space existed, finite
and successive motion in it would be impossible. In a vacuum there could
Infinity – An Outline of Conceptions 269

neither be resistance to motion, so bodies would fall with infinite velocity.

Nothing could be thrown through empty space, and there would be no way
to distinguish one point from another, nor possibility of determining place
or direction.
Neither was there anything infinitely small, infinitely large or infinitely
numerous. For him, the idea of something being actually infinite was a
privation, not a perfection or totality but only absence of a definite limit.
What was limited, was actual and real with respect to what surrounded it.
Infinity could only be potential, not actual and complete, like the sequence
of natural numbers among which there was no largest number. The se-
quence could always be continued but would never be completed nor exist
as an actual and finished thing.
Aristotles reasoning was rigorous and settled, for the time being, the
question of infinity in the negative. Satisfied with this view, natural philoso-
phers, mathematicians and mechanicians could turn their attention to more
earthly questions regarding astronomy, mechanics and mathematics. It also
brought about what became known as horror of infinity, which perhaps had
also negative effect especially to the development of mathematics.
The Greek and already earlier Oriental civilizations which developed
geometry and arithmetics conceived also the possibility of unimaginably
large and small in terms of measure and number. The idea of infinitely
large and small must have occurred also early as witnessed by special names
and attempts to give explicit designations to numbers as large as one could
think of, speak about or write
In the third century the Greek mechanician and mathematician
Archimedes (278–212), the ‘father of physics’, wondered how many sand
grains there might be on a beach and devised designations in his “Sand-
reckoner” for huge numbers or myriads and demonstrated that even they
can be counted. “Many people believe, King Gelon, that the grains of sand
are without number. Others think that although their number is not with-
out bound, no number can ever be named which will be greater than the
number of grains of sand. But I shall try to prove to you that among the
numbers which I have named there are those which exceed the number of
grains in a heap of sand the size not only of the earth, but even of the
universe”.
The largest number ever having been given a name of its own apart from
compound mathematical notations is the Buddhist number asankhyeya,
Skr. a, non + san, together, number, khyeya, know, say, innumerable,
uncountable, equal to 10140 . This is a fairly big number indeed, at least for
270 J. Seppänen

practical purposes, if compared with for instance the number of atoms in

the universe which is estimated to be of order 1085 ‘only’.
There are also legends involving large quantities. According to a legend,
the invention of the game of chess, Iran. shah, emperor, impressed and
delighted the Shah of ancient Persia so much that he asked to see the
inventor and to award him. As introduced to the Shah the poor but learned
inventor asked merely that one grain of wheat be put on the first square
of his chess board, two on the second, four on the third and so on until
the whole board was covered. The Shah, embarrashed with the modesty of
the request, ordered the wish to be fulfilled, but was soon struck to realize
that the wheat deposits of his empire would be depleted long before the
completion of the request.

Stoics – Infinite Extracosmic Void

In the Hellenistic period conceptions of infinity from the oriental religions
and cosmologies were assimilated with the Greek natural philosophy and
the cosmologies were reinterpreted in new light. The Stoic philosophers re-
considered Aristotle’s view of the world and heavens as a system of spheres
accepting it in outline agreeing that the intracosmic world was spherical
and finite but disagreeing with his conception about extracosmos. The
various Stoic philosophers were not, however, unanimous about the nature
of extracosmos themselves.
The Stoic philosopher Cleomedes (1st c. A.D.) denied Aristotle’s view
of finite space reasoning in his “De motu circulari corporum caelestium”
rather in favour of Archytas who had held that no body could exist be-
yond the physical world because no material could limit void and because
it was absurd to suppose that void could limit void and make it terminate
and finite. The existence of an infinite extracosmic void beyond the heav-
ens seemed thus inevitable. The infinite extracosmic void was seen as a
receptacle for the finite intracosmos and that appeared to be its only pur-
pose because interaction between cosmos and void was seen possible (Grant
1982).
Like Aristotle, the later Stoic philosopher Philoponus (490–570) denied
the existence of infinite extracosmic void rejecting also the interpretation
of spatial dimensions. Moreover, void could never be empty of matter.
The prominent Medieval commentator of Aristotle, Simplicius (6th c.), in-
stead, became one of the loudest proponents of Aristotle’s view in support
of the finite intracosmos but infinite extracosmos in his “Commentary on
de Caelo”, paving way for theological interpretations of the infinite extra-
Infinity – An Outline of Conceptions 271

cosmos as the abode of God. In their conceptions about astronomy and

physics the Stoics followed, however, Aristotle, agreeing with his view of
the world and heavens as a finite system of spheres. In the Middle ages
the Stoic conception of the universe and the idea of an infinite extracosmic
void were conveyed to the Latin West from a number of sources.
In spite of the great strides made by the Greek cosmology, astronomy,
natural philosophy, physics and mathematics, it is often claimed that the
development of science was hampered by the refusal to accept the notion
of apeiron and rather to adopt the notion of atom as the primary building
block of nature. Especially in mathematics the horror of infinity became
detrimental. Yet, there were ideas about the space and time being infinitely
divisible, the space extending indefinitely, and time and geometrical line be-
ing continuous and containing an infinity of points. The study of number
sequences and ratios and the method of exhaustion in the calculation of
areas and volumes witness, however, of the use of potential infinity in ge-
ometry, number theory and the theory of proportions anticipating their
revival in the Renaissance.

Roman Time – Heavenly Spheres

The Roman philosophers and commentators had less concern for natural
philosophy and little to add to the Greek philosophical thought. The Mile-
tan, Pythagorean, Eleatic, atomist, Aristotelian and Stoic conceptions of
cosmology were digested and perpetuated but with little original contribu-
tion.
The Roman poet and materialist philosopher Carus Lucretius (c.99–55)
rendered the work of Epicurus (341–270) into a poetic form in his work “De
Rerum Natura” and argued for the unboundedness of space reasoning in
the following way: “Suppose for a moment that the whole of space were
bounded and that someone made his way to the uttermost boundary and
threw a flying dart. Apparently the dart must either fly past the boundary,
whereby it is not a boundary, or else be stopped by it, in which case there
must be something beyond the boundary that hence cannot be the limit of
the universe”, rephrasing the argument of Archytas.
The idealist Roman philosopher Plotinus (205–270), the founder of
neoplatonism, the last Greek ‘pagan’ philosophy flavoured with Oriental
mysticism, taught that the world began with the incomprehensible divine
‘One’, which was the eternal source of all being. Reviving Plato’s objective
idealism, he disagreed, however, with his master’s denial of the infinite and
thought that at least God was infinite. According to his “Enneads” the
272 J. Seppänen

absolute ‘One’ has no measure, stands outside of number and has no limit
in regard to anything.
The universe consisted of a hierarchy of spheres of being, the highest
being the realm of the ‘One’ and the lowest the earthly time and space that
can be perceived by senses. In between lay the spheres beyond space and
time, each derived from the one above, of higher degree of unity and each
being a lesser image of the one above. As one moved down the scale, greater
degree of multiplicity, separateness and increasing limitation become evi-
dent and at the lowest level separation into atoms of the spatio-temporal
world occurred. The highest sphere from which everything was derived,
was itself derived from the ultimate ‘Good’ that transcended everything,
was beyond being, absolutely simple and devoid, could be imagined or de-
scribed and could be known only when the mind arose into union with
it.
The neoplatonist cosmologist exhibited spiritual elements from the an-
cient Oriental cosmologies and religions and became the prime philosophical
source of ideas for the Christian theology in the Middle Ages developed by
the patristic and Catholic theologians and scholastic philosophers in sup-
port of their speculations about the immensity, omnipresence and other
infinite qualities of God.

4 Middle Ages – Infinities in Theology

In the middle ages, as the Christian belief spread and the early church
became gradually established, natural philosophy and mathematics suffered
set-back and were cut off in the Greco-Latin world. The Biblic cosmogony
became canonized by the patriarchs and neoplatonism was seen as a suitable
source of philosophical ideology for the Christian theology while the Stoic
conception of infinite extracosmos was adopted as a worthy abode for God.
The Aristotelian thought was selectively digested by the Catholic the-
ologians and scholastic philosphers but his denial of the actual infinity could
in no way be reconciled with the immensity of God and the infinite virtues
ascribed to Him alone. The Aristotelian and other Greek traditions of nat-
ural philosophy, cosmology and science, including astronomy, physics and
mathematics survived, however, and continued to flourish in the Arab world
and allowed the Greek legacy of wisdom and knowledge to be mediated to
Europe in the Renaissance.
The Arab centres of wisdom in the Middle East maintained connections
through the Silk Road also to Oriental civilizations and mediated achieve-
Infinity – An Outline of Conceptions 273

ments in astronomy and mathematics, notably the decimal notation for

numbers and arithmetics, from India to Europe.

Indian Astronomy and Arithmetics

The oldest Indian work on astronomy and mathematics, Skr. ganita, dates
from the turn of the 5th century. The text of the Jaina school mathemati-
cian Aryabhata (c.476) known as “Aryabhatiya” (c.499) discusses arith-
metics, algebra, Skr. rasi, quantity, geometry, Skr. ksetra, field, and
trigonometry. The Ind. mathematician Bhaskara I (c.629) was the first
to give a systematic treatment of decimal arithmetics, the rules for all the
six arithmetical operations including negative and irrational numbers and
zero. There was also a rule for division by zero ‘0/0 = 0’, which today
would, however, be considered as mistaken.
His contemporary Brahmagupta (c.598-c.666) wrote “Siddhanta” (628)
which consisted of 25 mathematical texts and were translated into Persian
and Arabic. He called the result of ‘a/0’ zero-divisor, Skr. khakkheda,
but did not give value to the result. In the late Middle Ages Bhaskara
II (1115–1185) continued the work of Brahmagupta and took division by
zero to give infinite, Skr. ananta-rasi, infinite quantity. For the Hindus
the wisdom of arithmetics was of divine origin and developed in association
with the determination of the proportions and construction of temples and
altars as cosmological symbols, whereby the notion ‘ananta’, infinity, was
attributed to the Supreme God, Vishnu.

Arab Astronomy and Mathematics

In the 7th century the Syrian bishop Severus (465-538) wrote praisingly
about Indian astronomy and arithmetics which used nine different signs for
the digits, that is the Hindu numerals and not letters of the alphabet as
had been customary in the Middle Eastern and Greek civilizations. This
was significant, indeed, since the introduction into the Arab world of what
came to be known as the Hindu-Arabic numerals, led to the adoption also
in Europe into mathematics and astronomy the positional decimal number
system which made the development of mathematics, unthinkable in Roman
numerals, possible.
The Arab mathematician, astronomer and geographer al-Khwarizmi
(c.800–c.847), a member of the House of Wisdom founded by the Abba-
sidian caliph al-Mamun (d.833) in Baghdad, expanded on the work “Arith-
metica” of the Greek mathematician Diophantus of Alexandria (c.250) deal-
274 J. Seppänen

ing with number theory and equations of rationals with rational solutions
and became with his book “Algebra” the founder of the field. Diophantos
had introduced symbols to denote quantities, operations and relationships
as previously and long after numbers had been represented by geometric
arrangements of dots, number symbols and words only which had became
an obstacle to the development of mathematics. The Arab. astronomer
and mathematician Tabit ibn Qurra (c.835-901) applied algebra and the
theory of proportions to Euclid’s geometry and discussed his fith axiom
and the question of whether parallel lines meet in infinity or not.
The Arab astronomer and optician al-Haytham Lat. Alhazen (c.965–
1038) from Basra wrote about hundred works, among them “Opticae the-
saurus” which was translated into Latin in the 12th century and finally
published (1572). It was the first authoritative work to reject Aristotle’s
theory of light that the eye sends out rays to the object looked at and his
belief that the speed of light was infinite. Alhazen’s work on geometri-
cal optics became the foundation of analytical geometry developed by the
French mathematician and philosopher René Descartes (1596–1650)
The Baghdad astronomer and mathematician al-Battani (8th c.) gave
up the Greek system of chords of angles in angular geometry and adopted
the far more convenient proportion known as sine and its converse cosine,
laying foundation to trigonometry. He then applied trigonometry to pro-
jection of figures from the surface of a sphere on to a plane anticipating
projective and spherical geometry and the idea of point in infinity.
The astronomer Kamal al-Din (9th. c.) extended algebra to handling
irrational numbers such as square root and to higher degree equations while
the Islamic poet and mathematician al-Khayyami (d.1022) discussed meth-
ods for finding roots of fourth, fifth and higher powers. The Arab. physician
Yahya al-Samaw’al (12th c.) wrote a book “The Dazzling” when only 19
years. In it he discussed algebraic theory of powers and introduced multi-
plication and division of powers, series of powers and inverse powers. He
adopted also the unifying convention that 1 can be expressed as power zero
laying foundation to a general theory of series. Moreover, he was the first to
use negative numbers as distinct entities three hundered years before they
were understood and accepted in Europe. The Samarkand mathematician
al-Kashi (d. c.1429) applied series to the calculation of π obtaining its value
up to 17 decimals.
The legacy of Arab astronomy and mathematics was transmitted to
Europe in the beginning of the 13th century by the Ital. mathematician
Fibonacci of Pisa (c.1170–c.1250), known for the infinite sequence of num-
Infinity – An Outline of Conceptions 275

bers in which each member is the sum of the two preceding numbers. Fi-
bonacci obtained his series as the answer to the problem of the number of
rabbits bred in pairs in successive generations. The Latin translation of
al-Khwarizmi’s text “De numero indorum” appeared as part of Fibonacci’s
“Libri abbaci” (1202). There he explained the use and significance of the
number zero ‘0’, Skr. sunya, Arab. cifr, empty, Lat. zephirum, zero,
zephyris, cipher, and introduced the symbol ’–’ bar, Lat. solidus, for ratios.
The work became epoch-making in making possible the rebirth of math-
ematics in the Renaissance and its rapid development in the subsequent
centuries.

Patristic and Scholastic Philosophy

The Christian apostles and Medieval patristic theologians adapted the Jew-
ish and Platonic tradition and the scholasticians began to speculate with
the immensity, omnipotence and other infinite qualities of God. In the late
Middle Ages the scholastic philosophers adopted some of Aristotle’s con-
ceptions but had difficulty with others. Although medieval and scholastic
philosophers would follow Aristotle in many respects, they preferred to fol-
low the Stoics in assuming the existence of infinite extracosmos beyond the
heavens that fitted well with the Catholic theology.
The early Christian bishop of Hippo St. Augustine (354–430) held views
close to neoplatonism in suggesting that God not only was infinite but also
could think infinite thoughts. According to him, omiscient God must know
each and every natural number and that he even knew infinity in the form of
all the numbers taken at once, for otherwise they would exceed his powers.
God himself must thus lie beyond all numbers.
To the Pope St. Gregory the Great (c.540–604), founder of the Catholic
philosophy, the infinitude of God was unsurmountable to human compre-
hension. No matter how far our mind may reach in the contemplation of
God, it does not attain to what He is, but merely to what is beneath Him.
God’s immensity transcended the universe and the human imagination.
Among those who developed early scholastic thought about the phys-
ical space was the Arab. philosopher Ibn Rushd l. Averroes (1126–1198)
who lived in Spain during the Muslim faith he tried to prove the eternity
and uncreatability of matter and motion and the mortality of soul. As a
forerunner of the Renaissance and in upholding the philosophy of twofold
truth, Aristotelian and theological, he fell victim of brutal persecution both
by the Muslim and the Catholic church.
276 J. Seppänen

Distantia Terminorum and the Possibility of Motion

The antique paradox about the impossibility of motion was revived by the
scholastic philosophers under the notion of “distantia terminorum” which
became an argument in defence of the possibility of motion in the void. The
Span. mechanician and philosopher Ibn Bajja Lat. Avenpace (d.c.1138)
considered the celestial bodies and noted that their “motions are not im-
peded by anything but yet they move at a definite speed in definite time”.
The idea that not only distance but also time and velocity were equally
indefinitely divisible and proportional to one another provided the key to
the solution of the antique paradoxes. Motion was not only physically, but
also logically possible, a conclusion which gave way to further elaboration
of the notions of celestial and earthly mechanics.
The next argument sought to explain how finite motion could occur
in a physical vacuum or resistanceless medium. According to the Engl.
Franciscan philosopher Roger Bacon (c.1214–1292), one of the first to argue
on this matter, motion through void was possible because bodies have a
distance between their boundaries which, too, are indefinitely divisible.
The principle of distantia terminorum was considered and accepted also
by the Dominican scholastic philosopher Thomas Aquinas (1225–1274). Ac-
cording to him, any motion had a definite velocity which arose from the
ratio of motive power to the mobile, even when there should be no resis-
tance. To support of his reasoning Aquinas leaned on the example given by
Avenpace. The idea of motive power which caused the motions of earthly
and heavenly bodies and the assumptions of finite world and infinite extra-
cosmic void gave rise to further speculations about God and his powers. In
his “Summa Theologiae” Aquinas tried to give a sort of Aristotelian proof
that “although God’s power is unlimited, he still cannot make an absolutely
infinite things, no more than he can make an unmade thing”.
The Ital. poet Dante Alighieri (1265–1321) told in his epic poem “Div-
ina Commedia” (1307) the story of a journey to Hell and Purgatory, guided
by Virgil, and finally to Paradise (1300), guided by his lifelong beloved
Beatrice Portinari (c.1265–1290) passing through each of the nine spheres.
Beyond these lay the nine spheres of angels and beyond these the abode of
God, Empyrion, the highest heaven.
Scholastic concern for the question about infinite culminated in the
contradiction of Aristotle’s rejection in his “De caelo” of the infinite ex-
tracosmic space and the plurality of worlds in declaring categorically that
“neither place, nor void, nor time” can exist outside the heaven and of the
Stoic view of the existence of the extracosmic view. The dispute was fur-
Infinity – An Outline of Conceptions 277

ther aggravated by the Condemnation of 1277 issued by the bishop of Paris,

Étienne Tempier (d.1279) which on theological grounds forbid 219 proposi-
tions in Aristotelian teachings and other conceptions, including the eternity
of the world. The scholasticians were occupied also by the questions of con-
ditions that had prevailed prior to creation, namely the existence of infinite
void space before creation in which the world had to be places and with
the location of God himself.

Infinite Space and the Immensity of God

In the 14th century the idea of infinite void extracosmic space beyond the
finite spherical intracosmos and heavens took hold and won by the 16th
and 17th centuries widespread acceptance. Even astronomers and natural
philosophers who otherwise were opposed to scholastic speculations were
inclined to accept the idea. The nature of the extracosmic void space
was further elaborated by the scholastics and became assimilated with the
omnipresence of God. Although devoid of matter the imaginary void could
be identified with God and His attributes. Its reality was, however, contro-
versial and often characterized as a pure negation of the substantial world
and space.
The Mediaval theological philosophy was unable to deal with other kinds
of infinitude than that of God. No infinite thing, collection or quality apart
from divine could ever arise or be thought of in the earthly world.

5 Renaissance – Infinities in Cosmology

The ancient Mesopotamian, Greek and Aristotelian cosmologies had main-
tained that the universe was finite, spherical and geocentric whereas the
Stoics conceived an infinite extracosmic void beyond the skies. The lat-
ter view had been adopted by the Medieval theologians who associated
it with the immensity of God to be further refined in the speculations
of scholastic philosophy. In the Renaissance, instead, the idea of an
infinite physical space took root, and not only space but an infinite plurality
of other worlds became subjects of speculation. The formerly theological
questions were redressed in terms of natural philosophy, astronomy and
mathematics. Among the forerunners of the Copernican and scientific rev-
olution and of the new world outlook were dissident scholastic theologians
and early Renaissance natural philosophers who became the first to break
off from the official doctrins of the church. The introduction of the decimal
278 J. Seppänen

system and arithmetics to the Latin world made possible the development
of astronomy, celestial and earthly mechanics and mathematics.

Infinite Cosmos and Infinitely Many Worlds

The Engl. Franciscan philosopher Roger Bacon (c.1214–1292), an early
critic of Aristotle and advocate of the scientific method, defined the
extracosmic void as a space in which there is absolutely nothing. By al-
tering Aristotle’s definition he propounded the view that allowed for the
possibility of space beyond the world, still accepting that the absolute void
was not able to receive any material body.
The French physicist and nominalist philosopher Jean Buridan (1300–
1362) proposed the idea of divinely motive force or impetus that enabled
the planets to move on their orbits forever in the heavenly ether that was
assumed to offer no resistance to their motion. Moreover, he revived the
idea of Anaximander about the existence of an infinite plurality of other
worlds in the infinite cosmos.
The Germ. cardinal and natural philosopher Nicholas of Cusa (1401–
1461), he, too, a dissident thinker of his time, came to mark a transition
from scholasticism to humanism and to anticipate the transition from natu-
ral philosophy to the new science. Denying the central position of the Earth
in the world order he expounded in his “De docta ignorantia” (1440) that
the Earth and heavenly bodies were in motion in the cosmos and that the
cosmos was infinite and contained infinitely many other worlds. The cos-
mos must be infinite, since if it had limits, then beyond those limits there
would be something and the universe would not be the whole universe.
And since it was infinite, it could have neither a centre nor a circumfer-
ence. Any point could be taken as the centre like on sea and the horizon
would always be equally distant in all directions. Cusa also refashioned the
conception of God with his doctrine of concordance or unity of opposites,
Lat. coincidentia oppositorum. According to it, all opposites, finite and
infinite, the smallest and the greatest, single and multiple etc. coincided
in God. As a forerunner of mathematical treatment of infinity, Cusa pro-
posed the doctrine of maximum, the idea of which was the principle of not
being able to be exceeded. For instance, a line was a circle of infinite radius
which was the maximum of the quality of being circular. With these ideas
Cusa anticipated the introduction of infinite quantities and proportions and
infinitesimals into mathematics.
The Germ. astronomer and mathematician Georg von Peurbach (1423–
1461), instead, carried on the Aristotelian and Ptolemaic system with its
Infinity – An Outline of Conceptions 279

finite firmament as the prime mover of the heavens, while the Germ. as-
tronomer and mathematician Petrus Apianus (1495–1552) consented with
the Stoic and theological notion of infinite void as Empyrium, the abode of
God and angles.

God as the Infinite Universe

The Ital. Platonic scholastic philosopher Francesco Patrizi (1529–1597)
followed the Stoics in criticizing Aristotle’s cosmology in his “Nova de uni-
versis philosophia” (1591). Instead, he expounded the idea that the cosmos
itself was God, an infinite good, and that the world derives from Him as
a decreasing hierarchy of perfection, from the highest wisdom to life, in-
tellect, soul and lesser qualities and forms of things and finally the earthly
matter.
The Ital. dominican astronomer, cosmologist and platonist mysti-
cian Giordano Bruno (1548–1600) also rejected Aristotelian cosmology, the
Ptolemaic world order and the conception of the cosmos as a crystal sphere.
Instead, he adopted the Copernican view that the Earth and planets cir-
cles the Sun, that in the infinite universe thousands of other worlds circle
thousands of other Suns and that some of them might be inhabited. In his
“De l’infinito universo et mondi” (1584), written in dialogue form, Elpino
asks “How is it possible that the universe can be infinite?” to which Philo-
teo, Bruno in disguise, answers with a counter question “How is it possible
that the universe can be finite?”. “To a body of infinite size there can be
neither centre nor boundary ... Just as we regard ourselves at the centre ...
so doubtless the inhabitants of the Moon believe themselves to be at the
centre ...”
Still, Bruno upheld the pantheistic theological conception of the unity of
God with the infinite cosmos. In a poem “De immenso et innumerabilibus”
he writes “The One Infinite is perfect, simply and of itself nothing can
be created or better than it. This is the One Whole everywhere, God,
universal nature. Nothing but the Infinite can be a perfect image and re-
flection thereof, for the finite is imperfect, every sensible world is imperfect,
wherefore evil and good, matter and form, light and darkness, sadness and
joy unite, and all things are everywhere in change and motion. But all
things come in infinity to the order of Unity, Truth and Good, whereby it
is named universum”. In his writings “Della causa”, “Principio ed uno”,
Bruno developed his ideas about the dialectical unity of oppositions, uni-
versal motion, and the coincidence of contraries in the infinitely great and
the infinitely small.
280 J. Seppänen

Bruno travelled widely in Europe advocating his views until he became

betrayed by a wealthy Venetian Giovanni Mocenigo who had been commis-
sioned by the church to persuad Bruno to come to teach him “the art of
memory and inventio”, and was trapped and brought in front of the inqui-
sition. For eight years he was interrogated, kept imprisoned and tortured
but refused to revert his views until finally condemned of heresy and burnt
at a stake in Rome.

