You are on page 1of 328


Volume 2



JAAKKO HINTIKKA, Boston University


DIRK VANDALEN, Universityof Utrecht, TheNetherlands

DONALD DAVIDSON,Universityof California, Berkeley
THEO A.F. KUIPERS, Universityof Groningen, TheNetherlands
PATRICK SUPPES, Stanford University, California
JAN WOLENSKI, JagiellonianUniversity,Krakow, Poland


(Photo taken by Winston Scott Boyer and printedhere withhis kind permission)
Volume 2. Philosophy ofPhysics, Theory Structure,
andMeasurement Theory

CorcoranDepartment ofPhilosophy,
University of Virginia,
Charlottesville, VA, U.S.A.

Library of Congress Cataloging-in-Publication Data

A C.I.P. Catalogue record forthisbook is availablefromthe Library of Congress.

ISBN: 0-7923-2553-2
ISBN Set: 0-7923-2554-0

Published by Kluwer Academic Publishers,

P.O. Box 17, 3300 AA Dordrecht, TheNetherlands.

Kluwer AcademicPublishers incorporates

the publishingprogrammesof
D. Reidel, Martinus Nijhoff, Dr W. Junk and MTP Press.

Sold anddistributedin the U.S.A. andCanada

by Kluwer AcademicPublishers,
101 Philip Drive,Norwell, MA 02061, U.S.A.

In allother countries, soldanddistributed

by Kluwer AcademicPublishers Group,
P.O.Box 322, 3300 AH Dordrecht, TheNetherlands.

Printed on acidfree paper

AllRights Reserved
© 1994 Kluwer AcademicPublishers
No part of the materialprotected by this copyrightnotice maybe reproduced or
utilizedin any formor by any means,electronic or mechanical,
includingphotocopying,recording orby anyinformationstorageand
retrievalsystem, withoutwritten permission fromthe copyright owner.

Printedin the Netherlands

Volume 2: Philosophy of Physics, Theory Structure,
andMeasurement Theory


Barry Loewer /Probability and Quantum Theory /

Comments by Patrick Suppes 3
Arthur Fine / Schrodinger's Version of EPR, and
Its Problems./ Comments by Patrick Suppes 29
Jules Vuillemin / ClassicalField Magnitudes /
Comments by Patrick Suppes 45
BrentMundy / Quantity, Representation and Geometry /
Comments by Patrick Suppes 59
Paul Humphreys / Numerical Experimentation /
Comments by Patrick Suppes 103

Ryszard Wojcicki / Theories and TheoreticalModels /

Comments by Patrick Suppes 125
N.C. A.Da Costa and F. A.Doria / Suppes Predicates
and the Construction of UnsolvableProblems in the
Axiomatized Sciences / Comments by Patrick Suppes 151
JosephD.Sneed / Structural Explanation / Comments by
Patrick Suppes 195

R.DuncanLuce andLouis Narens / Fifteen Problems

Concerning the Representational Theory ofMeasurement /

viii table of contents

Comments by Patrick Suppes 219

Fred S.Roberts and Zangwill SamuelRosenbaum/
The Meaningfulness of Ordinal Comparisons for General
Order Relational Systems / Comments by Patrick Suppes 251
C. Ulises Moulines and Jose A.Diez / Theories as
Nets: The Caseof Combinatorial Measurement Theory /
Comments by Patrick Suppes 275
NameIndex 301
SubjectIndex 305
Table ofContents to Volumes 1and3 311


My historical view of the situation is that if

probability theory hadbeen developed to any-
thing like its current sophisticated state at the
timethebasicworkin quantum mechanics was
done in the twenties, then a very different sort
of theory wouldhavebeenformulated.

ABSTRACT. Patrick Suppes has writtenanumberofpapers in whichhe has emphasized

thenon-standardnature of probabilityinquantum mechanics. Specifically, thequantum
state of a system cannot beunderstood as characterizing a probability measure over the
space generatedby the possiblevaluesof observables.Hehas suggested that a rethinking
of quantum theory in which probability is treatedin a standard fashion would likely
solve foundational problems in quantum theory. Iargue that thenon-standardness of
probability is intimately related to the measurement problem. Once a stand is taken on
that problem probabilitycan be understood classically. Idiscuss two ways of doing
this; the orthodox account of quantum theory and Bohm's theory and argue for the
superiority of thelatter.

Quantum theory plays a significant although ambivalent role in Patrick

Suppes' philosophy of science. On the one hand, he appeals to it
to support a number of his favorite themes the fundamental role of
probability in science, the inadequacy of Laplacian determinism, and
the incompleteness of scientific theory. On the other hand, he has
emphasized the peculiar nature of quantum mechanical probabilities.
Heremarks that
Probabilityconcepts have a strange and awkwardappearance in quantum mechanics,
as if theyhavebeen brought within the framework ofthe theory only as an afterthought
and with an apology for theirinclusion.2
Suppes is certainly correct about the awkwardness of probability con-
cepts in quantum theory. A probabilistic theory is usually given by
specifying a probability space anda probability measure on that space.
But in quantum theoryprobabilities are introduced by Born'srule which

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 3-28.
© 1994 Kluwer Academic Publishers. Printed in theNetherlands.

assigns them only to the outcomesofmeasurements. Suppes has empha-

sized in a number of articles 3 that this leads to probability assignments
which apparently fail to conform to the usual mathematical laws of
probability. Specifically, if Q and Q* are non-commuting observables
pertaining to a systemin state $ thenBorn's rule determines probability
distributions for each separately but not for their conjunction. Further,
a series of important results has shown that 'natural ways' of extending
quantum mechanical probability distributions to joint distributions lead
to absurdities. The most significant results, Gleason's theorem and its
corollary, the Kochen-Specker theorem, establish that it is impossible
to embed quantum mechanical probabilities into a classical probability
space in such a way that these probabilities reflect the distribution of
values for observables.5 In addition to issues concerningthe mathemat-
ics of probability in quantum theory there is the question of exactly how
probability in the theory should be interpreted. Are quantum mechan-
ical probabilities best understood as frequencies, degrees of belief, or
objective chances? These are the issues which I will address in my
Suppes finds the formal results -especially thenon-existenceof joint
probability distributions - to be at the heart of the puzzles of quantum
theory. At one time he sought to adapt probability theory and logic
to orthodox quantum theory. In fact he argued from the non-existence
of joint probability distributions to a non-standard logic.6 But in more
recent writings he has reversed this line of thinking and instead has
sought to adapt quantum mechanics to classical probability theory and
logic. He says
Relative to the history of quantum mechanics over the past fifty years, it is a radical
thesis to maintain that orthodoxprobabilitytheory will win out over orthodox quantum
mechanics in providing a satisfactory theoretical framework. To a large extent, a
philosophicalviewpointthatderives from classicalrealismmotivatesmy ownincreasing
acceptance of the orthodox probabilisticviewpoint. Ifind it very mucheasier to think
about particles as havingcontinuous trajectories. The Copenhagen interpretation is
lacking intuitive persuasiveness and is too closely alliedto narrow positivistic views
that no longer seemplausible.7
Suppes's view here seems to be that one should try to formulate quantum
theory or a theory which accounts for quantum phenomena within the
framework of classical logic and probability. I very much agree with
this approach. In fact Iwill argue that despite the results previously
cited it is not all that difficult to interpret quantum theory so that its
logic and probability theory are perfectly classical. But seeing this

requires areorientation of the way in which we are accustomed to think

of quantum theory. Specifically, I will argue that prior to understanding
the role of probability in quantum theorem the so-called measurement
problem or Schrodinger's Cat paradox, must beresolved. Once a stand
is taken on the measurementproblemthe theory can be formulated with
perfectly classical logic and probability. Iwill illustrate this claim by
discussing two solutions to the measurementproblem, the orthodox yon
Neumann account and ahidden variable theory due to DavidBohm.
Let us start by listing the principles of elementary non-relativistic
quantum theory.
1. Every isolated physical system is described by a quantum state
$ which is represented by a vector |$) in an appropriate Hilbert
space and everyphysical quantity (observable) Q is representedby
a Hermitian operator Q in that space. \Q = q) is a state in which
Q possesses the value q.
2. Linearity: the quantum state of a system evolves in accord with a
deterministic linear law Schrodinger's equation:

ot 2m
3. Eigenstate-eigenvaluerule: thequantum state of asystem contains
a complete specification of its physical state andan observable A
has a value when and only when that state is an eigenstate of the
4. Born's rule: if S is in state $ then the probability that a perfect
measurementof Q on $ will yield theresult Q = 6 is |($ | Q = b)\ 2 .
In addition to these principles elementary quantum theory contains a
vast amount ofinformation concerning the connection between physical
quantities and operators and the quantumstates of realphysical systems.
The whole theory is enormously well confirmed within non-relativistic
Prior to formulating the measurement problem a few comments on
these principles are in order. First, it used to be, and to a lesser extent
still is, usual to try to understand quantum theory instrumentally as
merely an algorithm for predicting theresults of experiments. But like
Suppes I find 'classical realism' a more plausible philosophical view-
point. If the theory isunderstood realistically then it says that theentire
physical reality of an isolated system is characterized by its quantum
state. Second, Schrodinger's law is the sole dynamical law of this theory

and it is deterministic and linear. These features will be very impor-

tant to our discussion. Third, the eigenstate-eigenvalue rule is to be
understood literally. Anobservablehas a value when andonly when the
stateis an eigenstate of that observable. Since no state is an eigenstate
of non-commuting observables suchobservables cannot simultaneously
possess values. Finally, Born's rule introduces probabilities into quan-
tum theory but only for the outcomes of measurements. If Q and Q*
are non-commuting then there is no state in which they are both well
defined; i.e. no state corresponding to Q = q and Q* = q*. So, while
Born'srule providesprobability distributions for Q and Q* separately, it
gives no way of calculating a jointdistribution of these two observables.
To keep matters simple Iwill suppose that Q is an observable of
system S with two values q\ and q2 and that Mis a measuring device
whose pointer observable A has three values ao, a\, and ai (which
correspond to three pointer positions). The state \A = ao) is M's
ready-to-measure state. A perfect non-destructive measurement of Q
on S is an interaction between S and Mwhose Hamiltonian guarantees
that if theinitial state of Mis the ready state and the initial state of S is
= = = =
\Q 8q{) then the final stateof S and Mis\o qi) \A ai) (wherei
1, 2). Since the interaction between M and sis a physical interaction
itconforms to Schrodinger'sequation. That equationpredicts that if the
initial state of Sis not an eigenstate of Q but is ci |Q = q\ ) +c 2 \Q + q2 )
then it follows from the linearity of Schrodinger dynamics thatthe final
state of M + S is

MEAS cx \Q = qx) \A = ax ) + c2 \Q = ?2) |>1 = a2>.

The problem isobvious. MEAS is not an eigenstate ofeither Q or A so,
according to principle 3, neither Q nor A possesses a definite value in
MEAS.In other words, ifMEAS is the state of M+ S then the pointer
is notpointing in any of its three possible directions.
Schrodinger's famous cat paradox is a special case of this situation.
If the Mis a cat and the pointer observable is the cat's being alive or
dead then, according to 2 and 3, at the conclusion of the measurement
the cat is not alive andis not deadbut isin a superposition with both live
and dead components.9 There is a more general versionof the measure-
ment problem which Philip Pearle calls 'the reality problem. Pearle
observes that in ordinary circumstances (even circumstances in which
no measurements are being made) the states of physical systems will
evolve so that they are rarely eigenstates ofusual physical quantities.10
PROBABILITYand quantum theory 7
So quantum mechanical principles predict that most of the time most
of the properties which we think have values do not in fact have val-
ues. Anyone who thinks of quantum mechanics as providing a true
description of reality must come to terms with this problem.
The relationship between the measurementproblem andBorn's rule
should be clear. According to Born's rule the probability of a mea-
surement of Q yielding the result that Q=qi is equal to |ci|2 But.
Schrodinger's law is completely deterministic. Rather than assigning
the quantum mechanical probability it entails that this result will defi-
nitely not occur (since the final state is not an eigenstateof Q). So there
is a flat contradiction betweenBorn's rule, Schrodinger'sequation, and
the eigenstate-eigenvalue rule. The measurementproblem is more fun-
damental than the non-existence of joint probability distributions for
non-commuting observables since without a solution it does not even
make sense to talk about probability in quantum theory.11
Various solutions tothe measurementproblem havebeenproposed. I
cannotsurvey them allhere and so will, somewhat dogmatically, remark
that as long as quantum theory is understood non-instrumentally there
are only two strategies that can possibly work. One is to deny the uni-
versality of Schrodinger's equation (or replace it with a non-linear law)
and the other is to deny the completeness of the quantum mechanical
state. This strategy amounts to saying that a physical quantity may
possess a definite value even when the state is not an eigenstate of the
observable corresponding to that quantity.12 I will describe approaches
which follow each strategy.


The orthodox - that is, the usual text-book - way of dealing with the
measurement problem, which was given its canonical formulation by
yon Neumann13 replaces 2 with

2. As long asnomeasurementisbeingmade ona system Sitsquantum

state evolves deterministically in conformity with Schrodinger's
and adds
5. The Collapse Postulate: if a measurementis beingmade on S then
the quantum state of M+ S does not evolve in accordance with

Schrodinger's law but rather jumps so that the probability that it

evolves to \Q = qi) \A = ax) is equal to |q|2 .
This change of state is called the 'collapse' of the quantum state.
Itis clearhow yonNeumann's proposal intends to solve the measure-
ment problem. It entails that when an observable Q is being measured
the post-measurement state of the measuring apparatus and the quan-
tum system is an eigenstate of that observable. So, at the end of the
measurement considered above, the state is not MEAS but is either
\Q = qx) \A = ax) or \Q = q2) \A = a 2). The status ofBorn's rule is
clear on yon Neumann's account. The ruleis entailed by 5 and 3 since
the probability that the post-measurement state is an eigenstate of Q is
the one givenby therule.
I now want to show that if we take a realist view of orthodox quan-
tum theory then its logic and its probability theory can be understood
classically. Letuslook first at the logic. According to the theory,phys-
ical reality consists of a quantum state function which usually evolves
deterministically in conformity with Schrodinger'sequation but which,
every once in a while, when measurements are made, jumps in con-
formity with the collapse postulate. We define 'an orthodox quantum
possible world' @ as a world history which so evolves. This defini-
tion,of course, reflects the vagueness of the 'measurement. A precise
definition should include a precise physical characterization of those
quantum states in which measurementsoccur.
Anorthodox quantum mechanical propositionisa set of suchworlds.
A statement about a quantum mechanical observable, e.g. Qst = q,
expresses the proposition that the state of S at t in @ is an eigenstate
of Q with value q. Note that the quantum mechanical statement that
Q possessesno value also expressesa quantum mechanical proposition

since it says that the state is one of those which is not an eigenstate of
Q. If Q t q and Qis measured at t then the result of the measurement
will certainly be that Q = q but if Q possesses no value then the
measurementmay result in anypossible value of Q. The logic of these
statements is ordinaryBoolean logic. For example, the statement Q = q
& Q* — q* is perfectly well formed. If Q and Q* are non-commuting
then the conjunction is false since thereis no state whichis an eigenstate
of both operators.
The appearance that quantum mechanical statements obey a non-
standard logic may arise from not taking seriously the position that
on this view the physical reality of a system is completely given by
its quantum state. If we forget this then we may be tempted to think
that quantum mechanical statements shouldbe related to each other in
certain ways which suggest anon-standard logic. For example, let Ql
be an observable whose value is 1iff the state assigns an amplitude of
1 to the particle's being located in region L and Qr be an observable
whose value is 1iff the state assigns an amplitude of 1 to the particle's
being located in region R disjoint from L and Qrul an observable
whose value is 1 iff the state assigns an amplitude of 1 to the particle's
beinglocatedin theunion of RandL.Itis possible for Qlur — 1even
though Qr 1and Qi 1 This has suggested to some that the logic
of these quantum mechanical statementsis non-standard since it seems
that, according to quantum theory a disjunction - the particle is in L or
the particle is in R can be true even though both of its disjuncts are
false. But from the point of view of a theory which takes the quantum
state as the real and complete physical state this is misleading. The
proposition that the quantum state is an eigenstate of Qlur with value
1 is not the disjunction of the propositions that the state is an eigenstate
of Qr with value 1 and the stateis an eigenstate of Ql with value 1 Of.
course, this does not mean that theorthodox view in which the quantum
state is real and complete is satisfactory. It entails that particles may
lack definite location or move on definite trajectories and so on. That
is quite puzzling. My point is that the logic of propositions about the
quantum state is perfectly classical.14
Themathematical theory of probability for orthodox quantum theory
is alsoperfectly classical. Therelevantprobability space isan appropri-
ate set of subsets of the set of possible quantum histories. The histories
evolve in accord with the dynamics described by 2 and 5. This gives
rise to a branching tree structure where the branching points involve
measurements.The probabilities of thesebranches at abranch point are
givenby 5. For a given state $ at time t the probabilities of statements
of the form Qt = q are perfectly well defined. If $(£) is an eigenstate
of Q t then P(Qt = q) is either 1 or 0. If Q and Q* are non-commuting
observables then the joint probability P(Q = q & Q* = q*) = 0 since
no state is an eigenstate of both Q and Q*.
The probability at t that the system (or world) whose quantum state
is $f will evolve so that at Pt>(Q = q), where t' is after t, is givenby
the branching tree probabilities. It is the sum of the probabilities of
all the branches which are eigenstates of Q at t'. Once again the joint
probability Qt> =q & Q*i =q*is 0since the systemcannot evolve into

a state which is an eigenstate of both observables. There is noneed for

non-standard probability mathematics.
According to Born's rule when a measurement of Q is made on a
system S in state $ which is not an eigenstate of Q the probability
that the result is Q=q is |(Q =q | $)|2 In the orthodox account
this probability cannotbe an epistemic or ignorance probability. If it
were then there would be a matter of fact concerning the value of Q.
But there is no matter of fact about the value of Q prior to its being
measured since the state is not an eigenstate of Q. For a similar reason
the probabilities cannotbe understood as frequencies. It is not the case
that Sis one of an ensemble of systems in state $, \(Q = q | $)|2 of
whichare onesin which Q = q. If the state ofall these systems is $ then,
on the orthodox account, for none of them does Q = q. Instead, these
probabilities must be understood as objective chances. The quantum
mechanical probability is the objective chance of the system evolving
into an eigenstate of Q with value q when Q is measured.
If what I have said so far is correct then the logic and probability
theory of the orthodox accountis classical. The account clearly recon-
cilesBorn's rule with the dynamics andin doing so goes some distance
toward solving the measurementproblem and the realityproblem. But
now I want to point out some of the ways in which the orthodox account
is implausible and inadequate.
First, it should be noted that there is an inexplicitness in the ortho-
dox view which up until now we have mostly ignored. It is that the
view is not clearly specifieduntil a clear rule is given which says when
Schrodinger evolution occurs and when the collapse occurs and that,
of course, involves saying exactly which physical processes are mea-
surements. Defenders of the orthodox interpretation are rather casual
concerning this question since it turns out that as long as the collapse
does not occur in interactions involving only isolated microscopic sys-
tems (i.e. measurements always involve macroscopic systems) it will
be enormously difficult to empirically test whether or not a collapse
has occurred. For this reason the precise point at which Schrodinger's
equation gives way to the collapse is not practically testable.
Second, itis difficult to believe that there are two radically different
dynamical laws which govern interactions; the probabilistic collapse
for measurements and Schrodinger's law for all the others. There is
no smooth transition from one dynamic to the other. The proposals for
where to put a cut-off are not plausible; e.g. that Schrodinger's law

holds for systems with fewer than n particles (or degrees of freedom or
whatever) while the collapse holds for systems with nor more particles.
Third, the collapse of state is a physically peculiar process since it
is non-local. Einstein, Podolsky, andRosen pointed this out long ago.
David Bohm's version of their thought experiment involves a pair of
electrons in the singlet state which are spatially separated. Orthodox
quantum theory predicts that if the 2>spin of one of the electrons is
measured the state collapses and instantaneously results inan eigenstate
of ;c-spin for both particles. In effect, this means that the physical
reality - whether ie-spin has a definite value or not - of one particle is
instantaneously affected by a measurementon the other particle. Much
ismade of the point that the non-locality involved in thecollapse of state
is empirically compatible with special relativity since it is not possible
to transmit messages or energy by the collapse process. This is correct
but still the collapse is a non-local process and it is no easy matter to
formulate it without supposing thata preferred reference frame exists.15
If it is so formulated in a preferred reference frame then that amounts
to treating special relativistic claims about the structure of space-time
(that there is no preferredreference frame) not completely realistically,
or it involves treating the collapse of state not complete realistically.
Fourth, as we have already noted, the orthodox theory allows for
states which are not eigenstates of ordinary physical quantities, like
position, momentum, energy, etc. and, in fact, does not allow for states
whichare eigenstates ofboth position and momentum. Inconsequence
particles fail to possess definite trajectories.
Fifth, the orthodox view's attempted resolution of the measurement
problem is ultimately unsatisfactory. There are situations in which no
measurements are being made (on any reasonable construal or 'mea-
surement') which therefore evolve linearly in such a way that,except in
rare circumstances, their quantum states are not eigenstates of ordinary
macro-properties. That is, the quantum states of cats (or systems of
which cats are parts) typically evolve so that itis not an eigenstate of the
cat's beingalive or dead. Immediately after a position measurementthe
wave function beings to spread out so that soonitis not an eigenstate of
position. It follows that the orthodox account fails to solve the reality


I now want to consider a solution to the measurement problem which

follows the second strategy of denying that the quantummechanical state
is the complete physical state. Such interpretations are misleadingly
called 'hidden variable' theories.17 While there are anumber of hidden
variable approaches, by far the most satisfactory was developed by
David Bohm (based on an idea originally due to de Broglie) and was
given an especially nice formulation by John Bell. It is a scandal that
Bohm's theory has received so little attention from philosophers of
quantum mechanics.18 The reason for this is probably the existence of
a number of 'no go' theorems due to yon Neumann,Gleason, and Bell,
among others, whichhavebeen widely interpreted as demonstratingthe
impossibility of theories like Bohm's.
Suppes mentions John Bell's discussion of some of the 'no hidden
variable theorems' (specifically yon Neumann's resultand the Kochen-
Specker theorem) and seems to agree with Bell that they do not demon-
strate the non-viability of hidden variable interpretations. But he goes
on to say
In a beautiful series of papers beginning with Bell (1964 and 1966), a much more
reasonable and intuitive treatment of hidden-variable theories has been given, and
their impossibilityhas been demonstrated experimentally,at a rather satisfactorylevel
(Suppes, 1984, p. 25).

It wouldbeuncharitable to interpretSuppes as endorsing the commonly

held view that Bell's resultand Aspect's experiments literally show the
impossibility ofhidden variable results. Bellhimselfwasanenthusiastic
proponent of Bohm's hidden variable theory.19 In fact, he obtained his
famous result by reflecting on Bohm's theory and asking whether the
non-locality whichit manifests is anunavoidable feature of any hidden
variable theory capable of reproducing quantum mechanical statistical
results.20 Hisaffirmative answer does not show that suchhiddenvariable
theories are impossible, but rather that they must be non-local. On
any such theory events which are related in a space-like way may be
causally related. Perhaps this would count as an objection to hidden
variable theories from the perspective of orthodox quantum theory if
the latter were a local theory. Butit is not. As Imentioned previously,
Einstein,Rosen, and Podolsky had already established the non-locality
of orthodox quantum theory and Bell himself presents a beautifully
lucid argument that any theory -hidden variable or not - which lawfully
reproduces quantum mechanical statistics must be non-local.21
Whatever the reasons for its neglect, Bohm's theory should not be
ignored by philosophers of quantum theory. After briefly describing
the theory I will show that its logic and probability theory is perfectly
classical. Further, it provides a much more nearly adequate solution to
the measurement problem than does the orthodox account. However,
there are questions about exactly how probability shouldbeunderstood
within the theory and I will close with some reflections on this issue.
Bohm's theoryposits that, in addition to possessing a quantum state,
a system consists of a collection of particles (their number and nature
depending on the quantum state) which always occupy definite loca-
tions. It modifies standard quantum theory by dropping 4 (since par-
ticles have definite positions even when the wave function is not an
eigenstate of position) and adding to 1, 2, and3 the following:

6. Velocity Law: -^atdo = hIm ''

, ,(q).
, „ grad|s)
Pd(q) |($ | $)|2
7. Probability Law:
where q = (qx, ,an) specifies the positions of particles 1through n.
The velocity law says, in effect, that the quantum state assigns to each
point in space a velocity vector which determines the particle's velocity
if it occupies that position. If a system consists of n particles then the

wave function evolvesin3 n dimensional configuration space andthe
system of particles is located at a position qin configuration space. The
velocity law then describes the motion of the system in configuration
space andthis determines themotion ofeach particle in ordinary 3-space.
Notice that how a particle (or system of particles) moves depends only
on the value of the wave function where the particle is located (or the
system is located in configuration space). Non-locality arises since,
for example, the motions of a two particle system in 6-dimensional
configuration space may involve correlation of motions ofeach particle
in ordinary 3-space.
The centralresult of Bohmian quantum theory is that the probability
distribution given by 7 is preserved as the state of the system evolves
in conformity with Schrodinger's equation, andthe positions of the par-
ticles evolve in conformity with the velocity law. In fact, the velocity
law is selected precisely because it has this feature. If a measurement
interaction is understood as correlating the position of something (say

a pointer) with an observable being measured then Bohm's theory vali-

dates Born's rule for all observables. This means that, given the above
assumption concerning measurements, Bohm's theory is empirically
equivalent to standard quantum theory.22
It should be clear how Bohm's theory proposes to solve the mea-
surement and reality problems. Consider a measurement like the one
discussed earlier in the paper in which the state of M + S evolves into
the state MEAS. MEAS is a superposition of two components corre-
sponding,say, to pointer on the left and pointer on the right, whichare
effectively separatedin configuration space (this separation is required
of a measurement). The pointer's particles will evolve in such a way
so that they are located in configuration space with one or the other of
these components. In consequence the pointer will be pointing either
to the left or to the right. Further, as long as these components do not
subsequently overlap, the movements of the particles will be directed
only by the component in which they are located. This has the same
practical effect of a collapse of the wave function although the other
component continues to evolve. Should the components subsequently
overlap the other component may have a physically detectable effect.23
Similarly, in Schrodinger's cat experiment,by the end of the measure-
ment the 'live' and 'dead' components of the wave have separated in
configurationspace and thecat's particles are located inone or theother
component. So the cat will be either alive or dead. More generally,
by adding particle positions to the ontology Bohm's theory is able to
validate ordinaryproperties possessing definite values and sosolves the
Developinglogic and probability theory for Bohm's account is fairly
straightforward. Here is how we can develop the logic. A Bohmian
possible world is a world history of the evolution of the world's wave
function in conformity with Schrodinger's law and the evolution of
particle positions in accord with the velocity law. We represent it by
an ordered pair ($(0), g(0)), where 0 is some 'initial' time. Note that
the part of the history representedby $ is different from an orthodox
history since it always conforms to Schrodinger's law i.e. there is
no collapse of the wave function. Since the evolutions of both $(0)
and q(0) are deterministic these serve to determine their values at all
times. A Bohmian proposition is a set of Bohmian worlds. Ordinary
quantum mechanical statements areunderstood as expressingBohmian
propositions; e.g. Qs(t) = qexpresses the proposition thatthe quantum
state of S is an eigenstate of Q with value q. InBohm's theory thereare
also statements like X(t) = x which specify the location of the system
of particles (in configuration space). Truth functional combinations
of these statements are always well defined and their logic is classical
proposition logic.
The probability theory for Bohm's account is developedin this way.
For each world quantum state $(0) there is aclassical probability mea-
sure p(q{o)/$(0)) overthe particle positions which satisfies Born's rule.
There may alsobe a probability distribution,r(s), over the set of quan-
tum mechanical world histories, but this probability assignment is not
part of the theory (it represents subjective uncertainty about the true
quantum state). The probability of a statement about the value of an
observable; e.g. p(Q t = q/$o) is 1 if $o is an eigenstate of Q t and 0
otherwise. Notice that the probability (and the truth value) of a state-
ment about the quantum mechanical position observable may differ
from the probability (and truth value) of a statement about position. So
the probability that the position observable has some value qmay be
0 (or 1) even though the value of p(q(t) = g/$o) differs from 0 (or
1). Similarly, if e.g. momentum is defined in terms of rate of change
of q(t) then its value and probability given $(0) may differ from the
value and probability of the momentum observable given $(0) (which
will either be 0 or 1). Notice that joint probability distributions are
perfectly well defined. Just as in orthodox quantum theory, if Q and
Q* are non-commuting then P(Q = q & Q* = q*) = 0. If V(t) is
the velocity of the system at t defined in terms of rate of change of q(t)
then P(q(t) = x & V(t) = y) is also well defined andcalculated from
the Bohmian probability distribution. The jointprobability of obtaining
particular results for the simultaneous 'measurement' of Q and Q*, of
course, does not exist since itis physically impossible to simultaneously
'measure' non-commuting observables. Themoral is that as long as we
are clear about exactly what proposition a Bohmian statement expresses
there isno reason to think that probability mathematics isnon-standard.
An interesting feature ofBohm's theory is that perfect measurements
of position within Bohm's theory are faithful in that they reveal pre-
existing values of q(t). Suppose that the pre-measurement Bohmian
state of S + Mis


(where the $i are orthogonal position states). The measurement cor-

relates the values of a position observable A of M with values of the
position observable of S. Suppose also that qs(0) is located in $fc(0)
and that qm(0) is located in \R). Then the post-measurement state of
M+ Sis
[S\JM\si(t))\A = ai)i qs(t),qM(t)]
and it is guaranteed that qM.(t) is located within the region associated
with \A = dk). That is, the pointer faithfully records the position of the
particles. The situation contrasts with orthodox quantum theory since
on that accountmeasurements are faithful only when the system's state
is an eigenstate of the observable beingmeasured.
The situation in Bohm's theory with respect to observables whose
values are not determined by thevalue of the position observableis quite
different. Consider such a measurement of observable Q of system S
by observable A of systemM;i.e. the measurementcorrelates \Q = qi)
with \A = a^. Suppose that the initial quantum mechanical state of
S+ Mis |$(0)> = SUM Ci\Q(o) = qi) \R(0)) and that gM+s(o) (the
position of particles in S and the measuring device) are distributed in
accord with Born's rule. Then, as wehave noted, the post-measurement
state will be \s(t)) = SUM c% \Q(t) = qi) \A = a^ and the particles
in the pointer will still be distributed in conformity with Born's rule.
This means that the probability of obtaining the 'result' Q = qi is
given by Born's rule. We have to be careful in interpreting this as
a 'measurement' since neither at the beginning nor at the end of the
measurement interaction is the state an eigenstate of Q. So it would be
as much of a mistake on Bohm's theory as on the orthodox account to
think of the pointer asfaithfully recording a prior value of Q.
But the situation in Bohm's theory is even more peculiar. Exactly
which value of Q is recorded by the measuring device dependsnot just
on 5 andits statebut also on the state of M, including the positions of
its particles. This is easily seen by recalling that in Bohm's theory the
post-measurement positions of the particles in M - where the pointer
is pointing at the conclusion of the measurement - is determined by
the pre-measurement Bohmian state of S + M (assuming that this is
an isolated system) including the pre-measurement positions of the
particles in M. Finding that at the endof the measurement the pointer
is in the position associated with A = a^ not only does not record
a pre- or post-measurement value for Q on 5; it does not record an
intrinsic property of S at all. If the original positions of the particles in
the measuring device had been different the outcome might have been
The Kochen-Specker theorem says that any hidden variable theo-
ry in which the outcomes of measurements are determined contains
observables which are contextual.24 Contextuality means that the out-
come of a measurementof the observable dependson matters other than
the state (quantum andhidden variable) of the system being
- e.g. on what other observables are being measured or themeasured state of
the measuring apparatus. InBohm's theory contextuality is manifested
by the fact that the outcome of a measurement ofan observable (other
thanobservables which are a function of position) dependson the initial
positions of particles in M.
The inevitability of contextuality is often taken to be an objection to
hidden variable theories. But even the very strong kindof contextuality
exhibited by Bohm's theory is not problematic given the ontology of
the theory. Thatit seems so may arise from the habit of thinking of all
quantum mechanical observables as representingproperties intrinsic to
S and measurements as passive interactions which reveal those values.
It is natural to think that if Q is really a property ofS then its value,if
it has a value, should be counterfactually independent of the properties
of M and, in particular, exactly how the measurement is carried out.
But in Bohm's theory the momentum observable for example does not
characterizethe momentum(definedinterms ofposition) of the particles
but rather characterizes the quantum state. And 'measurements' of such
observables do not record values but are interactions which determine
the evolution of the Bohmian state of M + S. From this perspective
there is nothing mysterious or objectionable in its being the case that
exactly how a 'measurement' is carried out may affect the outcome of
the measurement.
How should probability be interpreted withinBohm's theory? Since
the theory is thoroughly deterministic,its probability statements cannot
be understood as referring to objective chances. It seems that these
probabilities must beconstrued epistemically, something like probabili-
tiesin statistical mechanics. There is, however,an important difference
between the two theories. In statistical mechanics itis at least in princi-
plepossible to obtain complete information about the physical state of
a system. In contrast, it is in principle impossible to obtain any more
information about the Bohmian state than that allowed by orthodox

quantum theory. The uncertainty relations, for example, put a limit on

how much information one can have about the outcomes of measure-
ments of non-commuting observables. While it may be possible to learn
the quantumstate of a system the mostone can know about the locations
of its particles is givenby Born's rule. So the probabilities in Bohm's
theory, while epistemic, have akind of objectivity that is grounded in
thelaws of the theory.
Thelimits on obtaining information about theBohmian state and, in
fact, all the empirical consequencesof Bohm's theory(and so the theo-
rem that Bohm's theory is empirically equivalent to orthodox quantum
theory), dependessentially on the assumption that particle positions are
initially distributed in conformity withBorn's rule. As I see it, the main
foundational problem with Bohm's theory concerns how to understand
this assumption. The easiest way is to imagine that the universe had
a beginning and at that initial moment some chance process God
tossing particles into the original quantum state? distributed particles
in accord with Born's rule. These initial chances serve to ground the
epistemic probabilities of Bohm's theory at all subsequent times. But,
it is hard to take this myth seriously.25
Without objective chances to ground Bohm's epistemic probabilities
it is difficult to understand their status. What is it about the world and
our relation to it whichrenders thesethe appropriate probabilities? I am
not sure thatthis question has, or even needs, an answerbut I would feel
thatI understoodBohm's theory betterifI knew some way of grounding
its probability claims.
Suppeshascalled for a rethinking of quantum theory in which ortho-
dox probability theory wins out overorthodox quantum mechanics. My
argument in this paper has been that the rethinking should begin by
confronting the measurement problem since, without a solution to that
problem, we cannot even make sense of quantum mechanical proba-
bilities. Bohm's theory provides a way of resolving the measurement
problem andalso provides a framework in which logic and probability
is perfectly classical. It even assigns particles continuous trajectories.
For these reasons Suppes may find it attractive. On the other hand, its
determinism andits focus on the positions of particles isreminiscent of
Laplace's world picture. These features are probably not congenial to
Suppes's philosophical views. I do not know whether it is the kind of
theory that Suppeshad in mind when he wrote that:
ifprobability theory hadbeen developedto anything likeits current sophisticated state
at the time thebasic work on quantum mechanics was done in the twentiesthen a very
different sort of theory wouldhavebeen formulated.26'27

Department ofPhilosophy,
Rutgers University,
New Brunswick,NJ 08903, U.S.A.