Omnipresent God in Infinite Physical Cosmos

The French physicist and astronomer Pierre Gassendi (1592–1655) also
adopted the Stoic view of cosmos as an infinite void but as a physical
and geometrical 3-dimensional infinite space, the ultimate container of all
things, capable of holding also physical matter, whereas Bruno had held
that it contained only ether and Patrizi that it containd only light. Un-
like the scholastics who maintained the view of a created non-dimensional
cosmos and identified God with the immensity of space, Gassendi assumed
non-created dimensional infinite physical cosmos coeternal with God inde-
pendent of cosmos but capable of omnipresence. Thus, Gassendi’s cosmol-
ogy and theology consisted of two kinds of independent actual infinities.
Gassendi accepted atomism, not as an infinity of eternal non-created
atoms and infinity of worlds, but as a finite number of atoms created by God
and making up a single and finite world placed in the infinite 3-dimensional
space. The dimensional spatial and atomistic doctrine of a created finite
world in an un-created infinite space of Gassendi, despite of his theological
mysticism, paved way for the cosmology of an infinite physical space, time
and matter that would shortly become accepted as the framework of the
scientific revolution and physical world outlook of the New Time.
This view raised, however, serious theological controversies that could
not be satisfactorily resolved without revision of the official Catholic doc-
trine. The dimensionality of space and of God came into contradiction as
the immensity of space and God had been identified as one. Either the
dimensionality of space would have to be rejected or God would have to be
thought as finite and 3-dimensional.
This radical move was taken by the English philosopher and theologian
Henry More (1614–1684), a central figure of the ‘Cambridge Platonists’,
in his “Enchridion metaphysicum” (1671). This meant rejection of the
explanation of the omnipresence of God by the doctrine of ‘whole in every
part’ that he called holenmerism, Gr. holos, whole, Engl. holy, sacred,
divine + Gr. mere, measure, Slav. mera, measure, Germ. mehr, more,
Infinity – An Outline of Conceptions 281

cf. Finn. määrä, measure. Instead, he insisted that everything, whether

corporeal or spiritual, including God, had finite dimension and extension
and occupied a place. Thus, also God, alike with space, albeit capable of
omnipresence, had dimensions and extension.
The views of Gassendi and More were accepted by the Engl. empiricist
philosopher John Locke (1632–1704), the founder of deism, Lat. deus, god,
Gr. theos, Zeus, Sanskr. deva, shining one, daiva, divinity, dagh, day, Finn.
taivas, sky, Germ. dag, Engl. day, the doctrine according to which God
had created the world but did not interfere with its subsequent events ever
since. This view allowed people to maintain their belief in God but allowed
them also to pursue goals themselves to change their condition and even
the society giving rise to the ideology of progress.
The Engl. physicist and mathematician Isaac Newton (1642–1727)
came to bring these ideas to fruition in science. As the chief architect
of classical mechanics he would construct a new physics and cosmology
within the framework of an infinite 3-dimensional space, time and matter,
and the laws of motion in his “Philosophia naturalis principia mathemat-
ica” (1687). Newton reflected also on the philosophical and theological
aspects of space and God, but left them unpublished. To him God was
an infinite omnipotent spirit that he preferred to call aether, “who could
move bodies without resistance”, as he noted in his “General Scholium”
(1715). In his cosmology the Solar system was set in infinite and absolute
Euclidean space. God was eternal and infinite and by existing always and
everywhere constituted duration and spatial extension Infinite suns were
spread throughout the infinite space and were held in dynamic equilib-
rium by their mutual attractions. The physical cosmology of Gassendi,
the Copernican world order and the Newtonian mechanics were to over-
throw the prehistoric and Biblic myths, natural philosophical cosmologies
of Antiquity and the Mediaval theological speculations.

6 Conclusion
From the times of Antiquity it was known that the heavens contained five
wandering planets, the Sun and the Moon. The other three planets Uranus,
Neptunus and Pluto were unknown as they were invisible to the naked eye
and could be detected only after the invention of the telescope in the 17th
century. The questions concerning the size of the universe and whether
the world was finite or infinite and unique or indefinitely plural remained
matters of speculation even after the telescope became available. In the
282 J. Seppänen

classical and modern astronomy and cosmology, infinity continued to play

a variety of roles until it was finally banished by the relativistic cosmology in
the large and by the quantum gravity in the small and became substituted
for by a conception of a universe of finit cosmology.
Yet, the notion of infinity was to undergo a Renaissance of its own in
mathematics – in geometry, number theory, analysis, set theory, classical
and mathematical logic, discrete mathematics and the theory of compu-
tation. Moreover, the dimension of complexity of the real world would
come to reveal potential infinities in the dimensions of the evolution of life,
mind, language, culture and human thought and civilization, including sci-
ence, technology and arts – and the notion of infinity as reduced merely to
a construction of the human mind.

Literature
Cultural History
1. L. Zippin, Uses of Infinity, Random House, New York, 1962.
2. J. L. Borges, Labyrints, New Directions, New York, 1962.
3. R. Péter, Playing with Infinity, Dower, New York, 1976.
4. E. Grant, Much Ado About Nothing, Cambridge University Press,
Cambridge, 1981.
5. E. Maor, To Infinity and Beyond. A Cultural History of the Infinite,
Birkhüser, Basel, 1987.
6. J. Seppänen, Empty words – Conceptions of nothing in language,
mythology, philosophy and science, Design for Architecture Sympo-
sium – Empty Spaces, Alvar Aalto Museum, Jyväskylä, 14-16 Aug.
1998.
Cosmology
1. N. S. Hetherington, Cosmology. Historical, Literary, Philosophical,
Religious, and Scientific Perspectives, Garland Publishing, Inc., New
York, 1993.
2. N. Ya. Vilenkin, In Search of Infinity, Birkhäuser, Basel, 1995.
3. A. M. Wilson, The Finite in the Infinite, Oxford University Press,
Oxford, 1996.
Infinity – An Outline of Conceptions 283

4. A. Koyré, From the Closed World to the Infinite Universe, Johns

Hopkins University Press, Baltimore, Md., 1974.

5. M. Blay, Les raisons de l’infini: Du monde clos à l’univers

mathématique, Gallimard, Paris, 1993.

Philosophy

1. S. Saarnio, Mitä tiedämme äärettömyydestä?, Wsoy, Porvoo, 1969.

2. J. Hintikka, Time and Necessity. Studies in Aristotle’s Theory of

Modality, Oxford University Press, Oxford, 1973.

3. R. Rucker, Infinity and the Mind. The Science and Philosophy of the
Infinite, Birkhäuser, Basel, 1982.

Encyclopedia

1. R. Barber, A Companion to Mythology, Ketrel Books, Har-

mondsworth, 1979.

2. M. Eliade, The Encyclopedia of Religion, Macmillan, New York, 1983.

3. S. Schumacher, G. Eoerne, eds., The Encyclopedia of Eastern Philos-

ophy and Religion, Shambala, Boston, 1994.

4. * * *, The Encyclopedia of Eastern Philosophy and Religion, Sham-

bala, Boston, 1989.
 Finite Versus Infinite Neural Computation
Hava T. Siegelmann
Information Systems Engineering
Faculty of Industrial Engineering and Management
Technion, Haifa 32000 Israel
E-mail: iehava@ie.technion.ac.il

1 Introduction
The Turing machine was suggested in 1935 as a model of a mathematician
that solves problems by a finitely specifiable algorithm using unlimited
time, energy, pencils, and paper. Although frequently used, there are other
possible computational models and devices, not all of them finitely speci-
fiable. The brain, for example, could be perceived as a powerful computer
with its excellent ability for speech recognition, image recognition, and the
development of new theories. The nervous system, constituting an intri-
cately interconnected web of 1010 –1011 neurons whose synaptic connection
strength changes in an adaptive and continuous manner, cannot be per-
ceived as a static algorithm; the chemical and physical processes affecting
the neuronal states are not specifiable by finite means.
In this work, we focus on the Analog Recurrent Neural Networks
(ARNNs). These constitute a finite number of continuous valued neurons
connected in a general fashion (not necessarily layered or symmetrical.) We
interpret the dynamical behavior of networks as a process of computation,
and study the effects of various constraints and different parameters on
their computational power. Altering the networks’ constitutive parameters
between finite and infinitely described numbers allows the model to coincide
with previously considered models that are computationally and conceptu-
ally different, from simple automata up to a model that encompasses and
transcends digital computation. In this sense, the neural model is a natu-
ral framework to test the connection between the level of infiniteness in a
computational machine and the associated computational power.
The analog neural network model is different from the digital computer
in some significant ways:

285
286 H. Siegelmann

(1) The classical (von Neumann) model of the computer consists of a

region of memory and a different region of a processing unit where
the program is executed. In the neural network’s model, memory
and processing are not separable. Each neuron is part of the pro-
cessing unit while the memory is encoded in the weights. Hence if
the network has an infinite number of memory registers (weights) like
in the Turing model, here it will also have infinite processing power
(neurons); and vice versa – a reasonable model of neural computation
that has bounded processing power is constrained to a finite number
of registers. This differs from the Turing model in which the finite
control has access to an unbounded tape of memory cells.

(2) Another inherent difference is the nature of the atomic data types in
the memory cells. Digital computers handle bits of information; in
analog neurons, the values are reals. If the neurons had finite precision
only, the finite interconnections would fall into the computationally
weak framework of a finite automaton.

(3) This difference has to do with continuity in the update function of

each neuron. In computational terms, there are no flow commands
which rely on infinite precision tests, such as

“If z is greater than 0, compute one thing, and otherwise

continue in another computation path.”

This principle of continuous flow is definitely a restriction on our

model.

(4) The last principle is the description length required to describe the
machine model. The Turing model is describable by a finite string
of bits. The neural network, having a finite number of neurons, but
incorporating exact real weights, should be described by a finite string
of real numbers. (This usage of weights with infinite bit description
is useless for the actual construction of the system, but it is still
appealing for the mathematical modeling of analog computation that
occurs in nature.)

Given the above, how does the analog network compare with the digital
computer in terms of computational capabilities? We show that although
the network is continuous, cannot branch on values, does not have a mem-
ory unit in the classical sense, and utilizes only a fixed number of neurons,
it still can compute anything that a digital machine does. Furthermore, due
Finite Versus Infinite Neural Computation 287

to its real-valued weights, it is inherently richer than the digital computa-

tional model and can compute functions which no digital computer can.
The network, although transcending the Turing model, is still sensitive to
resource constraints and still computes only a small subclass of all possi-
ble functions, thus constituting a well-defined super-Turing computational
model.
It is beyond the scope of this work, but we note that when the precision
in the weights and the neurons is not perfect or when noise is present,
the computational power reduces to the class of regular or even to definite
languages [2], [8], [7], [13].
The requirement of continuity in the computation, as specified in item
(3), translates to the interesting feature that although real weight neural
networks are defined with unbounded precision and although the neurons
are able to store infinite precision values, for time-bounded computation,
only bounded (linear) precision is required.
Adding discontinuities (e.g., some neurons that compute exact test for
0) to the analog network changes the resulting dynamical properties. Such
augmented networks were called “arithmetic networks” [3]. The first evi-
dence of the arithmetic networks’ computational superiority, relies on the
finding that arithmetic networks can recognize some recursive functions
arbitrarily faster than Turing machines and ARNN: they recognize arbi-
trarily complex recursive functions in linear time. The second evidence
concerns the amount of precision required to implement arithmetic net-
works on digital computers: Unlike the ARNN with their linear precision,
no fixed precision function is enough to simulate all arithmetic networks
running in linear time.
The work described here on ARNN can be found in full details in [12];
the work on arithmetic networks can be found in [3].

2 The Model
ARNN is a finite size network of neurons in which every neuron updates
by the equation

xi (t + 1) = σ(ΣN M
j=1 aij xj (t) + Σj=1 bij Ij (t) + ci ) (1)

where σ is the piecewise linear (semilinear or saturated-linear) function:


 0, if x < 0,
σ(x) = x, if 0 ≤ x ≤ 1, (2)

1, if x > 1,
288 H. Siegelmann

where a, b, and c are the weights and I is the external input [12]. There is
nothing particular about the σ function except that it is easy to treat math-
ematically. The network is uniform in the sense that its size, structure and
weights are fixed for all computations and inputs. The only unfixed values
are the activation values of the neurons that update in the computational
steps.
To consider the network as a computational device, we have to define
the input/output (I/O) map generated by the network. We can think of
the input as an initial state of the network; output will be defined then
similarly as the final state. An equivalent point of view which better fits
models from Control Theory uses input and output channels. In this view,
the input is a stream of letters which are transferred one at a time on a few
channels. A similar convention is applied to the output, which is a stream
of letters as well.
For simplicity, we constrain attention only to I/O maps which are de-
fined on bits. This is not at odds with the analog internal computation, and
by doing so, we separate the inherently different properties of the network
itself from interface (I/O) differences, and we enable a totally fair compar-
ison with the classical digital model. So the initial and final states will be
constrained to finite precision, or equivalently, the channels will be digital,
transferring bits. For the sake of clarity we next describe formal networks
for which all results reviewed here are obtained [12]. These networks com-
ply with the I/O protocol that follows. The input arrives on two binary
input lines. The first of these is a data line, which is used to carry a binary
input signal; when no signal is present, it defaults to zero. The second is
the validation line, and it indicates when the data line is active. It takes
the value “1” while the input is present and “0” thereafter. We use “D”
and “V ” to denote the contents of these two lines, respectively, so
I(t) = (D(t), V (t)) ∈ {0, 1}2
for each t. Similarly, there are also two output processors that take the role
of data and validation lines; they are denoted G(t) and H(t), respectively.
We now encode each input string
ω = ω1 · · · ωk ∈ {0, 1}+
by
Iω (t) = (Dω (t), Vω (t)) , t ∈ N,
where !
ωk , if t = 1, . . . , k,
Dω (t) =
0, otherwise,
Finite Versus Infinite Neural Computation 289

and !
1, if t = 1, . . . , k
Vω (t) =
0, otherwise.
A word ω ∈ {0, 1}+ is classified in time r by a formal net starting from the
initial state x(1) = 0, if the input lines (Dω , Vω ) take the values Dω = ω0∞
and Vω = 1|ω| 0∞ , and the output line component Hω (t) = 0 for t < r and
Hω (r) = 1. If Gω (r) is “1” then the word is accepted, and if Gω (r) is “0”
the word is rejected.

3 Preliminary: Computational Complexity

A thorough theory of computation has been developed to classify functions
into classes of equally difficult functions. Each class is associated with a
set of rules that describe it or, equivalently, with some type of theoretical
machines that can compute exactly this class.
Let’s see some common classes:
1. A finite state automaton is a machine that has a fixed number of
different internal states. A light switch has two states: on and off.
A counter mod 7 has 7 states, 0, 1, . . . , 6. All tasks executed by any
finite automaton are named regular.
One could naively think that, because the world is finite, all machines
can be modeled by finite automata. But consider this task: “Given a
binary sequence, detect whether the number of 1’s in the sequence is
greater than the number of 0’s there.” There is no machine that can
memorize the difference between the occurrences of 0’s and 1’s of any
binary string using a fixed predetermined finite memory only, so the
class of regular functions is pretty weak.
2. Another model is the general purpose computer as suggested by von
Neumann. There are different types of computer hardware, computer
languages, and ways to represent programs to the computer; but all
general digital computers have some common features. The programs
executable on them form the class of recursive functions, and the
model describing them is called the Turing machine. In a Turing
machine, a finite automaton is used as a control or main computing
unit, but this unit has access to potentially infinite storage space.
Formally, a Turing machine consists of a finite control and a binary
tape, infinite in one or two directions. The tape is accessed by a read-
write head. At the beginning of the computation, an input sequence
290 H. Siegelmann

is written in binary code on the tape, surrounded by infinite sequences

of blanks on both sides. The head is located at the leftmost symbol
of the input string. At each step the machine reads the tape symbol
(a ∈ {0, 1, #}) under the head, checks the state of the control (q ∈
{1, 2, . . . , |Q|}), and executes three operations:

(a) It writes a new binary symbol into the current cell under the
head (b ∈ {0, 1, #}).
(b) It moves the head one step to either the left or to the right
(m ∈ {L, R}).
(c) It changes the state of the control (q
∈ {1, 2, . . . , |Q|}).

The transitions of the machine are described by a function g(a, q) =

(b, m, q
). When the control reaches a special state, called the “halt-
ing state,” the machine stops. The output of the computation is
defined by the binary sequence written on the tape extending from
the read/write head to the first # symbol on the right side of the
head. Thus, the I/O map, or the function, computed by a Turing
machine is defined in terms of the binary sequences on its tape before
and after the computation.
Turing machines can compute many more functions than finite au-
tomata but they are still limited: Consider for example the halting
problem; this is the decision problem defined as follows. Any Turing
machine M can be encoded by a finite binary sequence τ (M ). Given
two words τ (M ) and ω, decide whether the machine M would halt
when starting with the input ω; or in other words, whether M would
ever reach its halting state. It is easily proven that no computer can
execute this halting problem.
The field of Computational Complexity looks not only for functions
defined by computational models, but also for functions which are
efficiently computable on them. In the digital model, a function is
considered efficient if it requires only polynomial time in the length
of the input; more formally, if there exists a Turing machine M and a
polynomial p1 such that the machine halts after p1 (n) steps on every
binary input string of length n. The class of efficient functions is
called P, and it constitutes the basis of complexity theory. Another
type of efficiency is defined in terms of space constraints. A function
is considered space efficient if the number of tape cells which have
been used during the computation is bounded to polynomial in the
Finite Versus Infinite Neural Computation 291

length of the binary input. The class of space efficient classes is

called PSPACE. Due to the deterministic characteristic of the Turing
machine, PSPACE functions will be computed in up to exponential
time. The equality P = SPACE is one of the most interesting open
questions in the field of computational complexity.

3. A new model was born in the 80’s, when Karp and Lipton introduced
a nonuniform model of computation that is computationally stronger
than the Turing machine [4]. Nonuniform computers are a general-
ization of the concept of monotonically increasing (e.g., polynomial)
response time, that allow an increase in available hardware with input
length: in this model, there exists a polynomial p2 (n) such that out-
put for an input of length n is calculated by an acyclic interconnection
of p2 (n) digital components (e.g., McCullough and Pitts neurons or
logical gates). Inputs of different length are computed by different
(nonuniform) hardware. Note that the hardware for calculating input
of length n requires p2 (n) bits of description, and the whole family
together cannot be described finitely. Lipton and Karp showed that
such a family of nonuniform acyclic digital circuits exceeds the com-
putational capabilities of the Turing machine. If exponential time is
allowed, the nonuniform families compute all binary functions. Under
polynomial computation time these families output the nonuniform
class P/Poly. The class P/Poly strictly contains P. It also computes
some nonrecursive functions, but is still a very small subset of arbi-
trary binary languages.

4 Summary of Results on ARNN

ARNN constitute a parametric model of computation. Their computational
power depends on the type of numbers utilized as weights.
For mathematicians, natural choices of number sets are the integers,
rationals, or reals. For computer scientists, a natural choice of function
classes is the regular functions, the recursive function class, and the class
of all (arbitrary) binary functions. The correspondence between networks
with different number domains and computational classes is summarized in
Table 1 [15].
The left columns of the table say that networks with integers, rationals,
or real weights compute regular, recursive, and arbitrary functions respec-
tively when the time is not constrained. The third column describes the
resulting classes for the case that available computing time is constrained to
292 H. Siegelmann

be polynomial in the length of the input. This class is still the regular func-
tions for integer weights, the class P for rationals, and when the weights
are real, the resulting class is AnalogP. Although our model is uniform,
AnalogP was found to be equivalent to the super-Turing class P/Poly.

Table 1: The computational power of recurrent neural networks

Weights Arbitrary Time Polynomial Time
Z Regular Regular
Q Recursive P
R Arbitrary AnalogP

Focusing on the two bottom lines of this table, one may wonder about
the essence of the difference between rational numbers and real numbers
that causes such a difference in the associated power. For rationals, the
framework adheres to the classical realm of computer science of Turing
machines and recursive functions while for reals, the networks belong to
the analog regime. This question is addressed by Balcazar, Gavalda, and
Siegelmann in [1]. They describe rationals, reals, and a whole hierarchy be-
tween them in an information theoretic manner, using resource-bounded
Kolmogorov-type characterization. This characterization is sensitive to
both the amount of input data and the time that is needed to construct
the numbers. They prove that there is a proper hierarchy of complexity
classes defined by networks, whose weights have increasing Kolmogorov
complexity. P and AnalogP are just the two ends of the hierarchy for ef-
ficient computation; recursive and arbitrary are the ends for unbounded
computation time.

4.1 Networks which are Turing Equivalent

The main theorem in this section is as follows [15], [12]: It is clear that an
ARNN of finite size and rational weights can be described in finite terms
and be fully simulated by a Turing machine (with no more than polynomial
slowdown). We next state that the reverse inclusion holds as well.

Theorem 4.1 For any function Φ : {0, 1} → {0, 1}+ that is computable
by a Turing machine M, there exists a formal network N with rational
weights that computes it. Furthermore,

1. If Φ is computable on M in time T (n) for inputs of length n, then N

requires only T (n) + O(n) time.
Finite Versus Infinite Neural Computation 293

2. The structure and size of the simulating network is independent of the

computation time. These depend only on the structure of the Turing
machine, that is, on the number of its tapes and the size of its finite
control.

3. The statement holds for nondeterministic computation as well.

Corollaries:

1. There exists a universal neural network that computes all recursive

functions.

2. A derivative of the Halting problem is that there is no algorithm

that decides whether a neuron ever gets the value 1, or a value in
the vicinity of 1; similarly, no algorithm can decide whether a given
network ever converges.

A secondary theorem asserts that Turing universality is a relatively

common property of ARNNs [5]:

Theorem 4.2 Let ς be any general sigmoidal function (fully described in

[5], [12], e.g., ς(x) = 1+e1−x ). For any ς there exists a universal network of
ς-neurons that computes all recursive functions and suffers of up to expo-
nential slowdown.

4.2 Networks which are Beyond Turing

This subsection considers the full network model when real weights are
allowed. We first note the interesting property that although real weight
neural networks are defined with unbounded precision, they demonstrate
the feature referred to as “linear precision suffices” [14]:

Lemma 4.1 For up to q steps of computation, only the first O(q) bits
of both the weights and the activation values of the neurons influence the
result; the less significant bits have no effect on the outcome.

This property is used to properly formulate the time-dependent resis-

tance (“weak resistance”) of the networks to noise and implementation error
[14].
The main theorem of this part states that real weight neural net-
works compute in polynomial time the class AnalogP which is equivalent
to P/Poly.
294 H. Siegelmann

Theorem 4.3 There exists an integer N such that for every circuit family
C = {Cn |n ∈ N} of size S that computes the language ψC there exists a
neural network N with N neurons and one real weight computing ψN so
that ψN = ψC and TN (n) ∈ O(n S 2 (n)).

We next conclude three characteristics of real weight neural networks.

Corollaries:

1. Because of the polynomial correspondence with nonuniform circuits,

we conclude that, if exponential computation time is allowed, then
one can specify a network for each binary language, including non-
computable ones. In polynomial time, the networks compute exactly
the language class AnalogP which is equal to P/poly. This class com-
putes all unary languages including the unary encoding of the Turing
machine halting problem. But, it includes a very small fraction of all
binary languages. Although AnalogP strictly includes all P, it is not
likely to contain functions of the classical set NP, since otherwise the
polynomial hierarchy will collapse to Σ2 [4].

2. A nondeterministic version can also be specified: AnalogNP is equiv-

alent to nonuniform families of nondeterministic circuits.

3. This conclusion is a corollary of the property “linear precision suf-

fices”. The amount of information necessary for the neural network
is identical to the precision required by chaotic systems. Therefore,
neural networks may constitute a framework for the modeling of phys-
ical dynamics.

Remark. Unlike the Turing machine, the analog recurrent neural network
is not universal for digital input but only for input in the form of a finite
string of real numbers.

4.3 Analog Computation

Up to now, we concentrated on the very particular model of the Analog
Recurrent Neural Network. We next shift viewpoints to see this network
as only one example of a class.
As a first step we generalize the networks to generalized analog networks.
In vector form, a generalized analog network D updates via the equation

x+ = f (x, I),
Finite Versus Infinite Neural Computation 295

where x is the current state vector of the network, I is an external input

(possibly a vector), and f is a composition of functions:

f = ϑ ◦ π,

where
π : RN +M → RN
is some vector polynomial in N + M variables with real coefficients, and

ϑ : RN → RN

is any vector function that has a bounded range and is Lipschitz.

That is, for every ρ > 0, there exists a constant C such that for all
x, x̃ ∈ Domain(ϑ), if |x − x̃| < ρ, then |ϑ(x, u) − ϑ(x̃, u)| ≤ C|x − x̃| for any
binary vector u. A similar property holds for f = ϑ ◦ π .
It is not hard to prove that this whole class of networks is not compu-
tationally stronger than the basic homogeneous neural network [14]
Our next step is to add stochasticity to the networks by means of a
heads/tails coin, where the probability to fall on head or tail is a real
number.

Definition 4.1 A stochastic network has additional input lines, called

stochastic lines, that carry independent identically distributed (IID) binary
sequences, one bit per line at each tick of the clock. The distributions may
be different on the different lines. That is, for all time t ≥ 0, the stochastic
line li has the value 1 with probability pi (0 ≤ pi ≤ 1), and 0 otherwise.
We say that the language L ⊆ {0, 1}+ is R-recognized in time T by a
stochastic network N if every input string ω ∈ {0, 1}+ is classified in time
T (|ω|) by every computation path of N on ω, and the error probability in
deciding ω relative to the language L is bounded: eN (ω) < R < 12 .

A relatively standard lemma shows that for error-bounded stochastic

networks, the error probability can be reduced to any desired value [9]. This
indicates that the complexity class of functions recognized by stochastic
networks is well-defined.
The stochastic model reveals intermediate interesting classes between
P and AnalogP. For example, stochastic networks with rational weights
and real probabilities form a similar model to the quantum Turing machine
and computes BPP/log. The full model with stochasticity still computes
AnalogP.
296 H. Siegelmann

We next form a more general framework of analog computational mod-

els, which do not have to adhere to a network structure at all. This one
will handle real atomic data items, use real constants, and it will update
by simple continuous functions. The time update can be synchronous, or
non-synchronous. The analog shift map is a good example [11]. It is a gen-
eralization of the well-known shift map by Smale and the generalized shift
map by Moore [6]. The strong analog shift map computes exactly analogP.
Noting that the basic ARNN encompasses many other models of analog
computation, Sontag and Siegelmann proposed an analogy to the Church-
Turing thesis of computability, but to the realm of analog computation.
Their thesis of time bounded analog computation suggests that all reason-
able analog computational models can be described by the simple ARNN:

“No possible abstract analog device can have more computa-

tional capabilities (up to polynomial time) than Analog Recur-
rent Neural Networks.”

This can be interpreted as suggesting that the neural network be considered

a standard model in the realm of analog computation, functioning in a role
similar to that of the Turing machine in the realm of digital computation.

5 Networks with Infinite Precision Tests

The class of analog systems introduced in the previous section relies on the
feature of continuous update. We next show that the introduction of even a
single discontinuity changes the computational properties of these systems.
The material of this section describes the work from [3].
Arithmetic networks are finite size interconnections with two types of
neurons. Some are continuous and some are not. In the simple model,
some neurons will compute σ(φ(x, Φ)) where Φ is any polynomial and σ is
the saturated linear function of equation (2); other neurons will compute
σH (φ(x, I)), where φH is the threshold function
!
1, x ≥ 0,
H(x) =
0, x < 0,

or the zero test function

!
1, x = 0,
Z(x) =
 0
0, x =

(or any other function with “gap continuities” [3]).

Finite Versus Infinite Neural Computation 297

A real number r is called polynomial-time computable if there is a poly-

nomial p and a Turing machine M such that on input n, M will produce
the first n digits of the fractional part of r in time p(n). All algebraic
numbers, constants such as π and e, and many others are polynomial-time
computable. To emphasize how small this class is, we note that there are
no more polynomial-time computable real numbers than Turing machines;
hence, there are countably many of them. Furthermore, when used as con-
stants in ARNN, the networks still compute the class P only, just like in
the case where all constants are rational numbers.
The next theorem contrasts the “linear precision suffices” property of
ARNN.
Theorem 5.1 There is no computable precision function r(n) such that
“precision O(r(t(n))) suffices” to simulate all arithmetic networks running
in time t(n). This is true even if only polynomial-time computable weights
are used.
The next theorem states the unlimited speed-up of arithmetic networks;
it contrasts ARNN because with polynomially computable real weights,
they compute only functions in P.
Theorem 5.2 There are arithmetic networks that run in polynomial time,
have polynomial-time computable weights, and yet they accept recursive lan-
guages of arbitrarily high time complexity (in the Turing machine sense).
Theorems 5.1 and 5.2 are both consequences of the following theorem.
Theorem 5.3 For every time-constructible function t(n) there is an arith-
metic network N such that:
1. The weights in N are computable in time O(n);
2. N runs in time 2n;
3. The language T accepted by N is recursive, but not decidable in time
O(t(n)) by any Turing machine.
4. Precision O(t(n)) does not suffice to simulate N , that is, if N is
simulated with precision O(t(n)), then a language different from T is
accepted, even in the soft acceptance sense.
Remark. Although ARNN in which φ is a linear combination or a high-
order polynomial are computationally equivalent, the high order property
is required for the theorems on arithmetic networks.
298 H. Siegelmann

Acknowledgments
This work was partially funded by the Israeli Ministry of Art and Sciences,
the Binational US/Israel Foundation, the fund for promotion of research at
the Technion, and the VPR fund at the Technion.

References
[1] J. L. Balcázar, R. Gavaldà, H. T. Siegelmann, Computational power of
neural networks: A characterization in terms of Kolmogorov complex-
ity, IEEE Transactions on Information Theory, 43, 4 (1997), 1175–
1183.

[2] M. P. Casey, The dynamics of discrete-time computation with applica-

tion to recurrent neural networks and finite state machine extraction,
Neural Computation, 8, 6 (1996), 1135–1178.

[3] R. Gavaldà, H. T. Siegelmann, Discontinuities in recurrent neural net-

works, Neural Computation, 11, 3 (1999), 715–745.

[4] R. M. Karp, R. Lipton, Turing machines that take advice, Enseignment

Mathematique, 28 (1982), 191–209.

[5] J. Kilian, H. T. Siegelmann, The dynamic universality of sigmoidal

neural networks, Information and Computation, 128, 1 (1996), 48–56.

[6] C. Moore, Unpredictability and undecidability in dynamical systems,

Physical Review Letters, 64 (1990), 2354–2357.

[7] W. Maass, E. D. Sontag, Analog neural nets with Gaussian or other

common noise distribution cannot recognize arbitrary regular lan-
guages, Neural Computation, 11, 3 (1999), 771–782.

[8] P. Orponen, W. Maass, On the effect of analog noise on discrete time

analog computations, Neural Computation, 10 (1998), 1071–1095.

[9] I. Parberry, Circuit Complexity and Neural Networks, MIT Press,

Cambridge, MA, 1994.

[10] R. Penrose, The Emperor’s New Mind, Oxford University Press, Ox-
ford, England, 1989.

[11] H. T. Siegelmann, Computation beyond the Turing limit, Science, 268

(April 1995), 545–548.
Finite Versus Infinite Neural Computation 299

[12] H. T. Siegelmann, Neural Networks and Analog Computation: Beyond

the Turing Limit, Birkhauser, Boston, MA, 1999.

[13] H. T. Siegelmann, A. Roitershtein, Noisy computational systems and

definite languages, Discrete Applied Mathematics, 1999, to appear.

[14] H. T. Siegelmann, E. D. Sontag, Analog computation via neural net-

works, Theoretical Computer Science, 131 (1994), 331–360.

[15] H. T. Siegelmann, E. D. Sontag, On computational power of neural

networks, Journal of Computer and System Sciences, 50, 1 (1995),
132–150.
 On Information-Theoretical Aspects
of Relational Databases
Dan A. Simovici, Szymon Jaroszewicz
Department of Mathematics and Computer Science
University of Massachusetts at Boston
Boston, MA 02125, USA
E-mail: {dsim,sj}@cs.umb.edu

Abstract. We introduce the notion of entropy for a set of at-

tributes of a table in a relational database starting from the
notion of entropy for finite functions. We examine the connec-
tions that exist between conditional entropies of attribute sets
and lossy decompositions of tables and explore the properties of
the entropy of sets of attributes regarded as an outer measure
on the set of subsets of a heading of a table. Finally, we suggest
a generalization of functional dependencies based on conditional
entropy.

1 Introduction
This paper examines some applications of information theory to relational
database systems. The starting point is our axiomatization of entropies of
finite functions based on the composition operation between functions that
was presented in [5].
Unless we state otherwise, all sets considered in this paper are finite.
Let F be the class of all functions between finite, nonempty sets. Define
the partial order @ on F by f @ g if f : A → B, g : A
→ B
, A ⊆ A
,
B ⊆ B
and f (a) = g(a) for a ∈ A. Note that if A, B, C, D are finite sets,
such that A ∩ B = C ∩ D = ∅, f : A → B, and g : C → D, then there exists
sup{f, g}. The function f  g = sup{f, g} is given by
!
f (x), if x ∈ A,
(f  g)(x) =
g(x), if x ∈ B.

301
302 D. A. Simovici, S. Jaroszewicz

It is easily verifiable, that, whenever it is defined, the “” operation is

commutative and associative.
The composition of functions f : A → B and g : B → C is a function
gf : A → C, such that gf (x) = g(f (x)), for every x ∈ A.
The Cartesian product of functions f : A → B and g : C → D is a
function f × g : A × C → B × D defined as (f × g)(x, y) = (f (x), g(y)), for
every x ∈ A and y ∈ C.
Let A1 , A2 , . . . , An be n disjoint subsets of a finite set A. The generalized
characteristic function of sets A1 , A2 , . . . , An is a function:

χA1 ,A2 ,...,An : A1 ∪ A2 ∪ . . . ∪ An → {1, 2, . . . , n},

such that χA1 ,A2 ,...,An (x) = i if x ∈ Ai , for 1 ≤ i ≤ n.

Another important notion for this paper is the notion of multiset. We
present a few useful facts (see [3] for further details).

Definition 1.1 A multiset or bag on a set S is a subset M of S × N such

that for every a ∈ S there is an n ∈ N such that (a, m) ∈ M if and only if
m < n. The set of all multisets on S is denoted by M(S).

If S is a set, M is a multiset on S, and a ∈ S, then the number n such

that (a, m) ∈ M if and only if m < n is unique since it is the least number
r such that (a, r) ∈ M . We call this number n the multiplicity of a in M
and we denote it by M (a).
A multiset on a set S is determined by the multiplicities of its elements.
Note that for any set S, ∅ ⊆ M(S). The cardinality of a multiset M on S

is |M | = a∈S M (a).
Let S be a set and let M and N be multisets on S. It is easy to see
that M ∪ N is a multiset on S and that if a ∈ S has multiplicity m in M
and n in N , then a has multiplicity max(m, n) in M ∪ N . Similarly, M ∩ N
is a multiset on S and if a ∈ S has multiplicity m in M and n in N , then
a has multiplicity min(m, n) in M ∩ N . In other words, we have

(M ∪ N )(a) = max{M (a), N (a)},

(M ∩ N )(a) = min{M (a), N (a)},

for every a ∈ S.
The sum of M and N (denoted M + N ) is the multiset on S such that
for each a ∈ S, the multiplicity of a in M +N is the sum of the multiplicities
of a in M and N .
On Information-Theoretical Aspects of Relational Databases 303

Let S be a set. For each subset T of S, we can define a multiset MT on

S by defining
MT = {(a, 0) | a ∈ T }.
(So, if a ∈ T , then a has multiplicity 1 in MT , and if a ∈ T , then a has
multiplicity 0 in MT .)
For any two subsets U and T of S we have MU ∪T = MU ∪ MT and
MU ∩T = MU ∩ MT .
If M1 and M2 are multisets on S1 and S2 respectively, then the Cartesian
product M1 × M2 of M1 and M2 is defined as a multiset on S1 × S2 such
that
(M1 × M2 )(a, b) = M1 (a) · M2 (b),
for every a ∈ S1 , b ∈ S2 .

2 Functional Entropy
If f : A → B, g : A → C are finite functions, let pf (b) and pf,g (b, c) be
given by

pf (b) = P (f = b) = |f −1 (b)|/|A|,
pf,g (b, c) = P (f = b ∧ g = c) = |f −1 (b) ∩ g −1 (c)|/|A|,

for b ∈ B, and c ∈ C. Note that

 
pf,g (b, c) = pf (b) and pf,g (b, c) = pg (c).
c∈C b∈B

Definition 2.1 Let f : A → B be a function between finite sets, and let

B = {b1 , b2 , . . . , bm }. The discrete random variable associated with f is
denoted by X f and defined as:
 
f b0 b1 ··· bn−1
X = .
pf (b0 ) pf (b1 ) · · · pf (bn−1 )

Let f : A → B and g : A → C be functions between finite sets. The

functions f and g are said to be independent if the random variables X f
and X g are independent, that is, if pf,g (b, c) = pf (b)pg (c) for every b ∈ B
and c ∈ C.

Let H : F → R be a function assigning a real number to every f ∈ F

that satisfies the following properties:
304 D. A. Simovici, S. Jaroszewicz

(E1) H(f α) = H(f ), for every function f : A → B and bijection α : A

→
A.

(E2) H(gf ) ≤ H(f ), for every f : A → B and g : B → C.

(E3) Let A, C be finite sets, such that |A| ≤ |C|, let α : A → B and
β : C → D be bijections. Then H(α) ≤ H(β).

(E4) If f : A → B and f : C → D are functions between finite sets and

A ∩ C = B ∩ D = ∅, then

|A| |C|
H(f  g) = H(f ) + H(g) + H(χA,C ).
|A ∪ C| |A ∪ C|

(E5) H(f × g) = H(f ) + H(g) for all functions f : A → B and g : C → D.

We proved in [5] that for any such function H we have


m−1
|Ai | |Ai |
H(f ) = − log ,
i=0
|A| |A|

where f : A → B, B = {b0 , . . . , bm−1 } and Ai = f −1 (bi ) for 0 ≤ i ≤ m − 1.

The conditional entropy of two functions was defined in [5] as a function
H : F × F → R that satisfies the following conditions:

(CE1) H(f |c) = H(f ) for every constant function c.

(CE2) For every f : A → B, g1 : A1 → C1 , g2 : A2 → C2 , where A1 ∩ A2 =

C1 ∩ C2 = ∅ and A1 ∪ A2 = A we have

|A1 | |A2 |
H(f |g1  g2 ) = H(fA1 |g1 ) + H(fA2 |g2 ).
|A| |A|

If B = {b0 , . . . , bm−1 }, C = {c0 , . . . , cn−1 }, then, as we have shown

in [5], the axioms (CE1-2) imply that the conditional entropy between the
functions f : A → B and g : A → C has the form:

 m−1
n−1  |Aij | |Aij |
H(f |g) = − log ,
j=0 i=0
|A| |Aj |

where Aij = f −1 (bi ) ∩ g −1 (cj ), Aj = g −1 (cj ) for 0 ≤ i ≤ m − 1 and

0 ≤ j ≤ n − 1.
On Information-Theoretical Aspects of Relational Databases 305

Theorem 2.1 Let f : A → B, g : A → C be two functions. There exists a

function h : C → B such that f = hg if and only if H(f |g) = 0.

Proof. Let B = {b0 , . . . , bm−1 } and C = {a0 , . . . , an−1 }. Suppose that

H(f |g) = 0. Using the previous notations, for every nonempty set Aij we
have |Aij | = |Aj | for every i, 0 ≤ i ≤ m − 1. Since Aij ⊆ Aj it follows that
Aij = ∅ implies Aij = Aj . Thus, if there exists a ∈ A such that g(a) = cj
then f (a) = bi . This shows that the mapping h : C → B given by g(cj ) = bi
is well defined and f = hg.
Conversely, suppose that f = hg. Then, if Aij = ∅ we have Aj = Aij .
Indeed, suppose that a ∈ Aj , that is g(a) = cj , which implies f (a) = h(cj ),
which proves that Aj ⊆ Aij , where bi = h(cj ). The converse inclusion is
obvious, and this implies immediately H(f |g) = 0. ✷

The mutual information between functions f : A → B and g : A → C,

denoted by I(f ; g), is defined as I(f ; g) = H(f )−H(f |g) and can be written
as:

I(f ; g)
  pf,g (b, c)
= − pf (b) log pf (b) + pf,g (b, c) log (1)
b∈B b∈B c∈C
pg (c)
  pf,g (b, c)
= − pf,g (b, c) log pf (b) + pf,g (b, c) log
b∈B c∈C b∈B c∈C
pg (c)
 pf,g (b, c)
= pf,g (b, c) log .
b∈B c∈C
pf (b)pg (c)

Lemma 2.1 We have I(f ; g) ≥ 0 for all functions between finite sets f :
A → B and g : A → C, with equality if and only if f and g are independent.

Proof. The statement follows immediately from formula (1). ✷

Below we give a few facts concerning entropies of functions between

finite sets.
Theorem 2.2 We have H(f |hg) ≥ H(f |g), for every f : A → B, g : A →
C and h : C → D.
Proof. Let D = {d1 , . . . , dm } and Ai = (hg)−1 (di ). From (CE2) and (CE1)
it follows that
m
|Ai |
H(f |hg) = H(fAi ). (2)
i=1
|A|
306 D. A. Simovici, S. Jaroszewicz

Since g(Ai ) ∩ g(Aj ) = ∅, for i = j, we can decompose g into a union

g = gA1  · · ·  gAm , where gAi = g|Ai is a restriction of g to the set Ai .
Using (CE2) we obtain:


m
|Ai |
H(f |g) = H(fAi |gAi ). (3)
i=1
|A|

From Lemma 2.1 it follows that

H(fAi ) ≥ H(fAi |gAi ), (4)

for 1 ≤ i ≤ m. From (2), (3) and (4) we obtain: H(f |hg) ≥ H(f |g). ✷

Let f : A → B and g : A → C be two functions. Define the function

(f ; g) : A → B × C as (f ; g)(a) = (f (a), g(a)) for every a ∈ A. We proved
in [5] that
H(f ; g) = H(f |g) + H(g) = H(g|f ) + H(f ). (5)

3 Database Tables and m-Tables

Let U = {A0 , A1 , . . .} be a set of symbols referred to as attributes; U itself
is referred to as the universal set. We assume that for each attribute A ∈ U
there exists a finite set Dom(A), the domain of A, that contains at least
two elements.
A finite set of attributes {A0 , . . . , An−1 } is denoted by A0 . . . An−1 ,
which is the standard notation for sets of database attributes. The union
of sets of attributes A and B will be denoted by AB. At times however,
for the sake of clarity, the standard mathematical notation A ∪ B will be
used.
If L = Ai0 · · · Ain−1 ⊆ U , then we denote by Dom(L) the Cartesian
product
Dom(Ai0 ) × · · · × Dom(Ain−1 ).
Uppercase letters denote sets of attributes, while lowercase letters denote
values from the domain of the corresponding set of attributes, e.g., x ∈
Dom(X) or y ∈ Dom(Y ).
If u ∈ Dom(L) and K ⊆ L, we denote by u[K] the restriction of u to
K; we refer to u[K] as the projection of u on K.
Let L, M be two finite subsets of the universal attribute set U . The
tuples u ∈ Dom(L) and v ∈ Dom(M ) are joinable if u[L ∩ M ] = v[L ∩ M ].
Their join is the tuple u ✶ v ∈ Dom(L ∪ M ) defined by (u ✶ v)[A] = u[A]
On Information-Theoretical Aspects of Relational Databases 307

if A ∈ L and (u ✶ v)[A] = v[A] if A ∈ M . The joinability condition insures

that the tuple u ✶ v is well-defined.
Definition 3.1 A table is a triple σ = (T, L, ρ), where T is a string of
characters referred to as the name of the table, L = Ai0 · · · Ain−1 is a finite
subset of U called the heading of σ, and ρ ⊆ Dom(L) is an n-ary relation
referred to as the extension of σ. The pair (T, L) is the format of σ. An
element t ∈ ρ is called a tuple of σ.
Let σ = (T, L, ρ) be a table and let X ⊆ L. The projection of σ on the
set X is the table σ[X] = (T [X], X, ρ[X]), where ρ[X] = {t[X] | t ∈ ρ}.
Observe that for every X ⊆ L we have a projection function πσ,X : ρ →
ρ[X], where πσ,X (t) = t[X] for every t ∈ ρ.
If σ = (T, L, ρ), σ
= (T
, L
, ρ
) are tables, then their join is the table
σ ✶ σ
= (T ✶ T
, L ∪ L
, ρ ✶ ρ
), where
ρ ✶ ρ
= {u ✶ v | u ∈ ρ, v ∈ ρ
and u, v are joinable}.
The notations and terminology introduced above can be found, for ex-
ample, in [9].
To allow duplicate tuples in database tables, we introduce the notion of
m-table as follows:
Definition 3.2 An m-table is a triple τ = (T, L, µ), where T, L are as in
Definition 3.1, and µ is a finite multiset on Dom(L).
An element t ∈ µ will be called a tuple of τ .
If X ⊆ L is a set of attributes, the projection of an m-table τ on X is
the m-table τ [X] = (T [X], X, µ[X]), where

µ[X](x) = {µ(t) | t ∈ Dom(L) and t[X] = x}.
Let t = (r, n) ∈ µ, denote t[X] = r[X].
A table σ = (S, L, ρ) can be regarded as a special case of m-table
τ = (S, L, µ), where !
1, if t ∈ ρ,
µ(t) =
0, otherwise,
for every t ∈ Dom(L).
Example 3.1 Consider the m-table τ = (T, ABC, µ), where
Dom(A) = {a0 , a1 , a2 },
Dom(B) = {b0 , b1 },
Dom(C) = {c0 , c1 , c2 }.
308 D. A. Simovici, S. Jaroszewicz

Suppose that µ(a0 , b1 , c2 ) = 2, µ(a1 , b0 , c1 ) = 3, µ(a2 , b1 , c2 ) = 1 and µ(t) =

0 for every other t ∈ Dom(ABC). Then, we visualize this m-table as
T
A B C
a0 b1 c2
a0 b1 c2
a1 b0 c1
a1 b0 c1
a1 b0 c1
a2 b1 c2
The projection function πτ,X for m-tables has a definition that is similar
to the one used for common tables.

Definition 3.3 Let τ = (T, L, µ) be an m-table and let X ⊆ L. The

function πτ,X : µ → Dom(X) is given by πτ,X (t) = t[X].

The relational algebra selection operation for m-tables is defined in a

similar way as for ordinary tables.
Let τ1 = (T1 , L1 , µ1 ), and τ2 = (T2 , L2 , µ2 ) be two m-tables. The join of
m-tables τ1 and τ2 is the m-table τ1 × τ2 = (T1 ✶ T2 , L1 L2 , µ1 ✶ µ2 ), such
that
(µ1 ✶ µ2 )(r1 ✶ r2 ) = µ1 (r1 ) · µ2 (r2 )
for any joinable r1 ∈ Dom(L1 ) and r2 ∈ Dom(L2 ).
The behavior of m-tables with respect to the join operation differs from
the behavior of regular tables, as the next example shows.
Example 3.2 Consider m-table τ = (T, ABC, µ) given below
T
A B C
a0 b1 c3
a0 b1 c4
Let us form an m-table τ
= τ [AC] ✶ τ [AB]:
T
A B C
a0 b1 c3
a0 b1 c4
a0 b1 c3
a0 b1 c4
On Information-Theoretical Aspects of Relational Databases 309

Thus, in case of m-tables, we may have |µ

| > |µ| even if the functional
dependency A → B holds.

4 Entropy of Attribute Sets

Let τ = (T, L, µ) be an m-table, and X ⊆ L be a set of attributes. The
active domain of a set of attributes X in the table τ is the set
adomτ (X) = {x ∈ Dom(X) | t[X] = x for some t ∈ µ}.
Since tables are special cases of m-tables, the above notation is valid also
for tables.
Now we introduce the definition of entropy of sets of attributes of an
m-table.
Definition 4.1 Let τ = (T, L, µ) be an m-table and let X ⊆ L. The
entropy of X in τ is defined as H(πτ,X ). We will denote this entropy as
Hτ (X). We assume that Hτ (∅) = 0, for any m-table τ = (T, L, µ).
The conditional entropy between the sets of attributes X, Y in τ is de-
fined as
Hτ (Y |X) = H(πτ,Y |πτ,X ).
We assume that Hτ (∅|X) = 0, for any m-table τ = (T, L, µ) and any
X ⊆ L.
The joint entropy of X and Y denoted by Hτ (X; Y ) is defined as
Hτ (X; Y ) = H(πτ,X ; πτ,Y ).
These definitions and the equalities (5) imply
H(X; Y ) = H(X|Y ) + H(Y ) = H(Y |X) + H(X). (6)
Let τ = (T, L, µ) be an m-table and let
x ∈ Dom(X), y ∈ Dom(Y ), z ∈ Dom(Z),
where X, Y, Z ⊆ L. Define the probabilities:
pτ (x) = |{t ∈ µ : t[X] = x}|/|µ|,
pτ (y, z) = |{t ∈ µ : t[Y ] = y ∧ t[Z] = z}|/|µ|,
pτ (y|x) = |{t ∈ µ : t[Y ] = y ∧ t[X] = x}|/|{t ∈ µ : t[X] = x}|,
pτ (y, z|x) = |{t ∈ µ : t[Y ] = y ∧ t[Z] = z ∧ t[X] = x}|
/|{t ∈ µ : t[X] = x}|.
If τ is clear from context, we will omit this subscript.
310 D. A. Simovici, S. Jaroszewicz

Definition 4.2 Two sets of attributes X and Y are independent if the

corresponding functions πX and πY are independent.