1 Suppes (1979),
p. 12.
Suppes (1962), p. 334.
3 Suppes (1966) and (1984).
4 Wigner (1932) andNelson (1967).
Gleason (1957), Kochenand Specker (1967), andBell(1989).
In Suppes (1966) he argues that every proposition or event should be assigned a
probability and concludes from the non-existence of joint probabilities that arbitrary
intersections of propositions are not propositions. Suppes does not endorse the argu-
ment and inlater writings seems quite skeptical of it.
7 Suppes
inBogdan (1979), pp. 210-211.
This is a necessary conditionfor a perfectmeasurement. Givinga sufficient condition
is a much trickiermatter.
1am assuming that alivenesscorrespondsto aQM observable.This seems reasonable
if we think that the cat is a physical system and its state of health supervenes on the
values of physical quantities ofits particles.
The same point is made by AlbertandLoewer (1991).
11 The habit of talking about probabilities prior to resolving the measurement prob-
very misleading and has caused some writers e.g. Prosperi and Longeri
- to isthink that the measurement problem is not a genuine problem or is solved by
noting that the statisticalpredictions of a superposed state likeMEAS are forall prac-
tical purposes indistinguishable from the statisticalprediction of a mixture of its two
components (which allegedlyis non-problematic). But the problemis that given the
eigenstate-eigenvaluelink, states like MEAS areones in whichthe statisticaloutcomes
are non-existent.
This claim can be appreciatedby noting that linearity leads to the state MEAS and
the eigenstate-eigenvaluelink entails thatin this state the pointer observableis not well
13 YonNeumann (1955).
14 Thereis a
traditionof responding to the foundational problems of quantum theory by
arguing for a non-classical logic with a structure that mirrors the algebra of quantum
mechanical operators. Iknow of no such proposalthat actuallysucceeds in resolving
the measurement problem.
SeeMaudlin (1993).
16 Thereis another way of pursuing the strategy
of modifying linearity, which does a

much better job of dealing with the measurement problem, due to Ghirhradi, Rimini,
and Weber anddevelopedby PearleandBell. According to this strategy thereis a single
non-linear dynamic whichgoverns the evolutionof thequantum state. The basicidea is
thata system's wavefunction usually evolves in conformity withSchrodinger's lawbut
everyonce in a while, with aspecificprobabilitythatdepends onthenumber ofparticles
in the system, the wave function is multiplied by anarrow Gaussian. The probability
thattheGaussianis centeredat a point x depends on the amplitudeof the wavefunction
at x. Since the chance of a collapse (multiplicationby the Gaussian) is very small
for systems with few particles this theory is practicallyindistinguishablefromstandard
quantum theory for such systems. When a quantum system becomes correlated with
a macro-system the chance of a collapse is very large so states like MEAS are very
unstable and soonprobabilisticallyjump into states which (approximate)states which
are eigenstatesof pointer position. For discussions of GRW see JohnBell(1989) and
Albert and Loewer (1992).
17 Although,
as Bell observes, the title 'hidden variable' is not apt since the hidden
variables are usually whatis directly observed whilethe quantum state manifests itself
only in the statisticsof thehidden variables.
It receives no mention inRedhead (1987), Hughes (1989), or vanFraassen (1991).
Many ofthe claims foundin thesebooks arerefuted by the existenceof Bohm's theory.
See Albert andLoewer(1989).
19 Bell, J. (1989),
The Impossible Pilot Wave.
Bell(1989). The situationis a bit more complicated. First, it has becomepopular to
argue that the non-locality in orthodox quantum theory is not asbad as the non-locality
of hiddenvariabletheories. Sometimes it is claimedthat thelatterallows for superlu-
minal signalingand sois incompatible(or more incompatible?) withspecialrelativity.
But it can be shown that no theory which reproduces quantum mechanical statistics
can allow for superluminalsignalling and, for example, Bohm's theory does not. The
other complicationis that there is, in fact, one interpretationof quantum theory which
solves the measurement problem and is thoroughly local. This is the 'many minds
interpretation' due to David Albert and myself (Albert and Loewer, 1988). There is
no conflict with Bell's argument since this interpretation does not validate one of the
assumptions whichBell makes.
This claimis not exactly correct. Any collapse theory differs inprinciple from anon-
collapse theory likeBohm's, but in orthodox quantum theorycollapses are assumed to
occur only when systems involve many particles and for such systems it is not practi-
cally possibleto empiricallytest whether a collapsehas occurred. See Albert (1992).
Noticeableoverlap for many-particle systems is enormously difficult to bring about
experimentallyand probably seldom occurs for any length oftime. But in few particle
systems it is whatis responsible for characteristically strangequantum phenomena. For
example,in the famous two-slitexperiment the wave functionof a particle is literally
split, with components going through eachhole. The particle travels with one or the
other component. The apparatusis setup in such a way that the components overlapat
the screen. This overlap is what accounts for the interferenceeffects. SeeBohmand
Hiley (1992) for a detailedanalysis.
24 A
physical quantity is said to be 'contextual'iff theresultof measuring it depends not
only on its value, if any, buton othercontextual featuresof the measurement situation,

e.g. whatother quantities are beingmeasured.

25 There are stochastic theories similar to Bohm's. An interestingdiscussion of proba-
bility in Bohm's theory is found in Goldsteinet al. (1992).
26 Suppes (1972), p.
27 Thanks
to David Albert.


Aharanov, V and Albert,D.: 1981, 'Can We Make Sense of the MeasurementProcess

inRelativistic QuantumMechanics?', Phys. Rev. D., 21(2), 359-370.
Albert,D.: 1992, QuantumMechanicsandExperience, Cambridge:HarvardUniversity
Albert, D. and Loewer,B.: 1988, 'Interpreting the Many Worlds Interpretation',Syn-
these,77, 195-213.
Albert,D. andLoewer, B.: 1989, 'TwoNo-ColpaseInterpretationsofQuantumMechan-
ics',NOUS,23, 169-186.
Albert, D. andLoewer, B.: 1990, 'Wanted Dead or Alive: Schrodinger's Cat', in: A.
Fine, M.Forbes, and L. Wessels (Eds.), Proceedings of Science Association 1990,
VolumeI, East Lansing: Philosophy of Science Association.
Albert, D. and Loewer, B.: 1991, 'The Measurement Problem: Some Solutions',
Synthese, 81, 140-152.
Albert, D. andLoewer,B.: 1992, 'Tails of Schrodinger's Cat' (unpublished).
Bell, J.: 1989, The Speakable and Unspeakable in QuantumMechanics, Cambridge:
Cambridge Univ. Press.
Bogdan, R. (Ed.): 1979, Patrick Suppes, Dordrecht: Reidel.
Bohm, D.: 1952, 'A Suggested Interpretationof Quantum Theory in Terms of Hidden
Variables: Part I', Phys. Rev.,85, 166-179.
Bohm,D. and Hiley, D.: 1992, The UndividedUniverse,London: Routledge.
Daneri, A., Longer, A., andProsperi, G. M.: 1962, 'Quantum Theory ofMeasurement
of Ergodicity Conditions',Nuclear Physics, 33, 297-319.
Einstein, A., Poldosky,8., andRosen, N.: 1935, 'Can Quantum Mechanical Descrip-
tionsof Physical RealityBe ConsideredComplete?',PhysicalReview,47, 777-780.
Gleason, A. M.: 1957, 'Measures on theClosed Subspaces of a Hilbert Space',Journal
ofMathematicsandMechanics, 6, 885-893.
Goldstein, S.,Detles,R., Durr,D., and Zangi,N.: 1992, 'QuantumEquilibrium andthe
Origin of Absolute Uncertainty',Journal of StatisticalPhysics, 67, 843-907.
Hughes, R. I. G.: 1989, The Structure and Interpretation of Quantum Mechanics,
Cambridge: Harvard University Press.
Kochen, S. and Specker, E.: 1967, 'The Problem of Hidden Variables in Quantum
Mechanics', Journal ofMathematicsandMechanics, 17, 59-87.
Maudlin, T: 1993, QuantumMechanics andRelativity, Oxford:Basil Blackwell.
Nelson, E.: 1967, Dynamical Theories of Brownian Motion, Princeton, NJ: Princeton
University Press.
Pearle,P.:1992, Talk atConferenceon theFoundationsof QuantumTheory at Columbia
University (unpublished).

Redhead, M.: 1987, Incompleteness,Nonlocality,andRealism:A Prolegomenon to the

Philosophy of Quantum Metaphysics, Oxford: ClarendonPress.
Suppes, P.: 1961, 'ProbabilityConcepts inQuantum Mechanics', Phil, of Science, 28,
Suppes, P.: 1962, 'The Roleof Probabilityin Quantum Mechanics', in: B. Baumrin
(Ed.), Philosophy of Science, the Delaware Seminar, Vol. 2, New York: Wiley,
pp. 319-337.
Suppes, P.: 1966, 'TheProbabilistic Argument for a Non-Classical Logic of Quantum
Mechanics', Philosophy ofScience, 33, 14-21.
Suppes, P.: 1979, 'Self-Profile', in: R.J. Bogdan(Ed.),Patrick Suppes, Boston: Reidel.
Suppes, P.: 1984, ProbabilisticMetaphysics, Oxford:Basil Blackwell.
vanFraassen, B.: 1991, QuantumMechanics: An Empiricist View, Oxford: Clarendon
yon Neumann, J.: 1955, TheMathematicalFoundationsofQuantum Theory,Princeton,
NJ:Princeton University Press.
Wigner, E. P.: 1932, 'On the Quantum Correction for Thermodynamic Equilibrium',
Phys. Rev.,749-759.


Thereis muchthat Barry andI are in agreementabout concerningquan-

tum mechanics, especially in our critical view of the weaknesses of the
orthodox interpretation of quantum mechanics and in our skepticism
toward the need for any nonclassical logic or probability - a matter on
which I have changedmy position, as he indicates. Moreover, we are
alsoin agreement about staying within the general confines of classical
realism. For reasons that Iwill state, Ido not share his enthusiasm
for Bohm's theory, although I certainly do agree that it is one of the
important hidden-variable theories andhas undoubtedly beenmuch too
neglectedby philosophers of quantum mechanics. There is much to be
saidabout almost allof theissues he discusses,at least, much to be said
from my own viewpoint toward the foundations of quantum mechanics.
However,I willrestrict myself to four generaltopics because they repre-
sent the matters on whichhe and I differ the most and therefore it willbe
most useful to discussin detail. I dohave the feeling that with sufficient
time for analysis we would come to have positions that are relatively
close, both of us probably modifying some of our current views. The
reason I say this is that Loewer and I begin with such an agreed-upon
general framework for quantum mechanics. The prediction isalsobased
uponmy experience withhimas a graduate student many years ago. He
has always been prepared to defend in detail unorthodox or unpopular
views,but his underlying outlook isreasonable andamenable to making
revisions tomeetdetailed criticism. Ihope these are characteristics that
I share, at least in part.

Difficulties ofBohm's Theory. There are at least two major aspects of

Bohm's theory that disturb me. One is the existence of n-dimensional
wave functions (for large n in many-particle systems) that never col-
lapse. Both the physical interpretation of such waveseven for a single
particle, andeven more the problem of their not collapsing, are matters
that I find difficult to have any sort of feeling for. Unlike Barry,Ireally
do think of the wave functions that satisfy Schrodinger's equation as
being wave functions that are probability amplitudes -the exact charac-
ter of these distributions is discussedmore thoroughly below. The point
is that physical existence,except as probability distributions,especially
in n dimensions for many particles, is not something that I have any
physicalunderstanding of at all.
Second, Bohm's introduction ofa quantum potential which leads to
expressions for the velocities of particles, is also something thatI find
hard to have a feeling for. It seems too much like a deus exmachina.
How are we to think of this potential, whichin many ofits characteristics
is like a classical potential? In the classical case, however, we have a
good feeling for the physical source of the potential. In the quantum
mechanical case, however,I at least do not.
Third,Iam skeptical as to how the Bohm theory can be extended to
accountfor thebehavior ofphotons andelectrons in their usual quantum
electrodynamical framework. More generally, it is not clear to me how
theBohm theory canbe successfully extended to arelativistic setting.
I do want to emphasize that I am not in principle against Bohm's
theory and maybe ifI understood itbetter I wouldbe more sympathetic.

Quantum Mechanical Probabilities as Averages. As I

have empha-
sized on several occasions,but most recently in Suppes (1990), I think
of the probability distributions that arise from the Schrodinger equa-
tion as average or mean distributions. Consequently, the application of
the results to the behavior of single particles is to be treated carefully,
though it is not inconsistent to do so. The point is that the information
contained in the distribution is very weak. We cannot, for example,
from the probability distributions arising from the Schrodinger equa-

tion, compute autocorrelations of positions for a given particle. What

we really get is for each time t a mean distribution for particles satisfy-
ing the givenphysical conditions;for example, that ofa free particle, a
particle that is a harmonic oscillator, a particle in a Coulomb field, etc.
Consider,for example, the free particle as the simplest case. Letus spec-
ify some distributions at time to, and we then get from the Schrodinger
equation themean distribution at any other time t,past or present. This
mean distribution just tells us the distribution at time t. Itdoes not tell
us about how the distributions across time are related. This would go
beyond the formalism of standard quantum mechanics. What is impor-
tant about thesemeanor average distributions is that they are capacious
in the sense that they can accommodate many different hidden-variable
theories. This is a familiar fact about theories that are formulated only
in terms of mean distributions. Many different detailed theories can be
developed in a way such that the amplification is consistent with the
original given mean distributions.
Also notice that in the relativistic theory of quantum mechanics
this mean view does not really hold up too long, for something like
Feynman's propagators arise naturally and are constantly used. But
a propagator is just a way of studying aspects of the trajectory of a
particle. Note also that in standard quantum mechanics what amounts
to the kindof conditionalization that takes place in a propagator is not
really admitted. We cannot meaningfully conditionalize on the mean
distributions that are solutions of the Schrodinger equation. It is good
that we cannot, for conditionalizing on mean distributions in general
does not yield a meaningful result, but any more detailed theory, i.e.
any hidden-variable theory, is crying out for such conditionalization
as a legitimate and appropriate method for thorough investigation of
theoretical details. If we can say something about where a particle has
been, we can say a lot more about where it will be in the future. It is
perhaps the aspect of quantum mechanics that seems the most peculiar
from the viewpoint of probability andclassical physics, but if we think
of it in terms of mean probabilities it is not peculiar at all. It is just a
weak theory that does not have anything to say about trajectories but,
contra Bohm in his more extreme philosophical moments, this does
not mean at all that such trajectories do not exist. It only means that
classical quantum mechanics can have little to say about them. In fact,
it does not take much experience with the talk ofphysicists to know that
the concept of particles having trajectories is uneliminable from their
intuitive and experimental discourse.
Iemphasize then that the complex wave functions that arise in clas-
sical quantum mechanics require no peculiar interpretation. When
squared, they are just mean distributions at a given time. I realize
also that this needs detailed development for cases not involving the
distribution of position, orinvolving particles that must be givenarela-
tivistic treatment,as in the case of photons. But it is not really possible
to present further details in the present context.

Other Hidden-Variable Theories. Although it is by no means my

favorite,asI have alreadyindicated,I considerBohm's theory an impor-
tant one. It must be taken account of by anyone who is serious about
hidden-variable theories. I want to now turn to some ofthe alternatives.
One that I was devoted to for some time is stochastic mechanics,
developedespecially by EdwardNelson (1967,1990). Nelson wants to
start with classical mechanics andadd Brownian motion, sohe derives
in a natural way a diffusion for particles and from this he is able to
develop a good partof classical quantum mechanics.
On the other hand, critical parts are not really satisfactory. For
example, the analysis of interference is not persuasive, and more gen-
erally the account he gives of wave phenomena is not worked out in
detail, and only the virtuosity of some of his explanations gives it an
air of plausibility at all. Moreover, it is even less clear how stochastic
mechanics canbe extended to therelativistic case andprovide anything
like a proper treatment of photons and their interaction with electrons.
Albert and Loewer (1989, 1991) introduce a novel 'many minds'
hidden-variable theory which takes off from Everett's 'many worlds'
interpretation of quantum mechanics. Certainly this is perhaps the most
novelof any of the hidden-variable theories to be seriously discussed,
but it is too new and too strange for me to yet comment on seriously.
My own taste is too much for trying toaccountfor quantum phenomena
in a much more austere and more standard physical setup. It would,
above all, be useful to understand how Albert andLoewer propose to
extend their theory to the standard relativistic phenomena of photons
and electrons.
Finally, I come to my own current favorite view. I have had a
conversion late in life,away from stochastic mechanics and the use of
diffusion, to theuse of random walks from which one can derive wave

equations rather than diffusion equations. Given the importance both

historically andconceptually of wave phenomenain quantummechanics
it seems to me that this is essential for anyone like myself who wants
to begin from classical probability and as much classical physics as
possible. This particular approach is as yet not sufficiently developed
to permit a really detailed evaluation of its strengths and weaknesses.
My own work on itis being done in conjunction with a youngBrazilian
physicist, J. Acacio de Barros, and our first paper is about to appear
(Suppes and de Barros, in press), but this first qualitative analysis has
already been modified along the following lines.
Here I will just briefly sketch our approach. We take as a critical
case diffraction andinterference phenomena,especially as exhibited by
photons. The four fundamental qualitative assumptions are these:
I. Photons have velocity c.
11. Photons movein straight lines following classical paths inabsence
of charge.
111. In thenear presenceofa charge,a photon has apositiveprobability
of being scattered at some angle.
IV. Photons are emitted uniformly in alldirections by a harmonically
oscillating source.
When these assumptions are spelledout in mathematical detail, stan-
dard results can bederived, but now the wave properties are themselves
derived as properties of mean probability distributions p(x, y, z, t),
which, in the absence of charge,satisfy Huygens' wave equation and,
with the presenceof charge,satisfy Maxwell's equations.
One thing that has motivated us in constructing such a hidden-
variable theory at this stage is that we shouldtake account ofthe standard
relativistic behaviorof photons andelectrons. Itis toomuch at this stage
to claim that wecan carry the day with this theory, but we do think that
its development, even if conceptual modifications will be needed, is
very much worth pursuing. Notice that the ideas are very different
from Bohm's and very much driven from the standpoint of standard
probabilistic ideas. In my view there is much to be said for a theory
that assumes particles have trajectories but not necessarily the random
walks ofclassical diffusion.

Determinism Is Transcendental. Again I want to confess a change

in mind that has been underway for several years but now has firmly
crystallized. It is a change from the view that I held in Probabilistic
Metaphysics (1984) andis thebelief that the choicebetween determin-
ismor indeterminism is transcendental,in Kant's sense of transcending
experience. Barry rightly points out that in my former views I would
have been against the determinism of Bohm's theory and might well
have leaped upon the necessity of objective chance, as he describes
it, as a refutation of the theory. Now Ido not feel that way at all. I
am comfortable with all kinds of probability, includingobjective prob-
ability, within a deterministic framework. Two kinds of fundamental
results have persuaded me, bothof whichI have expounded on several
occasions, but have particularly summarized in a recent paper on the
transcendental character of determinism (Suppes,1993).
It is in this respect that my ideas really do differ from Loewer's
concept of separation of determinism and probability. Deterministic
systems do not have to be content only with epistemic probabilities.
A good example is the generation of arbitrary random sequences by
the motion of selected deterministic systems of three bodies operating
only under gravitational forces. I discuss such an example in some
detailin Suppes (1987). The secondkind of theorems are those proved
by Donald Ornstein and his colleagues concerning the isomorphism of
indeterministic systems and mechanical systems of given complexity,
for example,billiard tables withconvex obstacles. These isomorphisms
or the randomness in three-body problems do not depend at all upon
ignorance of initial conditions or the many other kinds of epistemic
probabilities introducedeither into statistical mechanics or inPoincare's
method of arbitrary functions for analyzing roulette wheels or throws
of dice. These objective probabilities, or propensities if you will, live
nicely and happily within systems that are as deterministic as any we
ordinarily count as such.
Of course, this does not mean thatdeterminism is true. It just means
that we have a transcendental choice for a wide variety of phenomena
between determinism or indeterminism. Bohm'shidden-variable theory
is an opportunity to make a deterministic choice. My transcendental
metaphysical prejudices run the other way, and I suppose I would opt
for indeterminism just because of its less exotic and more minimalist
features. But that is not important from a scientific standpoint. What
is important is to provide a proper probabilistic formulation, whether
the underlying probabilities arise from deterministic or indeterministic

processes,a matter we shall not be able to choose between on thebasis

of empirical data.
So, in my own view,thedemonof determinismin quantumor human
affairs andits twin demonofindeterminism have finally beenput to bed,
to lie in peace forever covered by aneo-Kantian quilt thatcannever be
penetratedby experience.


Albert,D. andLoewer,B.: 1989, 'TwoNo-CollapseInterpretationsof Quantum Theo-

ry',Nous, 23, 169-186.
Albert, D. and Loewer, B.: 1991, 'The MeasurementProblem: Some "Solutions'",
Synthese, 86, 87-98.
Nelson, E.: 1967, Dynamical Theories of Brownian Motion, Princeton, NJ: Princeton
University Press.
Nelson, E.: 1990, QuantumFluctuations, Princeton, NJ:Princeton UniversityPress.
Suppes, P.: 1984, ProbabilisticMetaphysics, Oxford:Basil Blackwell.
Suppes, P.: 1987, 'PropensityRepresentationsof Probability', Erkenntnis,26, 335-358.
Suppes,P.: 1990, 'ProbabilisticCausality inQuantumMechanics' ,JournalofStatistical
Planning and Inference, 25, 293-302.
Suppes,P.: 1993, 'The TranscendentalCharacterof Determinism', MidwesternStudies
inPhilosophy XVIII, 242-257.
Suppes, P. and de Barros, A.: in press, 'A Random Walk Approach to Interference',
InternationalJournalof TheoreticalPhysics.


[TJhere is no probabilistic causality in quan-

tum mechanics, for the essence of causality is
to relate behavior at one time to behavior at
anothertime (Suppes, 1990).

ABSTRACT. Schrodinger'sversionof EPR involves measuringa differentobservable

on each of the two component systems and using the strict quantum correlations to
infer values of theunmeasured observables. This procedure assigns values to all four
observables. We explainhow a recent result by Peres shows thatthis cannot be done
withoutviolatingthe product rule for commuting observables. We connect this, briefly,
with speculations about restoring some sort of causality in the quantum theory.


Can behavior at different times be related? As Suppes notes, the quan-

tum theory calls that into question, even for something so simple as
the location of a system at two different times. Position operators for
different times do not commute. There are no joint distributions for
locations at different times. Still, can we not somehow hold the loca-
tions together,at leastin thought, soas to compare them? Conventional
wisdom answers, 'No. This essay describes the fate of a tempting way
of trying to get beyond this conventional wisdom.


Almost immediately after the publication of the EPR paper in May of

1935 (Einstein et al,1935), Schrodingerbegansoliciting reactions to it
from many of the leading quantum theorists. He shared his amusement
over their sometimes contradictory responses with Einstein, writing to
him on July 13,
[I]t works as wellas a pikein a goldfish pond andhas stirredeveryoneup. ...
Itis asif
one person said, "It is bittercold in Chicago;" andanother answered, "That is fallacy,
it is very hot in Florida." (Quoted inFine, 1986, p. 74.)

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 29-43.
© 1994 Kluwer AcademicPublishers. Printed in theNetherlands.

During thatsummer Schrodingerhimself wrotetwo papers(Schrodinger,

1935a,b) inspiredby EPR wherehe analyzedandrefined the analysis of
what he called quantum 'entanglement' (Verschrankung), and he later
wrotea third paper on the same theme (Schrodinger, 1936). Inboth of
these 1935 discussions (the substance of his (1935a) analysis appears
againin (1935b),Sections 12 and 13) Schrodinger notes that in theEPR
situation we could determine simultaneous positions and momenta for
each of the entangled systems, "one by direct observation,the other by
inference from an observation on the other system" (1935a, p. 561).
That is, if we measure position on one system and momentum on the
other then we can use the quantum correlations to infer, respectively,
momentum on the one systemand position on the other, soas to assign
values to all four observables. Irefer to this variant of EPR, involving
measurements onboth systems and inferences across them to yield val-
ues for allfour observables,as Schrodinger's version. (See also Popper,
Using the analogy of a student in an examination who responds cor-
rectly to any questionposed,Schrodingerconcludes that in his version
ofEPR allfour observables really dohave definite simultaneous values,
just as we would naturally assume our student to know all the answers.
This only deepens the puzzles of EPR, however,because these values
do not relatein the expected ways. For instance, for any number b > 0,
Schrodinger notes that the eigenvalues of the operator (p 2/b + bx2) are
odd integral multiples of Planck's constant h. But if px and xx were,
respectively, the values of pi and xi onone system,it is simply not pos-
siblefor the valueof the correspondingexpression (p2 /b+ bx\) to be an
odd integral multiple of hfor everypositive b. Thus the values assigned
to non-commuting operators cannot be combined in a way that corre-
sponds to therelations between the operators themselves. In particular,
the correspondence fails for combinations of sums and squares.
The situation, Schrodinger argues, is even more paradoxical. For,
despite this failure of correspondence,the valuesof functions definedon
the separate systems relate justas if'the assigned momentaandpositions
did combine in the expectedmanner. Thuslook at the operators
-x 2 and p=pi + p2,

where the subscripts refer to the two entangled systems. Operators x

and p commute, so they have a simultaneous ip in which they take
eigenvalues,say, x and p. Now consider the operators A = /(xi,pi)
and B = f(x2 + %,P
— P2), where /(",") is any 'nice' two-place

function. Schrodinger shows that (A B)ip = 0andhe concludes that
in the stateip the valueof A is equal to the value ofB.Hence the values
of these functions are connected just as if'the value p of p wereidentical
to the sum of the value of pi plus the value of p2,and similarly for the
value x of x- although weknow that these connections cannot actually
obtain in general.
Let me draw these various elements together. (1) Schrodinger pro-
poses a varianton theEPR situation where wemeasure different observ-
ables on each system andinfer values for the correlated observables on
the other system. He concludes that allfour observables that enterinto
EPR (the positions and momenta for both systems) can be assigned
simultaneous values. (2) He proves that the following rules relating
these assigned values cannot both hold in general,
val(A + B) = val(A) + val(B) (SUMRULE)
val(A2 ) = [val(A)] 2 (SQUARERULE)
where the observables A and B do not commute. (3) He shows that,
paradoxically, values of observables defined on separate systems, nev-
ertheless,do relate as though the sum rule did hold. Indrawing this last
conclusion he makes use of the following inference. In the case where
observables A and B commute, infer from the fact that the value of
(A — B) is zero, that the value of A is equal to the value of B. Note
that this is, roughly, to apply the sumrule to commuting observables.
Schrodinger's discussion may call to mind yon Neumann's famous
no-hidden-variables proof (yon Neumann, 1955, 1V.2). The similari-
ties are striking. In his no-go theorem, yon Neumann shows that the
sum rule (for non-commuting A and B) is inconsistent with the square
rule (which is the principle that, in yon Neumann's terms, the ensem-
ble is dispersion-free). Schrodinger's example for proving (2) above
is actually a simple (and improved) way of demonstrating exactly the
same result. In the July 13 letter to Einstein, written while he was
preparing these 1935 papers, Schrodinger calls Einstein's attention to
yon Neumann's book, which he says he has been studying. More-
over, Schrodinger's (1936) uses results onmixtures that yonNeumann
developed in the book in tandem with his discussion there of hidden
variables.l It is,therefore, no accident that in exploring the issuesraised
by EPR, Schrodinger makes use of these functional rules featured by
yon Neumann. In trying to deepen the paradox of EPR, however,

Schrodinger goes one step further andconsiders the implications of the

sum rule applied only to commuting observables. He is on the verge
here of an important result, although one that casts into doubt his own
understanding of the Schrodinger version of EPR.


With hindsightit is easy to see,by means of exampleslike Schrodinger's,

that the yon Neumann sumruleisan arbitrary andunreasonable require-
ment to place on the assignment of values to non-commuting observ-
ables. It might, nevertheless,seemlike areasonable condition for com-
muting observables, as Schrodinger apparently took it to be. Results
achieved independently by Bell (1966) andKochen andSpecker(1967),
however, can beused to show that if we restrict the possible values of
an observable to the spectrum of the corresponding operator, then there
is no way to assign values to a certain finite set of pairwise commut-
ing observables in satisfaction of the sum rule (Fine and Teller, 1978,
Proposition 4). Itisan easyconsequenceof this result onsums to prove
that the following product rule for commuting observables A and B is
val(A " B) = val(A) " val(B). (PRODUCTRULE)

For suppose that values satisfying thisrule are already assigned accord-
ing to some function val(-) from the spectrum of operators. Then
we can define a new assignment of values, VAL(-) by requiring that
VAL(X) = log2[val(2x )]. Since val(2x ) is in the spectrum of 2X by
assumption, log2 [val(2x )] will be in the spectrum of X as required.
Moreover, VAL(-) satisfies the sum rule since the val(-) function is
assumed to satisfy the product rule. Thus the inconsistency of the prod-
uct rule follows from the inconsistencyof the sumrule (Fine and Teller,
1978, Proposition5).
This simple demonstration of the inconsistency of the product rule
piggybacks on the considerably more complex proof of the Bell-
Kochen-Speckerresult.2 In aremarkable paper Asher Peres shows the
inconsistency of the productrule directly andmoreperspicuously (Peres,
1990). The starting point for Peres' proof is a Schrodinger version of
theBohm spin-^ realization of theEPR experiment schematized below,
where the crs are the spin-component operators (or, ambiguously, the
spin observables) defined on the two subsystems (System 1 and Sys-
tem 2).

The state ip of the two-particle composite systemis the 'singlet state',

which can be written as .

0= -frttt^ ® («)] "

775^1 («) ® </>2»l
where a may be any direction (including x, y, or z) and where

cr^ = the spin component in direction aon System 1

state 0]*"(a)
J +1
— 1 in
\ in state.0f(a)
cr2 = the spin component in direction aon System 2
j +1
— in state
\ 1 in state
In state -0 the total spin in any directionis zero; i.e., if the spin is found
to be 'up' (+1) in direction a on one system it will be -down' (-1) in
direction a on the other system, and vice versa. (Put otherwise, the
product of the spin components on the separate systems in the same
directionis always -1.) Suppose that weconsider a Schrodinger version
of the experiment and make a measurement of axl on System 1 and
(simultaneously) of <x 2 on System 2, finding, respectively, values a and
b. We can then cross-infer values for cr2 and for <x* ,to get the results
as follows.
l, cr 2, <r 2,
Observables: trx
a, 6, —a, —b
Since the only possible values for a and b are ±1, if we multiply them
together the product (ab) will alsobe ±1, and its square (ab)2 = 1. So,

we have
val (<r x <rl)(crl - <rl)

= val (<r\ 2 ) " val (<r l " cr2 )

■ (T
y y
= val (<ri) " val (cr 2) " val (<rj) " val (<r2 )
= a- &"(-&) -(-a)
= (ab)2.

(1) Val [(0"i-cr2)(crl.cr2)] =1

However, by using standard relationshipsbetween the various spin oper-
2 2
ators we can calculate the value of the product \(<rx " cr )(o~y )] in a
second way. Firstof all,the spin operators in orthogonal directions (like
*x* and V above) on a single system (say, System 2) anti-commute;
that is,

(a2x a2y ) = -(a\

■ ■ a2x ).

Also, spin in the z-direction is related to spin in the x- and

by the equations
<T\ Z(cr^ " (T
2 = l((T
"/ 2 " (T
z x

where i2 =— 1. Using [cr 2x " a 2) = —{cr 2 " cr2

) we can rewrite the
expression for cr^,
= ~i(<rl " °i)
Using the fact that (<r^ " cr 2 ) = (cr2 " ay ) we can readily calculate
[cr\ " cr 2,) from the above relations as follows.
(al " a2z ) = i(a\ a\) (~i)(a2y a2x )
■ ■ ■

= ~(f)(a\
a\)(a2y a2x) ■ ■

= (al-a )(a'y -ax). 2

Thus,reading from far left to far right.

(2) val (alz " cr2 ) = val [(al

■ a2y )(a\ ■ a2x ) .
According to (1), RHS (2) = +1.
Now we know that for direction z, like any other, the spin values on
the two systems are opposite to one another,i.e., that
(3) val (azl " cr2 ) = val (alz ) val (cr 2 ) = -1.

(Indeed, the singlet stateip is an eigenstateof (azl " <x 2) with eigenvalue
-1, so the only possible value for (cr* " cr 2 ) is -1.)
So from (3), LHS (2) = -1. Since -1 1 (!), it follows that we
cannot consistently assign values as above.
Anexamination of thecalculations shows that the principles weused
to assign values are just these:
(SPECTRUMRULE) The only possible values for an observable A of
a system in a state ip are the eigenvalues of A that have non-zero
probability in ip.
(PRODUCTRULE) If A andB commute, then val (A >B) = val (A) "
val (B).
Thus Peres provides a proof of the following no-go theorem for a
Schrodinger version ofEPR:
Thereis noassignment of exact values to the quantum observables that
satisfies the spectrum and theproduct rules.


Schrodinger's version of an EPR experiment yields a simultaneous

assignment of values to all four observables, two by direct observa-
tion and the other two by inference. Peres' proof shows that this is
not feasible provided, like Schrodinger, we also suppose that the values
so assigned to commuting operators follow the same algebraic rules as
do the associated operators. Following Bell (1966), Peres interprets
this as showing the inconsistency of the assumption that the result of
a measurement of an observable depends only on the observable itself,
together with the state of the measured system. Thus Peres interprets
the proof as demonstrating contextualism,in the sense that measured

values do depend on the choice of other simultaneous measurements

(maybe even non-local ones).
While a number of investigators have found some sort of contextu-
alism attractive when interpreting measurementresults in the quantum
theory, it is important to recognize that contextualism does not actual-
ly follow from the no-go-theorem. All that follows is that the values
assigned in Schrodinger's version of EPR fail the product rule. That
rule, like the sum rule and other algebraic constraints on assigned val-
ues, is not containedin the quantum theory proper. For quantum theory
has nothing to say about interpolated, unmeasured values. The theo-
ry, at least as it is usually understood, only concerns relations between
measured values. For example, consider the joint probability formu-
la for commuting observables A, B, their product (A " B); and their
respective values x,y, and z in any state ip:

P^(A = x,B = y,(A-B) =z)= 0, if z^x-y.

This assertion, whichholds for all ip,may seem to say that the value of
a product cannotbe different from the product of the values. Soit does,
provided that by 'value' we mean 'simultaneously measured value.
More precisely, the rule is that (with probability 1) in a simultaneous
measurementofall three observables A, B, and (A " B) ona system in
stateip, themeasured value of the product will always be the product of
the simultaneously measured values of the factors. This is certainly a
principle of the quantum theory, but itis not the product ruleused in the
no-go theorems. In thePeres proof above, for example, the rule applies
to products like (ay " cr 2 ) where the values assigned to the factors are
inferred andunmeasured. There is nothing in the quantum theory itself
that covers suchcases. So what the Peres proof and the other algebraic
no-go theorems show is that the quantumrules for simultaneously mea-
sured values cannotbe extended consistently to unmeasured, assigned
values. That is important to know for, like the earlier yon Neumann
and Schrodingerproofs, it rules out certain extensions of the quantum
theory. It is also important to realize,however, that (pace Bell,Peres,
and others) these no-go theorems do not single out any particular inter-
pretation ofmeasured valuesin theunextended quantum theory, neither
contextual nor non-contextual.
In his letter to Einstein of August 19, 1935, Schrodinger remarks
that to violate what Ihave called above the sum and square rules would
be "altering the connections with the concepts of ordinary mechanics"
(Fine, 1986, p. 79). Clearly Schrodinger does not regard this as desir-
able, even for sums of non-commuting operators. Inthese terms,Peres'
proof shows that it is possible to assign values in a Schrodinger ver-
sion of EPR only by making an even less desirable alteration of those


A fundamental question about the quantum theory, perhaps the funda-

mental question, is how to understand superpositions. If the state of
the system is not an eigenstate of an observable (nor a proper mixture
of such states), does the observable have a value in that state? The
accepted answer to this questionis 'No. Thatanswer isbased onacon-
sequentialist analysis. That is, onetries to extend the quantum theoryby
assigning values in superposed states and one shows that the proposed
extension leads to difficulties. Perhaps the best known analyses of this
type are Bohr's repeated treatments of the double slit experiment. The
Bell theorem too canbe looked at as an analysis of this sort. There we
suppose definite premeasurement values for a small set of observables,
values that satisfy further constraints, including a principle of locality,
and we show that the quantum statistics rule this out.3 Similarly, the
Peres theorem and the other algebraic no-go results also point to a neg-
ative answer. But all these results can dois to point,and not logically to
compel. For every proof and every experimental demonstration makes
use of assumptions beyond the assignment of values. So in the end
we are faced with a choice of whether to give up the assumption of
definite values, or some others. These are precisely the circumstance
that call for good human judgment (Duhem's bon sens) sinceneither all
our theory nor all our practice can determine which choices to make.
Finally, to return to the initial theme, what then of relating behav-
ior over time as required for causality? To do so involves programs
(like Bohm's, or perhaps a stochastic mechanics) that assign simulta-
neous values to non-commuting observables, thus allowing values in
superposed states. In this connection Suppes writes, "But all the possi-
bilities are not lost. If [the construction of a 'non-Markovian stochastic
mechanics'] can be carried through,probabilistic causality is restored
for quantum phenomena" (Suppes, 1990, p. 301). These programs run
counterto the direction of conventional wisdom, which is the direction

indicated by the no-go theorems. We will have to march the opposite

way if we insist on causality in the quantum theory. As just discussed,
it is not clear whether it wouldbe wise to perform such a countermarch.
More fundamentally, it is not clear whether judgment that insists on
causality in the quantum theory, even a probabilistic causality, is good

Department ofPhilosophy,
Northwestern University,
Evanston,IL 60208, U.S.A.