Theorem 4.1 Let τ = (T, L, µ) be an m-table and X, Y ⊆ L be sets of

attributes. Then Hτ (XY ) = Hτ (X; Y ).

Proof. Let RXY = {(x, y) | x ∈ Dom(X), y ∈ Dom(Y ), x[X ∩ Y ] = y[X ∩

Y ]}. Define the function (πX ; πY ) : µ → RXY by (πX ; πY )(t) = (t[X], t[Y ])
for t ∈ µ.
If f : Dom(XY ) → RXY and g : RXY → Dom(XY ) are given by f (v) =
(v[X], v[Y ]) for v ∈ Dom(XY ) and g(x, y) = (x ✶ y) for (x, y) ∈ RXY , then
(πX ; πY )(t) = f [πXY (t)] and πXY (t) = g[(πX ; πY )(t)] for every t ∈ µ. From
(E2) it follows that H(πX ; πY ) ≤ H(πXY ) and H(πXY ) ≤ H(πX ; πY ), so
H(πX ; πY ) = H(πXY ). ✷

The functional dependency between sets of attributes X and Y is de-

noted by X → Y . The definition of functional dependencies for m-tables
in analogous to the corresponding definition for tables.

Theorem 4.2 Let τ = (T, L, µ) be an m-table and let X, Y ⊆ L be sets of

attributes. The following statements are equivalent:
1. τ satisfies the functional dependency X → Y ;
2. Hτ (Y |X) = 0;
3. H(X) = H(XY ).

Proof. The m-table τ satisfies the functional dependency X → Y if and only

if there exists a mapping f : adomτ (X) → adomτ (Y ) such that πY = f πX .
Indeed, if such a mapping exists then u[X] = v[X] amounts to πX (u) =
πX (v), which implies πY (u) = f (πX (u)) = f (πX (v)) = πY (v) for u, v ∈ µ.
Thus, u[Y ] = v[Y ], which shows that τ satisfies X → Y . Conversely, if τ
satisfies X → Y , then the mapping f : adomτ (X) → adomτ (Y ) given by
f (x) = y if there is a tuple t ∈ µ such that t[X] = x and t[Y ] = y is well
defined and πY = f πX . Thus, by Theorem 2.1, τ satisfies X → Y if and
only if Hτ (Y |X) = 0.
To prove the equivalence between the second and the third statement
observe that the equalities (6) and Theorem 4.1 imply
Hτ (XY ) = Hτ (X|Y ) + Hτ (Y ) = Hτ (Y |X) + Hτ (X). (7)
Thus, Hτ (X) = Hτ (XY ) if and only if Hτ (Y |X) = 0. ✷
On Information-Theoretical Aspects of Relational Databases 311

Theorem 4.3 Let τ = (T, L, µ) be an m-table and let X, Y, Z ⊆ L. The

following statements hold:

1. if X ⊆ Y , then Hτ (X) ≤ Hτ (Y ); the equality Hτ (X) = Hτ (Y ) holds

if and only if τ satisfies the functional dependency X → Y ;

2. Hτ (Y |X) ≤ Hτ (Y Z|X), with equality if and only if τ satisfies the

functional dependency XY → Z;

3. if Y ⊆ Z, then Hτ (Y |X) ≤ Hτ (Z|X), with equality if and only if τ

satisfies XY → Z.

Proof. For the first part note that the equalities (7) imply Hτ (X) ≤
Hτ (XY ) = Hτ (Y ). Moreover, if Hτ (X) = Hτ (Y ) = Hτ (XY ), then
H(Y |X) = 0, which implies that τ satisfies X → Y .
For the second part, we can write

Hτ (Y |X) = Hτ (XY ) − Hτ (X) ≤ Hτ (XY Z) − Hτ (X) = Hτ (Y Z|X).

Thus, that the equality holds if and only if XY → Z.

The argument for the last part is similar to the argument for the first
part and is omitted. ✷

Theorem 4.4 Let τ = (T, L, µ) be an m-table and let X, Y, Z ⊆ L. The

following statements hold:

1. Hτ (Y |XZ) ≤ Hτ (Y |X);

2. if Z ⊆ X, then Hτ (Y Z|X) = Hτ (Y |X);

3. Hτ (Y Z|XZ) ≤ Hτ (Y |X).

Proof. For Part 1 note that πX = f πXZ for some function f . It follows
from Theorem 2.2 that H(πY |πX ) = H(πY |f πXZ ) ≥ H(πY |πXZ ), hence
Hτ (Y |XZ) ≤ Hτ (Y |X).
For Part 2 we have

Hτ (Y Z|X) = Hτ (XY Z) − Hτ (X) = Hτ (XY ) − Hτ (X) = Hτ (Y |X).

Finally, for Part 3 observe that from the first two parts it follows that
Hτ (Y |X) ≥ Hτ (Y |XZ) = Hτ (Y Z|XZ). ✷
312 D. A. Simovici, S. Jaroszewicz

Theorem 4.5 Let X, Y ⊆ L. We have Hτ (XY ) ≤ Hτ (X) + Hτ (Y ), with

equality if and only if X and Y are independent.

Proof. By Theorem 4.1 and Lemma 2.1 we have:

Hτ (XY ) = Hτ (X; Y ) = H(πX ; πY ) = H(πY |πX ) + H(πX )

≤ H(πY ) + H(πX ) = Hτ (X) + Hτ (Y ).

From Lemma 2.1 it follows that H(πY |πX ) = H(πY ) if and only if πY and
πX are independent, which completes the proof. ✷

The restriction πY |(µ where X=x) will be denoted by πY |X=x for any
sets of attributes X and Y .
Theorem 4.6 Let X, Y, Z ⊆ L. We have Hτ (Y Z|X) ≤ Hτ (Y |X) +
Hτ (Z|X), and the equality holds if and only if for every x ∈ adomτ (X)
the functions πY |X=x and πZ |X=x are independent.
Proof. Let Ax = {t ∈ µ : πX (t) = x}, By Theorem 4.5 we have:
 |Ax |
Hτ (Y Z|X) = H(τ where X=x) (Y Z)
x |µ|
 |Ax |
≤ H(τ where X=x) (Y )+
x |µ|
 |Ax |
H(τ where X=x) (Z)
x |µ|
= Hτ (Y |X) + Hτ (Z|X).

The condition on equality follows from the equality condition of Theorem

4.5. ✷

5 Relational Algebra and Entropy

Relational algebra operations may generate m-tables even if they are ap-
plied to tables. Let σ = (T, L, ρ) be a table and let X ⊆ L be a set of
attributes. The projection without removal of duplicates of σ on X is the
m-table σ{X} = (T {X}, X, µ), where µ is the multiset on Dom(X) given
by
µ(x) = |{t ∈ ρ | t[X] = x}|
for x ∈ Dom(X). This operation models the effect of a standard SQL query
On Information-Theoretical Aspects of Relational Databases 313

select A1 , . . . , Ak from T,
where A1 · · · Ak ⊆ L.

Theorem 5.1 Let X ⊆ L. Then, Hσ{X} (X) = Hσ (X) ≤ Hσ (L).

Proof. The proof of the equality is straightforward; the inequality is an

immediate consequence of Part 1 of Theorem 4.3. ✷

The table obtained by removal of duplicate tuples from the m-table τ

will be denoted by τ .

Theorem 5.2 Let τ = (T, L, µ) be an m-table. We have Hτ (L) ≥ Hτ (L).

Proof. Note that after removal of duplicate tuples, the number of distinct
tuples will not change; all of them, however, will occur equally often, so the
function πτ,L becomes a bijection and the inequality follows. ✷

However, the entropy of a proper subset of L may either increase or

decrease after the removal of duplicate tuples form τ .
Let τ1 = (T1 , L1 , µ1 ) and τ2 = (T2 , L2 , µ2 ) be m-tables and such that
L1 ∩ L2 = ∅. Their Cartesian product is defined as an m-table τ1 × τ2 =
(T1 × T2 , L1 L2 , µ1 × µ2 ). The following theorem holds.
Theorem 5.3 We have

Hτ1 ×τ2 (L1 L2 ) = Hτ1 (L1 ) + Hτ2 (L2 ).

Proof. It is easy to see that πτ1 ×τ2 ,L = πτ1 ,L1 × πτ2 ,L2 . It follows form (E5)
that H(πτ1 ×τ2 ,L ) = H(πτ1 ,L1 ) + H(πτ2 ,L2 ). ✷

6 Entropy and Table Decompositions

Let U, V ⊆ L be a pair of sets of attributes of a table τ = (T, L, ρ). If
U ∪ V = L, then we refer to (U, V ) as a decomposition of τ . Let ρ
=
ρ[U ] ✶ ρ[V ]. In general, we have ρ ⊆ ρ
for every decomposition (U, V ) of
ρ. The set of spurious tuples of the decomposition (U, V ) is the set

spr(U, V ) = ρ
− ρ = (ρ[U ] ✶ ρ[V ]) − ρ.

If spr(U, V ) = ∅, then we refer to (U, V ) as a lossless decomposition; other-

wise, (U, V ) is a lossy decomposition.
314 D. A. Simovici, S. Jaroszewicz

Let τ = (T, L, ρ) be a table. Denote by nx (ρ) = |{t ∈ ρ | t[X] = x}|

the number of tuples in ρ containing x for x ∈ Dom(X). Observe that
nx (ρ[XY ]) is the number of different values of y ∈ Dom(Y ) present in
tuples of ρ containing x. If the functional dependency X → Y holds,
then nx (ρ[XY ]) = 1 for all x ∈ adomτ (X). It is easy to see that |ρ| =

x∈adomτ (X) nx (ρ).
−1
Note that |πτ,Y (y)| = |{t ∈ Dom(L) | t[Y ] = y}| = ny (ρ) and a similar
−1 −1
equality holds for πτ,X . Also, πτ,Y (y) ∩ πτ,X (x) = {t ∈ Dom(L) | t[XY ] =
xy}. Therefore, the conditional entropy Hτ (Y |X) can be written as
  nxy (ρ) nxy (ρ)
Hτ (Y |X) = − log2 .
y∈adomτ (Y ) x∈adomτ (X)
|ρ| nx (ρ)

Let (U, V ) be a decomposition of τ and let X = U ∩ V . It is clear that


spr(U, V ) = nx (ρ[U ])nx (ρ[V ]) − |ρ|.
x∈adomτ (X)

Moreover, we have the following result.

Theorem 6.1 Let (U, V ) be a decomposition of τ . Define X = U ∩ V, Y =

U − X, and Z = V − X. Then, we have
 
2Hτ (Y |X)+Hτ (Z|X)
spr(U, V ) ≥ |ρ| −1 .
maxx∈adomτ (X) nx (ρ)

Proof. For the conditional entropy Hτ (Y |X) we can write

  nyx (ρ) nyx (ρ)
Hτ (Y |X) = − log2 .
y∈adom(Y ) x∈adom(X)
|ρ| nx (ρ)

If x ∈ adomτ (X), then nx (ρ[XY ]) is the number of elements of the active

domain of Y that occur in the the relation ρx = (ρ where X = x) under

Y and nx (ρ[XY ]) = y∈adomτ (Y ) nxy (ρ). Consequently, the value of the
entropy of ρx will not exceed log2 nx (ρ[XY ]).
Since
 nx (ρ)
Hτ (Y |X) = Hτ (Y |X = x),
x∈adomτ (X)
|ρ|
 nx (ρ)
Hτ (Z|X) = Hτ (Z|X = x),
x∈adomτ (X)
|ρ|
On Information-Theoretical Aspects of Relational Databases 315

we can write
 nx (ρ)
(Hτ (Y |X=x)+Hτ (Z|X=x))
2Hτ (Y |X)+Hτ (Z|X) = 2 x∈adomτ (X) |ρ|
.

The convexity of the exponential function implies:

 nx (ρ) Hτ (Y |X=x)+Hτ (Z|X=x)
2Hτ (Y |X)+Hτ (Z|X) ≤ 2
x∈adomτ (X)
|ρ|
 nx (ρ) log2 nx (ρ[XY ])+log2 nx (ρ[XZ])
≤ 2
x∈adomτ (X)
|ρ|
 nx (ρ)
= nx (ρ[XY ])nx (ρ[XZ])
x∈adomτ (X)
|ρ|
maxx∈adomτ (X) nx (ρ)
≤ · |ρ
|.
|ρ|

Therefore, we have

|ρ|
|ρ
| ≥ · 2Hτ (Y |X)+Hτ (Z|X) ,
maxx∈adomτ (X) nx (ρ)

which implies
 
2Hτ (Y |X)+Hτ (Z|X)
spr(U, V ) ≥ |ρ| −1 .
maxx∈adomτ (X) nx (ρ)

7 The Entropy as a Measure

Let τ = (T, L, µ) be an m-table. We assume that Hτ (X|∅) = Hτ (X).

Theorem 7.1 Let τ = (T, L, µ) be an m-table. The entropy of the sets

of database attributes of the m-table τ conditioned on some fixed set of
attributes X ⊆ L is an outer measure on the set of attributes L, i.e., it
satisfies the following properties:

1. Hτ (∅|X) = 0 and Hτ (Y |X) ≥ 0, for all Y ⊆ L;

2. if Y1 ⊆ Y2 ⊆ L, then Hτ (Y1 |X) ≤ Hτ (Y2 |X);

316 D. A. Simovici, S. Jaroszewicz

 
3. Hτ ( Yi |X) ≤ Hτ (Yi |X), for every countable collection of sets of
attributes {Yi }, such that Yi ⊆ L for all i.

Proof. Property 1 follows from the definition of conditional entropy between

sets of attributes and from the fact that the conditional entropy of functions
between finite sets is always non-negative. Property 3 is an immediate
consequence of Part 3 of Theorem 4.3.
Since L is a finite set, in the proof of Property 3 we only need to consider
finite collections of sets of attributes {Yi }. The argument is by induction.
For collections consisting of only one set of attributes Property 3 holds.
Suppose now that the property holds for all collections {Yi }, i = 1, . . . , n of
n sets of attributes. Then, for every collection {Yi }, i = 1, . . . , n, n + 1, of
n + 1 sets of attributes we have (using theorem 4.6):

n+1   n 
 
Hτ Yi |X = Hτ Yi ∪ Yn+1 |X
i=1 i=1
 n 

≤ Hτ Yi |X + Hτ (Yn+1 |X)
i=1

n
≤ Hτ (Yi |X) + Hτ (Yn+1 |X)
i=1

n+1
= Hτ (Yi |X).
i=1

Note that, as a special case, when X = ∅, unconditional entropy of

attributes is also an outer measure on the set of database attributes L.
For a set of attributes E ⊆ L, denote E
= L − E.
Definition 7.1 A set of attributes E is measurable under an outer measure
Hτ (·|X) if for every Y ⊆ L we have Hτ (Y |X) = Hτ (Y ∩ E|X) + Hτ (Y ∩
E
|X).

Theorem 7.2 If either Hτ (E|X) = 0 or Hτ (E

|X) = 0 for some set of
attributes X, then both E and E
are measurable under Hτ (·|X).

Proof. Suppose that Hτ (E|X) = 0. It is a general property of measures that

a set whose outer measure zero is measurable. Another general property
of measures is that a complement of a measurable set is also measurable,
On Information-Theoretical Aspects of Relational Databases 317

which proves the measurability of E

. The proof when Hτ (E
|X) = 0 is
analogous. ✷

Corollary 7.1 ∅ and L are measurable under Hτ (·|X) for every X ⊆ L.

Theorem 7.3 A set of attributes E ⊂ L, E = ∅ is measurable under an

outer measure Hτ (·|X) if and only if for every x ∈ adomτ (X) the functions
πE |X=x and πE
|X=x are independent.

Proof. Take any Y ⊆ L and denote Y1 = Y ∩ E, Y2 = Y ∩ E

.
By Definition 7.1 and Theorem 4.6, E is measurable if for every Y ⊆ L
and x ∈ adomτ (X), p(y1 , y2 |x) = p(y1 |x)p(y2 |x) for every y1 ∈ Dom(Y1 ),
y2 ∈ Dom(Y2 ).
Suppose that condition holds. Then, for every x ∈ adomτ (X) we have,
p(e, f |x) = p(e|x)p(f |x) for every e ∈ Dom(E), f ∈ Dom(E
).
Denote Ey1 = {e ∈ Dom(E) | e[Y1 ] = y1 } and Ey
2 = {f ∈ Dom(E
) |
f [Y2 ] = y2 }, where y1 ∈ Dom(Y1 ), y2 ∈ Dom(Y2 ). From Y1 ⊆ E and Y2 ⊆
 
E
it follows that p(y1 |x) = e∈Ey p(e|x) and p(y2 |x) = f ∈Ey
p(f |x).
1 2
Choose arbitrarily Y ⊆ L and x ∈ adomτ (X). Then, for every y1 ∈
Dom(Y1 ) and y2 ∈ Dom(Y2 ), we have:
   
p(y1 , y2 |x) = p(e, f |x) = p(e|x)p(f |x)
e∈Ey1 f ∈Ey
2 e∈Ey1 f ∈Ey
2
 
    
= p(e|x) p(f |x) = p(e|x)p(y2 |x)
e∈Ey1 f ∈Ey
2 e∈Ey1

= p(y2 |x) p(e|x) = p(y2 |x)p(y1 |x).
e∈Ey1

Since the choice of Y and x was arbitrary, the above equality holds for
every Y ⊆ L and x ∈ adomτ (X). This proves, that when the condition of
the theorem implies the measurability of E.
Suppose now that the condition of the theorem does not hold. We have:

Hτ (L|X) = Hτ (E ∪ E
|X).

From Theorem 4.6 and the fact that the condition of the theorem does not
hold, it follows that:

Hτ (E ∪ E
|X) < Hτ (E|X) + Hτ (E
|X)
= Hτ (L ∩ E|X) + Hτ (L ∩ E
|X).
318 D. A. Simovici, S. Jaroszewicz

This shows that when the condition is violated, E is indeed not measurable,
which completes the proof. ✷

As a special case, when X = ∅, a set of attributes E ⊂ L, E = ∅

is measurable under unconditional entropy if and only if πE and πE
are
independent.
Another characterization of measurability is given by the following the-
orem.
Theorem 7.4 E ⊆ L is measurable under unconditional entropy if and
only if Hτ (E) + Hτ (E
) = Hτ (L).
Proof. If E = L or E = ∅, then the result obviously holds. The proof when
E = L and E = ∅ is an immediate consequence of Theorems 7.3 and 4.5.
✷

Theorem 7.5 If E ⊆ L is measurable under unconditional entropy, then:

(i) the functional dependency X → Y does not hold for any X ⊆ E,
Y ⊆ E
, unless Hτ (Y ) = 0, i.e., πY is constant or Y = ∅;
(ii) the functional dependency X → Y does not hold for any X ⊆ E
,
Y ⊆ E, unless Hτ (Y ) = 0, i.e., πY is constant or Y = ∅.

Proof. Suppose, that E is measurable. There is a functional dependency

X → Y , and Hτ (Y ) = 0. Suppose that X ⊆ E and Y ⊆ E
. For the set of
attributes X ∪ Y we can write:
Hτ [(X ∪ Y ) ∩ E] + Hτ [(X ∪ Y ) ∩ E
]
= Hτ (X) + Hτ (Y ) = Hτ (X) = Hτ (X ∪ Y ).
Thus, E is not measurable. This completes the proof of part (i). The proof
of (ii) is similar. ✷

Note that when in the above theorem Hτ (X) = 0 and the functional
dependency X → Y holds, then Hτ (Y ) must also be equal to zero, so such
a dependency does not exclude the measurability of E.

Theorem 7.6 If the functional dependency X → E holds, then E is mea-

surable under Hτ (·|X).

Proof. If X → E, then Hτ (E|X) = 0, and since E has outer measure 0, it

is measurable. ✷
On Information-Theoretical Aspects of Relational Databases 319

Theorem 7.7 If K ⊆ L is a candidate key, then K is measurable under

unconditional entropy if and only if Hτ (K
) = 0.

Proof. The proof that K is measurable when Hτ (K

) = 0 is straightforward.
To prove the converse, note that since K is a candidate key, there is a
functional dependency K → K
, so by Theorem 7.5, K is not measurable,
unless Hτ (K
) = 0. ✷

Theorem 7.8 Every X ⊆ L is measurable under an outer measure

Hτ (·|K), where K is a candidate key.

Proof. If X = ∅, the proof is trivial. Suppose, that X = ∅. Note that all

tuples t for which πK (t) = k for some k ∈ Dom(K), are identical. Denote
by tk any tuple from the set {t ∈ ρ | t[K] = k}, where k ∈ Dom(K). Then
for every X ⊆ L, x1 ∈ Dom(X), x2 ∈ Dom(X
) and k ∈ Dom(K), the
following holds
#
0, if tk [X] = x1 ,
p(x1 |k) =
1, if tk [X] = x1 ,
#
0, if tk [X
] = x2 ,
p(x2 |k) =
1, if tk [X
] = x2 .

Therefore,
#
0, if tk [X] = x1 or tk [X
] = x2 ,
p(x1 , x2 |k) = = p(x1 |k)p(x2 |k),
1, if tk [X] = x1 and tk [X
] = x2 ,

and πX and πX
are independent. ✷

8 Entropy and Integrity Constraints

We believe that the study of integrity constraints can benefit from the ap-
plication of information-theoretical techniques. For instance, entropic de-
pendencies introduced below as generalizations of functional dependencies
can offer greater flexibility to database designers.

Definition 8.1 Let τ = (T, L, µ) be an m-table. We say that τ satisfies

an entropic dependency between the sets of attributes X and Y , denoted by
X →; Y , if Hτ (Y |X) ≤ R.
320 D. A. Simovici, S. Jaroszewicz

Theorem 8.1 Let τ = (T, L, µ) be an m-table and let X, Y, Z ⊆ L be sets

of attributes. We have:
1. X →; Y Z implies X →; Y ;
2. X →; Y implies XZ →; Y ;
3. X →; Y implies XZ →; Y Z;
4. X → Y if and only if X →0 Y ;
5. X → Y and Y →; Z imply X →; Z;
6. X →; Y and Y → Z imply X →; Z;
7. X →;1 Y and X →;2 Z imply X →;1 +;2 Y Z.

Proof. Statement 1 is a direct consequence of Part 2 of Theorem 4.3,

Statement 2 follows from the first part of Theorem 4.4; also, the third
statement is implied by the second part of Theorem 4.4, while Property 4
follows from Theorem 4.2.
To prove Statement 5 note that
Hτ (Z|X) = Hτ (XZ) − Hτ (X) = Hτ (XY Z) − Hτ (XY )
= Hτ (Z|XY ) ≤ Hτ (Z|Y ) ≤ R.
The last part of the theorem follows from
Hτ (Z|X) ≤ Hτ (Y Z|X) = Hτ (XY Z) − Hτ (X)
= Hτ (XY ) − Hτ (X) = Hτ (Y |X) ≤ R.
✷

Let σ = (T, L, ρ) be a table, an let X, Y be to nonempty subsets of

L. Denote by Z the set Z = L − XY . We say that there is a multivalued
dependency X → → Y between sets of attributes X and Y if for every tuples
t1 , t2 ∈ ρ, such that t1 [X] = t2 [X], there is a tuple t ∈ ρ such that t[X] =
t1 [X] = t2 [X], t[Y ] = t1 [y], and t[Z] = t2 [Z].
It is easy to see that the multivalued dependency X → → Y holds if and
only if for every x ∈ adomτ (X) we have
(σ where X = x)[Y Z] = (σ where X = x)[Y ] × (σ where X = x)[Z]
(8)
Multivalued dependencies can be expressed in terms of function inde-
pendence as follows.
On Information-Theoretical Aspects of Relational Databases 321

Theorem 8.2 Let σ = (T, L, ρ) be a table. If X, Y ⊆ L, denote by Z the

set of attributes Z = L − (X ∪ Y ). The multivalued dependency X → →Y
holds if and only if for every x ∈ adomτ (X), the functions πY |X=x and
πZ |X=x are independent.

Proof. Suppose X → → Y holds. Then, the equality (8) holds for every
x ∈ adomτ (X). Let n1 = |(σ where X = x)[Y ]| and n2 = |(σ where X =
x)[Z]|. Then,

1 1 1
p(y, z|x) = = · = p(y|x)p(z|x),
n1 · n1 n1 n2

for every y ∈ Dom(Y ), z ∈ Dom(Z).

To prove the converse take any y ∈ Dom(Y ) and z ∈ Dom(Z) such that
p(y|x) > 0 and p(z|x) > 0. From the independence property it follows that
p(y, z|x) = p(y|x)p(z|x) is also greater than zero. Thus regardless of the
choice of x, and y above, there exists a tuple t ∈ ρ such that t[X] = x,
t[Y ] = y, and t[Z] = z. ✷

We believe that the notion of entropic dependency merits further atten-

tion. It is clear, from Theorem 6.1, that any decomposition (XY, XZ) of a
table with few spurious tuples entails the existence of entropic dependencies
of the form X →; Y and X →;
Z with low R, R
. The reverse implication
remains as an open problem. Also, an axiomatization of entropic dependen-
cies (in the style of Armstrong’s axiomatization of functional dependencies)
would be very useful.

Acknowledgment
Szymon Jaroszewicz wishes to acknowledge the support received from the
Fulbright Scholarship Board as a visiting researcher at the University of
Massachusetts at Boston (grant no. IIE#:15984114-ICB).

References
[1] K. T. Cheng, V. Agrawal, An entropy measure of the complexity of
multi-output Boolean function, Proc. of the 27th Design Automation
Conference, 1990, 302–305.

[2] V. Cheushev, V. Shmerko, D. A. Simovici, S. Yanushkevich, Functional

entropy and decision trees, Proc. ISMVL’98, 1998, 257–262.
322 D. A. Simovici, S. Jaroszewicz

[3] P. A. Fejer, D. A. Simovici, Mathematical Foundations of Computer

Science, vol. I, Springer-Verlag, New York, 1991.

[4] C. H. Hwang and A. C. H. Wu, An entropy measure for power estimation

of Boolean function, Proc. of Asia and South Pacific Design Automation
Conference (ASP-DAC), 1997, 101–106.

[5] S. Jaroszewicz, D. A. Simovici, On axiomatization of conditional en-

tropy of functions between finite sets, Proc. of the 29th International
Symposium for Multiple-Valued Logic, Freiburg, May 1999, to appear.

[6] A. Lloris-Ruiz, J. F. Gomez-Lopera, R. Roman-Roldan, Entropic mini-

mization of multiple-valued logic functions, Proc. ISMVL’93, 1993, 24–
28.

[7] A. M. Mathai, P. N. Rathie, Basic Concepts in Information Theory and

Statistics – Axiomatic Foundations and Applications, Halsted Press,
John Wiley & Sons, 1975.

[8] D. A. Simovici, C. Reischer, On functional entropy, Proc. of the 23rd

International Symposium for Multiple-Valued Logic, 1993, 100–104.

[9] D. A. Simovici, R. L. Tenney, Relational Database Systems, Academic

Press, New York, 1995.
 A Version of Ω for which
ZFC Cannot Predict a Single Bit
Robert M. Solovay1
Department of Mathematics
The University of California
Berkeley, CA 94720 – 3840
E-mail: solovay@math.berkeley.edu

1 Introduction
In [2], Chaitin introduced the real Ω and proved the following:

Theorem 1.1 Assume that ZF C is arithmetically sound. (That is, any

theorem of arithmetic proved by ZF C is true.) Then ZF C can determine
the value of only finitely many bits of Ω. In fact we can explicitly compute
a bound on the number of bits of Ω which ZF C can determine.

Chaitin’s theorem is much more general than what we have stated.

ZF C can be replaced by any recursively axiomatizable theory in which
Peano arithmetic can be interpreted.
The real Ω depends on the choice of “universal Chaitin machine”. It
is natural to suspect that by tuning this choice one can improve Chaitin’s
result.
Here is the main theorem of this paper:

Theorem 1.2 We can choose the universal Chaitin computer U so that

ZF C (if arithmetically sound) can not determine any bit of the Ω associated
with U .

The rest of this paper is organized as follows. Section 2 contains a

review of the basic definitions of Chaitin’s theory that we use. In section
1
I wish to thank the Isaac Newton Institute for providing the environment where this
paper was written.

323
324 R. M. Solovay

3, we recall the notion of “1-consistent”. (The hypothesis that ZF C is

arithmetically sound can be sharpened, in both my theorem and Chaitin’s,
to merely asserting that ZF C is 1-consistent.) Section 4 gives a more
detailed discussion of Chaitin’s theorem. The remaining sections of the
paper are devoted to a proof of our theorem.
I am grateful to Greg Chaitin for giving me a copy of his book [1] which
started me thinking about this circle of ideas.

2 Preliminary Definitions and Notation

2.1 Bit Strings
ω is the set of non-negative integers. I follow von Neumann so that each
integer n is equal to the set {m ∈ ω | m < n} of integers less than n.
Σ∗ is the set of finite sequences of 0’s and 1’s. Thus an element of Σ∗
is just a finite function whose domain is in ω and whose range is included
in {0, 1}.
The concatenation of the two bit strings s and t will be denoted by
s T t. If j is one of 0 or 1, the bit string of length 1 whose sole component
is j will be denoted by (j). Of course ∅ is the unique string of length 0.
If s is a bit string, we write |s| for the length of s.
The usual theory of partial recursive functions is done considering func-
tions whose domain and range are subsets of ω. We want to import this
theory over to functions whose domain and range are subsets of Σ∗ and for
that, it is convenient to fix a canonical bijection between Σ∗ and ω. This
is done as follows:
We linearly order Σ∗ by putting s < t if either:
1. |s| < |t| or
2. |s| = |t| and s lexicographically precedes t.
With this ordering, there is a unique order isomorphism of Σ∗ with ω
(which will serve as the “canonical bijection” between the two sets.)

2.2 Prefix-Free Codes

A subset A of Σ is a prefix-free-code if whenever s and t are members of A
such that s ⊆ t then s = t.
Associated to any prefix-free code, A, is a real number ΩA defined by

ΩA = 2−|s| .
s∈A
A Version of Ω 325

This has the following probabilistic interpretation: Pick a real x in [0, 1]

at random using the Lebesgue measure on [0, 1]. Then ΩA is the probability
that some initial prefix of the binary expansion of x lies in A.

2.3 Chaitin Machines

A Chaitin machine is a partial recursive function whose domain and range
are subsets of Σ∗ and whose domain is a prefix-free code.
Let U be a Chaitin machine with domain A. Then we set ΩU = ΩA .
A Chaitin machine U is universal if it can simulate any other Chaitin
machine. More precisely, U is universal if for every Chaitin machine V
there is a bit string πV such that the equality

U (πV T s) V (s)

holds for any bit string s.

Here, as usual x y holds between two partially defined objects x and
y if (a) x is defined iff y is defined and (b) if they are both defined, then
they are equal.
It is proved in [2] that universal Chaitin machines exist. Moreover, if U
is universal, then ΩU has a strong randomness property. (It is now know
that ΩU is Martin-Löf random.) As a corollary, ΩU is irrational and does
not have a recursive binary expansion.

2.4 Gödel Numbering Chaitin Machines

We fix one of the usual Gödel numberings {ϕi | i ∈ ω} of all partial recursive
functions from Σ∗ to Σ∗ . Then the function Φ : ω × Σ∗ → Σ∗ given by

Φ(i, s) ϕi (s)

is partial recursive.
It follows that the domain of Φ is recursively enumerable. Since it
is clearly infinite, we fix a recursive enumeration without repetitions of
the domain of Φ: ( (ni , si ) | i ∈ ω). We can certainly arrange that this
enumeration is primitive recursive.
We are going to construct a new function Ψ : ω × Σ∗ → Σ∗ . Defining
ψi : Σ∗ → Σ∗ (for i ∈ ω) by

ψi (s) Ψ(i, s)

will yield the desired Gödel numbering of Chaitin machines.

326 R. M. Solovay

Ψ will be the restriction of Φ to a certain subset of its domain. We

proceed by induction on i to determine if the pair (ni , si ) will be placed in
the domain of Ψ.
Place (ni , si ) in the domain of Ψ iff for no j < i for which (nj , sj ) has
been placed in the domain of Ψ do we have nj = ni and sj compatible with
si . (That is, sj ⊆ si or si ⊆ sj .)
This construction has the following properties (whose proof is left to
the reader):

1. The domain of ψi is prefix-free.

2. If the domain of ϕi is prefix-free, then ψi = ϕi .

2.5 Timing
We let Dn be the domain of ψn . Let Ωn = ΩDn .
Intuitively, Dn [t] consists of those elements of Dn which have appeared
by time t. More precisely,

Dn [t] = {s | (∃j ≤ t)(nj = n and sj = s and (nj , sj )

was placed into Dom(Ψ) at stage j}.

We put Ωn [t] = ΩDn [t] . Intuitively, this is the approximation to Ωn

computable at time s.
The following facts are evident:

1. Ωn [t] is a rational number with denominator a power of 2.

2. Given n and t we can compute (by a primitive recursive function) the

finite set Dn [t] and the rational number Ωn [t].

3. As t increases to infinity, the Ωn [t] increase monotonically to the limit

Ωn .

3 1-Consistency
Throughout this section T is a theory with a recursive set of axioms in
which Peano Arithmetic (P A) is relatively interpretable. We fix a relative
interpretation of P A in T . Of course, the basic example we have in mind
is ZF C equipped with the usual relative interpretation of P A in ZF C.
For brevity in what follows, we say “interpretation” rather than “rela-
tive interpretation”.
A Version of Ω 327

Our main theorem will use the hypothesis that “ZF C is 1-consistent”.
In this first part of this section, we review known results without proof that
describe the relationship of the notion of 1-consistency to other notions of
soundness.
In the second part of this section, we derive from the assumption that
ZF C is 1-consistent that any determination that ZF C makes about one
of the binary digits of ΩU (for some universal Chaitin computer U ) is true.
This will be the only use we make of the 1-consistency hypothesis.

3.1 A Spectrum of Soundness Hypotheses

3.1.1 ω-Models
Since there is a fixed interpretation of P A in T , any model of T determines
a model of P A. We say that a model M of T is an ω-model if the associated
model of P A is isomorphic to the standard model of P A.
Our first soundness assumption is that T has an ω-model.

3.1.2 Arithmetic Soundness

Each sentence of the language of P A has a translation into a sentence of
the language of T , determined by the interpretation of P A in T . We shall
blur the distinction between a sentence of P A and its translation. We use
the phrase “sentence of arithmetic” to indicate a sentence of the language
of T that is the translation of some sentence of P A.
Our second soundness assumption is that T is arithmetically sound.
That is, if ϑ is a sentence of arithmetic which is a theorem of T , then ϑ is
true (in the standard model of P A).
Remark. Our metatheory is ZF C. So we know that P A itself is arith-
metically sound.

3.1.3 ω-Consistency
The notion of ω-consistency was introduced by Gödel in connection with
his incompleteness theorems.
It is easiest to define when a theory T is not ω-consistent (i.e., is ω-
inconsistent.) This happens if there is a formula of the language of T , θ(x)
(having only the indicated free variable x) such that the following happens:

1. T proves “There is an x ∈ ω such that θ(x).”

2. For each natural number n, T proves “¬θ(n)”.

328 R. M. Solovay

In theories like P A which have a canonical term to denote each natural

number, n is the canonical term that denotes the integer n. In theories like
ZF C that lack such terms the explication of what the formula θ(n) is, is
a little more subtle, but we presume the reader will be familiar with the
details.

3.1.4 1-Consistency
A theory T is 1-consistent, if whenever it proves a Σ01 sentence, θ, then θ
is true.
There is a notion of when a formula of P A is “primitive recursive”.
Basically these are the formulas that arise in implementing Gödel’s proof
that every primitive recursive predicate is expressible in P A.
A sentence of P A is Σ01 if it has the form (∃x)P (x) where P is primitive
recursive.
A special case of Σ01 sentences are the formalizations of assertions that
a particular Turing machine halts if started (in its canonical “start state”)
on an empty tape. Indeed, every Σ01 sentence is provably equivalent in P A
to such a “halting statement”.

3.1.5 Consistency
T is consistent if it does not prove the assertion “0 = 1”. Equivalently, T
is consistent if it does not prove every sentence in the language of T .

3.1.6 Positive Relations Between the Different Notions of

Soundness
These claims are all trivial: Every theory T that has an ω model is
arithmetically sound and ω-consistent. If T is arithmetically sound or ω-
consistent, then T is 1-consistent. If T is 1-consistent, then T is consistent.

3.1.7 Negative Relations Between the Different Notions of

Soundness
The claims that follow are not entirely trivial, but are all well-known. De-
tails will not be given. The proof of our main theorem does not depend on
these results.
There are theories T1 , T2 , T3 , and T4 , all in the language of ZF C and
all extending ZC (Zermelo set theory with choice) such that:
1. T1 is arithmetically sound but not ω-consistent.
A Version of Ω 329

2. T2 is ω-consistent but not arithmetically sound.

3. T3 is consistent, but not 1-consistent.
4. T4 is not consistent.
From now on, when we say that a theory T is 1-consistent, it is implied
that the theory has a recursive set of axioms and comes equipped with some
definite interpretation of P A in T .

3.2 Proving Facts About Ω

3.2.1 Binary Expansions
A dyadic rational is a rational number of the form r/2s where r and s are
integers and s ≥ 0.
If x is a real number which is not a dyadic rational, then x has a unique
binary expansion. If x is a dyadic rational, it has two distinct binary
expansions. In this paper, we shall always pick the one that ends in an
infinite sequence of 0’s.
With this convention in place, the following can easily be formalized
in P A: “The ith binary digit of Ωj is k.” (Here k < 2, of course, if the
assertion is true.)
We start numbering the digits of the binary expansion of a real with
the 0th digit. Thus the 0th digit of the binary expansion of 1/3 is 0; the 1st
digit is 1; the 2nd digit is 0, etc.

Lemma 3.1 Let ψj be a Chaitin machine which P A can prove universal.

Let T be 1-consistent, and let T prove the assertion “The ith binary digit
of Ωj is k”. Then this assertion is true.

Our proof of this lemma will proceed in two steps. We first show that
any Π02 sentence proved by T is true. We then show the sentence in question
in the lemma is provably equivalent in P A to a Π02 sentence.

3.2.2 Π02 Sentences

A Π02 sentence is a sentence of P A of the form ∀x∃yP (x, y), where P is
primitive recursive.
Suppose then, towards a contradiction, that T proves the translation of
such a Π02 sentence, and that the sentence is false. Then for some particular
integer n, the sentence ∃yP (n, y) is false and provable in T . But this latter
sentence is Σ01 , and this contradicts the assumption that T is 1-consistent.
330 R. M. Solovay

3.2.3 Proof of the Lemma

We work in P A. We know that Ωj is irrational. Hence, we can express the
fact that the ith digit of Ωj is k as follows:
(∀m)(∃nm) [the ith digit of Ωj [n] is k].
(The proof of this claim is easy and left to the reader.) But the assertion
just displayed is visibly Π02 . The lemma is proved.

4 Chaitin’s Results About Predicting Bits of Ω

The results of this section, which are stated without proof, are not needed
for the proof of the main theorem.
Throughout this section we fix a universal Chaitin machine U . (In the
later parts of this section, we even implicitly specialize to a particular such
U .)
We fix a j such that U = ψj . Formal assertions about U refer to this j.
We write Ω for ΩU . As discussed in Section 3.2.1, we can easily formalize
in P A the assertion “The ith binary digit of Ω is k”. (This formalization
uses the Gödel number, j, of Ω.)
Now let T be a 1-consistent theory of the sort discussed in Section 3.
Following Chaitin, we want to give an upper bound on the set of i ∈ ω such
that T proves a theorem of the form “The ith bit of the binary expansion
of Ω is k” for some k ≤ 1. We refer to this cardinality as “the number of
bits of Ω that T can determine”. Of course, a priori, this cardinality might
be infinite; however, it will turn out to be finite.

4.1 H(T )
We wish to give a definition of the number of bits it takes to describe the
arithmetical theorems of T .
We fix a Gödel numbering of the sentences of P A.
Now consider a theory T of the sort described above. We proceed to
associate an recursively enumerable set of strings WT to T . Let s ∈ Σ∗ .
Then s corresponds to an integer ns as discussed in Section 2.1. Then
s ∈ WT iff ns is the Gödel number of a sentence of P A whose translation
is a theorem of T .
Now let s ∈ Σ∗ . We say that s is a program for WT if
1. U (s) is defined and has the value t. Let n be the integer corresponding
to t. (Cf. Section 2.1.)
A Version of Ω 331

2. The domain of ϕn is WT .
Finally, let H(T ) be the length of the shortest program for WT .
We can now state Chaitin’s theorem (proved in [2]).

Theorem 4.1 Let U be a universal Chaitin computer. Then there is a

positive constant C (depending only on U ) such that (for T a 1-consistent
theory) T can determine at most C + H(T ) bits of Ω.

In [1], Chaitin describes a particular universal computer (whose imple-

mentation is done in a dialect of Lisp that Chaitin devised.) For a definition
of H(T ) which is similar in spirit to the one I have given above, Chaitin
proves the following:

Theorem 4.2 Let U be the particular universal Chaitin computer defined

in [1]. Let T be a 1-consistent theory. Then T determines at most H(T ) +
15328 bits of ΩU .

5 Precise Statement of the Main Theorem. Out-

line of the Proof
Theorem 5.1 Let T be a 1-consistent theory. Then there is a universal
Chaitin computer, U (equal to ψj ), such that:
1. P A proves the fact that U is universal.

2. T can not determine even a single bit of ΩU .

In particular, our theorem applies to ZF C provided that ZF C is 1-

consistent.
Of course, the U provided by the theorem depends on T .
Here is a sketch of the proof. (Some technical details have been omitted
from this sketch. They will be provided in the following sections where the
proof will be presented in detail.)
We fix a standard Chaitin universal computer V such that the univer-
sality of V is provable in P A.
Our computer U will be undefined on the string ∅. For strings of the
form (0) T s, we will have:

U ((0) T s) V (s)

This will ensure the universality claims made about U .

332 R. M. Solovay

We are still free to define U on strings of the form (1) T s as we wish.

We will use this freedom to prevent T from guessing a single bit of ΩU .
Thanks to the magic of the recursion theorem, we can assume that when
defining U we know the Gödel number of U . Our algorithm when given a
string of the form (1) T s first begins an enumeration of the arithmetical
theorems of T , looking for the first one of the form “The nth bit of ΩU is
k”. This search may go on forever without finding such a sentence. If it
does succeed, we note the particular values of n and k. If s does not have
length n, then U ((1) T s) is undefined.
Let r be the dyadic rational whose dyadic expansion begins with s T (k)
followed by an infinite string of 0’s. Let r
= r + 2−(n+1) . We search for a t
such that Ω[t] lies in the interval (r, r
). If we find such, we make U ((1) T s)
be defined with the value ∅.
It seems reasonable that the final value of Ω should be at least Ω[t] +
2−(n+1) since we have just added a new string of length n + 1 to the domain
of Ω. Thus the action we have just taken prevents Ω from being in the
interval (r, r
).
But clearly, if T has correctly predicted the value of the nth bit of Ω
then Ω will lie in an interval of the form (r, r
) for some length n bit string
s. Thus our assumption that T can predict a single bit of Ω has led to a
contradiction.
There are two points where we have to amplify the sketch to turn it
into a correct proof.

1. We must check that the self-reference in the sketch can indeed be

handled by the recursion theorem. (This is routine, but we shall
treat this carefully in the final proof.)

2. The phrase “It seems reasonable” probably could be turned into a

rigorous argument. But the detailed proof will proceed differently at
this point.

6 Description of the Construction

We will be defining a function U : Σ∗ → Σ∗ that depends on an integer
parameter j. (Intuitively, j is a guess at the Gödel number of U .) We will
specify the value of j presently.
U can be viewed as coming from a function U1 : ω × Σ∗ → Σ∗ . (So
U (s) U1 (j, s).) Our construction will be such that U1 is partial recursive.
A Version of Ω 333

As discussed in the sketch, we fix a universal Chaitin computer V such

that the universality of V is provable in P A.
We proceed to define U (s) by cases:
Case 1: s = ∅.
Then U (s) is undefined.
Case 2: s = (0) T t for some bit string t.
Then we set U (s) V (t).
Case 3: This is the case where s = (1) T t for some bit string t.
Our construction begins with an preliminary calculation to determine
certain constants n and k. This preliminary calculation may not converge.
In that case U will be undefined at s for any s falling under Case 3.
The preliminary calculation lists the theorems of T in some definite
order (not depending on t) searching for a theorem of the form “The nth
binary digit of Ωj is k”. If it finds such a theorem, then the value of n and
k for the rest of the construction are those given by the first such theorem.
We will only define U ((1) T t) if |t| = n.
Suppose then that |t| = n. We define dyadic rationals r and r
as follows.
r is the unique dyadic rational (in [0, 1)) whose binary expansion starts with
t T (k) and whose digits (after the nth one) are all 0. r
= r + 2−(n+1) .
We now proceed to search for the least integer m such that Ωj [m] lies
in the open interval (r, r
). Of course, this search might fail. If so, U (s) is
undefined.
Recall that Dj [m] is the finite set of strings in the domain of Ωj that
have contributed to the computation of Ωj [m]. (Cf. Section 2.5.) If s
appears in Dj [m], then U (s) is undefined. Otherwise, we set U (s) = ∅.
The recursion theorem assures us that there is a value of j such that
ϕj (s) U1 (j, s). We fix such a j and set U = ϕj . Thus in the definition of
U just given, the value of the parameter j was the Gödel number of U .

7 Analysis of the Construction

U is a Chaitin Machine
Suppose that s1 and s2 are two elements of the domain of U such that
s1 ⊆ s2 . We have to see that s1 = s2 .
Since U is undefined on the empty string, |s1 | ≥ 1. Let r = s1 (0). Let
si = (r) T ti . Clearly t1 ⊆ t2 . If r = 0, then t1 and t2 are in the domain of
the Chaitin computer V . Hence t1 = t2 . So s1 = s2 .
If r = 1, then for U (s1 ) and U (s2 ) to be defined, we must have the
integer n defined in the course of the construction. But then |s1 | = |s2 | =
334 R. M. Solovay

n + 1. So s1 = s2 as desired.
It follows that U = ψj and that the real Ωj used in the course of the
construction is ΩU . (Cf. Section 2.4.)

U is Universal
This follows from the definition of U on strings beginning with a 0. It is
also clear that U inherits from V the fact that its universality is provable
in P A.
Note that it follows that ΩU is irrational. (Cf. Section 2.3.)
Now, towards a contradiction, assume that T can determine some bit
of ΩU . Then in the course of the construction the integers n and k are
defined.
Let r be a dyadic rational with denominator 2n+1 such that r < ΩU <
r + 2−(n+1) . (We use here the fact that ΩU is irrational.) Let r
= r +
2−(n+1) .
Since T is 1-consistent, the assertion “The nth binary bit of ΩU is k”
is true. Hence the first n + 1 bits of the binary expansion of r have the
form t T (k) where t is a bit string of length n. For all sufficiently large
m, Ωj [m] will lie in the interval (r, r
).
Let s = (1) T t. Consider now the computation of U (s). The r and
the r
involved in that computation are the ones we have just defined. The
search for an m such that Ωj [m] ∈ (r.r
) will succeed.
Could it be that s ∈ Dj [m]? No, for then U (s) would not be defined.
But Dj [m] ⊆ Dj , so we would have s ∈ Dj . I.e., s would be in the domain
of U after all, a contradiction.
So U (s) is defined, and Dj contains in addition to the members of Dj [m]
the string s of length n + 1. It follows that ΩU ≥ r + 2−(n+1) = r
. But this
contradicts the definition of r. The proof of the main theorem is complete.

References
[1] G. J. Chaitin, The Limits of Mathematics, Springer-Verlag, Singapore,
1998.