This essay was inspired by a comment of Patrick Suppes. In the spring of 1980 I
circulateda series of notes in one of whichIproved thatif there were a factorizable,
stochastichidden-variables theory, there wouldalsobea deterministicone. Responding
to that note, Suppes remarked that Iseemed to be a 'crypto-determinist'. Inever
understoodPat's comment, but Ithink thislittleessay is a reflectionon it.
1 Thereis amore detaileddiscussion ofthe Schrodinger-Einsteincorrespondenceover
EPRin (Fine, 1986, Ch. 5).
2 See (Peres, 1991)
for elegant and relativelysimple proofs of the Kochen-Specker
theorem. Inhis introductionPeres remarksthat theBell proof of this resultinvolves a
continuum of vector directions. While thisis true of Bell's exposition,his actual proof
only requires finitely many directions, as shown in(Fine and Teller, 1978).
Indeed(Fine, 1974) showshow to interpret theBell theoremas a direct proofof the
inconsistency of the product and spectrum rules.


Bell, J. S.: 1966, 'On theProblemof HiddenVariablesinQuantumMechanics', Reviews

ofModern Physics, 38, 447-452.
Einstein, A.,Podolsky,8., andRosen,N.: 1935, 'Can QuantumMechanicalDescription
of Physical Reality Be Considered Complete?',Physical Review, 47, 777-780.
Fine, A.: 1974, 'On the Completeness of Quantum Theory', Synthese, 29, 257-289.
Fine, A.: 1986, The Shaky Game: Einstein, Realismandthe Quantum Theory, Chicago:
University of Chicago Press.
Fine, A. and Teller,P.: 1978, 'Algebraic Constraints onHidden Variables',Foundations
ofPhysics, 8, 629-636.
Kochen, S. and Specker,E.P.: 1967, 'The Problem of Hidden Variables in Quantum
Mechanics', Journal ofMathematics and Mechanics,17, 59-87.
Peres, A.: 1990, 'IncompatibleResults of QuantumMeasurements', Physics Letters A,
151, 107-108.
Peres, A.: 1991, 'Two Simple Proofs of the Kochen-Specker Theorem', Journal of
Physics A,24, L175-Ll7B.
Popper, X.: 1959, The Logic ofScientific Discovery, London: Hutchinson.
Schrodinger, E.: 1935a, 'Discussionof ProbabilityRelations between Separated Sys-
tems', Proceedings of the CambridgePhilosophicalSociety, 31, 555-563.
Schrodinger, E.: 1935b, 'Die gegenwartige Situation in der Quantenmechanik', Die
Naturwissenschaften,23, 807-812, 824-828, 844-849.
Schrodinger, E.: 1936, 'ProbabilityRelationsbetweenSeparatedSystems',Proceedings
of the Cambridge PhilosophicalSociety, 32, 446-452.
Suppes,P.: 1990, 'ProbabilisticCausalityinQuantumMechanics',JournalofStatistical
Planning and Inference,25, 293-302.
yon Neumann, J.: 1955, MathematicalFoundationsofQuantumMechanics, Princeton:
Princeton University Press.


I have learned a lot from Arthur Fine's various papers on quantum

mechanics and the present oneis no exception. I have probably learned
even more from our many conversations over the years. What seems
appropriate hereis toengagein yetanother conversation withArthur. He
pretty well throws down the challenge to show that probabilistic causal-
ity should be taken seriously in a fully satisfactory theory of quantum
phenomena. It is obvious enough that standard quantum mechanics
does not provide such causal analysis, and on this point Arthur and I
It is important to mention one standard move that is often made.
It is to claim that standard quantum mechanics is causal because the
solutions of the Schrodingerequation are deterministic,in the sense that
the state at one time uniquely determines the state at another time, but
these states do not give anything like a standard view of causal analysis
in physics, that is, an analysis in terms of forces acting on particles
and thereby determining individual trajectories. There is, it must be
admitted, something to this particular view, for as the Hamiltonians
used in the Schrodinger equation vary, so do the solutions, and the
Hamiltonians represent the waysin which the idea of force is imported
into quantum mechanics via the Schrodinger equation. But still,all we
get are distributions and not a causal analysis of individual trajectories.
AsI have emphasizedinother commentson papers in these volumes,
what is really missing in standard quantum mechanics is the concept
of a trajectory for a particle. It is around this concept that we would

expect to find the full development of causalideas. This does not mean
that every aspect ofa trajectory is givena causal analysis. It just means
that changes in trajectories away from uniform motion are explained
directly by the impact of forces. I choose the word impact deliberately
for one radical version of causal analysis in physics, a Cartesian view,
is that the only forces that really exist are very short-range ones.
Arthur says at the end of his paper, in response to my remark about
non-Markovian stochastic mechanics beinga possible way of providing
a probabilistic causalanalysis of quantum phenomena, "these programs
run counter to the direction ofconventional wisdom, whichis the direc-
tion indicated by the no-go theorems. We will have to march the
opposite way if we insist on causality in the quantum theory." It is
evident enough, however,that the consideration just of no-go theorems
will not in any sense settle the issueof whether we can expect to havea
probabilistic causal analysis of quantum phenomena. As Arthur heads
the final section of his paper, it is indeed a question of back-to-basics.
The fundamental considerations here go much more to the heart of
quantumphenomena than the no-go theorems by themselves. What I
havein mind especially are the elaborate methods devised in quantum
electrodynamics for studying the interactions ofelectrons and photons.
Especially as conceptualized by Feynman, in general terms we have a
very natural probabilistic causal analysis. Even if the probabilities have
a somewhat strange feature, the causality has a relatively intuitively
straightforward feeling. This includes everything from path integrals
to the derivation of propagators which furnish a version of trajectory
analysis in quantum electrodynamics. I am not suggesting that all I
would like to see is to be found in quantum electrodynamics. For
example, there are no assumptions of definite trajectories for photons
or electrons but there are clearly assumptions about causal interactions
between photons andelectrons.
The point thatInow want to make in various detailed waysis thatthe
furtherrelativistic development of the interaction between electrons and
photons,a subject not handled in any satisfactory way at all by classical,
nonrelativistic quantum mechanics,inevitably and necessarily deals in
causal concepts. To say this is not to say that a clear classical concept
of trajectory is used, but rather that the classical, in fact preclassical,
concept of causal interaction between objects isused repeatedly.
To beginlet me quote some sentences from the abstract of a famous
1949 paperby Feynman.
Electrodynamics is modifiedby alteringtheinteraction of electrons at shortdistances.
All matrix elements are now finite, with the exception of those relating to problems
of vacuumpolarization. The latter are evaluated in a manner suggested by Pauli and
Bethe, which gives finiteresults for thesematrices also. The results then agree with
those of Schwinger. A complete, unambiguous, and presumably consistent, methodis
thereforeavailablefor the calculationof allprocesses involving electrons and photons
(p. 769).
As another example that does not require a quotation because of its
widespread use in all detailed discussions of quantum electrodynamics,
consider the talk that photons are absorbedor emittedin a givenprocess
or in a given region. This talk of absorption or emission, and the
detailed analysis that accompanies it is inevitably causal in character
and is ordinarily recognized as such in informal discussions of the
formalism. But still to give one particular reference, note the way that
Kallen (1972) describes a propagator function.
In this way the function actually gives a 'causal' description of the collision. It is
oftenreferredto in the literatureas the 'causal propagatorof the photon',or simply as
the 'photonpropagator"(p. 106).
Again notethe use of theconceptof collision,acausal concept ifever
there was one in physics. Note this typical passage in the well-known
bookon non-relativistic mechanics by Landau andLifshitz (1958)
Born's formula can beappliednot only to collisionsof two elementaryparticles,but
alsoto an elasticcollisionbetween,say, an electronand an atom, ifthe potential energy
U(r) is suitably defined. The condition for the Born approximationto be applicable
to such a collision requires that the velocity of theincidentelectron should be large in
comparison with those of the atomic electrons(pp. 426-427).
This passage, asI have already mentioned, occurs in a famous text-
book on classicalquantum mechanics andis in the chapter on the theory
of elastic collisions, which contains also a discussion of the general
theory of scattering. Notice what is going on in this case and why
probabilistic causality is more fundamental in the analysis of physical
phenomena even than the existence of detailed trajectories. I in fact
lay down the challenge to Arthur that it is not even a part of anyone's
conventional wisdom to thinkin any other thancausal terms about these
collisions,collisions of the kind that dominate the theory of scattering.
Of course we can, as inclassical mechanics, distinguishbetween elastic
collisions andthose which are not. Here thatdifference is not important,
although the mathematical analyses are quite different. What is quite
important for my point is that the thinking about interacting particles
is inevitably causal in character. This does not mean that every aspect

of causality is brought over from classical physics. It does mean that

causal concepts, the rock-bottom basics of collisions or interactions,
cannot be avoided.
A nice relativistic example of scattering and therefore of causal
processesis photon-photon scattering. In this case the scattering of one
photon by another is a quantum electrodynamicalphenomenon which
does not occur inMaxwell's classical electrodynamics due to the fact
that Maxwell's equations are linear. The typical diagrams for these
interactions, as in the case of other fundamental particle interactions,
are nothing if not causal in their conception and in physicists' ways
of thinking about them. Of course, this is even more obvious in the
more accessible phenomena of electron-photoninteractions. Both kinds
of interaction exhibit well enough for my purposes natural notions of
probabilistic causality.
Finally,in my continuingdialogue withArthur about the foundations
of quantum mechanics, I want to throw back at him his conventional
wisdom and say it is datedand appliesonly to thefoundations ofclassical
quantummechanics, and not even there in its extensions to scattering.
Probabilistic causality is alive and well in the actual physics used to
analyze quantumprocesses of interaction,and especially in Feynman's
particle conception of quantum electrodynamics.
Itis obvious that in this discussionI have gone wellbeyond classical
quantum mechanics in its standard formulation, becauseI think this is
the direction in which the philosophy of quantum mechanics should
move. Any discussion of basics must look at the wider context of
quantum electrodynamics and quantum field theory. I do not mean to
suggest also that my views are entirely accepted by everyone. I am
just making the case that the robust discussion of causal processes is
alive and wellin contemporaryphysics andisuneliminable from almost
anyone's discourseabout these matters.Itis undoubtedlydeepintuitions
about these matters, arising long ago in preclassical physics, that led
to Feynman's valiant attempts to develop a space-time approach and
thereby his path-integral approach to quantum mechanics. The defect is
classical quantummechanics,not the wide-rangingintuitive concepts of
collisions andinteractions usedbyphysicists. The theory of the future, I
would predict, will move in the direction of amore explicit treatment of
causal processes and not in thedirectionofclassical quantummechanics.

Feynman, R.: 1949, 'Space-Time Approach to Quantum Electrodynamics', Physical

Review, 76, 749, 769-789.
Kall6n,G.: 1972, QuantumElectrodynamics,trans, by C. K.Iddings andM.Mizushima,
New York: Springer Verlag.
Landau, L.D. andLifshitz, E.M.: 1958, QuantumMechanics:Non-Relativistic Theory,
Vol. 3, Course of TheoreticalPhysics, trans, by J. B. Sykes andJ. S.Bell,Reading,
MA: Addison-Wesley.


ABSTRACT. Difficulties in the measurement of physical quantities arise in classical

physics as soon as classical physics has to deal with waves in a concrete sense. The
concept of a monochromaticplane wave is completely and precisely defined, but it
remainsabstract. Concrete waves as collective objects, on the contrary, are burdened
with spectralinequalities,characteristic ofclassicalfield magnitudes.

Quantum mechanics defines physicalmagnitudes as operators acting on

a vector,the vector of state of a given system. Classical mechanics,for
example Lagrange's analytical mechanics,is completely foreign to this
conception. The evolution of a mechanism is completely determined
when the values ofits generalizedcoordinates qi andof their derivatives
with respect to time q\ are given at any time. The state of such a
mechanism is nothing less than the assignation of these geometrical
magnitudes (distances and angles) and these kinematical magnitudes
Itis true to say that quantummechanics has completely transformed
both of the concepts of physical magnitude and state. Nevertheless,
we must not too hastily give quantum mechanics what belongs to the
classical theory of fields.
The proper concepts of physical magnitudes as operators and of
states as vectors are relevant to such a theory. Quantum mechanics
steps in whenEinstein's and de Broglie'sequations associate properties
of particles with properties of waves.
My aimhere is to analyze the conceptualchange of the concept of
physical magnitude within the limits of classical physics. I shall first
describe the main characteristics of the field magnitudes. This sketch
will then suggest some philosophical conclusions.

The celebrated representations of Nature by the atoms and their com-

position in Greece and the less well-known analysis of the tides by

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 45-57.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.

Posidonius the Stoic testify that ancient physics had already recognized
the necessity of using two conflicting conceptualizations in order to
describe matter.
On the other hand reality is conceived as made of discontinuous,
localized, exclusive bits which move by transportation according to
well-defined trajectories. Physics has essentially to state the rules of
conservation that governthe collisions ofthe atoms and their aggregates.
Collisions require laws ofconservation relative to physicalmagnitudes
which may be locally defined. But the forces which finally yield their
collisions and secure unity for a world act at a distance. Newton showed
how to work with this hypothesis of separated centres of gravity. When
Newton's laws are rewritten in terms of gravitational fields, this intro-
duction of the field conceptualization means that the atomist uses a
formal trick rather than that he makes concessions.
On the other hand there are genuine phenomena of propagation
without transportation, where the magnitudes involved are continuous,
defined at every point of space, and, therefore, everywhere present,
and, moreover, intersect their respective paths without disturbing them.
Because it is continuous, this conceptualization is more akin than its
rival to the requisites of geometry. This is thereason why the followers
of Descartes and Huygens did not give ground despite the successes
of Newtonianism in mechanics until their apparently final victory with
Young and Fresnel. Compare, for example, the laws of impact of one
body against another one with the law of thebeat of two sound waves
in the air. In the first case, we deduce the laws from the principle of
conservation of the quantity of motion, without being able to describe
what happens at the instant of impact. By contrast, given two sound
sources with slightly different frequencies, we have only to make an
algebraic addition at each point in order to construct the figure of the
interference in time with allits particulars.
Geometry shows,however, that alot of surprises may be stored in a
completely intuitive construction. Letus review two of them under the
head of the categories of causality and substance.
When from a law of conservation a physicist predicts how two bil-
liard balls will rebound, he simply equates two accounts before and
after the impact. Thereis succession, there is not causality. Prediction
is still possible, but one should speak of global determinism rather than
of strict causal laws, since the balance of observable quantities before
and after an event does not imply that the event itself be subjected to
a detailed and continuous computation. Hume's doubt is to the point.
But add two waves. Each development is entirely determined once the
amplitude, the rate of temporal evolution (pulsation) and the rate of
spatialprogression (wave number) are fixed. And their superposition
obeys the same complete determination. We follow it as we follow
them. Causality is here the product of algebra. The surprise comes,
nevertheless,if we draw the physicalconsequences from the superpo-
sition, namely that two vibrations, when added, mutually reinforce or
extinguish themselves,according as the additionis made at two troughs
or at twocrests on the one hand, or at a trough and a crest on the other
hand. The astonishing phenomena of interference are thus explained
away without introducing Kant's so-callednegative magnitudes.1 With-
out forgetting that monochromatic waves belong to abstractphysics, not
to concretephysics, we mightsay, figuratively, that,putting together two
illuminations may produce obscurity. In the same way, it will be seen
that all the disturbances bound with a superposition of waves infinitely
extendedover space may be cancelled everywhere except over the tiny
extensionof a packet's centre.
As to the categoryof substance,thesurprise came off in two episodes.
First, in his Analytical Theory ofHeat, Fourier analyzed the diffusion
of heat without taking sides with the upholders either of the caloric
or of the kinetic theory. Whereas the physicists had to understand the
mechanisms according to which columns of air are displaced by the
propagation of soundin order to state and to solve the wave equation of
sound, the stationary distribution of the temperatures that results from
the diffusion of heat in a wall or in a ring is obtained though no mod-
el in terms of fluid or molecules is offered to support the distribution.
Auguste Comte, the founder of positivism, well understood the impli-
cations of Fourier's method: physicalmagnitudes can be measured and
known, while the substances that they make manifest are kept in com-
plete ignorance. Therefore phenomenal magnitudes, not substances are
relevant to mathematical physics.
The second episode is more telling still. In Fourier's perspective,
physical magnitudes were abstracted from their substances, but their
substances, whatever they might be, were not denied a physical exis-
tence. On the contrary, when themechanical theory of light, asconstrued
by Fresnel (who assimilated the vibrations of light with the transverse
elastic vibrations of solids and gave the ether, which was supposed to
support them, the paradoxical properties of absolute incompressibility

and solidity) was forsaken and Maxwell's electromagnetic theory was

adopted, it was gradually realized that field magnitudes need no sub-
stantial medium to which they should adhere. The last tie between
magnitudes and substances was broken.
Field magnitudes are superposable and may live an autonomous
existence independentofamaterialmedium. Theseunexpectedphysical
phenomenaresult from the geometrical constraint of continuity. A pure
analytical surprise, if I dare say such, following Wittgenstein's ukase,
still waitsfor us, whenboundaryconditions areplacedupon the solutions
of the wave equation, as is required by the possibility of experience,for
example, a plucked vibratingstring, a vibrating membrane, a system of
two coupled pendula, or the evolution of a particle in a potential well.
Fourier's trigonometric series decomposesany periodic function which
is regular enough into an infinite sum of harmonic functions affected
by suitable coefficients that determine the relative participation of each
function in the whole. Two analyticalphenomenarelative to such series
claim our attention: the emergence of eigenfunctions and eigenvalues
and the coupling of the magnitudes.
Owing to boundary conditions, discontinuity, so to speak, emerges
from continuity. When the vibrating string is fixed at two points, the
space part of the wave function, whichis an ordinary differential equa-
tion, is constrained to vibrate at certainnaturalmodes.2 The wave num-
ber, and with it the frequency,is allowed to take as values only integral
multiples of a constant function of the period. These values are called
eigenvalues. To each of them there corresponds an eigenfunction. The
eigenfunctions have two importantproperties. They are orthogonal and
complete. This means that they define a generalized vector space of
an infinite number of dimensions and thusafford aconvenient basis for
developingany regular enoughfunction thatsatisfies the same boundary
Understood as a theorem of generalized vector analysis, where the
elements of the vector space are the real, continuous functions of a
real variable defined over an interval,Fourier's decomposition has far-
reachingconsequencesfor the concept of physical magnitude. In ordi-
nary geometry a length is defined by the coefficients which multiply
its projections upon a Cartesian orthonormed basis. In the same way
let us suppose that the state of a system is characterized by a given
periodic function of a continuous variable. Then, owing to Fourier's
analysis, the spectrum of theFourier coefficients or eigenvalues, if the
eigenfunctions havebeenpreviously normalized, willdefine the parts of
a physicalmagnitude. As the spectra of eigenvalues and eigenfunctions
are producedby the action of certain operators upon the periodic func-
tion, physical magnitudes become associated with operators. A simple
example is given by the energy theorem. The energy of a wave over a
period is proportional to the squareof its amplitude over the period. The
operator, or rather the functional since the transformation is from a
vectorto anumber- actingupon the function of time ishere the definite
integral over the period. According to the theorem, the total energy of
a wave is simply the sum of the energies of allFourier's components.
This is precisely Pythagoras's relation.
Fourier's analysis may beextended to complex functions and tonon-
periodic functions. Inthe first case provisions are made for getting real
eigenvalues: the operators must beHermitian. In the second case, con-
tinuous spectra, integrals and Fourier transforms, respectively, replace
discrete spectra, series and Fourier components. The integrals which
occur in Fourier transforms are reciprocal functions of continuous vari-
ables: the wave vector and the position (since we took the space form
of the wave equation). Analytical considerations exhibit a classical
relation between the widths of two Fourier transforms. Their product
has an inferior bound.
Let AA; be the width of the scale of undulation, the inverse of the
wavelength,and let Axbe thecharacteristic height or spread of the one-
dimensional position, i.e., the space extension containing the centre of
the wave packet. The function ip(x,o) is obtained by integrating its
Fourier transform over k. If Ax is greater than l/Ak, the transform
oscillates several times within the interval Ak and the integral over
k takes a negligible value: there is destructive interference. If x =
xo, Ax being less than 1/AA;, the function which is integrated does
not practically oscillate and its integral over k takes an appreciable
value; the wave packet centre, where the amplitude (x,0) is maximal,
is situated at x = xq. Therefore, to form a wave packet with a limited
extension - to simplify let us consider an abstract wave packet by
replacing the concrete superposition of an infinity of plane waves by a
discrete interference of such waves, finite in number we need several
harmonic components of wavelengths, the range of which is the more
extended the more the wave packet is localized.3
The spatialspectralinequality couples the two physical magnitudes:
wave vector and position. But since physical magnitudes have been

associated with operators, the question arises: what operators are asso-
ciated with wave vectorsand positions, and what properties of operators
are responsible for the coupling of these magnitudes? Therelevant oper-
ators are shown, respectively, to involve a first derivative with respect
to space and a multiplication by the space coordinate. Each of them
is characterized by a spectrum of eigenfunctions and eigenvalues. The
basis afforded by the system of the eigenfunctions of one of themallows
the development of the other. But coupled operators do not admit the
same eigenvectors and do not commute. The magnitudes with which
they are associatedare therefore not simultaneously measurable with an
accuracy superior to a finite givenquantity.4
Analyzing the temporal form of the wave equation would produce
similar conclusions. A temporal spectral inequality should take the
place of its spatial counterpart. Insteadofarelation between the widths
of the waveand the position vectors, therelation wouldbe between the
widthof the pulsation spectrum and the widthof the temporal extension.
These limitations have nothing to do with quanta. They obtain with-
in thebounds ofclassical physics and even within thebounds of the part
ofclassicalphysics that precisely deals with continuous magnitudespar
excellence. Specific laws rulethe propagationofcollective phenomena,
the form of which may beconfined while the superposed waves fill the
whole space. The interferences act constructively only within a thin
volume around the centre of the wave packet and destructively every-
where else. As the group of waves runs towards either circle that limits
the perturbation caused by a stone thrown in a pond, this form may
get a proper velocity, different from the phase velocity. The velocity
of the maximum of the wave packet is not the mean phase veloci-
ty, since the component waves in a dispersive medium have different
velocities because of their different wavelengths and the interferences
slowly change the determination of thismaximum. This group velocity
depends on what spectral extensions the pulsation and the wave vector
of the wave packet have.5
Thereforeitis no wonder thatcommunication theory, which isentire-
ly grounded on classical principles, presents us with the new concept
of physical magnitudes associated with operators, when pulsations and
gains are measured by linear transmission of messages.6
The mutual relationbetween signal and transmission channels gives
rise to the concept of impedance, a physicalmagnitude whichis gener-
ally representedby an operator andis measurable when the signal isan
eigenfunction of this operator, the result of the measurement being the
associated eigenvalue.7


Both of the basic concepts of classical physics raise difficulties that are
bound up with conciliating discontinuity and continuity. A material
point means a finite quantity of matter, but without extensionand there-
fore infinitely concentrated. In so far as a wave is infinitely extended
and everywhere defined by finite magnitudes, it is not burdened with
a problem of singularity. The application of the laws of science is
nowhere prohibited. Huygens' light waves, which progress through
the void, or rather through the ether, with the same velocity may even
be said to be unproblematic. The paradoxes that affect the classical
concept of physicalmagnitudesbegin with Young's orFresnel's waves,
with dispersion.
Reading an equation - or rather into an equation - is not an easy
task. The solutions of the wave function are circular functions of the
difference between two products: of the wave vector into the space
coordinate, andof the pulsation into the time coordinate. At the begin-
ning the physicists supposedthat thepulsation whichis a function of the
wave vector always expresses the product of this vector with a constant
phase velocity. Abandoning this supposition led to important progress
in physical field theory during thenineteenth century.
This progress in a sense answers, or in another sense, questions
philosophicalpositions which go back to Descartes and Huygens on the
one hand, and to Kant on the other.
Eager to avoidNewton's singularities Kant introduced,besides ordi-
nary or extensive magnitudes corresponding to quantities and to the
axiomsof intuition,intended specific magnitudes, calledintensive mag-
nitudes and correspondingto qualities and to the anticipations of percep-
tion. What hehad in mindis explainedin abook by him which hasbeen
'undulyneglected'(exceptby Patrick Suppes8),theMetaphysicalFoun-
dations ofNatural Science, where, in the division of science of motion
between kinematics and mechanics, he inserts dynamics. Dynamics
relies on the opposition of two forces - attraction and repulsion - the
balance of which accounts for the size of the objects in the universe. Let
us remark that quantum mechanics will answer the same question by

putting forward Heisenberg'sinequality that results from the coupling

of the magnitudes occurring in the wave equation, once itis interpreted
in terms of quanta, and by postulating for identical particles a specific
behaviour which,in the case of antisymmetry, obeys Fermi's principle
This remark and Kant's own weaknesses of construction advise us
to replace Kant's so-called intensive magnitudes by the specifically
'dynamical' concept of a wave, a substitution hinted by Kant himself
with respect to colors and musical sounds.9 This move would give
Kant's dynamics its proper scope. All the physical magnitudes should
be conceived of as field magnitudes, what remains of corpuscular ideas
having to be explained away as figures of speech. Kant often insists
on the fact that continuity is built into the possibility of experience: an
object, he says, could not be known, were its properties not given as
physical magnitudes. But it is precisely those properties that should
not meet the requirements of continuity imposed by space and time
as forms of possible experience that could not be given as physical
magnitudes. The supreme transcendental principle goes so far as to
identify the possibility of experienceand the possibility of the object of
In other words, waves and their magnitudes do define any possible
In this view, when it is said that matter is opposed to radiation as
corpuscles are opposed to waves, this opposition results from a sheer
appearance, since everyphysical description must fit in with the condi-
tions of possible experience. Maybeatomistic hypotheseshave aheuris-
tic value. They are, however,deprived of ontological import. Phenom-
ena are continuous. The question whether reality in itself is atomic
or continuous, or whether theuniverse is finite or infinite bypasses the
power ofour knowledge andinvolves reasonin dialectical antinomies.
Such a view was generally accepted until the beginning of this centu-
ry. Duhem, for example, and withhim allthe followers of energeticism,
though they did not profess either that space and time be the subjective
forms of our intuition or that pure reason be affected with insupera-
ble contradictions, agreed that physical theories should only 'save the
phenomena' without speculating upon the so-called reality of atoms.
Recalcitrant phenomena compelled the physicists to renounce this
kind of phenomenalism, to admit the wave-particle dualism within the
facts and therefore the possibility of experience. Itisremarkable that the
revolution brought aboutin our notions of physicalmagnitude and state
originated from the very domain ofradiation where the classical devel-
opments and difficulties that havebeen described hadarisen. Quantum
mechanics may even be constructed from merging the concepts of func-
tion and vector, as is requiredby Fourier's generalizedanalysis, on the
conditionthat pulsation and wave vectorbe respectively associated with
energy and momentum through Planck's constant.
But letus stay within thebounds ofclassicalphysics andask aboutthe
meaning of classical field magnitudes. It is here that a distinction used
by Descartes and by all the Cartesians,Malebranche,Huygens,Spinoza
and Leibniz will be relevant, namely the distinction between abstract
physics and concrete physics. Its best and ideal illustration (since the
mathematical and physical developments came later) is probably given
by analyzing the concept of wave.
A simple wave is an elementary harmonic wave, i.e., a sinusoidal
function whichmoves allinone piece with aphase velocity that depends
on thematerialmedium. Whatdefines theabstract individuality of these
elementary waves are the quantities by which the different periodici-
ties are determined (period and pulsation for time, wavelength and
wavenumber for space). A successive section shows equal amplitudes
each time the variable t has grown a period and an instantaneous sec-
tion shows equal amplitudes each time the space variable has grown a
wavelength. Such waves belong to abstract physics. According to the
temporal spectral inequality, the temporal extension of the sinusoidal
wave is infinite with a unique pulsation. But a monochromatic plane
wave has no physical reality. Only superpositions of harmonic waves
may be realized.
Concretephysics then deals with composed or collective objects. But
whileabstract individual objects were welldefined, difficulties andpara-
doxes arise when concrete individuals,that is groups, are considered. 10
Monochromatic waves were well defined in so far as one precise value
of the pulsation and of the wave number were associated with them.
On the other hand they are denied any localization in time and space.
The abstract character of such simple individuals directly results from
the spectral inequalities. Ithasno counterpart in classical mechanics of
corpuscules where a centre of gravity has at any time a well-determined
localization and momentum. For field magnitudes only groups may
be localized, but owing to the spectral inequalities, localization is cou-
pled with complementary magnitudes. A system relevant to classical

mechanics does not necessarily vibrate with a definite wavelength11 or

pulsation. Therefore, in so far as continuity prevails, physical reality
only belongsto systems whichare not susceptible of complete determi-
nation, though their internal evolution follows a strictly causallaw.


thank YvanCuche and Vincent Voirol for their criticisms.

College de France,
11, PlaceMarcelin Berthelot,
F-75231 Paris Cedexos, France


Versuch denBegriffder negativenGrbfien in die Weltweisheiteinzufuhren, by Johann
JacobKanter, Konigsberg, 1763.
2 See,
for example,Margenau, H. and Murphy, G. M.: 1956, The Mathematics of
Physics and Chemistry, 2ndcd.,Princeton: Van Nostrand, Ch. 7-8.
3 Cohen-Tannoudji, C, Div, 8., and Laloe,
F: 1973, Mecanique quantique, Paris:
Hermann,I, pp. 24-27; 11, AppendiceI, pp. 1446-1454.
Inradio-electricity the inequality relationis written:

r " Ay
where r is the time interval separating two successive annulments of the signal, and
Ai/ is the extension of the signal's spectrum (Blaquiere, A.: 1960, Calcul matriciel,
Paris: Hachette,11, pp. 123, 124). It is rare, in classical physics, that a magnitudeis
representedby an operator. See, however, for the impedance of a circuit: Blaquiere
(1960, p. 104).
5 Feynman, R.,
Leighton,R. H., and Sands, M.H.: 1963, The Feynman Lectures on
Physics, Reading,Mass.: Addison-Wesley, I, pp. 48^14.
Blaquiere (1960, 11, pp. 102-124). When a signal is linearly transmitted, let us call
gain theratiobetweenthe amplitudesof theinput and output ofFourier's terms having
the same impulse. For imperfect filters the gainis a function of the impulse. For a
(theoretical) filterthat should let only three impulses pass, if the operatoris the second
derivative with respect to time, it applies to the input function

fi (t) =ao+y^ an cos nu>t + N^ b n sinnut

n=l n=l
2 2=
to give the output function (d fi(t)/dt fo(t): correspondingto the general term of
rank n:

tn =an cos nut +bn sin nut

we have

t n =— n2 v2an cos nut —n 2v 2?bn smnut =—n 2v 2to-
With n = 3

(2 x 3) + 1terms
fa(t) =oxao — v (aicosut —b\ sinut —4u (a2 cos 2ut

+62 sin2ut) 9u 2(ai
cos 2>ut +63 sin 3ut).

7 Blaquiere (1960,
pp. 135-138); Feynman,Leighton, andSands (1963, 11, pp.22-21.
8 Suppes, P.: 1970, A Probabilistic Theory of Causality, Amsterdam: North-Holland,
pp. 86-89.
9 Vuillemin, J.: unpublished, Die Mbglichkeit der Erfahrung imLicht der zeitgenos-
sischen Physik, Trier (lecture).
10 Vuillemin, J.: 1990, 'Physique pantheiste et d&erministe: Spinoza et Huygens',
Studia Spinozana, 6, 231-250.
11 Feynman, Leighton, andSands (1963, 1, 49^15).


For many years I have benefitted from conversations with Jules Vuille-
min about the history of philosophy and the history of science. The
points about which he has instructed and corrected me are many and
large in number. He is equally at home in discussing the physics of
the Stoics or Aristotle, on the one hand, and on the other hand, the
great developments ofmathematical physics from the seventeenth to the
nineteenth centuries. The present article lays out in some considerable
detailthe wayin whichthebasic conceptualandmathematical apparatus
of quantum mechanics was first developed in the classical theory of
optics and electromagnetic fields, these developments in themselves
dependingupon the theory of waves developedbyHuygens andothers.
Without knowledge of this earlier history of classical physics it is too
easy to think that muchofclassical quantummechanics ismore original
in formulation than it is.

However, the really important point brought out by Jules is that in

major respects it is not classicalmechanics but classical electromagnetic
theory, and more generally the theory of wave phenomena, that is the
real intellectual predecessor of quantum mechanics. This applies to the
theory of operators, as well as the fact that if two operators in quantum
mechanics, such as position and momentum, are Fourier transforms of
each other, then an inequality of the form of Heisenberg's uncertainty
principle must hold, but initially just in the domain of the classical
theory of wave phenomena,especiallyelectromagnetic waves.
A point that Vuillemin does not emphasize, but I think he would
agree with,is that quantummechanics resembles electromagnetic theory
more than classical mechanics in other respects. The concept of a
particle's unique trajectory isabandoned. Itis wellknown that it is not
possible to compute in classical quantum mechanics even the simple
autocorrelation of the positions of a particle at two different times. As
I have argued recently in most explicit form, but also earlier as well,
quantum mechanics offers from acausal standpoint only a weak theory
of the mean probability distributions of particles, and nothing like a
full probabilistic theory of sample paths (Suppes, 1990). In classical
electromagnetic theory, as Jules points out, there is no concept of a
distinct particle path, but only of fields that extend continuously over
all of space, in the general case, and over a limited but continuous
domain in more restricted cases. Itis of greatimportance, however, to
insist on the point that physicists do not really think about electrons,
protons, and other particles just in terms of thesemean distributions or
in terms of the apparatus given tothem by classical quantum mechanics.
It is completely natural and indeed necessary for the serious detailed
discussion of experiments and of almost every kind of microscopic
physical phenomenon to think of electrons,protons, and other particles
as moving in space with trajectories. The concepts behind the exotic
apparatus of high-energyphysics are testimony to this. The purpose of
linear accelerators is justthat, to accelerate particles to very high speeds
but it would not make any sense to talk of suchacceleration without the
particles having trajectories.
There are two remarks Iwant to make about Vuillemin's analy-
sis of classical field magnitudes. The first is that the one thing that
quantum mechanics did add to the classical theory is the probabilistic
interpretation of wave phenomena. This means that what was classi-
cally interpreted as wave distributions,became under this interpretation

probability distributions of particles. As far asI know this probabilistic

interpretation does not occur anywhere in nineteenth-century physics
andis something genuinely new as part of quantummechanics. What is
ironical is that it is the interpretation, not the mathematical formalism,
that is the real contribution of quantum mechanics, even though the
probabilistic theory of quantum mechanics is a very weak one, in the
senseas already mentioned of beingonly a theory of mean distributions.
The second remark is a historical one. Jules remarks that Kant's
treatise The Metaphysical Foundations of Natural Science has been
undulyneglectedby most scholars, exceptby a few persons like myself.
But this is far too modest on his part. I consider myself an amateur
scholar of this treatise compared to Vuillemin. He has fortunately
caught several errors of mine before they have appeared in print; above
all Imention that he himself published the most detailed work (1955) I
know of on the MetaphysicalFoundations.


Suppes, P.: 1990, 'ProbabilisticCausalityinQuantumMechanics', JournalofStatistical

Planning and Inference,25, 293-302.
Vuillemin, J.: 1955 (2nd cd. 1987), Physique etMetaphysiqueKantiennes,Paris: Press
Universitaires de France.


ABSTRACT. Suppeshas developeda characteristicandconceptually unifiedapproach

to the theoriesof physical geometry and physical quantity, based on ideas and formal
techniques which he has also helped to develop within the representationaltheory
of measurement (RTM). These interrelatedideas, techniques and results are directly
relevant to the philosophyof physics, butremain mostly unknown within thatcontext.
Here Iaimto explaintheseideasand theirrelevance in a clearand systematic manner,
andalso to extendthis general researchprogramin several ways.
IncontrasttoSuppes I stress theneedfor a unified, realistandrevisionist approach to
the theory of physicalquantity,as opposedtohis piecemeal,empiricist andconservative
approach.I alsostress theneedforintrinsicformalization of wholequantitative theories,
as opposed to his hybrid extrinsic set-theoretic approach. As new contributions to
this ongoing program of research Ioutline an RTM-style representation theory for
differential geometry,leading toaxiomaticRiemanniangeometry. Thisalsoinvolves an
extensionof theRTM theoriesof uniquenessand meaningfulness to situationsinvolving
no natural symmetry oruniqueness group. Ialso describe somenew types ofrelationist
geometricaltheory, developedusing this sameRTM-based technical apparatus.


Suppes' work on geometry and space-time is closely linked to his sys-

tematic work on representational theory of measurement (RTM), in
ways explained below. This work is summarized in the three-volume
Foundations ofMeasurement (Krantz et al,1971; Suppes et al, 1989;
Luce et al, 1990) co-authored with Krantz, Luce and Tversky, cited
below as FM-1,etc.
In contemporary philosophy of physical science surprisingly little
attention is devoted to the generaltheory of physicalquantities,physi-
cal measurement, and the methodology of quantitative physical theory.
Historically these topics were much more actively discussed: recall for
example theEudoxian theory of quantity, orGalileo's turn from the qual-
itative method of Aristotle to the quantitative method of Archimedes,
and the associated philosophical theory of primary and secondaryqual-

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 59-102.
© 1994Kluwer Academic Publishers. Printed in the Netherlands.

Theseolder debates were drivenby substantive methodologicalprob-

lems in natural science. However, the quantitative methods of Galileo,
Huygens and Kepler were soon joined with the differential and integral
calculus ofNewton andLeibniz andthe analytic geometryofFermat and
Descartes to produce what I will call the 18th century methodological
consensus, epitomized by the analytical mechanics of Euler, Laplace
and Lagrange. Physical quantities were henceforth to be representedby
real-number variables, physical laws expressedby ordinary or partial
differential equations governing those variables, and physical science
(as stated in the famous dictum of Kelvin) was to express all physical
facts and laws using this quantitative conceptual apparatus. This con-
ceptual apparatus has dominated subsequentphysical science, allaying
the methodological doubts which had motivated earlier philosophical
The modern situation, however, has changed again. On the one
hand developments in mathematics and logic haveextended traditional
conceptions of quantity in many ways. The riseof abstract algebrahas
shownthat 'quantity' is not just one thing: there are many abstract num-
ber systems and quantitative structures, just as there are many abstract
geometries. The question which of these possible quantitative forms
figures in the physical world is therefore ultimately an empirical one,
just as the corresponding question for geometrical forms was finally
seen to be. Moreover, the traditional real-number continuum itselfhas
been found to conceal deep logical problems previously unsuspected
(unsolvability of the continuum problem within ZF set theory); con-
tradictions in the foundations of set theory itself)- Finally, of course,
modern developments in general formal logic give new and far more
powerful tools for attacking such problems of logical analysis.
On the other hand, modern theoretical physics has also moved
sharply away from the 18th century consensus in quantitative method-
ology. Both relativity and quantum theory involve radical shifts in the
underlying conceptions of physical quantity and of quantitative laws
(indefinite metric and variable curvature of Riemannian space-time;
infinite-dimensional state spaces and Hermitian operators of quantum
theory). In this modern contextit seems clear that traditional questions
about the nature, scope and content of quantitative methods in physi-
cal science must arise again in new and deeper forms, and should be
addressed within the context of a new general theory of the nature of
physicalquantities and quantitative theories,taking account of modern
developments in logic, mathematics and physics. This is an important
task for philosophers of physical science.


Paradoxically, physical measurement has recently been studied more

by psychologists than by philosophers of physics. The explanation
(FM-1, p. xvii) is of course that the social sciences still face real prob-
lems concerning the scope and significance of quantitative methods, as
the physical sciences did before Galileo but not after Laplace, so that
there the practical motivation for foundational analysis is still present.
Although the main goal of RTMis quantitative social science, the gen-
eral framework has been developed in part through detailed analysis
of physical measurement, yielding a body of work which essentially
belongs to the philosophy of physics,thoughproducedand read mainly
by psychologists. (Forexample, thechapter 'Dimensional Analysis and
NumericalLaws' ofFM-1is the most comprehensive existing treatment
of this topic). For a philosopher of physics it is abit disconcerting to
find psychologists discussing topics suchas velocity additionin special
relativity, and there is some temptation to dismiss this work as amateur-
ish dabbling in technical matters better left to specialists. But in fact
the RTM literature contains a more sophisticated treatment of physi-
cal quantity and measurement than anything in the current literature of
philosophy of physics.
Representationtheory iscentral for theory of quantity because it pro-
vides the link between the artificial numerical apparatus of the 18thcen-
tury analytical method (the numerical 'variables' postulatedas primitive
inallmodern quantitative theories) andthe underlyingintrinsic physical
reality. In the analytical method everything is already numerical: no
physical phenomenoncan be discussedunless it is already expressedin
the numerical language of classical mathematical analysis. Since the
analytical method assumes this translation to be given, i.e. assumes
physical facts to somehow already be representedby numerical vari-
ables, it cannot tellus what about the world makes that possible, orhow
the world appears when not so represented.
This is the first and most important theme in RTM: logical analysis
of the process ofnumerical representation, andof the conditions under
whichit is possible or useful. Todo this, RTM first gives a character-
ization of the subject matter in terms of intrinsic relations among the

objects, not dependent on or defined using any numerical representa-

tion. (RTM calls these empirical relations, but I reject the implication
of direct observability; rather they must simply be intrinsic to the phys-
ical subject-matter.) Then a formal characterization is given of the
representational conditions C which should be satisfied by a scale or
representingmap / from such a physical or empirical relational struc-
ture S to a correspondingnumerical structure M. Finally, a set of axioms
A constraining the intrinsic structure of S are stated, from which one
can derive the representation theorem that for any S satisfying axioms
A there willexist a function / from S toM satisfying conditions C. The
practice of numerical representation of elements s of S by numerical
scale values f(s) in M is then explained and justified by the assertion
that the underlyingphysical structure S satisfies theintrinsic axioms A.
Froma physical viewpoint, therefore,RTMaims to uncover andexpress
the intrinsic pre-numerical physical facts and laws which underlie the
traditional 18th century practice ofnumerical representation, and which
cannotbe expressedwithin that framework becauseit presupposesthem.
However,numerical representation is alsoarbitraryin various ways:
thereare different possible scales of measurementfor the same quantity.
This was recognized within the analytical method, and described using
a numerical group G of scale transformations: elements g of G gen-
erate alternative scales as /' = g o /. This is the older Klein-Stevens
approach to representation theory, in whicheach scale or representation
is postulated to have a numerical transformation group G associated
with it, the compositional action of which defines the range of possible
alternative representations. This topic of uniqueness of representation
is the second major theme in RTM. The group G is ultimately iden-
tified as the structural automorphism group of M, and one proves the
uniqueness theorem that if structure S satisfies axioms A then the maps
/ satisfying conditions C from S to M are 'unique up to' composition
with an element of G. This explains the role of the group G, which on
the Klein-Stevens approach was somewhat mysterious.
Finally, statements made using numerical scales (representational
propositions) may or may not be meaningful, i.e. have their truth inde-
pendent ofchoice of scaleorrepresentation. Asin the traditionalKlein-
Stevens theory, the RTMconditions of meaningfulness is invariance of
truth-value under the action ofG on the representationalpropositions or
sentences. However, in theRTMframework one can give amuchbetter
explanation and justification of this requirement, using the additional
information about the group G and the intrinsic structure S given by
the RTM theory of representation. This theory ofmeaningfulness is the
third major theme in RTM.1
Note that the same numerical transformation group G figures cen-
trally in all three aspects of RTM, playing three different roles: as
symmetry group of the representing space M; as uniqueness group of
the representation, giving the range of possible alternative representa-
tions ofSin M;andas meaningfulnessgroup,expressingthe invariance
condition which defines the meaningful representationalpropositions.
This pervasive role of a single group G has led to a generalfeeling that
transformation groups are central to representationalproblems, andhas
made it hard to approach representationalproblems havingnoassociat-
ed transformation groups. I will show in Section 7 that this feeling is
wrong: the truly centralideas of representationtheory are quiteindepen-
dent of transformation groups. When those centralideas are formulated
in a sufficiently general way and the special conditions for group-based
representation are recognized, the way becomes clear for development
of representationtheory in contextssuchas differential geometry where
no symmetry groups are present.
RTM lays out the methodological framework within which a theory
of physical quantity may be developed, by showing how to mathe-
matically ground extrinsic numerical representationalpropositions in a
non-numerical structure of intrinsic physical facts. However, it falls
short of actually presenting such a theory, in several ways. First there
are significant gaps in that methodological framework itself, connected
for example with the justification of the representational conditions C
imposed on the representingmaps /. RTMalwaysrequires / to beboth
homomorphic and faithful, i.e. to preserve the structural relations in
both directions,while the far more extensive literature of representation
theory in algebra is based on purely homomorphic representation. On
the other handI noted in 1986b the possibility of representations based
onfaithfulness alone, that beingintuitively more basic.
Second (Mundy, 1987a), physical representation theory should be
applied within a realist context,recognizing (unlike the narrow empiri-
cist or operationalist tradition of RTM) that fundamental physical quan-
tities need not be in any sense empirical or observable. Moulines and
Sneed (1979) raise this objection, and in reply Suppes (1979,pp. 207-
--208) concedes that we should not "look for fundamental measurement"
of these quantities. I say rather that weshould andmust do this,but with

the recognition (due to RTM itself) that 'fundamental measurement' is

a mathematical relation between an intrinsic physical structure S and
a numerical representation, not necessarily a directly implementable
empirical process.
Thus in seeking (for example) an RTM representation for time in
microphysics, weask in realist physical terms whether any structure of
intrinsic physical relations existsonanatomic level, sufficient to support
an RTM-style representation theory and thus to ground intrinsically the
standard microphysical use of a real-valued time coordinate. These will
certainly not be empirical relations,but theissue is whether they exist at
all, what they are, and what axioms or properties A of them ground this
representation. Unless this question can be answered, we simply do not
know what it means to say of any theory using such a coordinate (e.g.
standard non-relativistic quantum theory) that it is true of the actual
Third, while presenting a unified representational methodological
framework for a theory of quantity, RTM does not actually present
a single unified theory of physical quantities themselves (though the
RTM analysis of the dimensional algebra of scalar physical quantities
is quite detailed). Essentially independent axiom systems are presented
for different kinds of quantities, with no real attempt at unification,
for example, between scalar and vector quantities (discussed below),
or between structurally similar but physically distinct quantities such
as the geometric and dynamic vectors. Such issues require deeper
investigation of the physical content of the corresponding fundamental
physical theories thanis attempted within RTM.


RTMis essentially a theory of scaling, i.e. of numerical representation

of individual quantities, not amethod of analysis of whole quantitative
laws and theories. Like the operationalistbias of RTM, thisis doubtless
a vestige of its psychometric ancestry, wherein debate focused more on
the status of proposed numerical scales for particular variables than on
the discussion of full-fledged quantitative theories using them (FM-3,
p. 263).
As a philosopher of physics Suppes ofcourse recognizes the needfor
logicalanalysis of theories as well as scales,andunlike many has actu-
ally undertaken this, using the method of set-theoretical formalization
(often alsocalledsemantic), whichhe developedin contrast to theolder
method of syntactic formalization using a formal language. Semantic
approaches have since become fairly common. Ibelieve that the argu-
ments purporting to show the general untenability or impracticality of
the syntactic approach are misguided (Erkenntnis, 1987, p. 173; 1988,
p. 164; 1990, p. 345), and have outlined a syntactic formalization of
classical particle mechanics (Mundy, 1990). However,I am here con-
cerned not with the contrast between semantic and syntactic, but rather
between extrinsic and intrinsic formalizations.
Allofthe existing examplesof semantic formalization of quantitative
theories seem to make essential use of extrinsic numerical representa-
tional apparatus. This began with Suppes himself, and is a natural
consequenceof therole played by RTMin his conception of theory for-
malization (Moulines and Sneed, 1979). In effect he takes the problem
of logical analysis of quantitative structure to be solvedby RTM: once
the relevant numerical scales have been introduced by RTM methods,
Suppes will proceed to formulate physical laws as numerical represen-
tational equations relating those numerical scale values, in essentially
the standard 18th century analytic manner. A model of the theory will
then be a particular numerical structure (e.g. a particular solution to a
differential equation), and the laws of the theory (e.g. the differential
equation itself) will impose numerical conditions on those numerical
values,to be satisfiedin eachmodel of the theory.
However, those numerical values contain a mixtureofintrinsic infor-
mation about the physical system itself, and extrinsicinformation which
depends upon the numerical scales and coordinate systems. In other
words, not all of the mathematical orset-theoretical structure present in
these 'models' of the theory actually represents something physically
real. At this point, therefore, quite in accord with the RTM tradi-
tion, one must develop a theory of meaningfulness to distinguish which
propositions expressible using such a set-theoretic model are actually
'meaningful', i.e. correspond to something intrinsic in the world, and
which are not. This apparatus is closely related to the conditions of
dimensional and Galilean or Lorentz invariance imposed in the analyt-
ical tradition, and serves the same purpose of ruling out any statements
whose truth dependsupon the choice of scale ofcoordinate system.
If the required theory of meaningfulness is properly developed, as
in Suppes' own set-theoretic formalizations of physical theories, then
the resulting formalization may be formally adequate. (Some other

advocates of semantic approaches to theory-formalization, however,

have overlooked this problem of meaningfulness.) My point is rather
that an analysis of this kind is philosophically unsatisfactory, since it
does not succeed in presenting explicitly the true intrinsic content of
the theory as a set of conditions of some kind imposed directly upon
the world. Rather, it says that the world possesses an intrinsic structure
sufficient for the construction of various extrinsic numerical scales,and
that one or more of these scales will then be found to satisfy certain
numerically expressed laws. However, if these are indeed laws of
nature, it must be possible to express them directly as propositions
about that original intrinsic structure itself, with no detour through
this extrinsic numerical apparatus. Such an intrinsic formulation will
express explicitly some aspects of the content of those laws which are
only implicit in thenumerical formulation, and will deepen it in much
the same ways that the RTM analysis of scaling deepens the older
Klein-Stevens one.
In two special casesRTM seems, somewhat by accident,to have giv-
en intrinsic accounts of physicallaws or theories. One is the treatment
of dimensional algebra in FM-1, Ch. 10, which includes an intrin-
sic formulation of the simple laws of proportionality (e.g. Hooke's
law) which ground the dimensional algebra. However, the actual goal
there was not to find intrinsic formalizations of physical theories in the
present sense, butrather simply to secure the standard representationof
cross-dimensional relations by numerical multiplication of scalevalues.
Indeed, in FM-3 another wayis found of achieving that same goal with-
out intrinsic axiomatization of any individual laws,and this is regarded
as an improvement (FM-3, pp. 85-86, 315-316). The other context is
axiomatic geometry,discussed below.
The firstintrinsic formulationofa substantivephysical theoryoutside
of geometry was that of classical mechanics by Field (1980), extended
by Burgess (1984). However, I have argued (Mundy, 1989b) that these
formulations do not correctly express the physical contentof that theory,
because they dependessentially on assumptions about the structure of
physicalspace which are not part of classicalmechanics. Namely, they
assume the physical existence of space-time points, and of space-time
regionssatisfying strong existence conditions (a set-theoretic 'compre-
hensionaxiom'). Theintrinsic formulation of classicalparticle mechan-
ics described in my (1990) avoids these defects.
My fourth criticism (continuing from the end of Section 2) is that
RTMtakes an excessivelyconservative attitude towards physical theory.
On the onehand this appears inthe artificial divisionbetween scalingand
theformulation ofnumerical laws. RTMrightly questions the traditional
assumption of direct numerical representability ofindividual quantities,
constructing a detailed logical analysis of this procedure, while at the
same time accepting without discussion the traditional 18th century
analytical apparatus for the numerical formulation of quantitative laws
and theories.
On the other hand this conservatism shows up in the criterion for
success of a scalinganalysis: namely, that it should yield 'the expected'
representation and uniqueness results, i.e. should lead simply to a
formal justification of existing representational practice. In contrast,
I stressed earlier that modern developments in logic, mathematics and
theoretical physics cast substantial doubt on the correctness of many
aspects of traditional quantitative methodology. The ultimate purpose
of formal and axiomatic analysis of quantitative physical theory, in this
case as in all others, should therefore be not simply to explain and
reaffirm existingpractices, but rather also to identify respects in which
theymight usefully be modified.


An important theme of RTM,especiallyfor physical science,isitsassim-

ilation of geometry to theory of quantity and measurement. Axiomatic
geometry wasan important technical paradigm forRTMbecause it con-
tained the first explicit numerical representation theorems in the RTM
sense (FM-2, pp. 1-2). That is: Cartesian analytic geometry defines
a numerical relational system M having the structure of a qualitative
or non-numerical geometrical space S. The basic elements (points) of
M are n-tuples of numbers and the basic geometrical relations among
those points (congruence,colineality, etc.) are defined using numerical
properties of those numbers (congruence using the Pythagorean metric
formula, lineality usinglinear equations, etc.). In19th centuryaxiomat-
ic geometry qualitative Euclid-style axioms A governing a qualitative
geometrical space S were shown to imply the existence of a system
of numerical coordinates satisfying a condition C: that the coordinate-
based Cartesian numerical definitions of geometrical concepts agree
with the corresponding primitive qualitative concepts of the system S.

This is a representation theorembecause it shows that any model S

of axioms A canbe represented^the Cartesian numerical geometrical
space M: the representing map / is the constructed coordinate assign-
ment, whichcarries each point of the model to a 'point' in thenumerical
space, i.e. an n-tuple of real numbers, and satisfies the desired rep-
resentational condition C. There is an associated uniqueness theorem
showing thatthe alternative coordinate maps for such amodel, i.e. those
satisfying the same representational conditions C, are just those obtain-
able from the first one by composition with a numerical coordinate
transformation of a particular group G, characteristic of the geometryin
question (e.g. for Euclidean geometry the orthogonal group). Finally,
thereis a theory ofmeaningfulness(already formulated byKlein) iden-
tifying the meaningful numerical statements of the geometrical theory
as those invariant under the transformations of this group G.
We see here one source of the conservatism of RTM mentioned
above, since in this approach to foundations of geometrythe criterion of
interest or of correctness for a geometric axiom system is that it should
yield the standard numerical representation.
A basic test of the adequacy of an axiom system for a particular branchof geometry is
that it should lead to the desirednumerical representation and the desireduniqueness
theoremor automorphism group of the representation(FM-2, p. 81).
Such an approach obviously discourages lines of investigation (e.g.
non-Archimedean geometries) leading to theories whose models lack
nice numerical representationsin existingnumber systems. This special
status accorded to numerical coordinate representations in 19thcentury
geometryis perhapsanother manifestationof the 18thcentury consensus
in quantitative methodology mentioned above.
The RTM treatment of classical axiomatic geometry summarizedin
FM-2 does not add any fundamentally new ideas to this 19th century
framework; rather,it extends its significance by subsuming it under the
broader framework of representation theory. In particular, the RTM
analysis of geometry provides through its general theory of representa-
tion, uniqueness and meaningfulness a foundation for Klein's concep-
tion of geometry as based on groups G of numerical transformations
takenas primitive. Interestingly, a closely analogous Kleinian concep-
tionof scale types as based on fixed groupsof scale transformations was
given,apparentlyindependently ofKlein,by the psychologist Stevensin
1946. RTMgives a quite satisfying analysis andunification of these two
analogous theories of quantitative forms defined by numerical transfor-
mation groups,together with theirassociated theoriesof meaningfulness
as invariance under these transformations. (Seemy (1986b) for a simple
account; the full RTM theory of scale-types is in FM-3, Ch. 20.)
A significant defect in the geometric axiomatizations of FM-2, in
my view, is their failure to bring out fully the formal relations between
geometrical quantities and other scalar and vector physicalquantities.
I believe that the drafts for FM-2 which Isaw around 1980 included
attempts to formulate new axiom systems for the classical geometries
using the RTM apparatus of scalar extensive measurement. However,
the published axioms use only traditional geometrical relations such
as colineality and congruence amongpoints, unrelated to the extensive
primitives of order and addition used for scalar physical quantities in
By far the most naturalunified approach, in my view,is to interpret
the vector additionrelation of physical vector quantities as a generaliza-
tionof the additionrelation of scalar quantities,and themetric structure
of physical vector quantities as a generalization of the order relation of
scalar quantities. In the geometric context the relevant vectors are of
course the intrinsic displacement vectors relating pairs of points, and
the vector metric is the length-ordering of these segments. Indeed,
essentially this kind of intrinsic vectorial approach to the axiomatiza-
tion of geometry was mentioned as promising in Suppes'(l972) (1973,
pp. 387-388), though apparently not pursued. But this approach is not
evenlisted among the alternatives at FM-2, pp. 82-83.
On this approach one can obtain a uniform intrinsic axiomatization
of all scalar and vector physical quantities overthe same two extensive
primitives, showing their fundamental unity of algebraic structure as
commutative groups equipped with an order relation, representable as
vector spaces over anormed field together with a bilinear inner product
(unpublished work of the author). This same theory comprises the
Euclidean spaces ofclassical geometricand dynamic vectors(forces and
momenta), the Minkowski spaces of special-relativistic geometric and
dynamic vectors,and theHermitian spacesof quantum-mechanical state
vectors over the complex field, and their respective non-Archimedean
extensions to Pythagorean or Euclidean fields. When aiming also at
intrinsic representationofactual quantitative physical theories, asin my
(1990), itismore convenient to replace the order primitive by a stronger
intrinsic metrical primitive corresponding to scalar multiplication by
the dimensionless length-ratio of two similar quantities. This allows

polynomial equations to be expresseddirectly within the language of

theaxiomatic theory,as cannoteasily be donein thequalitative language
based on order and addition alone. (As anticipated by Suppes (1972),
thisisalsoaquantifier-free formulation in the sense ofMoler and Suppes
RTM always interprets geometry as simply another scaling theo-
ry, the only difference being that the representing 'numbers' are now
n-tuples (FM-2,p.81). As withscalar measurement, the workof formu-
lating true quantitative laws or theories as numerical statementsusing
scale values (e.g. thelaws of mechanics formulated using geometrical
coordinates) wouldthen not beginuntil this initial 'geometricalscaling'
stage was completed.
However, there is an important sense in which this interpretation
is incorrect: geometry by itself is already a true quantitative theory,
while the theory of a single scalar quantity is not. This is intuitively
recognizedby physicists, but also has a formal basis in the fact that
metrical geometry includes non-trivial quantitative laws: namely, the
laws of trigonometry relating the distances and angles in a triangle.
(All other geometrical laws in a classical geometry are consequences
of these.) In the scalar or 1-dimensional case, by contrast, these laws
reduce simply to the additivity of distances ('distance ab + distance be
= distance ac), which does not express a meaningfulphysical law but
rather merely defines the numerical measure of distance, beingactually
a representational condition C imposed on the length scale /, not an
intrinsic physical law. The trigonometric laws of a metrical geometry
of dimension > 2, however, carry additional intrinsic content after the
numerical scales for measurementof length and angle havebeen fixed.
Mypoint, then,is thatunder Suppes' extrinsic set-theoretic approach
to theformalization of quantitative theories (Section 3 above),onecould
give a complete formalization of a metrical geometry by simply stat-
ing the extensive measurement axioms for angle (FM-1, p. 76) and
length, and then asserting the trigonometric formulas of that geome-
try as numerical laws constraining those scalar quantities. Indeed, this
would produce a very pretty anduniform treatment of the various clas-
sical geometries,showinghow they involvethe same basic concepts but
assert different laws governing them. But Suppes andRTMnonetheless
emphasize the much more difficult traditional synthetic approach. I
think this must be because the authors recognize intuitively that these
intrinsic geometric axiomatizations give a better account of the actual
underlying contentof a classicalmetrical geometrythan wouldbe given
by a simple numerical statement of its trigonometric formulas. This is
an implicit acceptance of my earliercriticism of Suppes' method of for-
malization of quantitative theories: that one should replacenot only the
scales but also thenumerical laws by intrinsic descriptions of the under-
lying facts which ground those extrinsic numerical representations.


There is another very important difference between the geometrical

axiom systems of FM-2 and the scalar axiom systems of FM-1, not
discussed in FM-2but recognizedby Suppes elsewhere and stressedin
FM-3, e.g. p. 48. This is the difference between strong geometry-style
axiom systems yielding an isomorphism f from the representedspace
S to the corresponding numerical space M, and the weaker types of
axiom system developed in FM-1 for scalar quantities, yielding only
an embedding f (up to order-equivalence in S) of the space S into M,
uniqueup to some group G of transformations of M.
This disparity between ontorepresentationof geometry andinto rep-
resentation of scalar quantities may be acceptable in the RTM frame-
work, where each quantity is analyzed separately, but is surely not
acceptable within a unified theory of physical quantities. Since quanti-
tative laws normally relate different quantities (as stressed in the RTM
treatmentof dimensional algebra), itis implausible and probably incon-
sistent to impose far stronger existenceconditions on some thanothers.
This is precisely the type ofissue which transcends the piecemealRTM
approach and requires a truly unified theory of allof the physicalquan-
tities. (The intrinsic scalar multiplication primitive of my (1990) theory
makes all quantitiessatisfy the same existence conditions,over the same
Pythagorean orEuclidean field.)
Since the actual world presumably does not contain a continuum of
material objects, an onto representationof an actual physical structure
S in a continuous M is only possible if the physical world contains
immaterial objects of some kind. Immaterial physical objects might
include: points or regions of physical space or space-time, properties
or relations, possible objects, possible facts, and portions of physical
fields. Of course the physicalreality ofallof theseimmaterial objects is
controversial. I insist, however, that whether such objects exist isa real
physicalquestion,not avoidable by vague remarks about 'idealization.

Inparticular, a representationaltheory of quantity whose axioms require

the existence of more entities in the physical system S than its author
admits as physically real is simply not doing its job: it is not a theory
of the actual world at all, but of some imaginary world. Calling that
world an 'idealization' of the actual world cannot change the fact that
the stated theory is not true of the actual world, so that results about it
have no power to explain any actual facts.
There are then two basic responses to the disparity between into
and onto representation: the immaterialist response of accepting and
defending the existenceof some immaterial objects within one's theory
of quantity and geometry; and the materialist response of seeking a
workable theory of quantity whoseaxioms are satisfiable by amaterialist
physical structure. Of course immaterialism arises not only for onto
representations, but whenever the axioms require the existence of more
objects than exist materially. Thus e.g. afinitist materialist must resort
to immaterialism in order to use even so weak a property as closure of
S under addition, which forces its domain to be infinite. I willconsider
theimmaterialist responses first.
The obvious immaterialist interpretation of the elements of a quanti-
tative structure Sis as quantitative or geometrical propertiesof material
objects: geometrical points as location-properties (Horwich, 1978) and
elements of 1-dimensional extensive structures as possible values of
scalar physical quantities, e.g. lengths. Such a second-order or Pla-
tonist immaterialism is defended in my (1987b). The RTM literature
sometimes assumes such a second-order interpretation; for example,
FM-3 (p. 281) places within the represented structure S an element a\
described as "the (qualitative) length of spring a-i\ thus assuming a
physical world containing immaterial 'lengths' in addition to material
However, it must be stressed (Mundy, 1989b) that the immateri-
al entities of a second-order theory of quantity need not function in
physical theory as monadic quantitativeproperties of material objects;
they may also function as quantitative relations of any given degree.
In particular, geometricalquantities such as scalar distance and vector
displacement are most naturally analyzed as second-order quantitative
binary relations: one object alone has amass ora charge, but a distance
or a vector displacement is a quantitative relation between two objects.
This is especially natural inthe contextof a unified extensive theory of

physical quantities as described above, with physical geometry based

on the geometrical displacement vectors.
Insuch avectorial theory (e.g. my 1990) there arenopoints of space
or space-time, since these correspond to monadic location-properties
of objects. These are therefore relationist theories in precisely the
traditional sense: theories which deny the physical reality of points of
space or space-time (spatialism), and instead assert that geometrical
facts are grounded solely in geometrical relations. Moreover, there
is no appeal to possible objects or possible configurations of actual
objects, suchas most relationists have invoked and some authors have
thought they must invoke. The only immaterial entities required are
the geometrical relations themselves,and the second-order quantitative
structure (e.g. addition and order relations) over them. For any given
flat geometry one may construct such a second-order vector relationist
theory which accounts for the observable geometrical facts in a manner
provably equivalent to the monadic or spatialist theory. Moreover,
the unified theory of physical quantity obtainable using a second-order
vector relationist theory provides a strong argument for its physical
superiority to the monadic alternatives: this is a new argument for
second-orderrelationism derived from the theory of physicalquantities.
A second-order relationist theory need not use binary vector rela-
tions; this is merely the simplest and most natural approach for a flat
geometry. The quantitative relationist theory of my (1983), which has a
natural second-order interpretation, was based ona three-place extrinsic
scalar relation (corresponding to a six-place intrinsic scalar relation).
The quantitative formulations of classical metrical geometries using
their trigonometric laws asmentioned abovelead to second-order rela-
tionisttheories based on binary scalar distance relations and three-place
scalar angle relations among material objects. These theories too allow
the full empirical content of the corresponding spatialist geometries to
be deduced from finitely many actual instances of these relations as
instantiated by actual objects. This trigonometric approach furnishes
second-order relationist theories of hyperbolic and elliptic as well as flat
It is instructive here to glance at the 'quantity theory of geometry'
proposedby Teller. As a new alternative to spatialism and relationism
he suggests that

we shouldtake space-time as aphysical quantity differing fromquantities such as mass

and temperatureonlyin detailsof structure, and we shouldtake space-timepoints to be
the properties thatconstitute particular valuesof the space-time quantity (1987, p. 425).

Teller's basic idea of assimilating geometry to quantity is sound, but

poorly developed for lack of the relevant technical apparatus. Having
naively assumedthat physicalquantities must bemonadic properties, he
produces a theory which (as many have observed) is simply a restate-
ment of second-order location-property spatialism. In addition, by
disregarding those minor 'details of structure' he fails to notice that his
initial analogy between geometry and quantity becomes much stronger
when cast in terms of the intrinsic quantitative structures definable
within a geometric space -becoming strongest of all, as noted above,
in terms of the additive group of displacement vectors. However, we
also then saw that the resulting theory is not a new alternative between
spatialism and relationism at all, but rather a new kind of relationism,
namely second-order relationism,as in my (1983 and 1989b).
Both spatialism and second-order relationism are forms of imma-
terialism. They merely disagree over which immaterial entities exist:
unoccupiedpoints (uninstantiated monadic location-properties, with a
second-order structure of standard geometrical point-relations) or unin-
stantiatedgeometrical relations,withasecond-order structureof quanti-
tative operations. The only other precisely articulated form ofimmateri-
alist theory seems to be themodalimmaterialist relationism ofManders
(1982), in which theinfinite set of points or relations is replacedby the
infinite set ofpossiblefinitegeometricalconfigurations (finite structures
embeddable in the corresponding geometrical space). Manders (1982)
seems to havebeen the first formally precise relationist geometrical the-
ory articulated in print, followed closely by my (1983). Some authors
such as Earman (1989) still question the very possibility of relationist
RTM remains neutral (or rather, inconsistent) on the issue between
materialism and immaterialism. Suppes himself, however seems to
favor a materialist response. This emerges most clearly in his 1972
paper on space and time (1973a, pp. 391-395),leading to a quite dif-
ferent kindof relationist theory.
Hestarts from Whitehead's now-familiar ideaofa geometrical theory
whose basic elements are not points but extended entities related by
inclusion and overlap. Given strong enough existence conditions the
role of points can certainly be taken over by convergent sequences of
suchentities ('regions'),but such assumptionsare as dubious as those of
aspatialist theory itself. Suppesinstead seeks a theory whoseaxioms are
satisfiable by actual extendedmaterial objects ('bodies'). For example,
his axiom 8 implies that each body has only a finite number ofminimal
parts or 'atoms'. He proceeds (p. 394),
There are a numberof different ways to extend the axioms ,and byheavy-handed
methods, we can reach ordinary Euclidean geometry fairly rapidly. We simply have
to postulate enough bodies and atoms. We would of course not expect to get the full
Euclidean space,because of the finitedissectionproperty [i.e. axiomB], but we would
want to beable to imbed in three-dimensionalspace, andto get a representation ofthis
imbeddinguniqueup to the standard groupof rigid motions.
Thereis no doubt that thisprogram can becarried through. The techniques for the
one-dimensionalcase of measurement exploitedin many different directionsin Krantz
et al., 1971 [FM-1] provide more than adequatetools, but Ido not as yet see how to
pursue it in a simple and elegant fashion. At the same time Iam beginning to see
a philosophicallyinteresting aspect of this programif it can be satisfactorily carried
through. Properly carriedout, it should provide anew wayof looking at the nature of
Suppes' subsequent discussion of the philosophical significance of
this idea seems to me to be somewhat confused, drifting from finitist
materialism toward modal immaterialism in the manner of his (1974).
More important in my view is the technical suggestionin the first para-
graph; this defines what I call embedding relationism (see my 1983,
1986a) and is simply RTM representation theory applied to geometry
for the case of into rather than the traditional onto geometrical repre-
sentations. Note here the key elements of RTM:axioms,representation
theorem, and uniquenessup to the symmetry group G of the represent-
ing spaceM (which in turn willdetermine a theory of meaningfulness).
Suppes here suggests that the RTM theory of representation provides
the necessary technical apparatus for developmentof a materialist rela-
tionist theory: one which will capture the empirical content of a stan-
dard geometricaltheoryusing axioms overa structure S containingonly
material objects andproperties or relationsinstantiated by them, withno
additional structure involving any kind of immaterial entities, whether
they be points, relations,possibilities, or what have you.
I believe that no such materialist embedding relationist theory has
been stated in print, by Suppes or anyone else. Space here permits
only brief remarks on this line. Most important is an elementary but
fundamental limitation due to the fact that by a simple counting argu-
ment, the finite models of such a theory cannotreproduce exactly the
full range of finite possibilities describable by any standard geometrical

or quantitative theory. This is because (assuming a finite set of fini-

tary primitive relations, i.e. ones of finite degree), such a theory has
only finitely many structurally distinct models of any given finite size
n, while any standard quantitative theory allows infinitely many distinct
quantitative configurations of n bodies (e.g. triangles in geometry).
The point is that a materialist theory should allow finite models, since
there might be only finitely many material objects; the problem is then
that for cardinality reasons finite models alone cannot replace all of the
finite models of any standard quantitative theory.
This problem affects the materialist axiom system for scalar quanti-
ties given in FM-1 (pp. 81-85), in which closure under additionis not
assumed, thus permitting finite models. By the resulting representation
theorem each finite model is embeddable in (R ,<, +) uniquely up to

ratio transformations, thus determining n 1 independentreal-valued
ratios among the n objects in the model. By the preceding cardinality
argument,however, thesemodels cannot represent all of the size-ratio
combinations possible among n objects in a continuous scalar theory.
Inspection reveals thatthe axioms can only be satisfiedin a finite model
if all its elements are exact multiples of the smallest one, so that their
ratios to it are all positive integers.
A.finitist materialist embeddingrelationist theory over finitary prim-
itives can thus never achieve exact expressive equivalence (precisely
definable as in my 1986a, b) with any standard spatialist geometrical
theory. One line of response mentioned by Suppes (1979,p. 214) is to
stress the approximation to the continuous theory as the finite models
become larger.
Another response,developed in my 1989a for the case of scalar quan-
tities but also extendable to the geometrical case, yields finitist mate-
rialist theories with exact expressive equivalence to the corresponding
continuous theories. This is done by introducing anew kindof quali-
tative primitive relation: predicates of variable degree, which generate
an infinite set of distinct atomic facts even over a finite model, thus
avoiding the cardinality problem. The geometrical versions of this
approach yield a finitist materialist relationist theoryhaving full expres-
sive equivalence with the corresponding spatialist theories, but over a
non-standard logic.
A third response, taking therevisionist attitude tothe theory of phys-
ical quantity which I contrasted in Section 3 with the conservatism of
RTM,is to take seriously the departures from classical theory of quanti-

ty requiredby finitist materialism over finitary predicates. For example,

thebreakdown of space-time geometryin microphysics could conceiv-
ably be simply a manifestation of the coarse-grained structure of the
small finite models of such a finitist embeddingrelationist geometrical
I conclude by briefly mentioning another approach to embedding
relationism due to Friedman (1983). Friedman imagined, in contrast
with his own spatialist viewpoint, a relationist theory which takes
embeddability of the physical system Sin the mathematical space Mas
its sole axiom. I arguedbriefly in my (1986a) that Friedman had over-
looked the uniquenessproblem familiar from RTM: such a theory will
not have the expressiveequivalence he claimed for itunless the embed-
dings of S inM are unique up to a symmetry of M. I also noted that
RTMhad already developed a detailed methodology for dealing with
such embedding problems, recognizing the need for structuralaxioms
on the systemS and theneed for a uniqueness theorem, and that Suppes
has already applied these ideas to embedding relationism, as quoted
Inreply to this criticism Catton andSolomon (1988) defended Fried-
man's idea, suggesting thathe wasindeed aware of the uniquenessprob-
lem and had intended G-uniqueness of the embeddings as an implicit
condition on the primitives of the relationist theory (though the prim-
itives chosen by Friedman for his own example do not satisfy this
condition,as I had already pointed out). In my 1991 response (Mundy,
1991) I showed, using a variant of the counting argument given above,
that in general this condition cannot be satisfied by any finite set of
finitary primitives with embeddability as the sole axiom; I also pointed
out the general implausibility of such an approach to embedding rela-
tionism. This episode, like the naive suggestion of Teller mentioned
earlier, seem to me to show clearly how badly the philosophical litera-
ture on relationism is in needof both the general formal discipline and
the specific techniques and results of the RTM tradition (in agreement
with Suppes, 1973).