[2] G. J. Chaitin, A theory of program size formally identical to informa-

tion theory, ACM Journal, 22 (1975), 329–340.
 On the Power of Reading
the Whole Infinite Input Tape
Ludwig Staiger
Martin-Luther-Universität Halle-Wittenberg
Institut für Informatik
Kurt-Mothes-Str. 1
D-06120 Halle (Saale), Germany
E-mail: staiger@informatik.uni-halle.de

1 Introduction
Finite devices accepting infinite strings are the topic of several papers (see
the recent surveys [4], [8] or [10], [11]). Most of these papers deal with
finite automata. Thus finite automata as devices accepting infinite strings
are well-understood. The situation is a little bit more involved if one con-
siders more complicated accepting devices like, e.g., pushdown automata
or Turing machines.
A thorough treatment of a general class of accepting devices, so-called
X-automata, on infinite words has been given in [4]. If one compares the
results on the acceptance of infinite strings by Turing machines obtained
in the papers [2], [3], [4] and [12], [9] one observes slight differences in
the manner how the machines under consideration take into account their
behaviour on the input tape.

Type 1: The approach of [12], [9] (cf. also [7], [8]) does not take into
consideration the behaviour of the Turing machine on its input tape.
Acceptance is based solely on the infinite sequence of internal states
the machine runs through during its infinite computation.
Thus the machine may base its decision on a finite part of the whole
infinite input.
Type 2: For X-automata Engelfriet and Hoogeboom [4] require that, in
addition to the fulfillment of certain conditions on the infinite se-
quence of internal states in order to accept an input, the machine has

335
336 L. Staiger

to read the whole infinite input tape. Cohen and Gold [1] considered
this same type by for pushdown automata.
Thus, besides blocking1 as for Type 1, machines have a further pos-
sibility to reject inputs.

Type 3: The most complicated type of acceptance for Turing machines

was introduced by Cohen and Gold [2], [3]. They require in addition
to Type 2 that the machine scans every cell of the input tape only
finitely many times. This behaviour is termed as having a complete
non-oscillating run.

The aim of our paper is to explore those sets of infinite words accepted
by deterministic Turing machines with respect to Type 2 and compare them
to the ones accepted with respect to Type 1 which are widely investigated
(cf. [2]2 , [12], [9], [7], [8]). Following the lines of [12], [9], [7] we provide also
characterizations of Type 2 Turing accepted ω-languages in terms of the
arithmetical hierarchy and automaton-free descriptions using recursively
enumerable languages and limit operations.
Utilizing these characterizations, we show that in case of deterministic
Turing machines Type 2-acceptance is strictly more powerful than Type 1-
acceptance. The stronger acceptance power of Type 2, however, contrasts
with two elegant properties of Type 1-acceptance:
First, the classes of accepted ω-languages coincide fully with classes of the
arithmetical hierarchy [12], [9] (cf. also [7], [8]), and secondly, the hierarchy
of classes accepted by nondeterministic Turing machines does not collapse.

2 Notation
We start with some necessary notation. By IN = {0, 1, 2, . . .} we denote
the set of natural numbers. We consider the space X ω of infinite strings
(sequences, ω-words) on a finite alphabet of cardinality r := card X ≥ 2.
By X ∗ we denote the set (monoid) of finite strings (words) on X, including
the empty word e. For w ∈ X ∗ and b ∈ X ∗ ∪ X ω let w · b be their
concatenation. This concatenation product extends in an obvious way to
1
Most of the six state acceptance conditions considered in Section 3.2 can simulate
blocking.
2
Although Cohen and Gold [2] started their consideration of Turing machines on
ω-words with considering the more powerful Type 2-acceptance when turning to the
deterministic case they weakened their model stipulating the so-called Property C (con-
tinuity property) which requires that the Turing machine under consideration a priori
has to read the whole input tape in a non-oscillating manner.
On the Power of Reading the Whole Infinite Input Tape 337

subsets W ⊆ X ∗ and B ⊆ X ∗ ∪ X ω . As usual we denote subsets of X ∗ as

languages and subsets of X ω as ω-languages.
Furthermore |w| is the length of the word w ∈ X ∗ , hence X n = {w : w ∈
X ∗ ∧ |w| = n}. By b[m..n] := b(m) · · · b(n) we denote the infix extending
from the mth to the nth letter3 of a string b ∈ X ∗ , |b| ≥ n, or b ∈ X ω , and

A(b) := {b[1..n] : n ∈ IN ∧ n ≤ |b|} and A(B) := b∈B A(b) are the sets of
all finite prefixes of b ∈ X ∗ ∪ X ω and B ⊆ X ∗ ∪ X ω , respectively.
As usual we define Π1 -definable ω-languages E ⊆ X ω as

E = {ξ : ∀n(n ∈ IN → ξ/n ∈ WE )} (1)

where WE ⊆ X ∗ is a recursive language, and we define Σ2 -definable ω-lan-

guages F ⊆ X ω as

F = {ξ : ∃i(i ∈ IN ∧ ∀n(n ∈ IN → (i, ξ/n) ∈ MF ))}, (2)

where MF is a recursive subset of IN × X ∗ . Σ1 -definable and Π2 -definable

ω-languages are defined accordingly. A language W ⊆ X ∗ is called Π2 -
definable iff

W = {w : w ∈ X ∗ ∧ ∀i(i ∈ IN → ∃n(n ∈ IN ∧ (i, n, w) ∈ MW ))},

for some recursive subset MW ⊆ IN × IN × X ∗ .

3 Turing Machines
In order to be in accordance with the X-automata of [4] we consider Tur-
ing machines M = (X, Γ, Z, z0 , R) with a separate input tape on which
the read-only-head moves only to right, n working tapes, X as its input
alphabet, Γ as its worktape alphabet, Z the finite set of internal states, z0
the initial state, and the relation

R ⊆ Z × X × Γn × Z × {0, +1} × (Γ × {−1, 0, +1})n

defining the next configuration.

Here (z, x0 , x1 , . . . , xn ; z
, y0 , y1 , . . . , yn ) ∈ R means that when M is in
state z ∈ Z, reads x0 ∈ X on its input tape and xi ∈ Γ on its worktapes
(i ∈ {1, . . . , n}), M changes its state to z
∈ Z, moves its head on the input
tape to the right if y0 = +1 or if y0 = 0 does not move the head, and for
(i ∈ {1, . . . , n}) and yi = (x
i , mi ) with x
i ∈ Γ and mi ∈ {−1, 0, +1} the
3
If m > n or m > |b| then b[m..n] := e, if n > |b| then b[m..n] := b[m..|b|] .
338 L. Staiger

machine M writes x
i instead of xi in its i-th worktape and moves the head
on this tape to the left, if mi = −1, to the right, if mi = +1, or does not
move it, if mi = 0.
Unless stated otherwise, in the sequel we shall assume that our accepting
devices be fully defined, i.e., the transition relation R is to contain for every
situation (z, x0 , x1 , . . . , xn ) in Z × X n+1 at least one (exactly one, if the
device is deterministic) move (z, x0 , x1 , . . . , xn ; z, y0 , y1 , . . . , yn ).
For a detailed description of configurations (instantaneous descriptions)
and the behaviour (computing process) of Turing machines the reader is
referred to the literature (e.g., [6]).
To avoid clumsy notation, we will describe the construction of ma-
chines and their behaviour only in an informal manner leaving details to
the reader.
Let the input of the Turing machine be some sequence ξ ∈ X ω . We
call a sequence ζ ∈ Z ω of states a run of M on ξ if ζ is the sequence of
states the Turing machine runs through in its (some of its, if the machine
is nondeterministic) computation with input ξ.

3.1 Acceptance Without State Conditions

In this section we are interested in the question which ω-languages can
be accepted by Turing machines when we disregard the internal behaviour
(states) and consider only the reading behaviour of the head on the input
tape.
We say that an input word w ∈ X ∗ is read in full length by a Turing
machine T provided w is the empty word e or if w = x1 · . . . · x then T
starting on the leftmost letter x1 moves its input head at least once onto
the last letter x .
Let T be a Turing machine. Define the function ϕ : X ∗ × IN → X ∗ in
the following way


 w[1..n], if |w| ≥ n and T reads at least the first
ϕ(w, n) := n letters of w, and

 undefined, otherwise.

Obviously, ϕ is a partial recursive function. Hence ϕ(X ∗ × IN) is a recur-

sively enumerable language.
Thus we have the following relationship between prefix-closed, that is,
W = A(W ), recursively enumerable languages and Turing machines read-
ing words in full length.
On the Power of Reading the Whole Infinite Input Tape 339

Lemma 3.1 A language W ⊆ X ∗ is prefix-closed and recursively enumer-

able iff there is a Turing machine T such that W is the set of all input
words which T reads in full length.

We say that an ω-word ξ ∈ X ω is lim-accepted by a Turing machine M if

an only if there is a run of M on input ξ such that M reads every symbol
of ξ, and we refer to an ω-language F ⊆ X ω as lim-accepted provided
there is a deterministic Turing machine M such that F = {ξ : ξ ∈ X ω ∧
ξ is lim-accepted by M}.
We obtain the following characterization of the of the class of lim-
accepted ω-languages.

Theorem 3.1 An ω-language F ⊆ X ω is lim-accepted iff there is a recur-

sively enumerable language W ⊆ X ∗ such that F = {ζ : A(ζ) ⊆ A(W )}.

Proof. Let F be lim-accepted by a Turing machine M. Then from the

above lemma we know that the language LM of all words w ∈ X ∗ which
are read in full length by M is recursively enumerable. Thus F = {ξ :
A(ξ) ⊆ A(LM )}.
Conversely, let F = {ξ : A(ξ) ⊆ A(W )} for some recursively enumer-
able language W . Then A(W ) is also recursively enumerable, and utilizing
a Turing machine T which enumerates the language A(W ) it is easy to
construct a Turing machine M which for an input word v ∈ X ∗ reads only
the longest prefix w @ v, w ∈ A(W ). From the behaviour of M it follows
that M lim-accepts F . ✷

In [7] it is shown that the condition derived in Theorem 3.1 has several
equivalent formulations in terms of the arithmetical hierarchy and topology.
To this aim we consider X ω as a topological space with the basis (w ·
X ω )w∈X ∗ . Since X is finite, this topological space is homeomorphic to the
Cantor discontinuum, hence compact. We mention here still that a subset
F in X ω is closed if and only if F = {ξ : A(ξ) ⊆ A(F )}. For more details
the interested reader is referred to [8], Section 2.

Lemma 3.2 ([7]) Let F ⊆ X ω . Then the following conditions are equiva-
lent.

1. There is a recursively enumerable language W

⊆ X ∗ such that F =
{ζ : A(ζ) ⊆ W
},

2. There is a recursively enumerable language W ⊆ X ∗ such that F =

{ζ : A(ζ) ⊆ A(W )},
340 L. Staiger

3. There is a recursive language V ⊆ X ∗ such that F = {ζ : A(ζ) ⊆

A(V )},

4. There is a Π2 -definable language U such that F = {ζ : A(ζ) ⊆ U },

and

5. F is closed in X ω and A(F ) is Π2 -definable.

In the sequel we will refer to the class of lim-accepted ω-languages as

P := {F : F ⊆ X ω ∧ F is closed ∧ A(F ) is Π2 -definable}.

We mention still some closure properties of the class P. To this end we

consider homomorphisms h : X → X. We extend the homomorphism h in
the usual way to words. For ω-words ξ ∈ X ω we define the extension of h,
h, in the following way:
Let X := {x : x ∈ X ∧ h(x) = e} be the set of letters erased by the
homomorphism h.
# ω
undefined, if ξ ∈ X ∗ · X ;
h(ξ) :=
limw→ξ h(w), otherwise.

Lemma 3.3 ([7]) The class P is closed under union, intersection, non-
erasing homomorphisms and their inverse mappings.

3.2 State-Acceptance Conditions for Turing Machines

In this part we define the six conditions imposed on the runs of Turing
machines in order to specify a set of input ω-words (cf. [2]4 , [4], [8], [9]).
We say that an input sequence ξ ∈ X ω is accepted by M according to
condition (mode) Ξ if there is a run ζ of M on ξ such that ζ satisfies Ξ. In
the sequel we shall consider the following conditions using the notation of
[4].
Let α : Z ω → 2Z be a mapping which assigns to every ω-word ζ ∈ Z ω
a subset Z
⊆ Z, and let R ⊆ 2Z × 2Z be a relation between subsets of Z.
We say that a pair (M, Z) where Z ⊆ 2Z accepts an ω-word ξ ∈ X ω if and
only if

∃Z
∃ζ(Z
∈ Z ∧ ζ is a run of M on ξ ∧ (α(ζ), Z
) ∈ R) .

Here we shall be concerned with the following mappings α and relations

R: For an ω-word ζ ∈ Z ω let ran(ζ) := {z : z ∈ Z ∧ ∃i(i ∈ IN \ {0} ∧ ζ(i) =
4
Cohen and Gold consider only the five conditions introduced by Landweber [5].
On the Power of Reading the Whole Infinite Input Tape 341

z)} be the range of ζ (considered as a mapping ζ : IN \ {0} → Z), that

is, the set of all letters occurring in ζ, and let inf (ζ) := {z : z ∈ Z ∧
ζ −1 (z) is infinite} be the infinity set of ζ, that is, the set of all letters
occurring infinitely often in ζ.
As relations R we shall use =, ⊆ and B where Z
B Z

:⇔ Z
∩ Z

= ∅.
We obtain the six types of acceptance presented in the following table.
For the sake of completeness we add a simple description and their corre-
spondence to the five types originally defined by Landweber [5] and used
in [2], [3].

(ran, B) 1-acceptance at least once

(ran, ⊆) 1
-acceptance everywhere
(ran, =)
(inf , B) 2-acceptance infinitely often
(inf , ⊆) 2
-acceptance almost everywhere
(inf , =) 3-acceptance

4 Classes of ω-Languages Accepted by Determin-

istic Turing Machines
In this section we consider the ω-languages defined by Turing machines
according to Types 1 and 2 defined above and the six accepting conditions
introduced in the previous section. First we give a brief review on the re-
sults concerning the six classes accepted according to Type 1 obtained in
[12], [9] or [7]. Then we derive a relationship between the Type 1- and
Type 2-classes based on the results of Section 3.1. Here it turns out that
the class of lim-accepted ω-languages plays a crucial role. Finally, we use
this relationship to obtain automata-free characterizations of the classes of
ω-languages accepted according to Type 2. These descriptions yield im-
mediately that Type 2-acceptance is, in some cases, strictly more powerful
than Type 1-acceptance, whereas the proper Type 2-classes neither do co-
incide with classes of the Arithmetical hierarchy, as the Type 1 classes do,
nor do they enjoy the same closure properties.

4.1 ω-Languages Accepted According to Type 1

We start with some characterizations in terms of recursive or recursively
enumerable languages of the low level classes of the Arithmetical hierarchy
342 L. Staiger

of ω-languages (cf. [7], [8]).

Lemma 4.1
1. A set E ⊆ X ω is in Σ1 -definable iff there is a recursive (or, equiv-
alently, a recursively enumerable) language W ⊆ X ∗ such that E =
W · Xω.
2. A set F ⊆ X ω is in Π1 -definable if and only if F is closed in X ω
and X ∗ \ A(F ) is recursively enumerable, that is, A(F ) ⊆ X ∗ is
Π1 -definable.
3. A set E ⊆ X ω is in Π2 -definable iff there is a recursive (or, equiv-
alently, a recursively enumerable) language W ⊆ X ∗ such that E =
{ξ : ξ ∈ X ω ∧ A(ξ) ∩ W is infinite}.

We mention still the following well-known closure properties of the classes

of the Arithmetical hierarchy, analogously to the ones in Lemma 3.3.
Lemma 4.2
1. Each one of the classes Σ1 , Π1 and Σ2 of Σ1 -definable, Π1 -definable,
or Σ2 -definable subsets of X ω , respectively, is closed under union, in-
tersection, non-erasing homomorphisms and inverse non-erasing ho-
momorphisms.
2. The class Π2 of Π2 -definable subsets of X ω is closed under union,
intersection, and inverse non-erasing homomorphisms, but not under
non-erasing homomorphisms.
The class of all ω-languages F ⊆ X ω Type 1-accepted by deterministic
(nondeterministic) Turing machines with respect to mode (α, R) is denoted
by DT(X) (α, R) (or NT(X) (α, R), respectively). The following characteri-
zation of the six DT-classes is shown in [12], [9] (cf. also [7], [8]). We start
with the simple modes DT(X) (α, ⊆) and DT(X) (α, B).
Theorem 4.1
DT(X) (ran, ⊆) = Π1 = {F : F ⊆ X ω ∧ F is Π1 -definable}
DT(X) (ran, B) = Σ1 = {F : F ⊆ X ω ∧ F is Σ1 -definable}
DT(X) (inf , ⊆) = Σ2 = {F : F ⊆ X ω ∧ F is Σ2 -definable}
DT(X) (inf , B) = Π2 = {F : F ⊆ X ω ∧ F is Π2 -definable}
Since Turing machines allow for a parallel composition, we obtain a char-
acterization for the classes DT(X) (α, =). Here B(S) denotes the closure of
ω
the set S ⊆ 2X under Boolean operations.
On the Power of Reading the Whole Infinite Input Tape 343

Corollary 4.1
DT(X) (ran, =) = B(Π1 ) = B(Σ1 ),
DT(X) (inf , =) = B(Σ2 ) = B(Π2 )

B(Σ∈ )
❅
❅
❅
❅
Σ2 Π2
✟✟
✟
✟
✟✟
✟
B(Σ∞ ) P
❅
❅ ❅
❅
❅ ❅
❅ ❅
Σ1 Π1 S

Figure 1: Inclusion relations between various classes of ω-languages ac-

cepted by Turing machines

Figure 1 shows the relations between the six classes of Theorem 4.1 and
Corollary 4.1, the class P and its related subclass
S := {F : F ⊆ X ω ∧ F is closed ∧ A(F ) is recursively enumerable} ⊂ P.
Example 1.15 of [8] shows that S is incomparable to the classes of Π1 -
definable, Σ1 -definable and Σ2 -definable subsets of X ω . Thus, the inclusion
properties presented in Figure 1 are proper and other inclusions than the
ones represented do not exist.

4.2 Composition and Decomposition Theorem

In this section we analyze the behaviour of Turing machines having the
additional possibility to reject by not reading the whole input utilizing
the class of lim-accepted ω-languages and the classes accepted by Turing
machines not regarding the moving of the input head.
We agree on the following notation: By T (M
, Z) we denote the
ω-language accepted by the pair (M, Z) according to Type 1, and by
344 L. Staiger

lim -T (M, Z) the ω-language accepted by the pair (M, Z) according to

Type 2, and lim -T (M) is the ω-language lim-accepted by M.
We obtain the following easily verified decomposition theorem.

Theorem 4.2 (Decomposition Theorem) Let M = (X, Z, Γ, z0 , f ) be a

deterministic Turing machine and let Z ⊆ 2Z . Then for α ∈ {ran, inf }
and R ∈ {⊆, B, =} we have lim -TRα (M, Z) = TRα (M, Z) ∩ lim -T (M).

In some sense we can also prove the converse of the Decomposition Theo-
rem.
Theorem 4.3 (Composition Theorem) Let M = (X, Z, Γ, z0 , f ), Z ⊆ 2Z ,
and M
= (X, S, Γ
, s0 , g) be deterministic Turing machines. Then there
is a deterministic Turing machine M

such that for α ∈ {ran, inf } and

R ∈ {⊆, B, =} it holds lim -TRα (M

, Zα ,R ) = TRα (M, Z) ∩ lim -T (M

) for
suitable Zα ,R .
Proof. We give an informal description of how to construct an appropriate
Turing machine M

. M

is a parallel composition of M and M

in the
following sense:
M

consists of one copy of M and M

each and an additional copying and
step counting control.
M

starts working as M
. As soon as M
reads a new letter from the
input tape this letter is copied to an internal tape serving as the input
tape of M. Then M

switches to M, resets the input head of M to the

situation it had before copying and then simulates the work of M until M
requires an input symbol not yet copied to its input tape or otherwise the
number of steps given by the step counting control is exhausted. This step
counting control ensures M

to switch to simulate M
even in case M does
not require further input letters. ✷

As an immediate consequence of Theorems 4.2 and 4.3 we obtain that

(X)
the classes DT (α, R) of ω-languages accepted by fulfilling the acceptance
condition (α, R) and reading the whole input can be described as follows.
Theorem 4.4
(X)
DT (ran, ⊆) = P = {F : F is closed ∧ A(F ) is Π2 -definable},
(X)
DT (ran, B) = {E ∩ F : E is Σ1 -definable ∧ F ∈ P},
(X)
DT (ran, =) = {E ∩ F : E ∈ DT(X) (ran, =) ∧ F ∈ P},
(X)
DT (inf , ⊆) = {E ∩ F : E is Σ2 -definable ∧ F ∈ P},
On the Power of Reading the Whole Infinite Input Tape 345

(X)
DT (inf , B) = DT(X) (inf , B) = {E : E is Π2 -definable},
(X)
DT (inf , =) = DT(X) (inf , =).

Proof. The first four identities are obvious, and the last two follow from
the fact that every ω-language f ∈ P is Π2 -definable, and the class of Π2 -
definable ω-languages is closed under intersection and union. ✷

From the results presented in Figure 1 we obtain the following inclusion

relations involving the new classes and the ones described above.
Lemma 4.3
(X) (X)
P ⊂ DT (ran, B) ⊂ DT (ran, =) ⊂ Π2 ,
(X)
Σ2 ⊂ DT (inf , ⊆) ⊂ B(Σ2 ).
(X) (X)
In contrast to Lemma 4.2 the new classes DT (ran, B), DT (ran, =)
(X)
and DT (inf , ⊆), however, are not closed under union, they are only
closed under intersection. We provide an example.
Example 4.1 Let X := {a, b, c}. Take an ω-language F1 ⊂ {a, b}ω , F1 ∈
P which is not Σ2 -definable and let F2 := {a, b}∗ · c · {a, b, c}ω which
is Σ2 -definable, even Σ1 -definable, but not closed in X ω . Then F1 ∈
(X)
P ⊆ DT (ran, B) and, since F2 is Σ1 -definable, we have also F2 ∈
(X) (X)
DT (ran, B). Assume now F1 ∪ F2 ∈ DT (inf , ⊆).
The smallest closed subset F ⊆ {a, b, c}ω containing F1 ∪ F2 is {a, b, c}ω
itself. Hence, F1 ∪ F2 = E ∩ F with E ∈ Σ2 and F closed in X ω implies
F = {a, b, c}ω , whence F1 ∪ F2 = E ∈ Σ2 . But this contradicts the fact that
(F1 ∪ F2 ) ∩ {a, b}ω = F1 ∈
/ Σ2 . ✷

5 Nondeterministic Turing Machines

The classes of ω-languages accepted by nondeterministic Turing machines
can be obtained from the deterministic classes via Projection lemmas for
Type 1-acceptance and via the Homomorphism lemma [4], Lemma 3.11.
Let X, Y be finite alphabets, and let pr 1 : X × Y → X be a projection.5
Lemma 5.1 (Projection lemma for Turing machines [12], [9])
F ∈ NT(X×Y ) (α, R) implies that pr 1 (F ) ∈ NT(X) (α, R).
5
It is essential here that the mapping pr 1 is a non-erasing homomorphism, whereas
in Lemma 5.2 the homomorphism may erase letters.
346 L. Staiger

Conversely, for every E ∈ NT(X) (α, R) there is an F ∈ DT(X×Y ) (α, R)

such that E = pr 1 (F ).
This lemma yields in an elegant way an immediate characterization of the
ω-languages Type 1 accepted by nondeterministic Turing machines. Be-
cause Lemma 4.2 shows that the classes DT(X) (ran, ⊆), DT(X) (ran, B)
and DT(X) (inf , ⊆) are closed under projections, according to the projec-
tion lemma they coincide with their nondeterministic counterparts.
The nondeterministic counterparts of the other classes are obtained as
follows.
Theorem 5.1

NT(X) (ran, =) = NT(X) (inf , ⊆) = DT(X) (inf , ⊆),

NT(X) (inf , B) = NT(X) (inf , =) = {F : F ⊆ X ω ∧ F is Σ11 -definable}.