Following Einstein, Minkowski showed how space and time together

possess geometrical structure (e.g. concepts of distance and straight

line); today space-time geometry(STG) studies such geometrical struc-

tures and theories (STGs) allowing spatiotemporal physical interpre-
tation. While only pragmatically distinct from general mathematical
geometry, the physicalimportance of STG has created aliterature more
influenced by physics and philosophy than by traditional mathemati-
cal geometry. Thus, while Suppes' work on ST geometry falls mainly
within the sameRTM technical tradition as his work on general math-
ematical geometry, its external significance is different because the
relevant background literature is different.
Suppes' (1959a) paper on derivation of theLorentz transformations
does not fall within the RTM tradition,however, but rather within the
earlier Klein-Stevens tradition in which a class of numerical represen-
tations are assumed as given and the problem is to characterize the
family of numerical coordinate transformations relating them. This
Klein-Stevens framework meshes naturally with the standard physical
approach to space-time theory in terms of relativity, i.e. as a prob-
lem of relating the coordinate descriptions of the same event given
by different observers in different states of motion. On this tradition-
al approach the key to special relativity is to show these numerical
coordinate transformations to be Lorentzian rather than Galilean. The
standard derivation assumes the transformations to be linear and to
leave invariant the numerical values of the Minkowski metric form
s2 = c 2t2 —x2—y2
— z2 (i.e. itassumes allobservers to agree about the
speedof light, c, as determinedusing Einstein's optical definition of the
time coordinate, t). Suppes showed theLorentzian form to follow from
the weaker assumption of invariance of s2 for positive cases (timelike
intervals) only, without assuming linearity. Zeeman (1964) ultimately
showedit tofollow from theeven weaker assumption ofinvariance only
of the sign of s 2,i.e. of interval-type orcausal ordering.
However, all such Klein-Stevens approaches take the existence of
numerical representations (coordinate systems) as given, whereas the
point ofRTMis todeduce this as arepresentation theorem fromintrinsic
qualitative axioms over the empirical subject-matter, expressedin terms
of qualitative primitives. RTMthusleads to axiomatic or synthetic STG,
analogous to classical axiomatic spatial geometry. This idea was first
pursued by Robb (1911 and later), who gave qualitative axioms for the
Minkowski STGof specialrelativitity using the binary causal primitive,
and proved a coordinate representation theorem in the standard fashion
of axiomatic geometryandRTM. (My (1986d) contains a much simpler
development of essentially Robb's material.)
The importance of Robb's work was often stressed by Suppes, and
axiomatic STG was incorporated intoRTMalongside classicalaxiomat-
ic geometry. However, in contrast with the large mathematical literature
on that topic, the literature on synthetic STG wasrather sparse (FM-2,
pp. 128-130), andconsiderable effort was devoted to the development
ofnew axiom systems for STG which would fit naturally into theRTM
context. (This includes my 1982 thesis, directed by Suppes.) The
unpublished axiomatic work of Dorling onMinkowski STG also dates
from this period, as does the use in Field (1980) of RTM methods to
axiomatize classicalSTG.Allof this workdiffers fromRobb in develop-
ing STGusing classical geometricalprimitives (e.g. affine betweenness
andmetrical congruencerelations) more thancausal ones, thus empha-
sizing the strong formal relations among all of these geometries.
Axiomatic STG is particularly relevant to current philosophy of
physics because, unlike theories of other physical quantities, STGsare
recognized as important physical theories in their own right. (I noted
the underlyingreason for this earlier.) Axiomatic STGthsbears directly
ondebates concerningphilosophical and physicalinterpretation ofSTG
I willmention three examples,drawn from my (1986c). First, the tra-
ditional physical approach to space-time in terms of 'relativity' among
'observers' invites idealist interpretations of the theory as being about
our knowledge or observation rather than about any objective physi-
cal subject-matter which exists independently of us (e.g. Eddington,
1939). We canrebut such accountsby presenting an explicit axiomatic
analysis of the content of the theory as a set of statements constraining
some primitives whichare entirely physicalin character andinvolve no
reference to the presence of 'observers'. Second, Reichenbach (1924,
1928) andothers give causal analyses of Minkowski STGandinfer that
supraluminal causationis inconsistent with the theory, whereasaxioma-
tizationin terms ofnon-causalkinematic primitives leads to the contrary
conclusion. Third, Kuhn (1962) cites the transition from classical to
Minkowski STG as an example of the supposed incommensurability
of successive theories, while axiomatization shows the expressibility
of both theories in the same formal language, having the same physi-
cal interpretation, and differing only by acceptance or rejection of one
axiom (constancy of the speedof light) expressedin that common STG

language. These examples again support Suppes' (1973b) opinion of

the value for philosophy of spaceand time ofaxiomatic analysis in the
tradition of the foundations of geometry or RTM.
It should be noted that axiomatic STG does require axioms, not
merely definitions. Here we may contrast Suppes' long-standing stress
on Robb's axiomatic STG with the more recent flurry of discussion
of Robb sparked by Winnie in 1974 (see Winnie (1977) and several
other papers in the same collection). This discussion focused entire-
ly on the definability of the Minkowski metric structure from Robb's
causal primitive: this follows from Zeeman's invariance proof, and of
course also follows from Robb's earlier explicit axiomatization using
that primitive. This was philosophically relevant as an argumentagainst
the claims ofReichenbach andGrunbaum that causal structure is objec-
tiveor intrinsic while metric structure is conventional or extrinsic,and
also raised questions about the conformally flat STGs of general rel-
ativity, which possess the same causal structure as Minkowski STG
but different metric structure. None of this discussion involved any of
the details of Robb's axioms; it dependedsimply on the possibility of
some such axiomatization, whichis essentially the content of Zeeman's
invariance theorem. In effect Zeeman showed only the possibility of a
causal theory ofMinkowski STG, i.e. showed the causal primitive to be
sufficient for this purpose. However, in order to know what such a the-
ory actually says weneed to find some sufficient set ofintrinsic axioms
over that primitive. This is the goal of axiomatic STG in theRTM tra-
dition, following the generalpattern ofRTMaxioms and representation
The argument from the end of Section 4 applies again for STG. A
Suppes-style extrinsic numerical formulation of Minkowski STG can
be given simply by conjoining the extensive RTM axiomatizations for
length and timeintervals with the hyperbolic Minkowskian trigonomet-
ric formulas as numerical laws governing those quantities. Suppes'
long-standing emphasis on the importance of Robb's axiomatization
thus again implies a tacit recognition that such an intrinsic approach is
deeper,despite its greater complexity.

(a) TheProblem
The work in synthetic STGmentioned so far concerns only flat STGs,
whoseuniform underlying vector-space structure plays anessentialrole
in the axiomatic developments. The next natural case, apparently not
yetpursued, wouldbeSTGsof constantcurvature(used in cosmology),
presumably axiomatizable by analogy to the classical hyperbolic and
elliptic geometries. However,much greaterphysical interest attaches to
inhomogeneous differential geometries (DGs), suchas the Riemannian
STGsof general relativity (GR).
Axiomatizationof DGis a technicalproblem quite different in char-
acter from those hitherto treated with the RTM tradition. On the one
hand DGtheories are not categorical like theclassical geometries treated
in FM-2. Instead they comprise many differentmodels of very different
geometrical structure, and we seek only to axiomatize thebasic geomet-
rical properties (e.g. Riemannian structure) common to allof them. On
the other hand these different models are not allembeddable in a single
fixed Cartesian numerical geometry as with the non-categorical scalar
embedding theories of FM-1 .
In standard RTM the intended representing numerical- space M is
given in advance, and its known structure plays an important role in
the formulation of axioms. Every universal proposition true inMholds
in any S embeddable in it, so these necessary axioms must certainly
be included as axioms or theorems (FM-1,pp. 21-23). More deeply,
the associated uniqueness result constraining the range of alternative
representations is linked to the symmetries of the fixed representing
structure M. If no such structure exists, therefore, it becomes unclear
how toproceed. Indeed,Teller (1987) assumesthe task tobe impossible.
("If the curvature of space-time varies sufficiently irregularly, then no
representation theorem can be brought to bear. Such cases lack the
symmetries needed to characterize the family of representations ..."
p. 435.)
Such a conclusion is naturally suggested by the RTM literature.
Indeed, the extension of the RTM theories of representation, unique-
ness and meaningfulness to situations involving few automorphisms is
acknowledged at the end of FM-3 (p. 333) as an important openprob-
lem. ("Thereprobably are important connections between the problem
of defining meaningfulness and that of formulating satisfactory theo-

rems about the representation and uniqueness of structures when there

are few or no automorphisms" FM-3, p. 333.) In this section Iwill
outline (through application to DG) such a generalized form of repre-
sentation theory whichis applicable inall cases regardless of symmetry,
and which contains the RTM theory as a special case. However, in
this generalizedtheory theKlein-Stevens-RTMuniqueness group G no
longer appears.
I notedin Section 2how in standard RTM the single transformation
group G plays three distinct roles, as: (a) a symmetry group of the
representingspace M; (b) a uniqueness group of the representational
problem, determining by functional composition the class of alterna-
tive representations / meeting the representational conditions C; and
(c) a meaningfulnessgroup, determining meaningfulness as invariance
under its action on the representational propositions. The pervasive
role of the group G in standardRTMhas created a belief, illustrated by
the quote from Teller, that the formal theory of representation,unique-
ness and meaningfulness must in every case be dominated by a single
transformation group G playing all three of these roles.
However, there is no logical reason why these three roles should be
linked in this way, or indeed why roles (b) and (c) should be played
by numerical transformation groups at all. Roles (b) and (c) are of
courselinked by thebasic idea that meaningfulness is invariance under
change of representation, and thus the form of the uniqueness theory
for a given representational problem must always determine the form
of the corresponding theory of meaningfulness. However, there is no
a priori necessity that either of these should always feature some one
numerical transformation group G, whether given as the symmetry
group G(M) or otherwise. This is simply an expectation based on
the theory of representation as hitherto developed in RTM, using a
fixed symmetrical M. (A personal note: during my period of informal
contact with Suppesat StanfordI was particularly struck by his tendency
to reason by association between specific examples, without verbal
articulation of corresponding general principles. Isuspect that this
unusual 'semantic' as opposed to 'syntactic' style of thinking is the
source of many of his most characteristic traits as a philosopher and
This same tendency to identify the logically distinct representational
roles of a single group G has been even more pervasive and damag-
ing in space-time theory. In addition to the three roles (a)-(c) above
(symmetry, frame transformations,meaningfulness), 'the group' Gof a
space-time theorymay alsoexpress: (d) a relativityprinciple,constrain-
ing not only STGbut other physical theories as well (e.g. the familiar
condition of 'Lorentz invariance' imposed on dynamical theories). Itis
very often tacitly assumed that all four roles must be played by some
one group; for example in physics the derivation of theLorentz group of
frame transformations is almost always taken to justify itsuseas a sym-
metry and relativity group as well. The best-known error of this kind
is Einstein's confusion ofcovariance with generalized relativity, which
conflates roles (c) and (d) (see Friedman, 1983). My (1986d,Section 6)
exposes these confusions by presenting a version of Minkowski STG
which splits three of these four roles: it has the Lorentz group for its
frame transformation and meaningfulness groups but not its symmetry
group, and allows denial of the relativity principle.
I willnow sketch an account of the theory of coordinate representa-
tionin DG,leadingto a generalizedtheory of representation,uniqueness
and meaningfulness whichdoes notuse any transformation group G,but
which reduces to theRTM theory when appropriate special conditions
are satisfied.

(b) Generalized Theory of Uniqueness

Coordinaterepresentation instandardDG essentially follows the Klein-
Stevens format, in that the represented space is treated as a bare set
S, and the geometric structure imposed on it by the theory is fixed
entirelyby means ofnumerical representations / ofS taken asprimitive.
(Actually S is usually given an antecedent intrinsic topology, while the
differentiable and geometric structures are imposed using coordinate
representation,but it is more uniform to let the topology be induced by
the coordinates as well.) As in the 18th century analytical method, the
family ofnumericalmaps / is simply postulated,andthe theoryproceeds
entirely through specification of numerical connections among those
representations. In particular there is a theory of uniqueness, which
characterizes the range of possible alternative representations. This is
of course simply theKlein-Stevens-RTMuniquenesscondition,namely
that /'is apossible alternative representation to / justincase f = g°f
for some g in G, where G is the group defining the geometry or scale
type. G is always some group of transformations of M, usually its
automorphism group G(M).

However, DG has a different theory of uniqueness not fitting this

pattern, though it will be seen later to be a natural generalization of
it. We find by examination of the DG coordinate apparatus that there
is no one fixed embedding space M: different representations / have
different coverings of S by coordinate charts, and different values for
the geometrical coefficients such as the components of the metric tensor
or affine connection;this implies that the actualnumerical structure M
which serves to represent S under a particular coordinate representation
will be different for different /. In a more detailed presentation this
structure M can be defined precisely: it is itself aDG of the same type
as S, but with complex numerical entities for its points, like a Cartesian
Since the different / map into different numerical DGs M, there
cannotbe any Klein-Stevens uniquenessgroup G relating them all, and
hence the uniquenessproblem must be attacked in some other way. In
fact, in DG that problem is normally solved by a direct specification
of the conditions for two coordinate representations / and /' of S to
be equivalent, i.e. to impose the same geometric structure on S. This
definition is different for different types of DG space,but the basic idea
is always the same.
We first define equivalence or compatibility for two individual coor-
dinate charts cv , c v, within the intersection of their domains of defi-
nition. Since the charts are 1-1 maps into Rn, the 'bridge function'
kuv =cv o c~ l willbe a 1-1function fromRn intoRn,so that standard
analytical concepts are defined for it. The compatibility conditionrelat-
ing different charts is then simply that kuv should preserve whatever
properties or structure we wish to pull back onto S from Rn using the
cv: e.g. if Sisto be a topological manifold then the kuv need merely
be homeomorphisms, while for an m-differentiable manifold each kuv
must be ra-times partially differentiable in Rn These compatibility
conditions ensure that the various geometric properties pulled back to
S from Rn along the different cv will allbe consistent, since the bridge
functions kuv preserve them. Allof the chartsin any one representation
/ are required to be pairwise compatible wherever they overlap, and
two different representations / and /' are defined to be compatible if all
of their charts are pairwise compatible.
The only technicality is that the geometrical(Riemannian or affine)
structure pulled back from R along a given chart cv is not simply
the standard flat structure of Rn itself, but rather a modified structure
defined locally in Rn using additional numerical functions associated
with the chart cv Thus the localmetric structure pulled back from Rn
to Sby cv is not that determined by the standard Euclidean 'metric
form' ds 2 = Y^i(dx1)2, which would simply induce aflat geometry
on S, butrather that determined by theRiemannian metric form ds 2 =
Hii 9ij &x%&xJ'» where the n2 metriccoefficients gfj are functions taking
different values at different points in Rn, and are associated with any
chart cv for a Riemannian S. The compatibility condition for two such
charts is then that the bridge functions kuv should carry the lengths
defined by integration of the gfj quadratic differential form using cv to
the lengths defined byintegration of the g\j form using c v The 'length-
element' ds2 itself is then invariant, i.e. the same in all compatible
My firstpoint, then,is that by this definition the uniqueness problem
in DG is solveddirectly, without introduction of a uniquenessgroup G,
by a direct mathematical characterization of the relation which obtains
between two equivalent numerical representations / and /', i.e. ones
which represent the same structure on S. The standard representations
of S are thus unique up to equivalence,i.e. any two standard represen-
tations / and /' willbear thisrelation of equivalence. This in turn fixes
the form of the uniqueness theorem we should seek for an RTM-style
analysis of DG - i.e. we should seek to show that any two maps /
meeting our representational conditions C are equivalent in this sense.
Thus DG already has a uniqueness theory which does the same work as
the Klein-Stevens-RTM theory of G-uniqueness, while making no ref-
erence to anygroup G. Moreover, this uniqueness theory is welldefined
even though thealternative representations / whichitcharacterizes map
S into entirely distinct numerical structures Mand M.
It might seem that this 'direct' solution to the uniqueness problem
in DG is merely an accident, with no real connection to the Klein-
Stevens G-uniqueness theory. Butin fact there is anatural andintimate
connection, due to the compositional action of the function g on f
which generates thenew representation in the Klein-Stevens equation
/ =
Inthe generalDGcase we cannot compositionally generate
all of the alternative representations /'from any one / by the action of
a group G, since this will not take us out of the numerical representing
structure Mused by /. However, the same basic equation f = k o f
may well hold, where k is a mathematical function of some broader
class,not taken from one fixed transformation group G.

There is an obvious guess, confirmed by closer ' analysis. Imen-

tionedabove that the representingspacesMandM are themselvesDG
spaces of the appropriate type, e.g. Riemannian spaces, withnumerical
elements. Thus the standard concepts of isomorphism for a DG space
are defined for them, e.g. the conceptof a Riemannian isometry. Using
these we find in general that the definition of equivalence of two repre-
sentations / and /' canalso be put in terms of connectibility by a DG
isomorphism k. That is, two representations / inM and /' in M' are
compatible iff there is aDG isomorphism k of the corresponding type
from M to M' such that /' = k o /.
This enables us to recast theDG uniqueness theory in a composition-
al form which makes explicit its status as a direct generalization of the
Klein-Stevens G-uniqueness theory. Instead of a group G acting on a
single numerical Cartesianspace M,wenowhaveawhole category Xin
themathematical sense, i.e. the set ofallof the differentnumerical DGs
Mof the given geometric type (e.g. all numerical twice-differentiable
Riemannian manifolds), together with the structure-preservingmaps or
morphisms among them (among which the isomorphisms are a special
class). Our generalK-uniqueness condition then says that the repre-
sentations f of S are unique up to composition with an isomorphism k
of the category X, i.e. that for any equivalent or admissible /' thereis
sucha k in X with /' = k o /, and conversely any such k generates an
equivalent representation /'. If for somereason werestrict attention to
representations all in a single elementM of X, this K-uniqueness con-
dition will reduce to the Klein-Stevens G-uniqueness condition, with
G being the automorphism or symmetry group of M. But the present
theory is far more general,since the function category X ismuch larger
than any single transformation group G(M).
Note that in the DG case, as in most geometry, we have been con-
cerned only with onto representation. The concept of if-uniqueness is
in principle applicable also for into representation, where the condition
would be that for any two maps /, /' into structures M, M' of the
category X, both meeting conditions C, there is a morphism k in X
(of some specified type) from M to M' such that f — k o /. When M
is fixed this condition again reduces to uniqueness up to some type of
morphism of M. Note however that as X becomes larger this unique-
ness condition becomes harder to meet, and will generally require very
strong structural axioms on S andon the elements of K.
(c) Generalized Theory ofMeaningfulness
First recall from Section 2 that a theory of meaningfulness within the
RTMformat is essentially a condition on representationalpropositions
aiming to distinguish a subclass of them which express intrinsic infor-
mation about the represented structure from a larger class which need
not. Followingmy (1986b) we may write representationalpropositions
as 'p[/]'» where p is an actual sentence in some numerical represen-
tational language L. L contains terms of the form '/(s)' standing
for scale values determined by a representation, and 'p[/]' means the
proposition expressedby p when the values of those scale-dependent
terms of Lare determined using the particular representingmap /. Thus
p[f] and p[f] in general are different propositions expressible by the
same sentencep of L.
For p in L we have (as in (**),p. 420 of my (1986b)) the intuitive
characterization of meaningfulness as independence of representation:
(Ml) p[f] <-► p[f% for all / and /' satisfying G.
However, (Ml) is a semantic condition referring to the meaning of p
under the two interpretations / and /', andhencelacks clear criteria of
application. The basic goal of a theory of meaningfulness is to obtain
a precise syntactic criterion of meaningfulness defined directly for the
sentencesp ofL.
The Klein-Stevens theory of meaningfulness does this in two steps.
First weuse G-uniqueness to transform (Ml) into:
(M2) p[f] *-+ p[g o /], for all gin G and / satisfying C.
However,this too is still a semantic statement, concerning two different
interpretations of the same sentencep. The second step depends on the
further syntactic covariance assumption,
(SC) p[go/] <-> g(p)[f], for all gin G and / satisfying G.
The right side here involves a new statement g(p) of L, in which we
replace every termof the form '/(s)' in pbya term oftheform g(f{s))\

Here weassume that Lcontains a term (here written simply as '#') des-
ignating the element g of G, and in the sentence g(p) of L we prefix

that term to the functional expressions of L which express the values

f(s) of /inp.
This new sentence g(p) is a syntactic transform of the original sen-
tencep. Theintuitive justification for (SC), which canbemade rigorous
for a certain class of sentences p, is simply that each value f(s) of /
in p[f] will be acted on once more by g in g(p)[f], yielding the value
g{f(s)) at the same place in g(p)[f] where p[f] had the value f{s).
This is the same substitutionas is performed semantically when we use
g o f instead of / to interpret the original sentencep. Thus p[g o f] and
g(p)[f] willhave the same values g(f(s)) at the same places, andhence
'co-vary in truth-value. (SC) and (M2) yield the desired syntactic
criterion ofmeaningfulness:
(M3) p[f] <-► g(p)[f], for all gin G and / satisfying G.
Since both sides of (M3) are now interpretedby the same /,a sufficient
condition for (M3) is interdeducibility in L of the two sentencesp and
g(p). This is 'G-invariance' of a statementp in the ordinary mathemat-
ical sense: that by actual algebraic manipulation one can derive p from
g(p) and vice versa. Thiscriterion involvesno appeal to semantics,but
still implies the meaningfulness condition (Ml). A similar syntactic
invariance approach to representational meaningfulness was tried by
Suppesinhis early paper (1959b), but was later discarded in favor of
the RTM semantic approach.
This theory uses the Klein-Stevens uniqueness group G in both
steps: the appeal to G-uniquenessin the first step, and the appeal to a

syntactic action p ► g(p) of the group G on the sentencesp of L for
the second step. However, both of these steps can be carried over to our
extended framework. First we invoke the K-uniqueness condition of
the preceding section to rewrite the semantic meaningfulness criterion
(Ml) as
(M2') p[f] <-> p[k o /], for allkin X and / satisfying G.
Corresponding to (SC), we then must also seek some syntactic action
p —* k(p) of each k in X on the sentences p of L which mimics
syntactically the action ofa semantic change of representation/ = kof
on the propositions p[f].
This looks hard (and it is), since the action of k on / in the DG
categories X is much more complex mathematically than the simple
action of a numerical transformation group G on individual numbers.
But just such a syntactic action k(p) has already been worked out within
DG, being known as tensor analysis. Consider any pair of compatible
DG representations/ and /',and any if-isomorphism k with /' = kof.

Tensor analysis defines a syntactic action p > k(p) over L satisfying
the syntactic contravariance condition:

(SC) p[f] <-» k(p)[k o /], for all kin Xand / satisfying G.

The condition (SC) is opposite from (SC), because the tensor analysis
rules for 'transforming components' in sentences p of L are designed

to preserve the semantic content of an interpreted sentence p[f] under
a coordinate transformation / * ko /. That is, thenew sentence k(p)

will say the same thingunder the new interpretationkof thatp didunder
the interpretation /, so the syntactic action p ► k(p) of k counteracts

('contra- varies from') its semantic action / > k o /. This is expressed
in (SC).
However, either covariance or contravariance does equally wellfor
our purposes, since in either case (M2') can be applied to p or k(p) to

(M3') p[f] *-*■ k(p)[f], for all kin Xand / satisfying G.

As before, a sufficient condition for (M3') is the interdeducibility of

the numerical statements p and k(p) in the language L. Since (M3')
has been shown semantically equivalent to (Ml), this gives the desired
syntactic criterion ofmeaningfulness, as invariance under the tensorial
actions k(p) for all k in K. In tensor analysis it is shown, by direct
algebraic manipulations within thenumerical language L, that a certain
family of numerical sentences p known as tensor equations have this
invariance property,and are therefore meaningful.
Wehavethus shownthe generalpattern forconstruction ofa theoryof
meaningfulness basedon the generalX-uniquenesscondition of the pre-
ceding section as opposed to theKlein-Stevens G-uniqueness condition,
andhave seenhow toderive fromit asyntactic k(p)-invariancecriterion
for meaningfulness analogous to the Klein-Stevens G-invariance con-
dition,by a generalprocedure whichreduces to that in the G-uniqueness
case. The procedure is to construct for each k in X a syntactic transfor-
mation k(p) acting on the sentences of the language L, which satisfies
a syntactic co- or contra-variance condition. Either of these implies the
desired syntactic invariance criterion (M3').

There is no general recipe for constructing the syntactic operator

k(p): this will dependon the forms of the sentencesp ofLand the func-
tions k of K I remarked that the syntactic form of the tensorial action
k(p) is much more complicated than the simple substitutional action
g(p) of the Klein-Stevens groups G. This is because the sentences p

of DG languages always involve derivatives of the coordinate values,
so that on a coordinate change / » k o / new terms are generated
involving the derivatives of k. The Klein-Stevens theory, in contrast,
was essentially linear: first, the sentencesp rarely involved calculus,
and second, the numerical transformations g are themselveslinear for
the most common G, so that even when substituted into p under a
derivative or integral sign no additional terms involving derivatives or
integrals of g wouldbe generated. Thus eventhe G-uniqueness case can
become more complex: if Gis non-linear andL involves calculus then
the syntactic action g(p) willinvolve derivatives of g, like the tensorial


Using the generalized concepts of if-uniqueness and k(p) invariance

we have seen that two of the three central elements of RTM, unique-
ness and meaningfulness, are already present in standard DG.The first
element, qualitative axioms yielding a representation theorem, is cer-
tainly not. DG operates entirely at theKlein-Stevens level: coordinate
representation of S is simply postulated, and the theory is based on
invariance properties (uniqueness and meaningfulness) relating these
representations to one another.
This is also why we have had to postpone explicit statement of the
representationalconditions C on the maps / for DG. Suchconditions
can only be formulated within apre-representationalframework having
the capacity to make statements about the system S which are entire-
ly independent of numerical representation, and not merely invariant
under changes of numerical representation. (As I noted in my (1992),
this important distinction between statements about S which are truly
intrinsic and those which are merely invariant is generally overlooked
by physicists.) It is precisely this step which defines the line between
the earlier Klein-Stevens approach and the RTM one, as explained in
my (1986b). We now wish to take this step in the DG case.
This will require: (a) identification of a suitable set of intrinsic
primitives P in the DG spaces S; (b) articulation of representational
conditions C on the maps / in terms of the primitives P; (c) formu-
lation of axioms A on the spaces S using the primitives P; (d) proof
of a representation theorem that for any S satisfying A there is a rep-
resentation / satisfying G; and (c) proof of a uniqueness theorem that
the class of representations / satisfying G are K-unique, where X is
the corresponding DG category. This of course includes as a special
case the problem of axiomatization of the Riemannian STGs of GR,
This axiomatization problem was never articulated within the RTM
literature, probably because of the departure from the group-based
paradigm of representation which it requires. I know of only one pub-
lished line of attack on it, originated within GR to provide a more con-
crete interpretation of the physicalcontent ofSTG.Thecentralpapers in
this development are Ehlers, Pirani and Schild (1972) and Woodhouse
(1973). The motivations underlyingthis work are precisely those of my
Section 3 above: the authors recognize that extrinsic coordinate repre-
sentation conceals the intrinsic physical content of space-time theory,
and they believe that synthetic axiomatization will make that content
clearer (Ehlers etal, 1972, pp. 64-65; Woodhouse, 1973, p. 495).
However, these axiomatizations do not meet the technical criteria of
RTMandaxiomatic geometry. Both axiom systems include statements
which cannot be meaningfully expressed in terms of finitary primi-
tive relations alone, but rather require the antecedent specification of
some differentiable structure. For example, Axiom Dl of Ehlers et al.
begins: "Every particle is a smooth, one-dimensional manifold." The
term 'smooth' implies a differentiable structure, and is so used in the
subsequent derivations, whereas with finitary relations (e.g. order or
betweenness) one can only characterize the topological manifold struc-
ture ofa \-d space,but not any oneunique differentiable structure. Thus
there are no intrinsic finitary primitives in terms of which this axiom
can be meaningful. Whatever the authors mean by their primitive term
'particle', if a particle is to possess (as the proofs require) & particular
\-d differentiable structure overits constituent points, then it cannotbe
defined structurally using any finite set of finitary relations. A simi-
lar problem involving an illegitimate (because non-unique) shift from
topological to differentiable structure affects Woodhouse's Axiom 5.

These axiom systems only yield the desired results if they are read
as primitively postulating certain elements of differentiable structure on
individual particles. However, a differentiable structure is itself only
definable usingnumerical representation,sothese formulations ineffect
presuppose part of the standard apparatus of numerical representation
which a true geometric axiomatization should derivein its representation
theorem. Therefore theyreally only succeedin reducing one part of that
apparatus to another part, not in reducing the whole to something truly
However, it is in fact possible to present intrinsic axioms in at least
theRiemannian case. The possibility ofdoing sois shown by aZeeman-
type invariance proof given in the Appendix of my (1992). The qual-
itative primitives involved are natural Riemannian generalizations of
theclassical primitives of geodesic betweenness B(p,q,r) andmetrical
congruence C(p,q,r,s). These are true finitary primitives, relations
of degree three and four, respectively, over the points of the space
S, and the theorem says that any two Riemannian spaces which are
isomorphic as structures (S,B,C) are related by a standard smooth
Riemannian isometry. In terms of if-uniqueness, this means that the
(5,B, C) structure of a Riemannian space S determines its coordinate
representation as a numerical Riemannian geometryM, uniquely up to
Riemannian isometry.
We cannow alsofinally state explicitly the representationalcondition
G on the maps /, for the Riemannian case, which is of course simply
that the qualitative relations B and G in S be carried under / to the
correspondingrelations numerically definable in therepresentingspaces
M using the given coefficient functions: that triples s, t, v in S bearing
therelation B lie alonggeodesiesinMas defined using the gpq function
of M,and so on.
This invariance result is analogous to Zeeman's proof thatthe quali-
tative causal structure ofMinkowski STGdetermines its coordinate rep-
resentation uniquely up to aLorentz transformation, except that Zeeman
considered onlyrepresentations in one numericalMinkowski space M,
and hence stated only the conclusion of Lorentz G-uniqueness rather
than general if-uniqueness for a category K. Just as Zeeman's result
implies the sufficiency of the causal primitive for an axiomatization of
Minkowski STG, so this result implies the sufficiency of the two prim-
itives B and C for a synthetic axiomatization ofRiemannian geometry
(of any dimension and signature, hence including the STGs of GR). I
also have some explicit axiom systems using these two primitives, but
they are too complex to state here.
Ishould stress that the problem of axiomatization ofDG is different
for each different type or level ofDG structure: thereis no way to know
in advance what if any set of finitary primitives will provide a sufficient
basis for the axiomatization of a givenlevel of DG structure. An invari-
ance proof like that of my (1992) shows that a sufficient basis for affine
manifolds is provided by the geodesic betweenness relation B(p,q,r)
together with an affine congruence relation C(p,q,r,s), which gives
congruence relations along each geodesic rather than across different
geodesieslike the Riemannian congruence relation. For this case too I
alsohave explicit axioms.
At the other extreme, it is easy to prove that no basis of finitary
relations suffices for the axiomatization of any level of differentiable
structure alone, without additional geometrical structure. This follows
from the richness of the set of transformations relating these spaces,
which prevents any non-trivial finitary relations from beingpreserved
by all such transformations, as the basis relations for a geometry must
This last result shows how unphysicalis the procedure, common in
GR,of first postulating an underlyingdifferentiable manifold as ifit were
something independent and self-subsistent,over which some additional
geometrical structure may then be imposed as an afterthought. If one
believes that a physical theory shouldbe expressibleusing finitary rela-
tions, this result implies that the differentiable structure of space-time
must be determined by its higher levels of geometrical structure, and
could not exist physically without that. (This conclusion hasinterest-
ing connections to Mach's principle.) Conversely, the axiomatizatility
results quoted above show that the metric or affine structure definable
in terms of the primitives B and G is sufficient to yield that underlying
differentiable structure as well, just as in the classical geometries. (Of
course in classical geometry no one ever would have thought it neces-
sary to postulate an underlyingmanifold before proceeding to describe
affine or metric geometry. This odd and unphysical way of thinking
entered STGfrom modern differential geometry.)
Finally, to illustrate once more the philosophical and physical sig-
nificance of intrinsic axiomatization in the RTM tradition,I will briefly
mention thataxiomatic STGin this sense provides aclear and simple res-
olution ofa paradox which originally troubled Einstein and was recently

revived by Earman andNorton (1987), called the 'hole argument. This

problem is in my view entirely an artifact of the standard extrinsic
coordinate representation employed by physicists and the philosophers
who follow them, and disappears completely when the theory of STG
is properly expressed,in intrinsic axiomatic terms (Mundy, 1992).


This paper ought to contain one more section, combining the various
RTM-based relationist ideas of Section 5 with the intrinsic approach to
curved spatialist STGofSection 8 to explorethe prospects for relationist
theories ofcurvedSTG.On this important topic I canhere only refer to
thebrief hints of my (1989b, pp. 595-596) for the case of second-order
scalar relationism, and the observationof my (1983,p. 226) that such a
theory will have to incorporate elements of GR itself in order to have
any non-trivial content.
By way of generalconclusionI wish to stress that all of the preced-
ing material is rooted essentially in one idea and one technique, both
originating mainly with Suppes, though in my view not yet pushed far
enough. The ideais thatof informal theory ofphysicalquantity(mislead-
ingly called a theory of 'measurement'), including physical geometry
as a special case. The technique is the theory of representation: intrin-
sic axioms, representation theorems, uniqueness and meaningfulness.
Development and application of that technique in pursuit of that goal
leads to intrinsic axiomatization of quantitative theories, unified theory
of scalarand vectorquantities,relationist geometrical theories, axiomat-
ic differential geometry, and soon.

Department ofPhilosophy,
Syracuse University,
Syracuse, NY 13244, U.S.A.


TheprecedingidealizedcharacterizationofRTMandits accomplishments is developed
anddefendedin my (1986b), whereIalsoundertake to fill an important gap inthe RTM
literature by showing in a substantive and non-circular way that a representational
propositionp[f] is meaningful ifit isinvariantunder the actionof theuniqueness group
G. Istand by this claim, whichis not addressed in the discussionof meaningfulness in
The argument depends essentiallyon treating meaningfulness as a property of indi-
vidual representationalpropositionsp[f], as was done long ago in Suppes (1959b). In
contrast, RTM takes meaningfulness as a property of the relations occurring as pred-
icates in such propositions (FM-3, Ch. 22). My approach is more general because
a relation is meaningfuliff each instance of it is, whilenot every meaningfulpropo-
sition is an instance of a meaningful relation, since a predicate may be meaningful
for some instances but not others. This difference in approach may have hampered
communicationwhenIpresented thematerialofmy (1986b) at an RTM conference on
meaningfulnessin 1984; certainlyIdidnot then see its full significance, or theinherent
limitationsof the RTM approach. We willsee later that this differenceof approach is
crucial for the extension of the theory of meaningfulness to differential geometry and
other cases wherenouniqueness groupexists. These two successes of thepropositional
approach to meaningfulness illustrateits advantages over the RTM relational approach.


Burgess,J.: 1984, 'SyntheticMechanics', JournalofPhilosophicalLogic, 13, 379-395.

Catton,P. and Solomon,G.: 1988, 'Discussion: Uniqueness of Embeddings and Space-
TimeRelationism',Philosophy ofScience,55, 280-291.
Earman, J. andNorton, J.: 1987, 'What Price Space-TimeSubstantivalism? The Hole
Story', BritishJournal of Philosophy ofScience,38, 515-525.
Earman, J.: 1989, World Enough and Space-Time, Cambridge, MA: MIT Press.
Eddington, A.S.: 1939, The Philosophyof Physical Science, Cambridge.
Ehlers,J.,Pirani,F. A. E., and Schild, A.: 1972, 'The Geometryof Free Falland Light
Propagation',in: L. O'Raifeartaigh(Ed.), General Relativity: Papers in Honour
ofJ. L. Synge, Oxford: ClarendonPress, pp. 63-84.
Field,H.: 1980, Science withoutNumbers, Princeton: University Press.
Friedman, M.: 1983, Foundations of Space-Time Theories, Princeton: Princeton Uni-
versity Press.
Horwich, P.: 1978, 'On the Existence of Time, Space and Space-Time', Nous, 12,
Krantz,D., Luce,R., Suppes, P., and Tversky, A.: 1971, Foundations ofMeasurement,
VolumeI:Additiveand PolynomialRepresentation,New York: AcademicPress.
Kuhn,T: 1962, The Structure ofScientific Revolutions, Chicago: University of Chicago
Press. Second enlarged edition, 1970.
Luce, R. D., Krantz, D. H., Suppes, P., and Tversky, A.: 1990 Foundations ofMea-
surement, VolumeIII:Representation, Axiomatization, andInvariance, London and
New York: Academic Press.
Manders, X.: 1982, 'On the Space-Time Ontology of Physical Theories', Philosophy
of Science,49, 575-590.
Moler, N. and Suppes,P.: 1968, 'Quantifier-FreeAxioms forConstructivePlaneGeom-
etry', Compositio Mathematica,20, 143-152.

Moulines, C.-U.and Sneed, J.: 1979, 'Suppes' Philosophy of Physics', in: R. Bogdan
(Ed.), Patrick Suppes, Dordrecht: D. Reidel. pp. 59-91.
Mundy, B.: 1983, 'RelationalTheories of Euclidean Space and Minkowski Space-
Time',Philosophy of Science,50, 205-226.
Mundy, B.: 1986a, 'Embedding and Uniqueness in Relational Theories of Space',
Synthese, 67, 383-390.
Mundy, B.: 1986b, 'On the GeneralTheory of Meaningful Representation',Synthese,
67, 391-437.
Mundy, B.: 1986c, 'The Physical Content of Minkowski Geometry', British Journal
for thePhilosophy of Science,37, 25-54.
Mundy, B.: 1986d, 'Optical Axiomatizationof Minkowski Space-Time Geometry',
Philosophy of Science,53, 1-30.
Mundy, B.: 1987a, 'Faithful Representation,Physical Extensive Measurement Theory
and ArchimedeanAxioms', Synthese, 70, 373^-00.
Mundy, B.: 1987b, 'The Metaphysics of Quantity',PhilosophicalStudies,51, 29-54.
Mundy, B.: 1989a, 'Elementary CategorialLogic, Predicatesof Variable Degree, and
Theory of Quantity', Journal of PhilosophicalLogic, 18, 115-140.
Mundy,B.: 1989b, 'On Quantitative Relationist Theories', Philosophy ofScience, 56,
Mundy, B.: 1990, 'Mathematical Physics and Elementary Logic', in: A. Fine, M.
Forbes, and L. Wessels (Eds.), PSA 1990, Vol. 1, East Lansing,MI:Philosophy of
Science Association, pp. 289-301.
Mundy, B.: 1991, 'Discussion: Embedding and Uniqueness inRelationist Theories',
Philosophy ofScience,58, 102-124.
Mundy, B.: 1992, 'Space-Time and Isomorphism', in: D. Hull, M. Forbes, and K.
Okruhlik (Eds.), PSA 1992, Vol. 1, East Lansing,MI: PhilosophyofScience Asso-
ciation, pp. 515-527.
Reichenbach,H.: 1924, Axiomatizationof the Theory ofRelativity(in German),English
translationUniversity of California Press, Berkeley andLos Angeles, 1969.
Reichenbach, H.: 1928, The Philosophy of Space and Time (in German), English
translationDover,New York, 1957.
Robb, A. A.:1911, OpticalGeometry ofMotion: ANew Viewofthe Theory ofRelativity,
London: Heffer.
Robb, A. A.: 1936, Geometry of Time and Space, Cambridge: Cambridge University
Stevens, S. S.: 1946, 'On the Theory of Scales of Measurement', Science, 103, 677-
-680. Reprinted in: A. Danto and S. Morgenbesser(Eds.),Philosophy of Sciences,
New York: New AmericanLibrary, 1960, pp. 141-149.
Suppes, P.: 1959a, 'Axioms for RelativisticKinematics with or without Parity',in:
L. Henkin, P. Suppes, and A. Tarski (Eds.), The AxiomaticMethod with Special
Reference to Geometry and Physics, Amsterdam: North-Holland, pp. 291-307.
Reprintedin Suppes (1969, pp. 194-211).
Suppes,P.: 1959b, 'Measurement, EmpiricalMeaningfulness, andThree-Valued Log-
ic', in: C. W. Churchmanand P. Ratoosh (Eds.), Measurement: Definitions and
Theories, New York: Wiley, pp. 129-143. Reprinted in Suppes, P., Studies in the
Methodology andFoundationsof Science, Dordrecht: D. Reidel, 1969, pp. 65-80.
Suppes,P.: 1972, 'SomeOpenProblemsinthePhilosophyof SpaceandTime',Synthese,
24, 298-316. Reprintedin Suppes (1973a), pp. 383-401. Suppes,P. (Ed.): 1973a,
Space, Time and Geometry, Dordrecht: D.Reidel.
Suppes, P.: 1973b, 'Introduction', in Suppes (1973a), pp. ix-xi.
Suppes,P.: 1974, 'The EssentialbutImplicitRoleofModalConceptsin Science',in: K.
F. Schaffner andR. S.Cohen (Eds.), PSA 1972, Dordrecht: D.Reidel, pp. 305-314.
Suppes, P., Krantz, D. H.,Luce, R. D., and Tversky, A.: 1989, Foundationsof Mea-
surement, Volume II:Geometrical, Thresholdand Probabilistic Representations,
London andNew York: Academic Press.
Suppes,P.: 1979, 'Replies',in: R. Bogdan(Ed.), Patrick Suppes, Dordrecht: D. Reidel,
Teller,P.: 1987, 'Space-Time as a Physical Quantity', in: R. Kargon and P. Achinstein
(Eds.), TheoreticalPhysics in the 100 Years Since Kelvin's Baltimore Lectures,
Boston: MITPress.
Winnie, J. A.: 1977, 'TheCausal Theory of Space-Time',in: J. Earman, C. Gylmour,
and Stachel, J. (Eds.), Foundations of Space-Time Theories, MinnesotaStudies in
the Philosophy of Science, Vol. VII, Minneapolis: University of MinnesotaPress,
pp. 134-205.
Woodhouse, N.: 1973, 'On the Differentiable and Causal Structure of Spacetime',
Journal of MathematicalPhysics, 14, 495.
Zeeman, E. C: 1964, 'CausalityImplies the Lorentz Group', Journal ofMathematical
Physics, 5, 490^193.


Reading Brent Mundy's long paper on the representational theory of

measurementrecalls, for me, in a vivid way our many lively and argu-
mentative discussions about the theory of measurementand many other
matters in the philosophy of science during his years as a graduate stu-
dent at Stanford. For those who do not know Brent, the style of his
article conveys well his style of thinking. He writes well and clearly
about a variety of technical matters. He has in fact a considerable gift
for exposition of technical concepts. Interspersed with this exposition
are quirky and highly idiosyncratic views. Nowhere else in the litera-
ture on the foundations of measurement will you find anything like his
Section 5, entitled 'Materialism, Spatialism and Relationism'. On the
other hand, byanalysing critically and with care the writings of a num-
ber of philosophers about space-time andrelated topics, he has used the
concepts perfected in the last few decades in the theory of measurement
to criticize inadequacies in a variety of philosophical work. In spite
of this yeoman service,I find his views about materialism and realism

too dogmatic and sharply drawn for my taste. In his characteristic way
I am sure he would reply that this is just an expression on my part of
my piecemeal and pluralistic approach to philosophy, which too often
avoidsthe big issues. I have respondedto thisbroad sort of viewpoint in
various other commentsin these three volumes. What Mundy does not
discuss as an alternative to thekindof foundationalist view he advocates
is the instrumentalist and problem-solving approach that has its lineage
in Peirce and Dewey as an alternative fundamental way of thinking
about philosophical matters.
Inow turn to some detailed remarks on Mundy's paper. I list them
in their approximate order of occurrence in his exposition, not in terms
of their relative importance.

Eighteenth-Century Consensus. Ivery much agree with and like

Mundy's point about the eighteenth-century consensus on quantitative
methodology. From aphilosophical standpoint, one thing that is impor-
tant to noteabout this consensus is that it took place before the founda-
tions of analysis hadbeen put in proper order, orevenbefore the concept
of function had been satisfactorily clarified. What was apparent was
that the methods worked and provided new and deepinsights into the
application ofmathematics to natural phenomena.

Ancestry of RTM.At the beginning of Section 3 Mundy remarks that

RTMis essentially a theory of scaling with psychometric ancestry pri-
marily, but the fundamental papers on the representational theory were
written completely outside the psychometric tradition and even before
it was developed. I am thinkingespecially of the fundamental work of
Helmholtz (1887) in thenineteenth century, andaboveall the important
paper of Holder (1901). Moreover, without any question this tradition
as reflected in Helmholtz andHolder was very much influenced by the
slightly earlier geometricaltheory of representation as formulated clear-
ly and philosophically in a very satisfactory way by Pasch (1882) and
later given wide currency by Hilbert.

RepresentationTheoremsin Geometry. In Section 4 Mundy discuss-

esin an explicit and standard fashion representation theorems for given
geometries in terms ofCartesian orother numerical geometrical spaces.
It is important to recognize, however, that early in the geometrical tra-
dition more general representation theorems were proved over fields
that were not numerical fields and in fact satisfied no Archimedean or
continuity condition. Good modern examples are representations for
weak affine spaces over ternary fields or quasi-fields, which do not even
have standard commutative and associative operations of addition and
multiplication. Early examples can be found already in the nineteenth
century as well. This is, in fact, a very rich subject geometrically with a
vast modern literature of representation theorems that are not properly
numerical in character.

Intrinsic Formalization of Physical Theories. I agree with many of

Mundy'sremarks about the desirability of giving an intrinsic formaliza-
tion of physical theories, just as we do for geometry, but his objections
to such extrinsic set-theoretical formulations are too strong, or at least
are not sufficiently developedin the present paper. Certainly for many
physicalphenomena it would be very awkward to give suchan intrinsic
formalization and from the standpoint of a very extensivefoundational
tradition it would seem rather bizarre in the case, for example, of sta-
tistical mechanics, just to name one significant instance. The same can
be said for other parts of physics in which very extensive use is made
of probability theory. It is extremely awkward and, in fact, not done
anywhereinpractice - to formulate complicatedphysical processes that
are also stochastic processes in a purely intrinsic qualitative way from
a foundational standpoint.

Finitism in Physics. Again I agree with many of Mundy's remarks

aboutfinitism whichI findmore interesting than the discussion of mate-
rialism, but Ido wantto emphasize my own viewpoint on these matters.
Increasingly I look upon the choice between discrete and continuous
models as one driven by computational, and only computational con-
siderations. There is no real way of deciding if physical quantities are
actually continuous or actually discrete. Once a sufficiently fine dis-
crete mesh is considered, discrimination is not possible. Theone thing
that could turn out to be the case is proof that there is indeed a lower
limit and discreteness need not go beyondthis lower limit to provide a
completely satisfactory account. Letus suppose that this lower limit is
roughly on the order of Planck's constant and we shall postulate that in
terms of lengths it is of the order of 10 -40 meters. Even though we are

convinced of this by experiment andbelieve it in our hearts, we would

stillcontinue throughout large domains of physics to use in unimpeded
fashion computation based upon assumptions of continuity, not because
we believe the world was constructed this way, but because this was
the only sensible way to do any elaborate quantitative analysis. One
of the oldest and best examples of this is De Moivre's proof that for
large n thebinomial distribution is approximated very well by the nor-
mal or Gaussian distribution of the same variance. (De Moivre's proof
(1733/1738) was only forp = 1/2.) This kind of approximation would
still beused without even a shrug of the shoulders in a great range of
physical applications. These computational aspects of matters are not
emphasized at all byMundy, andin my view are too much neglectedby

Generalized Theory of Meaningfulness. I like Mundy's development

of differential geometry and the way he uses it to challenge standard
concepts of uniqueness and meaningfulness. I do think the story about
uniqueness and invariance in differential geometryis in practice a good
deal more complicated than his brief overview here suggests. The
study for example of something like infinitesimal affine transformations
in differential geometry has been a major topic and is driven by the
sameconsiderations of invariance that have driven the standard theory.
Even though I think the developmentsin Chapter 20of Foundations of
Measurement, Vol. 11l represent an excellent analysis of the standard
theory, as Luce and Narens also point out in their article in this volume,
there remain a number of problems of meaningfulness that are not yet
satisfactorily resolved. Mundy has put his finger onone set of problems,
namely, when there is not a natural group expressing uniqueness of the
I want to point to something rather different, driven by more physi-
cal and less mathematical sorts of problems. In the standard theory of
extensive measurement without anyplace for error, it is meaningful to
ask the question whether the ratio of the mass of the sun and earth is
rational or irrational or whether, if it is irrational, it is transcendental
or not. Such questions seem by ordinary physical standards of experi-
mentation and the like, nonsensical. The reason for that view is clear.
The representation as given in detail is not in terms of a unique number
relative to a given scale,butin terms of a distribution representingerror
in the measurement. This seems to me a philosophically better and
more important direction to go in generalizing the theory of meaning-
fulness than that taken by Mundy, without my needing to quarrel with
the correctness of many of his points.
If werepresent physicalquantitiesby random variables,as for exam-
ple in Suppes and Zanotti (1992), or in Chapter 16 of Foundations of
Measurement, Vol. 11, then each physical quantity is representednot by
a unique number but by a probability distribution. We can of courseask
the same nonsensical mathematical questions about the transcendental
character of the parameters of the distribution representing a physical
object but we really do not know these parameters in thismathematical
detail. The parameters themselves are estimated from finite samples
of data with error terms of their own, and we have no strong convic-
tion about the transcendental character, for example, of any estimated
parameter of anerror distribution. I do not mean to suggest thatmoving
from the representation of physical quantities by numbers to random
variables solves all the problems of meaningfulness, but it does take
care of some the problems that do not arise in the purely mathemat-
ical or geometrical theory of meaningfulness but that are natural to
raise from a physical standpoint. Notice that such matters as the ratio
of two masses has a natural expression in terms of error distribution,
although often not made explicit. That is, once we think of the masses
of two objects represented as random variables then we will take as
first approximation the ratio of the means of these distributions as the
ratio of their masses, but we willalso attachan error distribution to this
estimation. Of course, what I havecalled error willin the case of many
natural quantities be areflection of their natural variation over time, not
errors in their measurement. So, for example, it is natural to represent
the diurnal and secular change in temperature of a given location as
a stochastic process and thus as a family of random variables, or for
simple representation, as a single random variable, changing with time.

Axiomatic Space-Time Geometry. Again I find myself agreeing with

much that Brent has to say about theharder problems of axiomatizing
properly space-time geometry. He brings out well the complexities of
providing an intrinsic axiomatization for generalrelativity. I think the
direction of the various unpublished results of his own that he cites is a
practical one to obtain specific results for particular cases. Ialso think
that his strictures about beginning with a differentiable manifold alone
being a physically unsound way to begin are correct. Itis just not clear

whether anything like a genuinesynthetic tradition will ever get a very

solid foothold in general relativity, or more generally, in differential
To end on a note that is dear to my heart and in certain ways more
radical than Mundy's suggestion, let me take up again the idea of dif-
ferentiable structures. If we take seriously the concept that we cannot
hope to distinguish physically between discrete spaces with a very fine
mesh and continuous spaces, then to take seriously details of a dif-
ferentiable structure seems physically unsound,unless a case is to be
made purely on computational grounds. But if the case is to be made
purely on computational grounds then we would want to end up with
theorems showing that we have selected a differentiable structure as a
possible approximate structure just for computational purposes. This
requires a broadening of the concept of representation anda skepticism
about taking seriously much of the mathematical apparatus developed
for manifolds. This seems like a heterodox view that goes too far, and
not too long ago I would have said so myself, but the modern move
to digital computation of solutions of all kinds of physical problems
formulated in terms of partial differential equations, etc., suggests that
the mathematics of the future will be more serious in not accepting
as an unbridgable chasm the difference between the discrete and the
continuous,or even more, the differentiable.


De Moivre, A.: 1733, Approximatio terminorum binomii(a + b)n in seriem expansi,

pamphlet,London.Translated andcommented on inhis TheDoctrine ofChances,
2nd cd.,London, 1738.
Helmholtz, H. yon: 1887, 'Zahlen und Messen erkenntnis-theoretisch betrachtet',
Philosophische Aufsdtze Eduard Zeller gewidmet, Leipzig, 1887. Reprinted in
Gesammelte Abhandl., Vol. 3, 1895, pp. 356-391.
Holder,O: 1901, 'Die AxiomederQuantitat und dieLehre vomMass', Ber. Verh.Kgl.
Sachsis. Ges. Wiss., Leipzig, Math.-Phys. Classe, 53, 1-64.
Pasch, M.: 1882, Vorlesungen iiber Neuere Geometric, Leipzig: Verlag yon Julius
Suppes, P. and Zanotti, M.: 1992, 'Qualitative Axiomsfor Random Variable Repre-
sentationsof Extensive Quantities', in: C. W. Savage andP. Ehrlich (Eds.), Philo-
sophicalandFoundationalIssues inMeasurement Theory, Hillsdale,NJ: Lawrence
Erlbaum, pp. 39-52.


ABSTRACT. Iargueherethatthe computationalmethodsofnumericalexperimentation

constitute a distinctively new kind of scientific method, intermediatein kindbetween
empiricalexperimentationandanalytictheory. Aparallelisalso drawnbetweenextend-
ing our senses with scientific instrumentsand extending our mathematicalpowers by
usingcomputationalinstruments. A specific applicationofthese methodsto Isingmod-
els is describedin detail.

Throughout his career, Patrick Suppes has emphasized the centrality

of probabilistic methods in modern science. He has also insisted on
maintaining a commitment to a rigorous scientific empiricism within
the philosophy of science. Equally important is his repeated insistence
that we keep in mind the fact that real science is complex, messy,
and highly sophisticated. We should not, as he often put it, indulge
in 'philosophical fantasies' about the nature of science but address
ourselves to the content of actual science. I think that Pat never really
cured me of my tendency towards philosophical fantasies, but I hope
that this paper at least combines these three components ofhis approach
in a novel way.

Overthecourse ofthe pastcouple of decades, computational sciencehas

become an increasinglyimportant methodin many areas of the physical,
biological, psychological, economic and other sciences. Because of
the wide variety of methods subsumed under the general framework
of computational science, it is not easy to characterize the methods
succinctly, but for the purpose of this paper Ishall use the following
working definition:
Computationalscienceis thedevelopment,explorationand applicationof mathematical
modelsof non-mathematicalsystems using concrete computational devices.
To flesh out this working definition, Iinclude in the class of con-
crete computational devices both digital and analogue computers (i.e.

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 103-121.
© 1994 Kluwer Academic Publishers. Printedin the Netherlands.

analogue devices that are used for computational purposes and not just
as physical models). Humans are one special kind of computational
device, but the kind of novel philosophical issues that Iwant to dis-
cuss here occur only when we move beyond the category of human
computers. Moreover, it is important that the computation is actually
carried out on a concrete device, for mere abstract representations of
computations do not count as falling within the realm of computational
science. To be counted as part of computational science proper, the
computations must have a dynamic physical implementation. (I do,
however,include computations run on virtualmachines in the class of
This working definition is simple, but it already subsumes such pro-
cedures as applied finite difference methods, Monte Carlo methods,
molecular dynamics, Brownian dynamics, semi-empirical methods of
computational chemistry, and a wholehost of other, less familiar meth-
ods. For those of us who are interested in how theories are applied
to nature, the most important immediate effect of using thesemethods
of computational science is that a vastly increased number of models
can be brought into significant contact with real systems, primarily by
circumventing the serious limitations that our present restricted set of
analytical mathematical technique imposes on us. These constraints
should not be underestimated, for once we move past the realm of
highly idealized systems, the number of mathematical theories that are
applicable using only analytic techniques devoid of approximations is
very small. I shall not dwell on this aspect here, but refer the reader to
Humphreys(1991) for examples of this analytic unsolvability.
Many of the other features of computational science have no real
interest for philosophy, important though they are to the working sci-
entist. But some aspects of this methodology give rise to questions
that have clear philosophical content. Amongst these Ishall focus
on two specific issues. First, it is often claimed that computational
science provides us with a new kind of scientific method, one that is
complementary tothe existing methods of theory and (empirical) exper-
imentation. Such claims are often rather vague about just what is new
in these methods,ranging through a variety of claims such as: (1) that
computational science allows a reduction of the degree of idealization
needed in the models as well as a check on those remaining idealiza-
tions that are made; (2) that they allow far more flexibility and scope
for changing boundary and initial conditions than do real experiments;
(3) that they are precisely replicable; (4) that many parameters can be
varied that could not be altered by real experiments, perhaps because
such a change would violate a law of nature, perhaps because as in the
case of astrophysical systems the very idea of large scale experiments
is absurd; and soon.
Ishall argue here that at least one specific aspect of computational
science does indeed introduce a genuinely novel kind of method into
science. That method is numerical experimentation, andIshall argue
for this claim with reference to a particular class of mathematical models
- thelattice models of statistical physics;using a particular type ofcom-
putational method - the Monte Carlo -
method; and applying a specific
solution procedure to that method the Metropolisalgorithm. Because
of the wide applicability of MonteCarlo methods within computational
science, this example constitutes a lesser degree of special pleading
than might appear at first sight. Fritz Rohrlich (1991),followingNaylor
(1966) and others, has similarly argued for the importance of comput-
er simulations as a kind of 'theoretical model experimentation. The
examples used here reinforce his perspective while exhibiting a some-
what different aspect of computational experimentation, and they have
the additional advantage that they tie together two topics of perennial
philosophical concern, probability and empiricism, in a rather unusual
The secondfeature of philosophical interest involves aparallel that is
sometimes drawn between the useof computational science and the use
of scientific instruments. Whereas the latter enable us to transcend the
limitations ofour unaided sensory capacities,the former, so itis claimed,
enable us toextend ourlimitedmathematical abilities,especiallyincases
where computational processes play an important role. The degree to
which this parallel holds, if at all, is important at least because the
instrument sideof the parallel has hadprofound and well-known effects
on the development of contemporary scientific empiricism. We are all
familiar with the arguments that were developed in the middle third
of this century which produced serious difficulties for certain kinds
of foundationally inspired empiricisms. Many things too small or too
far away to be seen with the naked eye and things emitting radiation
outside the realm of the visible are now routinely considered to be
observable. For example, observing a cold virus under an electron
microscope seems to many of us to be not only a perfectly legitimate
part of scientific practice, but one that ought to be acceptable to a liberal

empiricist. There is no need to rehearse here the consequences that

extending the concept of 'observable' to include what is detectable by
scientific instruments has had on empiricism. It is enough to remind
oneself that they have been profound. So we must ask: if the parallel
between computational science and instrumentation is soundly based,
what consequences are there for how we should construe the use of
mathematical models in science? Ishall restrict myself here to the
issue of how the parallel affects numerical experimentation, and in
particular, whatkinds of epistemological criteria are appropriate to the
new methodology. But enough of the preliminaries. Letus turn to the
details of the model we shallexamine.


Ferromagnetism involves spontaneousmagnetization ofmaterials such

as iron andnickel when the temperature is lowered below the Curie or
critical temperature Tc The exchange energy involved in ferromag-
netism is a specifically quantum mechanical phenomenonmanifesting
itself at the macroscopic level,and one of the central applications of the
lattice models is to study this phenomenon.2 These models consist of
anra-dimensional lattice (m < 3), with thenodes occupiedby particles
having spin values Si. The Hamiltonian for the Ising model, which
describes a ferromagnet with strong uniaxial anisotropy is givenby:
Rising =—J z2 &iSj

Here Jis the exchangeenergy constant,restricted initially to interaction

between nearestneighbours. If J > 0, then the pairs (up,up), (down,
down) are favoured for neighbourpairs and this gives rise to the spin
alignments that produce ferromagnetism. If J < 0, then the pairs (up,
down),(down, up) are favoured, resulting in antiferromagnetism. His
the external magnetic field, ft is the magnetic moment of a spin and the
second term on the right is the Zeeman energy. When generalized to a
fully isotropic ferromagnet in the Heisenberg model we obtain

—— J 2_^(§i ' Sj) - pHz S?
ij i

(Sf)2 + (Sf)2 + (Sf)2 = l.
Here, as opposed to the Ising model in which the spins have values
+1 or -1 along the preferred axis,in the Heisenbergmodel the spins can
take on continuously many orientation values. Among the simplifying
assumptions used in these models are: (a) the kinetic energy of the
particles associated with the lattice site is neglected; (b) only nearest
neighbour interactions are considered (although this can be relaxed to
include nth neighbour interactions); (c) spins have only two discrete
values;(d) J, Hare considered to be uniform (this can also be modified
to allow random exchange constants J{j and different magnetic fields at
each point in the lattice).
Contact with empirical data comes through the calculation of aver-
age values for macroscopically accessible quantities, such as the mean
magnetization. In order to calculate these averages, the model is sup-
plementedby achoice of thermodynamical ensemble withits associated
probability distribution over states of the system. A common choice
here is the canonical ensemble, i.e. an ensemble of closed systems in
thermal equilibrium with a heat reservoir. Then for a (macroscopic)
observable A and a system with continuously many degrees of free-
dom, the thermal average is givenby

(1) (A(x))T = Z~ x [ exp[-H(x)/kBT]A(x) dx


Z= / exp[-#(x)/fc fiT] dx.

In the isotropic Heisenberg model, we have continuous degrees of
freedom leading to a properintegral,but inmodels where there are only
finitely many degrees of freedom, such as the Ising model, the integral
in (1) will, of course, be replacedby a summation.
The integration or summation is over the configuration space X,
which in the lattice model has as points Af-dimensional vectors X =
(5i, ,Sn) specifying the spin state for each particle in the lattice.
This introduces in an explicit way one probabilistic aspect of the mod-
el in that the normalized Boltzmann distribution exp[— H(x.)/kBT]/Z
describes the probability of the configuration xin thermal equilibrium.
The representation giveninEquation(1) is familiar, tidy, and explic-
it. Yet even given the severe simplifications involved in (a)-(d) above,

Equation (1) is impossible to apply in any realistic context. The inte-

grals involved in the Heisenberg model are unsolvable in any explicit
way, and although there exist analytic solutions to the Isingmodel for
one- and two-dimensional lattices, there is no such analytic treatment
for the three-dimensional lattice. Moreover, even in the discrete case,
for very simple models oflattices with 102 nodes in each spatial dimen-
sion,there are 210 possible states to sum overin the three-dimensional
case. Analytically intractable integrals produce one standard contextin
which Monte Carlo methods are used, andin employing these methods
we switch from a treatmentin which adeterminate solutionis produced
byanalytic methods to onein whicha statisticalestimate of thatsolution
is provided. To be clearabout how probabilities operate in this context,
itis essential to separate the fact that we began with a probabilistic mod-
el involving a probability distribution for the canonical ensemble over
configurationspace [x] from the probabilistic techniques that areused to
produce a solution to the multiple dimensional integral (1). We are here
concerned with the second of these probabilistic aspects. Traditional
integration methods,including those using approximation methods, do
not treat these averages as stochastic quantities to be estimated byprob-
abilistic methods. In contrast,Monte Carlo methods do just that, andin
fact Monte Carlo methods can be applied equally well to solvemodels
that are based on processes that are themselves deterministic. A brief
outline of how these methods are applied is neededhere.


The bridge between traditional methods of integration andMonte Carlo

methods is providedby the mean value theorem, given here for the one
dimensional case: If v, v are continuous functions on [a, b] and v does
not changesign in the interval [a,b], then thereis an £ in [a,b] such that
6 b
/ u(x)v(x) dx = u(£) / v(x) dx.
a a

Thus, when v(x) = 1everywhere, we have

(2) / u(x)dx = u(Z){b-a).
Here w(£) is themean value of v on [a, b].
Hence the point of Monte Carlo methods is to statistically estimate
the mean value u(£). then to calculate
u(x) dx

using (2). Now suppose we consider the quantity u(X) to be arandom

variable,and we sample values U{ = u(Xi) at random. Then
(3) u(x) =
Finally, by appeal to the strong law of large numbers, one can justify
probabilistic convergenceof the statisticalestimator u(x) to the analytic
mean value u(£).
Now this argument clearly has nothing to do with computational
devices; it is a result that derives solely from familiar facts about the
calculus and probability theory. But there are two facts that make this
otherwise routine piece of numerical mathematics potentially interest-
ing. The first is that because of the large dimension of the configuration
space involved in the lattice models (andin most other models with any
pretence to realism) the computation involvedin (3) has tobe carried out
in practice by resorting to computational devices using the technique
that is our principal focus of attention here, numerical experimenta-
tion. This technique can be illustrated by means of the Metropolis
algorithm. Within the general procedures for Monte Carlo estimators,
drawing sample points from configuration space for multidimensional
integrals in a non-random way often results in large portions of the
configuration space being sampled negligibly often. Using a uniform
probability distribution on the points of configuration space is usually
inefficient,because many points of the space are highly improbable and
correspondto negligibly small contributions to theestimator. Given the
large dimensional space over which the estimation takes place, this will
produce a rate of convergencethat is far too slow for the computation to
be carried out in practice. If,instead, the points in configuration space
are selected by a distribution that mimics the profile of the function
that is already being used to weight the quantity whose mean we are
interested in, then this makes much more efficient use of the sampling

Tosee this, suppose wehavechosen anensemble (here the canonical

ensemble) that has a probability distribution f(H(x)) over the energy
states. Then, in simple sampling, we would approximate the ensemble
average for an observable A by
(A) «X>(xO/(#(xO)/£/(tf(x,))
i=i ;=i

where x; is some element from phase space (e.g. the Af-dimensional

spin space [x; = (5i, ,Sat)]. But to avoid wasting time on ele-
ments of phase space with low probability, we choosethe elements with
probability P(xi) to get
(A) j
A{x )p-\x )f(H{x ))lY P-\xi)f{H{xi)).
l l l j

t=l t=l

(Note: this comes from choosing points according to the measure

P(xi) dx instead of theuniform measure dx.)
Now in the case of Ising models, suppose we choose P(xi) to be
proportional to the equilibrium distribution f(H(x)), which here is the
Boltzmann distribution. So, for example, if in the Ising model we
choose a sampling distribution
P(x) = Z~ x exp[-H(x)/kBT]
then the weightedaverage


reduces to

But in many cases the exact values of the equilibrium distribution

cxp[-H(x)/k BT]
are not known for a specific Hamiltonian, and so we would seem to
havemade no progress.
To circumvent the difficulty, the Metropolis algorithm constructs a
random walk in configurationspacehaving alimit distribution identical
to that of the sampling function P. If this were merely a piece of
abstract probability theory aboutMarkov processesit would, of course,
be nothing exceptional. The key fact about this method is that the
limit distribution is not computed by calculating a priori the values
of the limit distribution. Rather, by running concrete random walks
within numerical models of configuration space, the limit distribution
so generated is proportional to the equilibrium distribution.
The Metropolis algorithm is easily described. First, pick an arbitrary
pointin configurationspace,Xq. Then,if weare at point X{, to generate

amove in the random walk, choose anew point in configuration space,
Xj. Compute the transition probability w(Xi —► Xj) = r,for which
one common choice is the Metropolis function:
w(Xi Xj) = mm{l,exp(-[H(Xi) -H(Xj)]/kBT}}.
Next, generate a uniformly distributed random number md, and if
r > md let Xn+\ =V, otherwise let Xn+\ = Xn Iterate the pro-
cess for sufficiently many steps until equilibrium occurs. Then repeat
therandom walk choosing a different initial starting point. Iterate until
a frequency distribution of equilibrium states hasbeen generatedby the
equilibrium end states of the repeated random walk. (For a proof that
Peq(Xi) w(Xi), see Metropolis etal. (1953)).
Coupled with the transition probability is a physically interpretable
dynamics in configuration space that generatesnew states fromold. Two
common choices for the Ising model are the single spin flip dynam-
ics, within which the spin at the selected point is reversed, and the
spin exchange dynamics, within which the spin at the selected point
is exchanged with a neighbour. Putting this in the concrete case of
spin flip dynamics for the Ising model, an initial configuration for the
lattice is chosen, and then a random lattice site is picked and the spin
there is flipped. If this results in a decrease in total energy, then the
Metropolis transition probability (which in that case willbe unity) will
choose the new configuration and thenext step in the random walk will
be generated. If,in contrast, the spin flip results in an increase in total
energy, then the Metropolis transition probability is compared to the
uniformly distributed random number, and if the transition probability
is greater, the new configuration is accepted. If not, the process retains
the old configuration. What the Metropolis algorithm does, then, is

to get from an initial configuration to one in which (a) the energy is

globally, rather than locally, minimized and (b) the probabilities are
distributed according to the equilibrium distribution.