Here, as usual, an ω-language is referred to as Σ11 -definable provided

F = {ξ : ∃η(η ∈ X ω ∧ ∀n∃m ((n, η[1..m], ξ[1..m]) ∈ MF ))} ,

for some recursive relation MF ⊆ IN × X ∗ × X ∗ . Observe that the class

of Σ11 -definable ω-languages equals the class of projections of Π2 -definable
ω-languages.
Turning to Type 2 acceptance we have the following general relationship
between the determninistic and nondeterministic classes.
Lemma 5.2 (Homomorphism lemma for Turing machines [4])
(X) (Y )
F ∈ NT (α, R) implies that h(F ) ∈ NT (α, R).
(Y ) (X)
Conversely, for every E ∈ NT (α, R) there is an F ∈ DT (α, R) such
that E = h(F ).
Since a Turing machine can simulate even a counter, Theorem 3.11 of [4]
(X)
proves that the hierarchy of classes NT (α, R) collapses. Using recursion-
theoretic arguments one obtains.
Theorem 5.2 For all α ∈ {ran, inf } and all R ∈ {⊆, B, =} the class
(X)
NT (α, R) coincides with the class of Σ11 -definable ω-languages over X.
Thus, similar to the case of projections, where the class of Σ11 -definable
ω-languages over X can be obtained from the class of of Π2 -definable ω-
languages over X via projections, all Σ11 -definable ω-languages over X are
homomorphic images of (closed) ω-languages in P.
We can prove even more.
On the Power of Reading the Whole Infinite Input Tape 347

Lemma 5.3 For every Σ11 -definable F ⊆ X ω there are a Π1 -definable E ⊆

Y ω and a homomorphism h : Y → X such that F = h(E).
Proof. Since every Σ11 -definable ω-language is the projection (image un-
der non-erasing homomorphism) of a Π2 -definable ω-language, it suffices
to prove that every Π2 -definable F ⊆ X ω is the image under arbitrary
homomorphism of a Π1 -definable E ⊆ Y ω .
To this aim let Y := X ∪ {d} where d is a letter not in X. We define
h : X ∪ {d} → X ∪ {e} to be the homomorphism erasing the letter d, that
is, !
y, if y ∈ X,
h(y) =
e, if y = d.
We use Lemma 4.1 to obtain e recursive language W ⊆ X ∗ such that
F = {ξ : ξ ∈ X ω ∧ A(ξ) ∩ W is infinite}.
Next we pad the language W by inserting strings of ds in such a way
that we can reconstruct from every padded word all (proper) prefixes of
the original word:
Define for w = w1 ·w2 · · · wn ∈ W where A(w)∩W \{e} = {w1 ·w2 · · · wi :
1 ≤ i ≤ n} its padding as follows.

π(w) := {d|w1 | · w1 · d|w2 | · w2 · · · d|wn | · wn

: wn
@ wn }, and let

π(W ) := π(w) .
w∈W

Observe that π(W ) is a recursive language provided W is recursive, and,

moreover, π(W ) = A(π(W )).
Let E := {η : A(η) ⊆ π(W )} = {η : A(η) ∩ π(W ) is infinite}.
By definition h(η) is defined if η contains infinitely many prefixes ending
with a letter of X. Thus, h(η), η ∈ E is defined iff η has infinitely many
prefixes of the form d|w1 | · w1 · d|w2 | · w2 · · · d|wn | · wn
where e = wn
∈ X ∗ .
From this we can conclude that h(η) has infinitely many prefixes in W ,
that is, h(η) ∈ F .
Conversely, let ξ ∈ F . Then ξ = w1 ·w2 · · · wi · · · where A(ξ)∩W \{e} =
{w1 ·w2 · · · wi : i ∈ IN\{0}}. Then η := d|w1 | ·w1 ·d|w2 | ·w2 · · · d|wi | ·wi · · · ∈ E
and h(η) = ξ. ✷

Reading the previous proof more carefully, we observe that, since π(W )
is recursive, E := {η : A(η) ⊆ π(W )} is not only Π1 -definable, but also in
S, and we obtain the following corollary showing the power of erasing.
Corollary 5.1 For every Σ11 -definable F ⊆ X ω there are an ω-language
E ⊆ Y ω , E ∈ Π1 ∩S and a homomorphism h : Y → X such that F = h(E).
348 L. Staiger

References
[1] R. S. Cohen, A. Y. Gold, Theory of ω-languages I: Characterizations
of ω-context-free languages, and II: A study of various models of ω-
type generation and recognition, J. Comput. System Sci., 15, 2 (1977),
169–184 and 185–208.

[2] R. S. Cohen, A. Y. Gold, ω-computations on Turing machines, Theoret.

Comput. Sci., 6 (1978), 1–23.

[3] R. S. Cohen, A. Y. Gold, On the complexity of ω-type Turing accep-

tors, Theoret. Comput. Sci., 10 (1980), 249–272.

[4] J. Engelfriet, H. J. Hoogeboom, X-automata on ω-words, Theoret.

Comput. Sci, 110 (1993) 1, 1–51.

[5] L. H. Landweber, Decision problems for ω-automata, Math. Syst. The-

ory, 3, 4 (1969), 376–384.

[6] H. Rogers, Theory of Recursive Functions and Effective Computability,

McGraw Hill, New York 1967.

[7] L. Staiger, Hierarchies of recursive ω-languages, J. Inform. Process.

Cybernetics, EIK, 22, 5/6 (1986), 219–241.

[8] L. Staiger, ω-languages, in vol. 3 of Handbook of Formal Languages,

(G. Rozenberg, A. Salomaa, eds.), Springer-Verlag, Berlin, 1997, 339–
387.

[9] L. Staiger and K. Wagner, Rekursive Folgenmengen I, Zeitschr. Math.

Logik u. Grundl. Mathematik, 24, 6 (1978), 523–538.

[10] W. Thomas, Automata on infinite objects, invol. B of Handbook of

Theoretical Computer Science (J. Van Leeuwen, ed.), Elsevier, Ams-
terdam, 1990, 133–191.

[11] W. Thomas, Languages, automata, and logic, in vol. 3 of Handbook of

Formal Languages (G. Rozenberg, A. Salomaa, eds.), Springer-Verlag,
Berlin, 1997, 389–455.

[12] K. Wagner, L. Staiger, Recursive ω-languages, in Fundamentals of

Computation Theory, (M. Karpiński, ed.), Lecture Notes in Computer
Science, 56, Springer-Verlag, Berlin 1977, 532–537.
 Finite and Infinite in Logic
Victor Vianu
Department of Computer Science and Engineering
U. C. San Diego
La Jolla, CA 92093-0114, USA
E-mail: vianu@cs.ucsd.edu

1 Introduction
Classical logic generally focuses on infinite structures, such as the inte-
gers. However, in many applications, including computer science, finite
structures are the primary focus. In this article we consider the following
question: what happens to logic when structures are restricted to be finite?
We survey several central results of finite model theory – the study of logic
on finite structures. We start with the observation that many results that
provide crucial tools in the general case, such as the Compactness theorem,
fail in the finite case. We then present several key results of finite model
theory that provide specific tools to replace some of the failed classical
theorems. We focus on Ehrenfeucht-Fraissé games and 0-1 laws.
These provide very effective tools for proving expressiveness results on
finite structures, and some of the most elegant results in finite-model theory.
Some computer science applications involve structures that are con-
ceptually infinite. For example, geographic information systems deal with
regions in space, which in principle consist of infinitely many points. In
such cases it is of interest to find finite summaries of the infinite structure
that provide just sufficient information to answer the queries relevant to the
application. We illustrate this approach with an example from geographic
information systems, where the relevant set queries consists of topological
properties, i.e., properties invariant under homemorphisms of the space.
We show that such properties can be summarized using a finite first-order
structure called a topological invariant.
While conventional wisdom suggests that only finite representations are
of interest in computer science, we conclude by presenting an alternative

349
350 V. Vianu

point of view: sometimes it makes more sense to model finite structures as

infinite. As a concrete and somewhat amusing example, we consider the
World Wide Web and show some advantages of modeling it as an infinite
rather than finite structure.

2 Basics
We present an informal review of first-order logic. A formal presentation
can be found in a standard logic textbook, such as [10].

2.1 Basic Logic Terminology

For simplicity, we only consider here relational structures, without explicit
functions. A relational vocabulary R is a finite set of relation symbols
with associated arities. We always assume that R contains the equality
symbol. A structure over a relational vocabulary R (also called a relational
structure) is a mapping I associating to each relation symbol R in R a
relation of the same arity as R (the symbol = is always interpreted as
equality). A relational structure is finite if all of its relations are finite.
The set of elements occurring in a structure I is called the domain of I,
denoted dom(I).
Relational first-order logic (FO) formulas are statements about rela-
tional structures. The statements are expressions formed using relation
symbols P, Q, R, . . ., variables x, y, z, . . ., constants from some infinite do-
main dom, and the symbols (, ), ∧, ∨, →, ¬, ∃, ∀. The simplest formula in
FO, called an atom, is of the form R(x1 , . . . , xk ). It states that the tuple
x1 , . . . , xk of variables or constants belongs to relation R. Other formu-
las are obtained recursively from atoms by using the logical connectors
∧, ∨, →, ¬, and quantifiers ∃, ∀, with the usual semantics.
For example, consider a vocabulary consisting of a single binary relation
G (the edges of a directed graph). The following sentence states that there
exists a path in G of length three from node a to node b:
(†) ∃x∃y(G(a, x) ∧ G(x, y) ∧ G(y, b)).
Classical logic has developed a powerful array of techniques for proving
things about first-order logic. Gödel’s Completeness Theorem provides a
sound and complete set of axioms for proving validity of FO sentences
(the property of being true in all structures, finite and infinite). As a
consequence, the set of valid FO sentences is recursively enumerable, that
is, can be enumerated by a Turing Machine.
Finite and Infinite in Logic 351

Another consequence provides a technical tool which is a household

name to all logicians: the Compactness Theorem, stating that a set Σ of
FO sentences is satisfiable iff every finite subset of Σ is satisfiable. As
an example of its effectiveness as a tool in understanding the expressive
power of FO, consider the question of whether graph connectivity is FO
definable. Here is a simple proof that it is not, using compactness. Suppose
there exists an FO sentence σ defining graph connectivity. For each i > 0
consider an FO sentence σi stating that the distance between two nodes a, b
in the graph is more than i (for example, σ3 is the negation of the sentence
(†) above). Let Σ = {σi | i > 0} ∪ {σ}. Clearly, every finite subset of
Σ is satisfiable. However, there is no graph satisfying Σ: in such a graph
there would be no path of finite length from a to b, but a and b would be
connected. The contradiction shows that no such σ can exist.

2.2 The Impact of Finiteness

Restricting structures to be finite has a dramatic impact on the landscape
of classical logic. We mention two major examples.
First, Gödel’s Completeness theorem fails in a very strong sense: not
only is the set of validity axioms no longer complete, but there can be
no sound and complete inference mechanism for proving validity on finite
structures. This follows from another remarkable fact: the set of FO sen-
tences valid on finite structures is not recursively enumerable. For suppose
it is. Clearly, the set of invalid sentences is also recursively enumerable
(an algorithm generating the invalid sentences would alternatingly gener-
ate sentences and finite structures, and would output a sentence when a
finite structure which does not satisfy the sentence is found). But then the
set of valid sentences has to be recursive. However, this is known to be false
– it is undecidable if a sentence is valid on finite structures [8]. This shows
that the set of FO sentences valid on finite structures is not recursively
enumerable.
Perhaps as dramatically, the Compactness theorem also fails on finite
structures. This is easily seen: consider a vocabulary consisting of a unary
relation R (a set) and for each i > 0, let σi be an FO sentence stating that
R has at least i elements. Let Σ = {σi | i > 0}. Clearly, every finite subset
of Σ is satisfiable by a finite structure. However, there is no finite structure
satisfying Σ.
The failure of the Compactness theorem deprives logicians of a crucial
tool. For example, how might one prove that connectivity of finite graphs is
not definable in FO? We next consider two important results that provide
352 V. Vianu

elegant, effective tools for answering such expressiveness questions about

FO on finite structures.

3 Ehrenfeucht-Fraissé Games
This section presents a technique based on a two-person “game”, due to
Ehrenfeucht [9] and Fraissé [12], that can be used to prove that certain
properties, including evenness1 and connectivity, are not FO definable.
While the game we describe is geared towards the first-order queries, games
provide a general technique that is used in conjunction with many other
languages.
The connection between FO sentences and games is, intuitively, the
following. Consider as an example a FO sentence of the form

∀x1 ∃x2 ∀x3 ψ(x1 , x2 , x3 ).

One can view the sentence as a statement about a game with two players,
Spoiler and Duplicator, who alternate in picking values for x1 , x2 , x3 . The
sentence says that Duplicator can always force a choice of values that makes
ψ(x1 , x2 , x3 ) true. In other words, no matter which value Spoiler chooses
for x1 , Duplicator can pick an x2 such that, no matter which x3 is chosen
next by Spoiler, ψ(x1 , x2 , x3 ) is true.
The actual game we use, called the Ehrenfeucht-Fraissé game, is slightly
more involved, but is based on a similar intuition. It is played on two
structures. Suppose that R is a relational vocabulary. Let I and J be
structures over R, with disjoint sets of elements. Let r be a positive integer.
The game of length r associated with I and J is played by two players,
Spoiler and Duplicator, making r choices each. Spoiler starts by picking an
element occurring in I or J and Duplicator picks an element in the opposite
structure. This is repeated r times. At each move, Spoiler has the choice
of the structure and an element in it, and Duplicator must respond in the
opposite structure.
Let ai be the ith elements picked in I (respectively, bi in J). The set of
pairs
{(a1 , b1 ), ..., (ar , br )}
is a round of the game. The substructure of I generated by {a1 , ..., ar },
denoted I/{a1 , ..., ar } consists of all tuples in I using only these elements,
and similarly for J, {b1 , ..., br } and J/{b1 , ..., br }.
1
Evenness is the property that a finite set has even cardinality.
Finite and Infinite in Logic 353

Duplicator wins the round {(a1 , b1 ), ..., (ar , br )} iff the mapping ai → bi

is an isomorphism of the substructures I/{a1 , ..., ar } and J/{b1 , ..., br }.
Duplicator wins the game of length r associated with I and J if he has a
winning strategy, i.e., Duplicator can always win any game of length r on I
and J, no matter how Spoiler plays. This is denoted by I ≡r J. Note that
the relation ≡r is an equivalence relation on structures over R. Intuitively,
the equivalence I ≡r J says that I and J cannot be distinguished by looking
at just r elements at a time in the two structures. We will need the notion
of quantifier depth of a FO formula: the maximum number of quantifiers
in a path from the root to a leaf in the representation of the sentence as a
tree. The main result concerning Ehrenfeucht-Fraissé games states that the
ability to distinguish among structures using games of length r is equivalent
to the ability to distinguish among structures using some FO sentence of
quantifier depth r.

Example Consider the sentence ∀x (∃y R(x, y) ∧ ∃z P (x, z)). Its syntax
tree is represented in Figure 1. The sentence has quantifier depth 2. Note
that, for a sentence in prenex normal form, the quantifier depth is simply
the number of quantifiers in the formula.

∀x

∧
✑ ❜
❜
✑ ❜
✑ ❜
∃y ∃z

R(x, y) P (x, y)

Figure 1: A syntax tree

The main result about Ehrenfeucht-Fraissé games, stated in Theorem

3.1 below, is that if I and J are two structures such that Duplicator has a
winning strategy for the game of length r on the two structures, then I and
J cannot be distinguished by any FO sentence of quantifier depth r. Before
proving this theorem, we note that the converse of that result also holds.
Thus, if two structures are undistinguishable using sentences of quantifier
354 V. Vianu

depth r, then they are equivalent with respect to ≡r . Although interesting,

this is of less use as a tool for proving expressibility results, and we leave it
as a (non-trivial!) exercise. The main idea is to show that each equivalence
class of ≡r is definable by a sentence of quantifier depth r.

Theorem 3.1 Let I and J be two finite structures over relational vocabu-
lary R. If I ≡r J then for each FO sentence ϕ over R with quantifier depth
r, I and J both satisfy ϕ or neither does.

Proof. Suppose that I |= ϕ and J |= ϕ for some ϕ of quantifier depth r.

We prove that I ≡r J. We provide here only a sketch of the proof on an
example.
Let ϕ be the sentence: ∀x1 ∃x2 ∀x3 ψ(x1 , x2 , x3 ), where ψ has no
quantifiers, and let I and J be two structures such that: I |= ϕ, J |= ϕ.
Then

I |= ∀x1 ∃x2 ∀x3 ψ(x1 , x2 , x3 ) and J |= ∃x1 ∀x2 ∃x3 ¬ψ(x1 , x2 , x3 ).

We will show that Spoiler can prevent Duplicator from winning by forc-
ing the choice of elements a1 , a2 , a3 in I and b1 , b2 , b3 in J such that
I |= ψ(a1 , a2 , a3 ) and J |= ¬ψ(b1 , b2 , b3 ). Then the mapping ai → bi cannot
be an isomorphism of the substructures I/{a1 , a2 , a3 } and J/{b1 , b2 , b3 },
contradicting the assumption that Duplicator has a winning strategy. To
force this choice, Spoiler always picks “witnesses” corresponding to the ex-
istential quantifiers in ϕ and ¬ϕ (note that the quantifier for each variable
is either ∀ in ϕ and ∃ in ¬ϕ, or vice versa).
Spoiler starts by picking an element b1 in J such that

J |= ∀x2 ∃x3 ¬ψ(b1 , x2 , x3 ).

Duplicator must respond by picking an element a1 in I. Due to the universal

quantification in ϕ,
I |= ∃x2 ∀x3 ψ(a1 , x2 , x3 )
regardless of which a1 was picked. Next, Spoiler picks an element a2 in I
such that
I |= ∀x3 ψ(a1 , a2 , x3 ).
Regardless of which element b2 in J Duplicator picks,

J |= ∃x3 ¬ψ(b1 , b2 , x3 ).
Finite and Infinite in Logic 355

Finally, Spoiler picks b3 in J such that J |= ¬ψ(b1 , b2 , b3 ); Duplicator picks

some a3 in I, and I |= ψ(a1 , a2 , a3 ). ✷

Theorem 3.1 provides an important tool for proving that certain prop-
erties are not definable by FO. Indeed, it is sufficient to exhibit, for each
r, two structures Ir and Jr such that Ir has the property, Jr does not, and
Ir ≡r Jr . In the next proposition, we illustrate the use of this technique by
showing that finite graph connectivity is not expressible in FO.

Proposition 3.1 Let R be a relational vocabulary consisting of one binary

relation. Then the query conn defined by:

conn(I) = true iff I is a connected graph,

is not definable in FO.

a1 a3
✗❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ✔
B
✖❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ✕
a2
✛ ✘
B1 b1 b3
✚❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ✙

✛❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ❤ ✘
B2 b2
✚ ✙
Figure 2: Two undistinguishable graphs

Proof. Suppose that there is a FO sentence ϕ defining graph connectivity.

Let r be the quantifier depth of ϕ. We exhibit a connected graph Ir and a
disconnected graph Jr such that Ir ≡r Jr . Then, by Theorem 3.1, the two
structures satisfy ϕ or none does, a contradiction.
For a sufficiently large n (depending only on r) the graph Ir consists
of a cycle B of 2n nodes and the graph Jr of two disjoint cycles B1 and
B2 of n nodes each (see Figure 2). We outline the winning strategy for
Duplicator. The main idea is simple: two nodes a, a
in Ir that are far
apart “behave” in the same way as two nodes b, b
in Jr that belong to
different cycles. In particular, Spoiler cannot take advantage of the fact
that a, a
are connected but b, b
are not. To do so, Spoiler would have
356 V. Vianu

to exhibit a path connecting a to a

, which Duplicator could not do for b
and b
. However, Spoiler cannot construct such a path because it requires
choosing more than r nodes.
To illustrate, if Spoiler picks an element a1 in Ir , then Duplicator picks
an arbitrary element b1 , say in B1 . Now, if Spoiler picks an element b2 in
B2 , then Duplicator picks an element a2 in Ir far from a1 . Next, if Spoiler
picks a b3 in B1 close to b1 , then Duplicator picks an element a3 in Ir close
to a1 . The graphs are sufficiently large that this can proceed for r moves
with the resulting subgraphs isomorphic. The full proof requires a complete
case analysis on the moves that Spoiler can make. ✷

The above technique can be used to show that many other properties
are not expressible in FO, for instance evenness, 2-colorability of graphs, or
Eulerian graphs (i.e., graphs for which there is a cycle that passes through
each edge exactly once).

4 0-1 Laws
We now present a powerful tool that provides a uniform approach to resolv-
ing in the negative a large spectrum of expressibility problems. It is based
on the probability that a property is true in structures of a given size. This
study, initiated by Fagin [11] and Glebskiĭ [13], shows a very surprising
fact: all properties of finite structures definable in FO are “almost surely”
true, or “almost surely” false.
Let σ be a sentence over some relational vocabulary R. For each n,
let µn (σ) denote the fraction of finite structures over R with entries in
{1, ..., n} that satisfy σ, i.e.,

|{I | σ(I) = true and dom(I) = {1, ..., n}}|

µn (σ) = .
|{I | dom(I) = {1, ..., n}}|

Definition 4.1 A sentence σ is almost surely true (respectively, false) if

limn→∞ µn (σ) exists and equals 1 (respectively, 0). If every sentence in a
language L is almost surely true or almost surely false, the language L has
a 0-1 law.

To simplify the discussion of 0-1 laws, we continue to focus on constant-

free sentences.
We will show that FO has a 0-1 law. This provides substantial insight
into limitations in the expressive power of FO. It can be used to show that it
Finite and Infinite in Logic 357

cannot express a variety of properties. For example, it follows immediately

that evenness is not expressible in FO. Indeed, µn (evenness) is 1 if n is
even and 0 if n is odd. Thus, µn (evenness) does not converge, so evenness
is not expressible in any language that has a 0-1 law.
While 0-1 laws provide an elegant and powerful tool, they require the
development of some non-trivial machinery. Interestingly, this is one of
the rare occasions when it is needed to consider infinite structures while
proving something about finite structures!
For simplicity, we consider only the case of vocabularies consisting of
a binary relation G (representing edges in a directed graph with no edges
of the form (a, a)). It is straightforward to generalize the development to
arbitrary vocabularies.
We will use an infinite set A of FO sentences called extension axioms,
that refer to graphs. They say, intuitively, that every subgraph can be
extended by one node in all possible ways. More precisely, A contains, for
each k, all sentences of the form
. .
∀x1 ...∀xk (( (xi = xj )) ⇒ ∃y( (xi = y) ∧ connections(x1 , ..., xk ; y))),
i=j i

where connections(x1 , ..., xk ; y) is some conjunction of literals containing,

for each xi , one of G(xi , y) or ¬G(xi , y), and one of G(y, xi ) or ¬G(y, xi ).
For example, for k = 3, one of the 26 extension axioms is:
∀x1 , x2 , x3 ( (x1 = x2 ∧ x2 = x3 ∧ x3 = x1 ) ⇒
∃y (x1 = y ∧ x2 = y ∧ x3 = y∧
G(x1 , y) ∧ ¬G(y, x1 ) ∧ ¬G(x2 , y) ∧ ¬G(y, x2 ) ∧ G(x3 , y) ∧ G(y, x3 )))
specifying the pattern of connections represented in Figure 3.
A graph G satisfies this particular extension axiom if for each triple
x1 , x2 , x3 of distinct vertices in G, there exists a vertex y connected to
x1 , x2 , x3 as shown in Figure 3.
Note that A consists of an infinite set of sentences, and that each finite
subset of A is satisfied by some infinite structure. (The structure is obtained
by starting from one node and repeatedly adding nodes required by the
extension axioms in the subset.) Then, by the Compactness theorem there
is an infinite structure satisfying all of A, and by the Löwenheim-Skolem
theorem (see [10]) there is a countably infinite structure R satisfying A.
The following lemma shows that R is unique up to isomorphism.

Lemma 4.1 If R and P are two countably infinite structures over G sat-
isfying all sentences in A, then R and P are isomorphic.
358 V. Vianu

Proof. Suppose that a1 a2 ... is an enumeration of all elements in R, and

b1 b2 ... an enumeration of those in P. We construct an isomorphism be-
tween R and P by alternatingly picking elements from R and from P. We
construct sequences ai1 ...aik ... and bi1 ...bik ... such that aik → bik is an
isomorphism from R to P. The procedure for picking the k th elements aik
and bik in these sequences is defined inductively as follows. For the base
case, let ai1 = a1 and bi1 = b1 . Suppose that sequences ai1 ...aik and bi1 ...bik
have been defined. If k is even, let aik+1 be the first element in a1 , a2 , ...
which does not occur so far in the sequence. Let σk be the sentence in A
describing the way aik+1 extends the subgraph with nodes ai1 ...aik . Since P
also satisfies σk , there exists an element b in P which extends the subgraph
bi1 ...bik in the same manner. Let bik+1 = b. If k is odd, the procedure
is reversed, i.e., it starts by choosing first a new element from b1 , b2 , ...
This “back and forth” procedure ensures that (i) all elements from both R
and P occur eventually among the chosen elements, and (ii) the mapping
aik → bik is an isomorphism. ✷

y
❤
✒ ❈❖◗◗
❈ ◗
 ❈
 ❈
❤x ❤x  ❤x
1 2 3

Figure 3: A connection pattern

Thus, the above shows that there exists a unique (up to isomorphism)
countable graph R satisfying A. This graph, studied extensively by Rado
[21] and others, is usually referred to as the Rado graph. We can now prove
the following crucial lemma. The key point is the equivalence between (a)
and (c), called the transfer property: it relates satisfaction of a sentence by
the Rado graph with the property of being almost surely true.