There are certain distinctive features of numerical experimentation and

the Metropolis algorithm thatneed emphasis. First,the equilibrium dis-
tribution that the algorithm has as a limit is not in general analytically
computable in the sense that its values canbe calculated explicitly for a
closed form representationof the probability. The only way to compute
these probabilities is by means of the numerical random walk generated
by the Metropolis algorithm. Although thisrandom walk isamodel ofa
physicalprocess, where the generation of the random numbers couldin
principle be carried outby physically indeterministic means, thusresult-
ing in a traditional kind of empirical experimentation using stochastic
processes, themethod describedhere is carried out purely numerically.
It is this simulation by numerical methods of physical process that has
lead to the term 'numerical experimentation. There is no empirical
content to these simulations in the sense that none of the inputs come
from measurements or observations on real systems, and the lattice
models are just that, mathematical models. It is the fact that they have
a dynamical content due to beingimplemented on a real computational
device that sets these models apart from ordinary mathematical models,
because the solution process for generating the limit distribution is not
one of abstract inference,but is available only by virtue of allowing the
random walks themselves to generatethe distribution. It thus occupies
an intermediate position between physical experimentation and numer-
ical mathematics, being identical withneither.
A quick comparison with the well-known Buffon's needle process is
illuminating. Although this process is perhaps best known to philoso-
phers for havinginduced Bertrand to invent his paradoxes of equiproba-
bility, Buffon's originalprocess isoften considered to be the progenitor
of empirical Monte Carlo methods. In its simplest form, the process
consists of casting a needle of length L onto a set of parallel lines unit
distance apart. (Here L < 1.) Then,by an elementary calculation, one
can show that the probability that the needle will intersect one of the
lines is equal to ILJ-k. As Laplace pointed out, frequency estimates
of this probability produced by throwing a real needle can be used
to empirically estimate the value of n, although as Gridgeman (1960)
notes, this is a hopelessly inefficient method for estimating -k, since if
the needle is thrown once a second day and night for three years, the
resulting frequency will estimate -k with 95% confidence to only three
decimal places. Be this asit may, the Buffon-Laplace methodis clearly
empirical in that actual frequencies from real experiments are used in
theMonte Carlo estimator, whereas the standard methods for abstract-
ly approximating the value of 7r using truncated trigonometric series
are obviously purely mathematical in form. In contrast, the dynamic
implementation of the Metropolis algorithm lies in between these two
traditional methods. It is in this sense at least that I
believe numerical
experimentation warrants the label of a new kind of scientific method.

This point needs further discussion because it can easily be miscon-

strued. What of the fact that this is, after all,an algorithm within which
even the random numbers are generated according to a deterministic
formula? Does this not mean that in principle, this is no different from
any traditionalmathematical model and that ourinability to carry out the
requisite computations ourselves, and thus delegating them to an artifi-
cial computational device,is nothing but a practical concern andhence
devoidof philosophical interest? Here the parallel withinstrumentation
is relevant. It is natural to respond to this fact by appealing to an 'in
principle' argument,and to use the contingency ofourlimited computa-
tional powers to argue that the inability to carry out such computations
in practice is philosophically irrelevant, no matter how important it
might be for the purpose of applied science. Buthere it is necessary to
make a sharp distinction between on the one hand philosophical issues
concerned with mathematics conceived as an autonomous discipline
divorced from its scientific applications and on the other hand philo-
sophical issues concerned with that subset of mathematics needed for,
and sometimes developed in explicit recognition of, its applications in
scientific models. One could conceive of this subset as merely being a
part of pure mathematics,and that the philosophically important issues
occur when considering therelation between themathematics andreal-
ity. But that way of viewingcomputational science has the potential for

transferringphilosophical positions that are appropriate for the analysis

of pure mathematics to the area of appliedmathematics, where they are
not so obviously appropriate. For us, the key concept is that of 'com-
putable. When philosophers use this term, they almost always have
in mind the purely theoretical sense centred around Church's Thesis.
But the dispositional aspect of the concept 'computable' can also be
construed as involving what can be computed in practice on existing
computational devices. Consider now the parallel with the concept of
'observable' mentioned inSectionI.Many of thedisputes within empiri-
cism about the concept of observability hingedon whetherit should be
taken as fixed or as changeable. Those minimalist empiricists who
took 'observable' to be whatever is detectable by theunaided senses of
humans rejected the idea that what was observable depended upon the
current state of technological enhancement of our sensory apparatus.
For these minimalists, the epistemic acceptability of thingsbeyond the
immediately accessible was a matter of whether the terms referring to
those perceptually inaccessible entities were reducible by definition to
terms referring to things that were immediately accessible to the senses.
In contrast for the non-minimalists, what was observable was (at least
in part) a matter of what contingently available instrumentation was
The situation with the concept of 'computable' is in some ways the
inverse image of this situation regarding what isobservable.Minimalist
empiricists had a verynarrow conception of whatis observable, whereas
the realm of what was viewed as observable by the technological non-
minimalists was much wider. With therevised concept of 'computable'
the wide domain of recursive functions is drastically reduced in size to
obtain the concept of computable that is appropriate for technologically
accessible computational science. But within the domain of this tech-
nologically enhanced conception there remains a parallel, for what is
considered computable goes far beyond whatis actually calculable by
the limited powers of ahuman.


Itshould be noted that there is a double element of contingency in this

appeal to actual computability. It is obvious contingentupon what is
currently available to usin terms ofartificial computational devices, but
it is also contingent upon the complexity of the world itself. For, if
the number of degrees of freedom involved in physical systems was
always very small, we should not need augmentation of our native
computational abilities to employ the Monte Carlo methods described
above. (We might need such augmentation to deal with other kinds
of analytically intractable problems,but weare dealing here only with
the case at hand. For some other cases where analytical intractability
forces computational science on us, see Humphreys (1991).) But the
world is not simple in that particular way and in order to apply our
theories to systems with large number of degrees of freedom, such
augmentation is forced upon us. This double element of contingency
is not a reason to reject the position for which I
have argued. In fact it
is exactly the kind of consideration that should place constraints on an
empiricist's epistemology, because what is acceptable to an empiricist
has to be influenced by the contingencies of our epistemic situation
rather than by appeal to superhuman epistemic agents free from the
constraints to which we are subject. For in fact, an exactly parallel
double contingency holds in the case of observability. How far the
concept of 'observable' gets extended beyond the realm of what is
accessible to theunaided senses depends not only upon the current state
of technological developmentin instrumentation,but upon whatis in the
world. Ifthe world were composedentirely of relatively close medium
sized objects devoid of microscopic internal structure, and it possessed
onlypropertiesperceivable with our five senses, then we could not argue
with theminimalist conception of observability. But of course the world
is not that way, and so the minimalists do have to defend their limited
conception of the observable.
I note in passing here that van Fraassen's recent (1980) attempt
to return to a minimalist conception of what is observable is subject
to objections along the above lines. For van Fraassen, the moons of
Jupiter countas observablebecause we,as presently constituted humans,
could see them directly from close up. But this possibility depends
on technological advances just as much as does the observability (in
thenon-minimalist sense) of a charged elementary particle, which van
Fraassendenies is observable. But why allow the use of technology in
one case and not in the other? Evenifit were said that the technological
enhancement provided by space craft is an enhancement of a non-
sensory faculty we have, that of locomotion, andis thus different inkind
from the enhancement providedby bubble chambers, we could respond

that our ability to observe something directly is a relation between us

and the object, and it simply happens to be a linear spatial relation in
the moons of Jupiter case, and a relative spatial size relation in the
elementary particle case, if we are observing the particle, that is, rather
than its charge. But these are not clearly relations of a different type.
So, there isindeed a significant parallel thatcanbe drawnbetween the
issues involvedin instrumentational enhancement andthoseinvolved in
computational science. The most immediate conclusion to draw is that
in dealing with issues concerning theapplication ofmathematical mod-
els to the world,as empiricists we shoulddrop theorientation ofan ideal
agent who is completely free from practical computational constraints
of any kind, but not restrict ourselves to a minimalist position where
whatis computable is always referred back to the computational com-
petence of human agents. If, as many minimalist empiricists believe,it
is impermissible to argue that humans could have had microscopes for
eyes or could evolve into such creatures, and it is impermissible to so
argue because empiricism must be concerned with the epistemological
capacities of humans as they are presently constituted, then it ought to
be equally impossible to argue that in principle humans could compute
at rates 106 faster than they actually do. But that is just not plausible.
The position for whichI have argued shouldbe keptseparate from an
apparently similar issue that has been much discussed in the literature
on connectionist architectures in artificial intelligence. One of the first
applications of the Metropolis algorithm was to theprocess of simulated
annealing, where a liquid that is cooled slowly ends up in an ordered
state corresponding to an energyminimum. (In contrast, if the liquid is
cooledquickly, itends up in a state with ahigher energy level.) The sim-
ulated annealing methodhas also been employedby connectionists,for
example in Hopfield nets, to generaterelaxation schedules for multiple
constraint problems.3 Because it is often asserted that one of the major
differences between connectionism andclassical (rule-based) artificial
intelligence lies in the fact that the latter employs algorithms that are
explicitly computable, whereas the former employsnon-linear functions
that are not, and in consequence systems have to be allowed to settle
themselves into the appropriate output state, it might seem that there is
a similar lack of computational transparency in the Ising models. But
this is not so. Although there are interesting problems posedby the use
ofcontinuous analog computational devices, the Monte Carlo methods
described here are all computationally transparent, implementable on
digital machines and give finite approximations to thelimit distribution.


Iam grateful to Fritz Rohrlich for comments that helped improve this
paper. The preliminary work for the paper was done under NSF grant
# DIR 8911393, and an earlier version was read in June 1992 at the
Venice conference onProbability and Empiricismin theWork of Patrick

Department ofPhilosophy,
University of Virginia,
Charlottesville, VA 22903, U.S.A.

Those uninterested in the technical details of these models can move directly to
Section IV, referring back as necessary. It is not possible to fully understand why
numerical experimentationis forced on us without a grasp of the physical models,
The expositionof the formalaspects of latticemodels givenhererelies heavilyon the
treatment in(Binder andHeerman, 1988).
3 See e.g. (Rumelhart andMcClelland, 1986, Ch. 7).


Binder, K. andHeerman, D. W.: 1988, Monte CarloSimulation inStatisticalPhysics,

Berlin: Springer-Verlag. '
Gridgeman,N.: 1 960, 'GeometricProbability andtheNumber tt ,ScriptaMathematica,
25, 183-195.
Humphreys,P.: 1991, 'Computer Simulations', in: A. Fine,M.Forbes, andL. Wessels
(Eds.), PSA 1990, Vol. 2, East Lansing: Philosophy of Science Association.
Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller, A., and Keller, E.: 1953,
'Equationof StateCalculationsforFast Computing Machines',/.ChemicalPhysics,
6, 1087ff.
Naylor, T.H.: 1966, Computer Simulation Techniques,New York: John Wiley.
Rohrlich, F: 1991, 'Computer Simulationsin the Physical Sciences', in: A. Fine,M.
Forbes, and L. Wessels (Eds.), PSA 1990, Vol. 2, East Lansing: Philosophy of
Science Association, pp.507-518.

Rumelhart, D. and McClelland, J.: 1986, ParallelDistributed Processing, Vol. 1,

Cambridge, MA: MIT Press.
Van Fraassen,B.: 1980, The Scientific Image,Oxford:The ClarendonPress.


Paul Humphreys is undoubtedly right in emphasizing the increasing

importance of numerical experimentation, an importance sufficiently
great as to warrant the label for a new kindof scientific method that has
been introduced. It also seems apparent that it is going to be some time
before we have a stable typology of the different ways of approach-
ing numerical experimentation. On the one hand,numerical methods
for solving differential and integral equations already had a respectable
beginning inthe nineteenth century, but the serioususeof Monte Carlo
methods hardly predates the significant use of such methods by yon
Neumann andUlam in the 19405. The generalrecognition of the impor-
tance of finite element methods for solving complexphysicalproblems
is even more recent. Still more recently, there is hope that therapeutic
drugs of the future canbe designeddirectly bylarge scale computational
analysis ofmolecular structures. In stillanother direction, the statistical
analysis of large bodies of data has been completely transformed by
the possibilities of large scale computations. In fact, it is fair to say
that no large scale statistical analysis of data takes place today without
extensiveuse of numerical computational facilities.
I also like the parallel that Paul draws between the use of computa-
tional methods and the use of scientific instruments. Modern science
is wholly dependent upon high technology for both instruments and
computations. Current theoretical and experimental work would be
unthinkable without them. It is worth noting, however, that in philo-
sophically simple concepts of rationality much more attention is paid
to observation than to computation which is, I think, an emphasis that
is mistaken and has been only partly addressed by ahandful of serious
articles. Extensive computation, as anyone familiar with the American
tax code well knows, canbeas important to individual rationalbehavior
as it is to modern scientific research.
Thereare three conceptualremarks I want to make about numerical
experimentation. They do not go against the grain of what Paul has to
say, but areintended to amplify ourperspective on future developments.

Computation and Complexity. The groundbreaking work of Turing,

Church andothers established a sharp dividing line between that which
is computable and that which is not. This is one of the genuinely nov-
el mathematical results of the twentieth century, but the distinction is
not the right one for almost all numerical experimentation. We must
draw conceptual distinctions within that which is computable and this
theory in itselfis closely related to Kolmogorov's theory of complexity.
A familiar example to philosophers is Tarski's decision procedure for
elementary algebra and geometry. This procedure is not feasible, in
the sense that it is exponentially explosive as the size of formulas to
be decided as true or false increases. Within computer science we now
have the standardnotion of a feasible computation, that is,one that can
be made in polynomial time or space. But even this general notion of
feasibility is probably not really satisfactory for down-in-the-trenches
computation, when in many cases something close to linear, not just
polynomial, constraints is needed. There is no question about the con-
ceptual importance of the distinction between polynomial as opposed
to exponentialconstraints as lower bounds, but whether the constraints
below polynomial will play a correspondinglycritical conceptualrolein
our thinking about numerical experimentation is probably not yet clear.

NumericalUnease. Itseems to me that the kind of detailed model pre-

sented by Humphreys in the secondpart ofhis paperon ferromagnetism
is typical of what we shall see in the future and it undoubtedly has
already created and will continue to create a great sense of conceptual
unease about the nature of our scientific theories of complex phenom-
ena. We simply will not be able to compute even the most elementary
cases in an analytically closed form. Almost everything willconsist of
numerical computations in one form or another. The highly tentative
and experimental character of much of the results will consequently
bring about rather sharp psychological changes in the attitude toward
theory. This can be seen very clearly in Sparrow's (1982) numerical
analysis of the now famous Lorenz equations in chaos theory. This is
not numerical analysis in the traditional sense but experimental mathe-
matics, or to use Humphreys' phrase, numerical experimentation in the
modern sense. The Lorenz equations define a three-dimensional sys-
tem of ordinary differential equations depending on three real positive
parameters. What is important is that different values of the parameters

radically change the behavior of the solutions. The Lorenz equations

are especially of interest because for certain values of the parameters
we get what we now call chaotic behaviour. What is important about
this kind of example is that in spite of the apparent simplicity of the
equations we are able only to understand and study a relatively small
range of values of the parameters with any thoroughness. Even those
explorations are necessarily tentative in character.
If Sparrow's work and others' of a similar sort, much more oriented
toward practical problems, are any guideline, it may well be that the
greattradition of nineteenth-century mathematical physics - to focus on
a smallnumber of carefully-selectedproblems whose solutions can be
written down in closed form - may be coming to an end. The future,
at least within the present range of mathematical and computational
techniques, may be much messier and inconclusive in the analysis of
natural phenomena. If so, history will record the period running from
Euler to Lord Rayleigh as one of the greatperiodsofintellectual history,
comparable to the development of astronomy in Hellenistic times. We
shall not return to that paradise of elegant mathematical analysis but
certainly we can look back upon it with nostalgia.

Will Theories Change? But more than questions of nostalgia are in-
volved. Itis easy to trace the decay of geometrical methods used stillin
suchbeautiful form by Newton, as the power of the calculus took over.
The samemay be true of the differential equations and integralequations
that have been the mainstay of mathematical methods in the physical
sciences since the eighteenth century. If the actual solutions are really
going to be computed in terms of discrete models or by probabilistic
approximations of a similar sort, then it would be a natural prediction
thatas thesemethods take over, increasingly the old ways offormulating
theories will disappear. New ways more congenial to the modern style
of computing rendered necessary by the complexity of the problems,
will change the shape of theory itself. As I have already remarked
in discussion of the future of theories of measurement, the continuum
may disappear as an object of great interest, and the same may be true
for the use of the continuum of real numbers in the theory of physical
But I do not mean to suggest that the useofcontinuous mathematics
will vanish. The interplay will surely be more subtle than that. Let
me just give one simple example, the approximation of the binomial
distribution by the normal distribution for n of any size. The asymp-
totic correctness of this approximation for p = 0.5 was proved at the
beginning of the eighteenth century by DeMoivre andhas been extraor-
dinarily useful ever since. Even though the normal density is itself
not integrable, it is most convenient to compute the normal deviates
rather than to compute the binomial distribution for large n, which is
technically unfeasible. In similar ways the host of special functions in
mathematical physics will have acontinued active use. What has real-
ly changed is thekind of philosophical or ontological commitment to
naturereally beingnot only continuous but almost everywheredifferen-
tiable. Therole of the continuum of realnumbers andall the associated
mathematics will continue to be of great importance in various special
ways, but only for computational purposes. This is a conjecture that
goes against the grain for many people, but it is at least reasonable to
makein view of the massive changes in the way scientific theories are
actually applied to empirical data.
One addendum to this story. Of course the applications ofmathemat-
ically formulated theories to data have always been more complicated
than philosophers of science would like to recognize. This is easily ver-
ified from close examination of much scientific work in the nineteenth
century. Letme givejust one example. The reallycomplicatedcharacter
of computations, both analytic and numerical,required to solve any real
problems in the classical theory of wave optics was alreadyrecognized
and studied carefully in the last century. LaPlace's massive calculations
in his great work CelestialMechanics provide a muchearlier example.
This list is easily extendedin many different directions.


Sparrow,C: 1982, The LorenzEquations: Bifurcations, Chaos andStrange Attractors,

New York: Springer-Verlag.


ABSTRACT. The central question discussed in this paperis how empirical theories
are related to the empirical phenomena for which they are supposed to account. I
argue that each applicationof a theory to an empirical event or an empiricalstate of
affairsrequires forming a theoretical(mathematical)modelforthe problemsought tobe
solved. Different problems may require formingdifferentmodelsrelevant to the same
phenomenon. The wayin whichtheoreticalmodelscan berelated to semantic models
is examined.
In the context of the above ideas Isurvey some of Suppes' views on theories,
set-theoreticalmodels, and theroleof both in empirical science.


A shortreminder of some tenets of logicalempiricism can beuseful for

the following considerations.
For the logicalempiricist to interpret a theory T in order to makeit
applicable to empirical phenomena was to define how touse the princi-
ples of the theory in order to derive from them the right conclusions on
the observable states of affairs. Since the descriptive terms ofa theory
may not refer to observable entities, the derivation may require appeal-
ing to rules of correspondence,alsocalled coordinating definitions,i.e.
extratheoretical principles whose task is to relate the statements of the
theory to the hypotheses which can be decided by observation.
Ifthenotionof observability is interpreted inanorthodox way, famil-
iar from the writings of representatives of logical empiricism, notably
Carnap, in the task of reducing theoretical knowledge to observational
knowledge one encounters serious difficulties which, as the empiricists
themselves werebound to admit, cannotbe overcome.
Rather than discuss the ideas of logical empiricism in their origi-
nal form, concerning therelationbetween theoretical andobservational
sentences, I shall focus on the relation between theoretical claims and
experimental hypotheses, the latter being any sentences which a com-
petent experimenter can decide by means of some reliable empirical
procedures. How sophisticated those procedures are and how much

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 125-149.
© 1994 Kluwer Academic Publishers. Printed in theNetherlands.

involvedare the theories theypresupposes, does not matter. Measuring

a distance with the aidof alaser beamcan be a great deal more reliable,
not to say more accurate, than measuring it with the help of a meter
stick. The final judgment on whether the experimenter is competent
and the procedure is reliable belongs to the scientific community.1
SinceI have decided to replaceobservability byexperimental decid-
ability, a rule of correspondence has to be defined to be a postulate
that relates the theoretical statements to the experimentally accessible
(rather than observational) states of affairs.
The switchfrom observability to experimental decidability has some
far reaching consequences which in no way could be accommodated
to the orthodox empiricists' doctrine. Modern empiricisms, notably
van Fraassen's constructive empiricism, are much more flexible in this
respect. Most importantly while the observationally accessible world
wasassumed to be fixed onceand forever (to be observable was defined
to be observable for any 'normal' observer, regardless of what could
be his or her knowledge, training or past experience) the experimental
accessibility of specific states of affairs or events depends in the most
intimate way onboth the competence of the experimenter and onher or
his technical possibilities.


Even though some of the analyses carried out by the logical empiricists
can be viewed as semantical, the predominant orientation of logical
empiricism was syntactical. The investigations were carried out in
terms of sentences and deductive relations among them. For a logical
empiricist, the selection of the syntactical option was a matter of both
doctrine andhis theoretical possibilities. Modern semantics begins with
Alfred Tarski'scelebrated treatise on the notion of truth.2 But although
the semantic approach gained momentum rather quickly, it became a
viable alternative to the syntactical one only after a series of further
investigations which eventually resulted in transforming general ideas
of logical semantics in a full-fledged mathematical discipline the
theory ofmodels. This remark applies both to metamathematics and
philosophy of science. Suppes' research program for philosophy of
science based on ideas drawn from the theory of models was the first
fully mature and viable alternative to the received view.3
Unlikehis predecessors who defined theories to be deductive systems
formalized within the idiom of first order logic, Suppes argues that an
empirical theory, in fact any theories whatsoever, should be viewed
as being determined by the class of its realizations rather than that of
its valid sentences. Thus,if for the logicalempiricists the right way to
define an empirical theory T was to define a set of axioms from whichall
the other sentences valid in T are logically derivable, Suppes suggests
that to define T is to define a set-theoretical predicate that denotes all the
set-theoretical structures of which T is true in the Tarski sense. Recall
that a realization (also referred to as a semantic model or justmodel) of
a theory is just a structure of which the theory is true.
The question of whether theories are classes of semantic models
(non-statement view) or sets of sentences attracted much attention and
hasbeenboth widely and vividly discussed.4 Foran outsider,theessence
of this discussion may not be clear. After all, according to an old but
still robust doctrine, theories are neither sets of sentencesnor classes of
set-theoretic structures. They are sets of propositions which one may
represent either by the sentences applied to communicate them, or by
classes of possible worlds (thus semantic models) of which they are
true. If one deals with formalized theories, i.e., ones for which both
the notion of truth and that of logical derivation are well defined, one
can equally well define a theory either to be the set of all sentences
derivable from a set of some initial principles (axioms), or to be the
class of all the realizations determined by exactly the same principles.
As Suppes himself states clearly (cf., Suppes, 1967, Ch. 2, p. 24) the
essence of his set-theoretical approach "is to add axioms of set theory
to the framework of elementary logic, and then to axiomatize scientific
theories within this set theoretical framework." This by no means
presupposes that the axiomatization should result in a non-statement
reconstruction of the theory. One who sticks to the traditional linguistic
option can use exactly the same axioms by which the set-theoretical
predicate has been defined to define the corresponding class of valid
sentences. What matters is the set-theoretical framework not the choice
of either 'statement' or 'non-statement' option.
All these points granted, Suppes' set-theoretical approach was a
radical turn in philosophy of science. To think ofa theory as anetwork
of sentenceslinked by therelation of deducibility or to thinkof it as a set
ofclaims that concern a specific class of structures is to chosebetween
two dramatically different perspectives. The semantic shift originated

bySuppes was, to appeal to a familiar Kuhnian idea,a shift of paradigm

with all the consequencesof such a step.
One might expect that the semantic turn originated by Suppes would
automatically focus our attention on the problem of the intended inter-
pretation of empirical theories. Actually, Suppes' approach does not
require any investigations into theissues characteristic of such an inter-
pretation. In order to define the class of set-theoretical structures of
which a theory is true - thus true relative these structures but not neces-
sarily true in the absolute sense of the word oneneed not know what
the theory is intended to be about. As was made explicitby Tarski, the
relative notionof truth does not coincide with the absolute one. Onehas
to know theintendedinterpretation inorder to define thelatterbut not the
former. Thus,regardless of how surprising it can be, Suppes' semantic
considerations are irrelevant to the problem of factual interpretation of
empirical theories.
One of the basic tenets of Suppes' analysis of empirical theories
was to the effect that both empirical theories and mathematical ones are
determined by their realizations and nothing else. According to Suppes
(1967, Ch. 2, p. 29)
..there is no theoretic
.mathematicsand way of drawing a sharp distinctionbetween a piece of pure
a piece of theoreticalscience. The set-theoreticaldefinitions of the
theory of mechanics, the theory of thermodynamics, and the theory of learning, to
give three rather disparate examples, are on all fours with the definitions of purely
mathematicaltheories of groups, rings, fields, etc. From a philosophical standpoint
thereis no sharp distinctionbetweenpure and applied mathematics, in spite of much
talk to the contrary.
This standpoint calls for some comments. To begin with, it may not
be as radical as it looks like at the first glance. After all, Suppes does
not maintain that thereare no differences between empirical disciplines
and mathematics. What he denies is the existence of any systematic
difference between the two which can be grasped in rigorous terms.
Even so, he is skeptical about the expediencyof any attempt to separate
empirical theories from purely mathematical ones and to view them as
belonging to two different areas of theoretical studies.
Suppes' position implies (and I do not seehow one could avoid this
conclusion) that philosophy of sciencereduces to selected problems of
metamathematics. Indeed, if we are not able to examinein a systematic,
rigorous manner the differences that separate empirical theories from
pure mathematics, then any theorizing about the former must collapse
to theorizing about their mathematical aspects, and thus, in fact, it must
collapse to metamathematics. The view that philosophy of science is
impossible as a separate discipline was explicitly held by Tarski (see
thepreface to hisIntroduction to Mathematical Logic).
The coincidence of Tarski's and Suppes' tenets is not accidental.
Even though it departs from many dogmas of logical empiricism,
Suppes' philosophy of science belongs to the logical trend originat-
ed by the logical empiricists. Now, for logical empiricists, regardless
of the quarter they belonged to - Vienna Circle or, like Tarski, Lwow-
Warsaw School, or any other the only acceptable way of theorizing
was that which could be expressed in a fully precise language, and
the only precise language was that of logic and mathematics. On this
point neither logicalempiricists nor their successorshave beenready to
For someone who restricted his interest to formal problems, the
idea of a factual interpretation of an empirical theory appears as a
certain vague heuristic notion which, escaping any formal treatment,
escapes any theoretical investigations. Certainly, he would argue, we
form empirical theories toaccountfor the empirical phenomena, but our
intentions cannotbe subjectedto any formal studies. What really counts
is the product of our intentions not the intentions themselves. And the
latter are theories. Now, in order to characterize a theory adequately,
one has to learn what are its realizations. Only after this is done may
one ask whether, besides abstract set-theoretical structures, the class
of the theory's realizations contains also some empirical systems and
thus whether one may use the theory to account for the states of affairs
characteristic of the system. Thus,instead of dealing with an enigmatic
idea of a factual interpretation of a theory, one is advised to look for the
empirical applications the theory admits.


In what follows the idea of a model of a theory which has some real
objects as its parts, andhence can be a domain of a factual application
of the theory, will be our special interest. The following passage from
Suppes (1960, p. 291-292) can serve as both a concise and instructive
introduction to the topic.
To define a model formally as a set-theoretical entity, which is of a certain kind of
ordered tuple consisting of a set of objects andrelationsand operations onthese objects
is not to rule out the kindof physical model whichis appealing to physicists, for the

physical model may be simply taken to define the set of objects in the set-theoretical
model. .We may axiomatizeclassical particle mechanics in terms of five primitive
notionsof a set P of particles, an interval T of realnumbers correspondingto elapsed
times, a position function s defined on theCartesianproduct of the set of particles and
the time interval, amass function m defined onthe set of particles, and a force function
f defined on theCartesian product ofthe set of particles, the time interval, and the set
of positiveintegers (the set of positive integers enters into the definition of the forces
function simply in order to provide a methodof naming forces). A possible realization
of the axiomsof classical particle mechanics, that is, of classicalparticle mechanics, is
then an orderedquintupleP= (P, T, S, m, f). Amodelofclassicalparticle mechanics
is such an orderedquintuple. For example,in the caseofthe solarsystem wesimply can
take the set of particles to bethe set of planetarybodies. Another slightly more abstract
possibility is to take the set of particles to be the set of centers of mass of planetary
bodies. This generally exemplifies the situation. The set-theoreticalmodelof a theory
will have amongits parts a basicset which willconsists of theobjects originallythought
to constitute the physical model.

Needless to say, as long as an empirical theory is not provided with

any factualinterpretation, itremains merely acertain formal system. But
in spite of Suppes' skepticism on thematter, which we mentionedbriefly
in the previous section,onemay wonder whether the differentia specifica
allowing us to tell an empirical theory from a piece of pure mathematics
does not consist in the fact that the former has some intended empirical
applications. This is exactly the line of thought which was pursued by
Adams (1959).
He suggested that while a purely mathematical theory can be ade-
quately representedby the class of all its realizations, a theory meant
to have a factual contentmust be representedby two elements. One of
them is again the class of all the theory's realizations. It forms, to use
the expression coined by Sneed (1971) the formal core of the theory.
The other, whichis characteristic of empirical theories only, is the class
of all the intended applications of the theory which is meant to be the
class of all empirical structures (physical systems) of which the theory
is expectedto be true. In general, the theorymay fail these expectations;
an intended application may not be a realization of the theory.
Adams' ideaof intended applications was modified by Sneed. While
for Adams,an intended application of a theory T is apossible realization
of this theory, i.e., it is a structure of the same set-theoretical type as
any of the realizations of T, Sneed requires that no component of an
intended application can be 'theoretic relative to T\ that is, one whose
use presupposes appealing to some of the principles of the theory. For
example, according to Sneed, mass is a theoretical concept relative to
Newtonian mechanics, for in order to measure mass, one must resort to
Newton's SecondLaw.
Clearly, neither Adams nor Sneed have denied that the users of
an empirical theory can revise their views on which of the empirical
systems are legitimate intended applications of the theory. One can try
to improve an empirical theory both by changing its principles and by
defining anew its scope of applicability.
Another idea relevant to the present discussion is Sneed's idea of
an empirical claim. Suppose the formal core M(T) of an empirical
theory T is defined by a set of axioms. These conditions are some
formulas (e.g., mathematical equations) which, as long as we abstract
from the intentions of its users, are not statements; they do not refer to
any specific state of affairs. The situation changes if one claims that
the conditions which determine M(T) are true of a particular empirical
system cr, in other words one maintains that a is one of the structures
in M(T). An empirical claim is just a statement to this effect. Note
that an empirical claim need not concern an intended application of an
empirical theory; it may refer to any physical system whatsoever. But,
clearly, defining a theory to be the couple (M(T),J(T)) is tantamount
to accepting all the empirical claims of the form 'cr is in M(T)' where
a is in /(T).
Onecan find theAdams-Sneedamendment to Suppes' conception of
the formalization of empirical theories both self-explanatory and desir-
able. And, indeed, they were wholeheartedly accepted by the struc-
turalists - philosophers who, grouped around Sneed and Stegmuller,
undertook the task of scrupulous investigations into a 'formal archi-
tectonic for science'.5 Still, lam going to argue that the amendment,
as well as the structuralist research program based on them, diverge
rather substantially from themain ideas of the Suppes approach to the
problems of philosophy of science.


No scientific theory is just true about the empirical systems to which

it applies. The reason why it must be so is fairly obvious. Empirical
systems are, to appeal to Peircean metaphor, 'nebular. They are nev-
er fully separable from their environments; moreover they consist of
objects whose properties are not uniquely determined and may not be
exactly such as is required by the theory. Thereby, if a theory applies

to any such system it applies to it in an 'approximate' way. Apparently

then, what the scientists should attempt to reach is approximate truthof
their claims rather than truth in any plain sense of the word. The latter
can merely be viewed as an idealized counterpart of the former.
An immediate conclusion from the above observation seems to be
that inorder to make the doctrine ofintended applications more realistic,
the idea of an empirical claim to which it appeals should be defined in
terms of approximate truthrather than plain truth. Exactly this step was
taken by the structuralists,who in much of their investigations devoted
themselves to the issue of approximate correspondence between the
realizations and the intended applications of theory.
The idea of approximate truth may seem not to be very involved.
Actually it takes some effort to define approximate truth in a sufficiently
rigorous way.6 Still, in the present discussion, we can safely restrict
ourselves to examining acommon sense version of this concept. From
the intuitive point of view a statement is approximately true, if the
difference between what it asserts and the actual state of affairs is
'small. The question which immediately comes up is 'small under
whichrespect?';needless to say, evena small errormay yielddrastically
false conclusions and thus fatal consequences.
On the other hand, it is a fact of life that we systematically base our
activity onsome 'nearly true' hypotheses. We can dothis because, inany
ordinary circumstances, we are able to tell the acceptable conclusions
which a nearly true sentence justifies from unacceptable ones. What
really matters is not the 'distance' of a hypothesis from truth but the fact
that it allows for true conclusions. Even a patently false sentence can
be usefulin this way.
But if so, wehad better giveup the misleading idea of approximate
truth. The much debated fact that scientific theories dramatically fail to
be totally true of the states ofaffairs to which they refer, does not imply
that they are totally false. In other words, theories are partially true.
More accurately, they are true in certain selected respects; if applied in
the right way, they allow us to arrive at true conclusions. This is not the
deceptive notion of approximate truth, but rather that of 'partial truth',
orstill better 'relative truth', which weneedin order to account for how
empirical theories refer to empirical phenomena. Inthe sections which
follow I examine the above idea in some detail.
An application of an empirical theory T, by itself or combined with
some auxiliary hypotheses, to an empirical phenomenon IIin order to
solvea specific problem Q concerningn may require the formation of
a theoretical model (one may prefer to say mathematical model) of the
phenomenon, thus it may require defining an abstract system M. meant
to satisfy the following two conditions:
(ml) M is a realization of T;
(m2) M is a faithful representationof n under all the respects that are
relevant to Q. Thus if A is the solution to the question Q, then A
is factually true, i.e., true of n if and only ifA is true of M.
Note the phrase 'meant to satisfy' whichappears in the above definition.
It may be desirable to state the definition of a theoretical model in
purely formal terms, thus avoiding any pragmatical references. But,
in the context of the present discussion, the pragmatical option has
an immediate advantage we need not discuss separately what the
theoreticalmodels are and what they are for.
If, instead of the question Q, one is interested in answering another
question Q' concerning the same phenomenon, it may turn out that
the model M does not meet the faithfulness requirement with respect
to the new application; it does not yield an adequate solution to Q',
and therefore another model M! of n should be constructed. Thus, a
theoretical model of n is not somuch amodel of n, but amodel of an
aspectU/Q, 11/ Q, of thephenomenon or justa model for a problem
Q,Q',. Also note that there may not be any obvious way to associate
n with a specific singular empirical system.
An example can help clarify the above points. Suppose someone
wants to determine the position of Mercury at a specific time t, say at
some fairly remotefuture time. Denote the problem thus defined by Q t
Actually, besides stating theproblem Qt,one has to state the postulated
parameters of accuracy and reliability the solution totheproblem should
satisfy. But, for the time being, thisneed not be our worry; I shallreturn
to this point later.
One may rush to point out that the problem Q t concerns the solar
system of which the planet is a part. But,in fact, what we know for sure
is that Qt concerns Mercury's movement. Consequently, rather than
presuppose in advance that Qt concerns a specific system of physical
bodies, one must take into account all the factors whatsoever of which

one has some rational reason to believe may affect the movementof the
planet. The proviso 'of which one has some rational reason to believe'
cannot be dropped out. We neither see the point of examining, nor
we are able to examine, all the adhoc hypotheses one can put forward
seeking to account for the movementof the planet.
Suppose our knowledge of the matters relevant to Qt is roughly the
same as that of the nineteenth-century philosophers of nature. Thus,
we have some fairly sophisticated, even though inadequate,idea of the
nature of space andtime. We are aware of various physical propertiesof
celestialbodiesand wehave someideaof their spatialconfiguration. We
have gathered a large quantity of evidence concerning their movements,
especially those of some planets in our solar system. And on top of that
wehave some powerful theories: Euclidean Geometry, EG, meantto be
the theory of physical space, Newton's Law of Universal Gravitation,
LG, and Newton's Particle Mechanics, NM. This fairly large body of
beliefs, combined with some ontological ideas on the structure of the
universe,form the world view within whichour search for the solution
to Qt must be carried out. Still, the discussion which follows will be
ahistoricalin various respects. Most importantly, I shall presupposethat
the construction of the theoretical model for Qt involves a fairly large
amount of statistical considerations based on some notions unfamiliar
to the nineteenth-century scientists.
To begin, we note that weare bound to accept the default hypothesis
to the effect that the only source of Mercury's movements is the set of
gravitational forces. This is not to say that we are not able to figure
out that besides gravitational forces there are other factors which are
relevant to the planet's motion, but we have no evidence which could
allow us to conjecture their nature and then state the conjecture in the
form of a reasonable hypothesis.
Bya similar default argument wearrive at the conclusion that only the
bodies in the solar system can influence the movement of Mercury in a
detectable way;the gravitational effects of allthe remaining ones are too
small for the solution of Qt to dependon them. Moreover, the effects of
gravitational interaction between Mercury and various elements of the
solar system, say planetoids, satellites of Mars, and even most remote
planets, can also be negligible small. Thus, we eventually arrive at the
conclusion that in order to solve Qt, it suffices to take into account the
gravitational interactions among a few bodies of the solar system which
are the most massive and nearest to Mercury. Which of such bodies
have actually to be taken into account depends on how accurate and
how reliable the solution is postulated to be.
The two requirements cannotbe separated from one another. The
more accurate the solution is postulated to be, the narrower is thelimit
of error we admit for the solution and the lessreliable is the solution we
are able to find out, i.e., thelower the probability that the actualposition
of Mercury is within the admitted limit of error.7 Clearly, one is able
neither to experimentally find out the recent position of Mercury, nor to
theoretically predict any future one with unlimited accuracy; the idea of
unlimited accuracy is justempirically senseless.
It has to be noticedthat the two requirements are formal in the sense
that an evaluation of whether the solution satisfies them or not is a
matter of statistical analysis of all the steps of the procedureby means
of which the solution has beenachieved. From the actual point of view
the solution iseither factually adequate(thus true) or not.
Suppose, after defining the postulated parameters of accuracy and
reliability for the solution to Qt, and after examining how strongly the
movement of Mercury is affected by the bodies of the solar system,
we have eventually arrived at the conclusion that the model for Qt
can be fairly simple: it can be defined to involve only the Sun a and
Mercury p, both being treated as 'particles' or, as we prefer to say
today, 'mass points'. Since the gravitational interaction between the
two bodies is assumed to fully determine the movementof Mercury, the
motion equations should be derivable from the union N = NM + LG
of the principles of Newtonian particle mechanics, NM, and the law of
universal gravitation, LG.
Let NM denote the equations of Mercury's motion derived from N.
Observe that the formulas in N^ differ from those in N in that they refer
to two specific 'mass points', the Sun and Mercury, while the principles
of N are statedin the form ofa general schemaof equationsapplicable to
any set of mass points whatsoever.In order for the equations NM to have
a unique solution, the values of some of the parameters the equations
involve should be either established experimentally or deduced from
the available experimental data with the help of the hypotheses we
consider to be well confirmed, the laws of N included. Let E be the
set ofconditions one has to add to N^ in order to assure the uniqueness
of the solution; let us call them factual conditions. As will be clear
somewhat later, only a part of E states the initial conditions, those
whichcharacterize the state of the two bodies at a given initial time to.