Lemma 4.2 Let R be the Rado graph and σ a FO sentence. The following
are equivalent:

(a) R satisfies σ;

(b) A implies σ; and

(c) σ is almost surely true.

Finite and Infinite in Logic 359

Proof. (a) ⇒ (b) : Suppose (a) holds but (b) does not. Then there exists
some structure P satisfying A but not σ. Since P satisfies A, P must
be infinite. By the Lowënheim-Skolem theorem we can assume that P is
countable. But then, by Lemma 4.1, P is isomorphic to R. This is a
contradiction, since R satisfies σ but P does not.
(b) ⇒ (c): It is sufficient to show that each sentence in A is almost surely
true. For suppose this is the case and A implies σ. By the Compactness
theorem, σ is implied by some finite subset A
of A. Since every sentence in
/
A
is almost surely true, the conjunction A
of these sentences is almost
/
surely true. Since σ is true in every structure where A
is true, µn (σ) ≥
/

µn ( A ), so µn (σ) converges to 1 and σ is almost surely true.

It remains to show that each sentence in A is almost surely true. Con-
sider the following sentence σk in A:
. .
∀x1 ...∀xk (( (xi = xj )) → ∃y( (xi = y) ∧ connections(x1 , ..., xk ; y))).
i=j i

Then ¬σk is the sentence

/
∃x1 ...∃xk ( ( i=j (xi = xj )) ∧
/
∀y( i (xi = y) → ¬connections(x1 , ..., xk ; y))).

We will show the following property on the probability that a structure

with n elements does not satisfy σk :
1 (n−k)
(†) µn (¬σk ) ≤ n · (n − 1) · ... · (n − k) · (1 − ) .
22k
Since limn→∞ [n · (n − 1) · ... · (n − k) · (1 − 212k )(n−k) ] = 0, it follows that
limn→∞ µn (¬σk ) = 0, so ¬σk is almost surely false, and σk is almost surely
true.
Let N be the number of structures with elements in {1, ..., n}. To prove
(†), observe the following:
1. For some fixed distinct a1 , ..., ak , b in {1, . . . , n}, the number of I sat-
isfying some fixed literal in connections(a1 , . . . , ak ; b) is 12 · N .

2. For some fixed distinct a1 , ..., ak , b in {1, . . . , n}, the number of I sat-
isfying connections(a1 , . . . , ak ; b) is 212k · N (since there are 2k literals
in connections).

3. The number of I not satisfying connections(a1 , . . . , ak ; b) is therefore

N − 212k · N = (1 − 212k ) · N .
360 V. Vianu

4. For some fixed a1 , ..., ak in {1, . . . , n}, the number of I satisfying

.
∀y( (ai = y) → ¬connections(a1 , ..., ak ; y))
i

is (1 − 212k )n−k · N (since there are (n − k) ways of picking b distinct

from a1 , ..., ak ).

5. the number of I satisfying ¬σk is thus at most

1 (n−k)
n · (n − 1) · ... · (n − k) · (1 − ) ·N
22k
(from the choices of a1 , ..., ak ). Hence (†) is proven.
(c) ⇒ (a): Suppose that R does not satisfy σ, i.e., R |= ¬σ. Since
(a) ⇒ (c), ¬σ is almost surely true. Then σ cannot be almost surely true,
a contradiction. ✷

The 0-1 law for FO now follows immediately.

Theorem 4.1 Each sentence in FO is almost surely true or almost surely

false.

Proof. Let σ be a FO sentence. The Rado graph R satisfies either σ or ¬σ.

By the transfer property ((a) ⇒ (c) in Lemma 4.2), σ is almost surely true
or ¬σ is almost surely true. Thus, σ is almost surely true or almost surely
false. ✷

The 0-1 law for FO also holds for recursive extensions of FO such as
fixpoint [5], [23] and while [17]. The proof uses once again the Rado graph
and extends the transfer property to the while sentences. It follows imme-
diately that many queries, including evenness, are not while sentences. The
technique of 0-1 laws has been extended successfully to languages beyond
while. Many languages which do not have 0-1 laws are also known, such as
existential second-order logic. The precise “border” separating languages
that have 0-1 laws from those that do not has yet to be determined, and
remains an interesting and active area of research.

5 Turning the Infinite Into Finite

Most computer science applications involve structures which are finite. In
some cases, however, the structures of interest are conceptually infinite.
Finite and Infinite in Logic 361

For example, geographic information systems deal with regions in space,

which in principle consist of infinitely many points. In such cases it is of
interest to find finite summaries of the infinite structure that provide just
sufficient information to answer the queries relevant to the application.
We exemplify this approach with geographic information systems when the
target set queries consists of topological properties, i.e., properties invariant
under homeomorphisms of the space.
We use a spatial model that speaks about regions in the two-dimensional
plane. Regions are specified by inequalities involving polynomials with
rational coefficients (such regions are called semi-algebraic). Topological
properties of regions are those that are invariant under homeomorphisms
of the plane. This means, intuitively, that continuous deformations and
reflections of the spatial structure do not affect satisfaction of the prop-
erty. For example, the property “the intersection of regions P and Q is
connected” is a topological property. On the other hand, the property “P
lies above Q” is not topological. In this section we present a result of [20]
showing that topological properties of semi-algebraic regions in a spatial
database can be completely summarized by a finite first-order structure,
called the topological invariant of the database. Moreover, the topological
invariant of a semi-algebraic database can be constructed very efficiently –
in NC (polylogarithmic time using polynomially many processors).
We use the following model for spatial databases. We assume given
an infinite set names (consisting of names of regions). A spatial database
vocabulary is a finite subset Reg of names. A structure I over a vocabulary
Reg is a mapping from Reg to subsets of R2 . For each r ∈ Reg, I(r)
provides a set of points called the extent of r. We generally refer to a set of
points in the plane as a region. In practice, each I(r) is finitely specified,
although this may be transparent to the user. All regions considered here
are compact, and specified by a disjunction of conjunctions of polynomial
inequalities with rational coefficients (such regions are usually referred to
as semi-algebraic). A region is connected if its boundary is connected. In
the following, the term region will be used in the restricted manner just
described, unless otherwise specified.
As discussed earlier, we are interested here in topological properties of
spatial structures. Two structures I, J over a spatial vocabulary Reg are
topologically equivalent iff there exists a bijection λ : R2 → R2 such that
both λ and λ−1 are continuous and for each r ∈ Reg, λ(I(r)) = J(r).
A property τ of spatial structures is topological if it is invariant under
homeomorphisms, that is for all topologically equivalent structures I, J
362 V. Vianu

over a given vocabulary, I satisfies τ iff J satisfies τ .

Topological invariants. In [20] it was shown that one can efficiently
compute from a given semi-algebraic spatial structure I a finite relational
structure top(I) called the topological invariant of I, that describes com-
pletely the topological properties of I. We briefly describe the construction
of the invariant and the results.
The invariant is constructed from a maximum topological cell decom-
position of the spatial structure. A topological cell decomposition of I is a
partition of R2 into finitely many subsets called cells, such that for every
homeomorphism λ of R2 , if λ(I) = I then λ(c) is a cell for every cell c.
It is easily verified that for each spatial structure I there exists a unique
maximal (in terms of number of cells) topological cell decomposition. The
maximum topological cell decomposition can be constructed from a semi-
algebraic spatial structure in NC, using results on cell complexes obtained
in [4], [18]. We summarize their approach, slightly adapted to our context.
Given a semi-algebraic spatial structure I over a vocabulary Reg, a sign
assignment is a mapping σ : Reg → {o, −, ∂}, and the sign class of σ is the
0
set I σ = r∈Reg r
σ(r) , where r o is the interior of r, r ∂ is the boundary,

and r− is the exterior. A cell complex for I is a partition of R2 into finitely

many, non-empty, pairwise disjoint regions, called cells, such that:

1. each cell is homeomorphic to R0 , R1 or R2 − {a finite set of points}.

The dimension of a cell is defined in the obvious manner.

2. the closure of each cell is a union of other cells;

3. each cell is included in some sign class I σ .

It is shown in [18] that a cell complex can be constructed from a given semi-
algebraic spatial structure in NC. One can further show that the maximum
topological cell decomposition can be constructed from the cell complex
obtained in [18], and the overall complexity remains NC.
The topological invariant for a spatial structure I is build up from the
maximum topological cell decomposition for I. Cells of dimension 0,1 and 2
are called vertices, edges, and faces, respectively. The topological invariant
associated to spatial structures over a vocabulary Reg is a finite structure
consisting of the following relations (their meaning is explained intuitively):

1. Unary relations Vertex, Edge, Face, and Exterior-face providing the

cells of dimension 0, 1, 2, and a distinguished face of dimension 2
called the exterior face.
Finite and Infinite in Logic 363

2. Endpoints is a ternary relation providing endpoint(s) for edges.

3. Face-Edge is a binary relation providing, for each face (including the

exterior cell), the edges on its boundary.

4. Face-Vertex is a binary relation providing, for each face (including

the exterior cell), the vertices adjacent to it.

5. For each region name p ∈ Reg, a unary relation p providing the set
of cells contained in region p.

6. Orientation is a 5-ary relation providing the clockwise and counter-

clockwise orientation of edges incident to each vertex. More precisely,
(←V, v, e1 , e2 , e3 ) ∈ Orientation iff v is a vertex, e1 , e2 , e3 are edges of
faces incident to v, and e2 lies between e1 and e3 in the clockwise
order on the incidents cells of v, and (W→, v, e1 , e2 ) ∈ Orientation iff
v is a vertex, e1 , e2 , e3 are cells incident to v, and e2 lies between e1
and e3 in the counterclockwise order on the incident cells of v.

Let inv(Reg) denote the above vocabulary and let top denote the mapping
associating to each spatial structure I over Reg its topological invariant
over inv(Reg).
The main result on the topological invariant is the following.

Theorem 5.1 Let Reg be a spatial vocabulary.

(i) The mapping top associating to each spatial structure over Reg its topo-
logical invariant is computable in polynomial time (and NC) with respect to
the size of the representation of I.
(ii) For all spatial structures I, J over Reg, I and J are topologically equiv-
alent iff top(I) and top(J) are isomorphic.

How can topological invariants be used? Suppose that a topological

query is posed against the spatial database. In principle, the query can be
answered by another query posed against the topological invariant. Since
the topological invariant is in most cases much simpler than the full spa-
tial data, this strategy is likely to be more efficient. In order for this to
work, topological queries in the spatial query language need to be effectively
translated into queries in some query language for topological invariants.
We present an example that provides some intuition into the difficulties
involved in translating FO sentences on the spatial representation into FO
sentences on the invariant. Consider the query on schema Reg = {P, Q}:
364 V. Vianu

“Regions P and Q intersect only on their boundaries.”

Clearly, this is a topological property. It can be expressed in FO with order

over the reals by the sentence:

(†) ∀x∀y[(P (x, y) ∧ Q(x, y)) →

(boundaryP (x, y) ∧ boundaryQ (x, y))],

where boundaryP (x, y) (and similarly boundaryQ (x, y)) is the formula:

P (x, y) ∧ ∀x1 ∀y1 ∀x2 ∀y2 [(x1 < x < x2 ∧ y1 < y < y2 ) →
∃x
∃y
(x1 < x
< x2 ∧ y1 < y
< y2 ∧ ¬P (x
, y
)].

Clearly, the same property can be expressed by the FO sentence over

inv(Reg):

(‡) ∀u[(P (u) ∧ Q(u)) → (V ertex(u) ∨ Edge(u))].

However, how to get from (†) to (‡) is mysterious. The difficulty is

to algorithmically extract the topological meaning of a sentence like (†)
that uses non-topological statements involving reals and <. This problem
is considered in [22]. It is solved for the case when Reg contains a single
region name (however, the region can be quite general); it has been recently
shown by Grohe and Segoufin that the translation is not possible with
several region names.
It is natural to wonder if these techniques can be extended beyond di-
mension two. Unfortunately, the picture is largely negative. The existence
of a topological invariant for 3-dimensional semi-algebraic databases im-
plies a positive answer to an open problem in Knot Theory: the existence
of an invariant for topologically equivalent knots [7]. In dimension four
(and higher) there is no finite topological invariant, because topological
equivalence itself is undecidable. This is shown by adapting the proof of an
undecidability result on topological equivalence of manifolds [19]. The lat-
ter proof is by reduction of the word problem for finitely generated groups
to isomorphism of the fundamental groups of two topological spaces, which
in turn is equivalent to their being homeomorphic.
In summary, we have shown how to produce, from a semi-algebraic spa-
tial database which is conceptually infinite, a finite structure that precisely
summarizes the topological properties of the original structure. This exam-
ple is typical of applications where structures that are in principle infinite
are turned into finite structures that can be efficiently manipulated.
Finite and Infinite in Logic 365

6 From Finite Back to Infinite?

We lastly consider a surprising idea: is it useful to sometimes model objects
which are finite as infinite structures? Admittedly, this is not altogether
new: after all, computers, which in practice are always finite, are better
modeled by a Turing Machine with infinite tape than by a finite-state ma-
chine. In this section we pose the same question about another well-known
finite object: the World Wide Web.
In [3], we propose a model that views the Web as an infinite labeled
graph. We believe this captures the intuition that exhaustive exploration of
the Web is –or will soon become– prohibitively expensive. The infiniteness
assumption can be viewed as a convenient metaphor, much like Turing
machines with infinite tapes are useful abstractions of computers with finite
(but potentially very large) memory. As a consequence of the infiniteness
assumption, exhaustive exploration of the Web is penalized in our model by
a nonterminating computation. Thus, our model draws a sharp distinction
between exhaustive exploration of the Web and more controlled types of
computation.
Note that this approach is fundamentally different from other attempts
to model infinite data (e.g., [14], [15]) which focus on finitely representable
structures. In contrast, we do not assume the Web is finitely represented.
Instead, we view it as a possibly nonrecursive infinite labeled graph which
can never be entirely explored. Intuitively, a node in the graph is an ab-
straction of a Web page, and the labeled edges represent html links between
pages. The links allow navigating through the Web, in hypertext style. Our
model leads to a focus on querying and computation where exploration of
the Web is controlled.
We consider queries which are mappings whose domain consists of pairs
(o, I) where I is an infinite labeled graph and o is a designated starting
node of I. For each pair (o, I), the result of the query consists of a subset
of the nodes of I.
We begin by exploring the notion of computable query in the context
of the Web. Our model is along the lines of the computable queries of
Chandra and Harel [6]. We briefly describe a machine model of computation
on the Web that we call a Web machine. This works much like a Turing
machine, but takes as input an infinite string and may produce an infinite
answer. Based on the Web machine, we can define the notions of finite
computability and eventual computability of queries. The latter notion arises
from the fact that infinite answers to queries are allowed. A query is finitely
computable if its answer is always finite and computable by a halting Web
366 V. Vianu

machine. A query is eventually computable if there is a Web machine,

possibly nonterminating, which eventually outputs each node in the answer
to the query (a node cannot be retracted once it is output).

Examples The notions of finitely computable and eventually computable

queries are illustrated by the following queries on input (o, I):

1. Finitely computable:

• Find the nodes reachable from o by a path labeled a.b.c (an a-

labeled edge, followed by a b-labeled edge, followed by a c-labeled
edge).
• Find the nodes o
such that there is a path of length at most k
from o to o
.
• Find all nodes lying on a cycle of length at most 3 which contains
o.

2. Eventually computable with possibly infinite answers (so not finitely

computable):

• Find the nodes reachable from o.

• Find the nodes referencing o.
• Find the nodes belonging to a cycle.

3. Eventually computable with finite answers, but not finitely com-

putable:

• Find the nodes on the shortest cycle containing o.

• Find the node(s) at the shortest distance from o that reference
o.

4. Not eventually computable:

• Find all nodes that do not belong to a cycle.

• Find all nodes which are not referenced by any other node.
• Output o if and only if all nodes reachable from o have non-nil
references2 .
2
Nil references can be modeled by references to a special node named nil.
Finite and Infinite in Logic 367

In particular, it is clear from the above examples that finitely computable

and eventually computable properties are not closed under complement.
We next describe informally two specialized machine models that cap-
ture directly the main styles of computing used on the Web: browsing and
searching. The browser machine model allows for navigational exploration
of the Web. The browse/search machine additionally allows searching in
the style of search engines.
The idea underlying the browser machine is to access the Web nav-
igationally, by following references starting from the source node o. A
browser machine has an infinite browsing tape, an infinite work tape, and
a right-infinite output tape. It is equipped with a finite state control which
includes a special state called expand. The computation of the machine on
input (o, I) is as follows. Initially, the browsing tape contains the encoding
of the source node o. If the expand state is reached at any point in the
computation and the browsing tape contains the encoding of some node o
,
this is replaced on the browsing tape by an encoding of all nodes in I with
incoming edges from o
.
Obviously, browser machines have limited computing ability, since they
can only access the portion of I reachable from the source node o. However,
this is an intuitively appealing approach for controlling the computation.
The following result confirms the central role of this style of computation
in the context of the Web.

Theorem 6.1 Each finitely computable Web query is finitely computable

by a browser machine.

We next augment browser machines with a search mechanism. The

search is essentially a selection operation such as: (i) select all edges with
label “Department”; (ii) select all edges with label A and node 556 as desti-
nation; and (iii) select all edges. In general, a search triggers an eventually
computable subquery, whose result may be infinite. This leads to the prob-
lem of integrating nonterminating subcomputations into the computation
of a query. We adopt the following model.
A browse/search machine is a browser machine augmented with a right-
infinite search-answer tape and a separate search-condition tape. There is
a distinguished search state. The computation of the machine is nonde-
terministic. A search is triggered by writing a selection operation on the
search-condition tape, then entering the search state. The search-answer
tape functions similarly to the answer tape of an eventually computable
query. Answers to previously triggered searches arrive on the search-answer
368 V. Vianu

tape at arbitrary times and in arbitrary order. More precisely, suppose the
set of selections triggered up to some given point in the computation is
{σ1 , . . . , σn }. In any subsequent move of the machine, a (possibly empty)
finite subset of the answers to some of the σi ’s is appended to the search-
answer tape. This is nondeterministic. The order in which answers are
produced is arbitrary. Each tuple in the answer to σi is prefixed by σi (ev-
erything is encoded in the obvious way). It is guaranteed that all answers
to a triggered search will be eventually produced. However, note that there
is generally no way to know at a given time if all answers to a particular
search have been obtained. The rest of the computation occurs as in the
browser machine. A Web query q is finitely computable by a browse/search
machine if there exists a browse/search machine W such that each com-
putation of W on input (o, I) halts and produces an encoding of q(o, I)
on the answer tape3 . The definition of query eventually computable by a
browse/search machine is analogous.
What is the power of browse/search machines? This is elucidated by
the following result.

Theorem 6.2 (i) A Web query is eventually computable if and only if

it is eventually computable by a browse/search machine.

(ii) A Web query is finitely computable if and only if it is finitely com-

putable by a browse/search machine (if and only if it is finitely com-
putable by a browse machine).

We close with a brief discussion of the ability of query languages to

express queries on the Web. We consider in [3] the classical languages FO,
Datalog, and Datalog¬ (see [1] for a presentation of these languages). The
questions of interest for each language are the following: (i) Are the queries
in the language finitely computable or eventually computable? (ii) Which
fragments of each language can be implemented by browsers and which by
a combination of browsing and searching? We provide syntactic restrictions
that guarantee computability by browsers or by browse/search machines in
FO and Datalog(¬) .
One of the interesting results of [3] is with respect to negation. The
“positive” fragment of FO is eventually computable. The addition of re-
cursion yields no problem. However, negation brings trouble, and some
simple FO queries are not eventually computable. The Datalog¬ languages
3
However, it should be clear that a browse/search machine that uses the search feature
in a nontrivial way cannot terminate.
Finite and Infinite in Logic 369

yield some surprises: the standard semantics, stratified and well-founded

[24], are ill-suited for expressing eventually computable queries, whereas
the more procedural inflationary semantics [2], [16] turns out to be natu-
rally suited to express such queries, and thus has a fundamental advantage
over the first two semantics.

7 Conclusion
The relationship between the finite and the infinite raises fundamental is-
sues in the context of logic. We examined the impact of the finiteness
assumption, and noted the failure of many classical results when struc-
tures are restricted to be finite, such as the Compactness theorem. We
discussed two very effective and elegant alternative techniques for proving
expressiveness results on finite structures: Ehrenfeucht-Fraissé games and
0-1 laws. We also illustrated ways to finitely represent information of in-
terest about infinite structures, using the example of topological invariants
of two-dimensional semi-algebraic regions. Conversely, we argued that in
some situations it is beneficial to model finite objects as infinite structures.

References
[1] S. Abiteboul, R. Hull, V. Vianu, Foundations of Databases, Addison-
Wesley, Reading-Massachusetts, 1995.

[2] S. Abiteboul, V. Vianu, Procedural and declarative database update

languages, Proc. ACM Symp. on Principles of Database Systems, 1988,
240–250.

[3] S. Abiteboul, V. Vianu, Queries and computation on the Web, Proc.

Intern. Conf. on Database Theory, Springer-Verlag, 1997.

[4] M. Ben-Or, D. Kozen, J. Reif, The complexity of elementary algebra

and geometry, J. of Computer and System Science, 32 (1986), 251–
264.

[5] A. Blass, Y. Gurevich, D. Kozen, A zero-one law for logic with a fixed
point operator, Information and Control, 67 (1985), 70–90.

[6] A. K. Chandra, D. Harel, Computable queries for relational data bases,

Journal of Computer and System Sciences, 21, 2 (1980), 156–178.
370 V. Vianu

[7] R. Crowell, R. Fox, Introduction to Knot Theory, Springer-Verlag,

GTM vol. 57, 1963.

[8] R. A. DiPaola, The recursive unsolvability of the decision problem for

a class of definite formulas, J. of the ACM, 16, 2 (1969), 324–327.

[9] A. Ehrenfeucht, An application of games to the completeness problem

for formalized theories, Fund. Math., 49 (1961), 129–141.

[10] H. Enderton, A Mathematical Introduction to Logic, Academic Press,

1972.

[11] R. Fagin, Probabilities on finite models, Journal of Symbolic Logic,

41, 1 (1976), 50–58.

[12] R. Fraissé, Sur les classifications des systèmes de relations, Publ. Sci.
Univ. Alger, I:1 (1954).

[13] Y. V. Glebskiĭ, D. I. Kogan, M. I. Liogonkiĭ, V. A. Talanov, Range and

degree of realizability of formulas in the restricted predicate calculus,
Kibernetika, 2 (1969), 17–28.

[14] D. Harel, T. Hirst, Completeness results for recursive data bases, Proc.
ACM Symp. on Principles of Database Systems, 1993, 244–252.

[15] P. Kanellakis, G. Kuper, P. Revesz, Constraint query languages, Proc.

ACM Symp. on Principles of Database Systems, 1990, 299–313.

[16] P. G. Kolaitis, C.H. Papadimitriou, Why not negation by fixpoint?

Proc. ACM Symp. on Principles of Database Systems, 1988 231–239.

[17] P. Kolaitis, M. Vardi, 0-1 laws for infinitary logics, Information and
Computation, 98 (1992), 258–294.

[18] D. Kozen, C-K. Yap, Algebraic cell decomposition in NC, Proc. IEEE
Symp. on Foundations of Computer Science, 1985, 515–521.

[19] A. A. Markov, Unsolvability of the problem of homeomorphy, Proc.

Intern. Congress of Mathematics, 1958, 300–306 (in Russian).

[20] C. H. Papadimitriou, D. Suciu, V. Vianu, Topological queries in spatial

databases, J. of Computer and System Sciences, 58 (1999), 29–53.

[21] R. Rado, Universal graphs and universal functions, Acta Arith., 9

(1964), 331–340.
Finite and Infinite in Logic 371

[22] L. Segoufin, V. Vianu, Querying spatial databases via topological in-

variants, Proc. ACM Symp. on Principles of Database Systems, 1998.

[23] V. A. Talanov, V. V. Knyazev, The asymptotic truth value of infinite

formulas, All-union seminar on discrete mathematics and its applica-
tions, 1984, 56–61.

[24] A. Van Gelder, K. A. Ross, J. S. Schlipf, The well-founded seman-

tics for general logic programs, Proc. ACM Symp. on Principles of
Database Systems, 1988, 221–230.