Tomake a long story short, wehave to do two things. First, we must

empirically determine the motion of Mercury for some initial period
of time At, and then deduce the conditions E from the assumption
to the effect that this determined fragment of the planet's movement
satisfies the principles N. Itshouldbe noticed that to some extent weare
free in choosing the factors which are to be determined by the factual
conditions E; all we have to be sure is that N^ + E admits only one
solution to Q^.
Suppose we select E to define six parameters: the mass, the 'initial
position' and the 'initial velocity' of both the Sun and Mercury. Infull
accordance with thenineteenth century world view, two of them,namely
the initial position andthe initial velocity of the Sun,canbe assumed to
be miliary. Actually, in order to make this assumption consistent with
N, we must view the Sun to be 'almost' immobile,i.e., we must assume
that the movementof the Sun, whatever it may be, does not result in any
effect relevant to the solution to Qt Furthermore, note that instead of
calculating both the mass of the Sun and that of Mercury, it will suffice
that we can calculate one of them, say that of the Sun, relative to the
other. In this way we are left with three constants to find out via an
analysis of theinitial motion: the position, xo ± Ax, and the velocity,
vo ± Ay, of Mercury at the 'initial' time to and the relative mass of the
Sun, ma ± Am. The limits of error Ax, Ay and Am must correspond
to some presupposed reliability of the relevant hypotheses. Let us dwell
on this issue for a while.
Note that the data whichcharacterize the motion of the planet within
the examined initial period At cannot be just a direct inductive gen-
eralization of astronomical observations. The reasonis fairly obvious.
The frame of reference relative to which we observe the movement of
celestial bodies is the Earth. Consequently,inorder to know how any of
these bodies, Mercury in particular, movesrelative to the Sun, we must
calculate their motion with the help of therelevant transformations from
the data concerning the apparent movement of the Sun. Certainly, I am
not going to survey the procedure; the technical aspects of itare oflittle
interest for us. Rather, I shall briefly discuss a certain methodological
Denoteby Qa* the problem of determining Mercury's motion within
the period At. There seems to be somethingparadoxicalin that in order
to solve the problem Qt one has to solve a problem which actually is
a variant of it. In fact, the procedure is by no means paradoxical. Itis
fairly obvious that in order to apply the laws that govern the dynamics
of an empirical phenomenon to a specific instance IIand a specific
problem Q concerning this instance, one must learn the conditions
characteristic of 11, i.e. ones which may serve to distinguish IIfrom the
other instances of the phenomenon. In this way one arrives at another
question Q' concerningn. Since the factual conditions which are to be
determined by solving Q' are relevant to Q, the two questions Q and
Q' are interdependent, and often fall under the same general category.
This is exactly the case of Q t and Q&t-
It is fairly obvious that unless we are going to find ourselves in the
situation of regressumad infinitum,the methodology of solving the Q At
cannot be the same as that for Qt When dealing with the former, we
cannot apply the laws of N in any way which requires appealing to
factual conditions E.In fact, this is the only restriction to be observed;
note that it does not imply that the laws of N cannot be useful in
dealing with Q& t in any way whatsoever. Actually, since we assume
that Mercury's motion satisfies N, we require the solution to Q& t to
be consistent with NM This amounts to imposing a rather essential
constraint of the conditions of adequacy of the solution to Q&t, thus
narrowingthe spectrum of all the acceptable solutions at whichone can
arrive by the statistical analysis of the empirical data. In what follows
the solution will be denoted by E& t \ it goes without saying that the
equations of Mercury's motion for the period At of which it consists
canbe neither fully accurate nor fully reliable.
The reader familiar with Suppes' (1962, 1974) notion of model of
the data, as well as his analyses concerning this notion, has certainly
noticed that the procedures discussedabove for solving Q& t are typical
for forming a model of the data. Themodel of the data thus produced is
the two-body system whoseelements are theSunand Mercury,assumed
to satisfy the conditions Q&t- One can use the following rather non-
standard but self-explanatory notation

{{a,p.] : EAt )
to stand for this defined system. The fact that formation ofa theoretical
model presupposes formation of a model of the data as well as the
fact that formation of a model of the data can be controlled by the
requirement of consistencyof the model with the corresponding theory
areof keysignificance for properunderstanding of theinterplay between
the data and the theories, and thus for proper accounting for both the

corrigibility of the data and falsifiability of the theoretical claims. I

not going to examine these issues, however; they go beyond the scope
of this paper.
Suppose the factual data E are given. Then the model for Qt is the
structure Sdetermined by the following two conditions:
(i) all the elements of S are a and fi;
(ii) S is a realization of theconditions N^, E.
In what follows,rather than explicitly stating conditions of the form (i)
and (ii), Ishalluse the already familiar notation:
S = {{cr,»}:N +E).
Needless to say, ({cr, //,} : N +E) isbut one of numerous theoretical
models of the movement of Mercury one can form, and which have
been formed. Recall that for a long period of time, the planet was of
special interestbecause of the peculiarities of its perihelion movement.
Seeking to explain them, scientists not merely analyzed models which,
besides the Sun, involved some of the remaining large bodies of the
solar system, they alsoconstructed models presupposing the existence
of some hypothetical factors affecting the movement of the planet.
Thus, for example, Leverrier postulated a group of planets revolving
inside the orbit of Mercury. Another hypothesis related the observed
irregularities of the planet's movement to the alleged oblateness of the
Sun.8 Incidentally, an oblateness of the Sun's mass distribution rules
out the possibility of representing the Sun by a single mass point.
I hope I have succeeded in presenting theidea ofa theoretical model
in a sufficiently clear way. Still some further comments on it can be
useful. To begin with let me point out that theoretical models are not
empirical systems, unless we consider them to be empirical systems
of some hypothetical or possible worlds. The reason is quite obvious.
Even though the objects a model involves are supposed to be real (such
as e.g., the Sun and Mercury in the model discussed above), still these
objects are postulated to satisfy certaintheoretical assumptions (such as,
e.g., the laws of NM) which they may not actually satisfy. A theoretical
model is just a certain theoretical construct, an abstract entity, even if
some of its parts appear to be parts of physical reality.
Since the aboveconsiderations were preceded by a critical appraisal
of the Adams-Sneed doctrine of intended applications, I should men-
tion that theoretical models are not the right candidates for the role of
intended applications in the Adams-Sneed sense. It is not so much
because they are not empirical systems - they can certainly be viewed
as some idealized representations of the latter - but because, by the
very definition, they are realizations of the theories intended to refer
to those systems. On the other hand, an intended application in the
Adams-Sneed sense is postulated to be a structure which may turnout
to be a realization of the theory in question. This possibility is the gist
of the doctrine of intended applications, for it is a formal counterpart
of the fact that an empirical theory may happen to fail to account ade-
quately for empirical states of affairs within the scope of its intended


While the notion of a theoretical model cannot be accommodated to

ideas of structuralist philosophy, it certainly fits directly and in a most
natural way the Suppes research program. The passage from Suppes
(1960) I quoted at the beginning ofSection 3 can serve as a straightfor-
ward piece ofevidence supporting this claim. Asone may easily notice,
thephysical models discussedin the quotation are just theoretical mod-
els in the sense of this paper.
In this section I digress from the main topic of this paper. Still,
I believe, Ishould not avoid discussing the problem Iam going to
address, for discussing it may contribute to a better exposition of some
methodologicalpoints relevant to Suppes'conception ofmodel. I wish
to comment upon Suppes' claim that of the two notions of a model,
namely the set-theoretical and the physical one (or ifI am to stick to my
terminology - the semantical and the theoretical one) the former can
be viewed as more fundamental than thelatter. One of the passages in
which thisposition is expressedis the following (Suppes, 1960,p. 291).
In the preceding paragraphs we have used the phrases 'set-theoretical model' and
'physical model. There would seem to be no use in arguing about which use of
the word 'model' is primary and more appropriatein the empirical sciences. My
own contention is that the set-theoreticalusageis the more fundamental. The highly
physically minded or empirically mindedscientists who may disagree with this thesis
and who believe that the notion of a physicalmodel is the more important thing in a
givenbranchof empirical science may still agree with my systematic remarks.
Rather than challenge the opinion that 'the conceptof modelused by
mathematical logicians is the basic and fundamental concept ofamodel

needed for exact statementin any branch of empirical science' (Suppes,

1960,p. 294),I am going to argue thatthe question of which of the two
concepts of model is more fundamental can be examined from quite a
different angle than that taken by Suppes.
There is no question that, on many occasions, the search for a the-
oretical model for a question Q is the search for a semantical model
of a specific theory T such that this particular model will be the right
one to examine the question Q. This is the situation which obtains
whenever one believes that T is the theory which allows for solving Q.
If the formation ofa theoretical model is from the very beginning meant
to consist in selecting the right semantical model then, of course, the
notion of a semanticalmodel is prior to that of a theoretical one. The
discussion in the previous section concerned the abovementioned case;
Newtonian Particle Mechanics was viewed as the right theory to solve
the question Qt concerningMercury's movement.
However, there are numerous situations in which the search for a
model for Q starts whenone does not consider any theory to be directly
applicable to Q. Bohr's model of the atom, whichin fact was a model
for some specific problems on the dynamics of the particles of which
the atom was believed to be composed, was certainly not meant to be
a semantic model of any official theory; rather it was meant to be a
certain tentative theory of the phenomenon. A long list of examples
can be produced to support the above observation. And, of course,
there is a rather substantial difference between the situation when a
model is, as the scientists would say, 'derived from' a theory and the
situation in which a model is a step towards formation of a theory.
In the latter case there is no theory (among the theories known at a
giventime) from which the theoretical model in question is derived, so
the model cannot be intended to be a semantic model of any official
theory. The construction ofa theoreticalmodel precedes any semantic
considerations andin this senseit is prior to the latter.
The above discussion can be summarized as follows. The idea of a
modeliscentral for two, ina sense,opposite activities: applying theories
and forming them. The interplay between the two is what largely
determines the dynamic of science, thus its successive transformations
in time. Now, while the concept of semantic model is central for the
former, the theoretical model iscentral for the latter.
Suppes seems to believe thatthe triadic relation 'theoretical model -
theory - semantic model' canbereduced to the dyadic relation 'theory -
semantic model. Thisis whatI doubt, andI doubtit because the 'archi-
tectonic of science' to use the expression coined by the structuralists,by
no means can be adequately accounted for in terms of formalized theo-
ries and formal relations between their semantic models. The structure
of science isin acontinuous process of formations,and thus any attempt
toaccountadequately for it must bebased onan analysis of the function
which specific units are supposed to play, rather thanon any taxonomy
presupposed in advance. If we take this perspective, the question of
whether a theoretical model has been intended to be a semantic model
of a theory, or has been intended to be an embryo of a theory to be
formed does really matter.
The present discussion provides a good occasion to note that the
model for a singular concrete problem concerning singular concrete
phenomenon is but one of many kinds of models one encounters in sci-
ence. Questions one may seek to solve byconstructing amodel for them
can be both singular and general,both factual and purely hypothetical.
Another thing to be noticedis the distinction between theoreticalmodels
and theories may not be as sharp as it is suggestedby some tentative
definitions offered in the course of the above considerations. Let us
dwell on the last point.
Consider themodel ({<7, fi} : N^), one which results from themodel
({cr,n} : NM +E) for the question Qt byremoving the initial conditions
E. Themodel thus formedisamodel for the two-bodyproblem. Clearly,
the conditions imposed on this model define it merely partially; all the
questions concerning it which cannot be solved without resorting to
the factual conditions E are left open. The question Q t we have been
discussing is but one of this kind.
A structure defined by the conditions which do not determine the
behavior of its elements is a partially defined structure, and hence it
actually is the class of all those structures which satisfy the conditions
in question. As such, it determines a theory just in the set-theoretical
sense proposed by Suppes. What is more, the theoretical model can
usually be identified with the conditions by means of whichit has been
defined. Consequently it can be viewedas a theory inthe linguistic sense
of the word. Itis worth noticing that this possibility is fully exploited in
scientific discourse. In all scientific disciplines in which the concept of
theoretical modelis in use, it is customary to shift freely from the set-
theoreticaluse of the term 'model' to its linguistic counterpart and vice
versa. Thus, dependingon circumstances, a model is viewed as either

acertain mathematical entity (e.g., a curve in Euclidean space meant to

be a model of motion of a physical body) or the class of corresponding
mathematical equations. This explains why the term 'model' is often
used interchangeably with 'theory. Thus, for example, the physicist
willconsider of minor significance whether one prefers to speak about,
say Boyle'stheory ofideal gas or ratherabout Boyle's model of an ideal
One may consider this state of affairs deplorable, for certainly from
the logicalpoint of view a set of sentencesanda set-theoretical structure
that the sentences are supposed to characterize are of entirely different
logical type. But the difference between them matters only in some
rather special contexts which are of interests for the logicians, but they
are oflittle interest for working scientists.


Theproblem ofhow empirical theoriesare related to empirical phenom-

ena has two aspects. Only one of themhasbeen discussedin this paper.
I want to mention the other in order to localize the ideas of this paper in
a somewhat more general context.
An integral part of any empirical theory is its factual interpretation
determined by the relevant world view. One cannot understand an
empiricaltheory and thus onecan know neither what the theory is about
nor how to form a theoretical model for specific problems relevant to
the theory unless one has some idea what is the part (or aspect) of the
world to which the theory refers, how this part is related to the others,
which is the ontology of all these parts, and how both the claims of
the theory and the empirical data on which it is based arerelated to the
entities whoseexistence the ontologypresupposes.
Itshould beclear that the factual interpretation as understood above
is global, in the sense that it relates the theory to the universe as a
whole. As such,it differs radically from any local interpretation, which
is defined to be a specific theoretical model formed to be amodel of the
The fact that theories are not autonomousunits of science,and thus
the fact that both to understand and apply them one must conceive
of them as parts of some larger setting, has been fully appreciated by
contemporaryphilosophy of science. Thisis not to say that philosophers
of science are of the same opinion on how this larger setting is to be
defined. For Kuhn (1970) this is a paradigm, for Lakatos (1978) this
is a research program, Laudan (1977) calls it a research tradition, to
mention only a few.
This paper is not the right occasion to discuss the idea of the global
interpretation of empirical theories. I havementioned this topic mainly
because ofits obvious andimmediate relevance to themain issueof this
paper. Still, part of my concern was also Suppes' claim to theeffect that
there is no systematic difference between theoretical science and pure
mathematics. Such a difference, I believe, cannotbe grasped in terms
of the local interpretations of the theories. If it can be grasped at all, it
canbe graspedby resorting to theidea of factual interpretation.
Needless to say, the factual interpretation of a theory determines
its factual applications. On the other hand, in learning the factual
applications of a theory (or rather some paradigmatical examples of
them) one acquires some knowledge of its factual interpretation. Thus
the two characteristics of an empirical theory are related to one another
in a direct and strong way. Still, in this paper I have dealt with factual
applications only. In the concluding remarks that follow, I comment
on those few notions of factual application whichhavebeen my special
For logical empiricists a factual application of a theory consistedin
deriving some observational hypotheses from the principles of a theory,
the relevant rules of correspondence, and the available observation-
al evidence. Nobody has ever attempted to reconstruct in a full and
systematic manner the rules of correspondence which allegedly were
associated with empirical theories. Or rather, every attempt to recon-
struct these rules resulted in a successive modification of the idea of
empirical meaningfulness, and thus theidea of what these rules should
look like. The swan song of logical empiricism was Carnap's The
Methodological Character of Theoretical Concepts.
Theenormous work done by the structuralists resulted in some most
valuable case studies of the formal structure of scientific theories and
formal aspects of interrelations between them. But, as I have been
arguing, the doctrine of intended applicationsunderlying the structural-
ist approach cannot work. It provides not merely an oversimplified
account of the relation between theories and empirical phenomenabut
it is based on some ideas which are substantially wrong. At this junc-
ture, I have to mention that exactly the same criticism shouldbe directed
against my Topics in the Formal Methodology ofEmpirical Sciences, the

monograph in which I examined some ideas based just on the doctrine

of intended applications. Investigations into the theory of approximate
truth were a part of this monograph. Thus the view I present andadvo-
cate in this paper is a radical departure from some ideas to which I
subscribed earlier.9
A concise, though somewhat metaphorical, exposition of Suppes'
position on how theories are related to empirical states of affairs is
contained in the following quotation (Suppes,1967, Ch. 1,p. 12):
The concrete experiencethat scientistslabel an experiment cannot itselfbeconnected
to a theory in any complete sense. The experiencemust be put through a conceptual
grinder thatinmany cases is excessivelycoarse. Once the experienceis passed through
thegrinder, whatemergesare the experimental datain canonicalform. These canonical
dataconstitute a modelof experiment,anddirect coordinating definitions are provided
with thismodelof the experimentrather for a modelof a theory.Itis also characteristic
that themodelof the experimentis of relatively different logical type fromthat ofthe
model of the theory. It is common for the models of a theory to contain continuous
functions orinfinitesequences,butfor themodelofthe experiment to be highly discrete
and finitistic in character.
The assessment of the relation between the model of the experiment and some
designated modelof the theory is the characteristic fundamental problem of modern
Ibelieve I have succeeded in making itclear that theidea of a theo-
retical model around which this essay has been organizedcorresponds
directly to 'somedesignatedmodel ofthe theory' from thelast paragraph
of the above citation.
The last remark is personal. There hardly is any philosopher of
science of whom I could say that he has inspired my own work so
deeply as it has been done by Pat. In spite of all the reservations about
his views I kept expressing, this paper, Ido hope, is in full agreement
with the main ideas of Pat's research program. It certainly was meant
to be acontribution to it.


For a discussion of theseissues see e.g., Woodward(1989).
2 version,
The original Tarski (1933) was publishedin Polish. For an English version
of it, see Tarski (1956).
Somealternativeideas wereput forwardbyEverthBeth (1949) (see also vanFraassen,
1970) and RomanSuszko (1957, 1968).
For an account of this discussion thereaderis referred to Pearce (1987).
allude here to the titleof the most recent monograph (Balzer,Moulines, and Sneed,
1987) on the structuralist approachin philosophy of science.
6 Cf. Wojcicki (1979).
It goes without saying that this paper is not the right place to discuss fairly sophisti-
catedtechnical counterparts ofthe ideaof accuracy and thatof reliability.
8 For
a discussion of methodological aspects of various hypotheses concerning Mer-
cury's movement see Griinbaum (1971).
The first publicationin whichIhaverevisedmy earlierviews was Wojcicki (1990).


Adams,E. W.: 1959, 'TheFoundations of RigidBody Mechanics and the Derivation

of Its Laws fromThose of ParticleMechanics', in: L. Henkin, P. Suppes, and A.
Tarski (Eds.), The AxiomaticMethod, Amsterdam:NorthHolland.
Balzer, W., Moulines, C. W., and Sneed, J.D.: 1987, An Architectonicfor Science; The
StructuralistProgram, Dordrecht: D. Reidel.
Beth, E.: 1949, 'Towards an Up-to-Date Philosophy of Natural Sciences',Methodos,
1, 178-185.
Carnap, R.: 1956, 'The MethodologicalCharacter of Theoretical Concepts', in: H.
Feigl andM. Scriven (Eds.), Minnesota Studies inPhilosophyof Science, Vol. I,
Minneapolis: University ofMinnesota Press.
Griinbaum, A.: 1971,'CanWe Ascertainthe Falsity of a ScientificHypothesis?',in: M.
Mandelbaum(Ed.), Observationand Theory inScience, Baltimore:Johns Hopkins
Kuhn, T. S.: 1970, The Structure ofScientific Revolutions (second edition), Chicago:
University of Chicago Press.
Lakatos, I.: 1978, The Methodologyof Scientific Research Programmes, Cambridge:
Cambridge University Press.
Laudan,L.: 1977, Progress andItsProblems, London: Routledge and Kegan Paul.
Pearce,D.: 1987, Roads to Commensurability, Dordrecht: D.Reidel.
Sneed, J. D.: 1971, The Logical Structure ofMathematicalPhysics, Dordrecht: D.
Suppes,P.: 1960, 'A Comparison of the Meaning and Uses ofModels in Metamathe-
maticsandthe Empirical Sciences', Synthese, 2/3, 287-301.
Suppes, P.: 1962, 'ModelsofData',in:E. Nagel, P. Suppes, andA. Tarski (Eds.), Logic,
Methodology andPhilosophy of Science: Proceedings of the 1960International
Congress, Stanford, CA: StanfordUniversity Press, pp. 253-261.
Suppes, P.: 1967, Set-TheoreticalStructures in Science, Stanford, CA: Stanford Uni-
versity Press.
Suppes, P.: 1974, 'The Structure of Theories and the Analysis of Data', in: F. Suppe
(Ed.), The Structure ofScientific Theories,Urbana, IL: University ofIllinoisPress.
Suszko, R.: 1957, Logikaformalna a niektore zagadnienia teoriipoznania ('Formal
Logic and SomeIssues of Epistemology', inPolish), MyslFilozoficzna,2 and3.
Suszko, R.: 1968, 'Formal Logic and the Evolution of Knowledge',in: L. Lakatos
(Ed.), Problems in the Philosophy ofScience, Amsterdam:NorthHolland.

Tarski, A.: 1933, Poj§cie prawdy wjezykachnauk dedukcyjnych ('The Notionof Truth
in the Languages of Deductive Sciences', in Polish), Warszawa: Towarzystwo
Naukowe Warszawskie.
Tarski, A.: 1956, Logic, Semantics, Metamathematics; Papers from 1923 to 1938,
Oxford: ClarendonPress.
VanFraassen, B.C: 1970, 'On theExtensionofBeth'sSemanticsofPhysicalTheories',
Philosophy of Science,37, 325-339.
Van Fraassen,B.C: 1980, The Scientific Image, Oxford: ClarendonPress.
Wojcicki, R.: 1979, Topics in the FormalMethodologyofEmpirical Sciences, Dor-
drecht: D. Reidel and Wroclaw: Ossolineum.
Wojcicki, R.: 1990, Teorie w Nauce ('Theories inScience', inPolish), InstytutFilozofii
iSociologii,PAN, Warszawa.
Woodward, J.: 1989, 'Data andPhenomena', Synthese, 79, 393^172.


It is not surprising that I like Wojcicki 's sympathetic treatment of my

views on theories, set-theoretical models and experiments. The only
minor point on which I directly disagree is his statement in Section 2
that philosophy of science, on my view,is a special branchof metamath-
ematics. It is exactly the point of using set-theoretical methods that we
throw things into a standardmathematical contextrather than the meta-
mathematical context familiar in much of logic. My point was that we
clarify scientific theories the way we clarify mathematical theories by
usingappropriate set-theoreticalmethods whichare not metamathemat-
ical in character. This is not however a major point and I turn to some
comments that are not really in disagreement with Wojcicki's but either
supplement or modify in certain ways the views he sets forth.

Theoretical Versus Semantical Models. Although in Section 6 of

his paperWojcicki's remarks about the contrastbetween theoretical and
semanticalmodelsof theories are meantto propose certainmodifications
of my views,I find myselfmostly in agreementwith what hehas to say.
The important point is that he is not introducing theoretical models
as a separate class of formal models as opposed to semantic models.
The emphasis is rather on at what stage one is creating a model as a
method for finding anew theory or as a test of agiven theory, theformer
constituting a theoretical model and the latter a semantical model.
I certainly would agree that models are usedin many different ways.
Moreover,I also agree, and I have indicated this in various prior publi-
cations, that scientistsuse the term 'model' in a much looser way than
do logicians, but without serious harm.
Something that I have omitted in my own discussions of these mat-
ters and that is not referred to by Wojcicki explicitly is the important
role played by the derivation of differential equations in allof science,
but especially in physics. Deriving a differential equationis not at all
a matter of setting boundary valueconditions as, for example, in his or
my earlier discussions of the two-body problem. Here it is a question
of understanding how to conceptualize the forces at work that create
change. A differential equation can cover, as we know from modern
studies,a great variety of phenomena, but the given differential equation
isnot ordinarily thoughtof as part offundamental theorybut one derived
for a class of phenomena that broadly fall under some general theory.
There is from a logical standpoint, I suppose, a kind of hierarchy of
theories involved. The important point is that the differential equation
in complex situations will be such that its class of models, i.e., the class
of objects characterizing solutions of the differential equation, will not
at all be understood. We can have a very poor feeling for the class of
solutions,especially as theclass of boundary conditions becomes com-
plicated, but we can have a very detailed scientific knowledge about
many features of the differential equation. We can,of course, formalize
the differential equation as a set-theoretical object, but this is not very
interesting. The real point is that such differential equations represent
an important way of characterizing classes of models and therefore, if
you will,subtheories,but they have notbeen as much recognizedas they
shouldhavebeen by meorothers workingin the set-theoretical tradition.

PureMathematics Versus TheoreticalScience. Wojcicki quotes pas-

sages in which I indicate that from a set-theoretical standpoint there is
not a significant difference between pure mathematics and theoretical
science. I still stand by those passages but I havealso come to empha-
size the many ways in which science and mathematics are different. As
theoretical physicists whoare good friends of mine like to say "If there
are not real conceptual differences between mathematics and physics
then we theoretical physicists are just to be thought ofas primitive math-
ematicians." Those of you who know the elanof theoretical physicists
know how unlikely they are to accept this idea.
Thisis not the place to expound in detail the many ways thatI think
mathematics and theoretical physics differ. I do just wantto mention one

or two salientpoints. Successful physicists need to havean enormously

good intuitive feeling for physical phenomena, to know just which
properties to reach for in deriving new results. Moreover, almost every
problem in physics is now too intractable to be solved by really simple
methods. The wise choice of just what to take account of and what
to try to compute in a serious way is a large part of being successful.
Unlike pure mathematics,a very large percentageof physicalproblems
now require extensive computation. Most mathematicians still avoid
explicit numerical computations like the plague. Physicists may not
like the way computers have takenovertheoretical work but it certainly
is inevitable and it is here to stay.
Aneven more critical point concerning the difference is that physi-
cists put togetherjusttheright mix of assumptions to make an analysisof
the given physical phenomena work. From amathematical standpoint
we might say that they cleverly choose just the right axioms. Mathe-
maticians on the other hand get kudos for doing clever things with a
fixed set of axioms, that is, clever derivations, clever arguments, etc.
Of course, mathematicians also getcredit for introducing new concepts,
but still this is a working difference that I think is of great importance
in contrasting the two disciplines.
Finally, for better or worse,theoretical physicists are driven by exper-
iments. Theydo notperform them,but theyeagerly await thelatest word
from the current run. A really significantempirical resultin physics can
be followed in a matter of a few months by severalhundred theoretical
papers. Nothing like this concern for empirical matters is to be found
anywherein pure mathematics.

HierarchyofModels for Experiments.I want to amplify my ownideas

about experimentsandthemodels relevant to them, which Wojcicki dis-
cusses briefly. The more I think about scientific practice andreflect on
how to give an accurate account of the complicated processes that go
into experimentation, the more I am persuaded that there are a large
number of distinctions needed to describe experimentation thoroughly,
especially as data are purified for quantitative, and even more statisti-
cal, analysis. Itis a long way from running around the laboratory doing
one thing and then another, to having a set of data as printout or on a
computer screen ready for analysis. That process stillneeds much more
thorough attention than it hasreceived. Much of the current interest in
thephilosophy of scienceindiscussingexperimentshas shied away from
the gruesome details of exactly how data are purified and selected for
analysis, not to speak of details of how they are generated, whichitself
may involve, as equipment becomes increasingly complicated, many
different independent tests of reliability and accuracy of the equipment.
Itis acommon complaint of engineers in high-tech laboratories thatthe
scientists do not understand well enough the actual performance char-
acteristics of the equipment they use. We will have come a long way
in the philosophy of science when a really thorough analysis of these
important phenomenaappears for some particular discipline.

Unlimitability of Ordinary Informal Language. Once we turn to

experimentsit is important to recognize that weare not in any meaning-
ful way going to eliminate ordinary languageusedinits informal natural
fashion. The attempt to formalize all the way down what happens in
experiments is an enterprise that can only be pursued in vain. It is an
important point, in fact, in recognizing and constructing formal models
of experiments in terms of the purification of data for statistical analysis
that the analysis does not go in any formal way all the way down. Itisin
thiskindof respect thatit seems to me that Quine's eliminative program
for the use of language is mistaken. Inno serious way will intentional,
causaland otherkinds of ordinary idioms be successfully eliminated in
describing the rich activities that make up actual experiments. It is not
just ordinary experience, but scientific experience that cannot be sub-
ject to any serious reductionist program in all respects. Those parts of
science that are reducible are thin, austerepieces of theory,farremoved
from the rich experimentaljunglein which any true science necessarily
This does not mean thatI amnot in favor ofreduction when possible.
What we do not need, however, are reduction romances, stories too often
told byphilosophers without recognition of their fairy tale qualities.



ABSTRACT. We firstreviewour previous workon Suppespredicates andthe axiomati-

zationofthe empiricalsciences. We thenstate someundecidability and incompleteness
results in classical analysis that lead to the explicit construction of expressions for
characteristicfunctions in all complete arithmetical degrees. Out of thoseresults we
show that for any degree there are corresponding 'meaningful' unsolvable problems
in any axiomatizedtheory thatincludes the language of classical analysis. Moreover
wealso show that withinour formalizationthere are 'natural' unsolvableproblems and
undecidablesentences whichare harder than any arithmeticproblem. As applications
we discuss a 1974HilbertSymposium problemby Arnold on the existence of algo-
rithms for the decision of properties of polynomial dynamical systems over Z, prove
the incompleteness ofthe theory of finishNash games, anddelve on relatedquestions.
Neither forcing nor diagonalizations are used in those constructions.


Suppes predicates were the starting point in the recent development

of a technique for the construction of both algorithmically undecid-
able sets of objects in physics and undecidable 'meaningful' sentences
about physical objects. (See for details (da Costa,1988; Suppes,1967,
1988).) That technique allowed us to prove the undecidability and
incompleteness of most of classicaland quantum physics, providedthat
they are given a first-order axiomatization (through Suppes predicates)
that includes the language of classical elementary analysis (da Costa,
1991a, 1991b, to appear, 1994a; Stewart, 1991). Actual examples dealt
with the proof of the incompleteness of chaos theory (da Costa, 1991a)
and with the related question of the existence of problems in dynam-
ical systems theory whose solution is equivalent to solving very hard
Diophantine problems, such as Fermat's Conjecture (da Costa, 1994a).
The same technique reached beyondphysics andled to the proof of the
incompleteness of the theory of Hamiltonian models for the dynamics
of economical systems (Lewis, 1991b), while providing a partial result
related to the recent proof by Lewis of the noncomputability of Arrow-
Debreu equilibria (da Costa, 1992d; Lewis, 1991a). We can still list

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 151-193.
@ 1994 Kluwer Academic Publishers. Printedin theNetherlands.

among its consequences two examples that are discussedin the present
paper: a result onone of Arnolds problems concerning algorithms for
properties of polynomial dynamical systems over the integers (Arnold,
1976a) and theproof of the incompleteness of the theory of finite games
with Nash equilibria (da Costa, 1992d).
Our main theorems depend in an essential way on a lemma of
Richardson (Richardson,1968);noncomputability inaxiomatized phys-
ical theories was anticipated by Scarpellini (Scarpellini, 1963) and by
Kreisel (Kreisel, 1976). Also, we wish to emphasize that no forcing is
needed for our independence results.
We obtain here a whole plethora of new intractable questions in
Suppes-axiomatizedtheories. The present results are new in thefollow-
ing sense: all our previousexamples for undecidability and incomplete-
ness withinaxiomatized physics canbe formally reduced to elementary
arithmetic problems. However, that reduction cannot always be made
in the present case, as some of our new examples are not elementary
number-theoretic problems in disguise; they stand beyond the pale of
There are even weirder situations: we obtain formal expressions
that describe physical systems such that nothing but trivialities can be
provedabout them. For wecan explicitly construct undecidable families
of objects within a classical first-order language Lt such that:
No nontrivial properties of those families can be algorithmically
No assertion about the system can be reduced to an arithmetic
assertion, thatis to say, the systemlies fully outside thearithmetical
hierarchy and belongs to the nonarithmetical portion of set theory
(if weare working, say, within Zermelo-Fraenkel set theory).
Those results are consequences of general incompleteness theo-
rems that apply to any nontrivial propertyP in the theory T; those
theorems extend a previous one (Proposition 3.28 in (da Costa,
1991a) that originated in a suggestedbySuppes.
Again wehavea correspondingincompleteness theorem as thereare
formal expressions for systems all of whose nontrivial properties must
be formulated as undecidable sentences. (Again no property of those
systems canbereduced to an arithmetic property.) Thoseundecidability
and incompleteness results are to be found below in Propositions and
Corollaries 3.28, 3.30, 3.37, 3.41, 3.47, 3.49, 4.1.
Section 2 of this paper reviews the theory of Suppes predicates for
empirical theories in the da Costa-Chuaqui version (da Costa, 1988,
1990b, 1992a). Section 3 deals with the undecidability and incom-
pleteness of classical analysis in its several aspects; the main points
are the explicit construction of expressions for characteristic functions
of any complete degree in the arithmetical hierarchy, and the gener-
al incompleteness theorems about expressions of elementary functions.
Section 4discusses our examples, whileSection 5 comments on possible
implications of our results.

Preliminary Concepts andNotation

We are here interested in formal languages strong enough to represent
all the usual mathematical theories. For a review of the main ideas
see (da Costa, 1990b, 1991a). Those formal languages are built out
of a finite alphabet, and its sentences 'meaningful assertions' - are
finite sequencesof letters from thebasic alphabet. We therefore reduce
everything to finite sequencesof letters. (As an example, an intuitively
infinite set suchas a straight line on theCartesian plane R2 is represented
by the finite sequenceof symbols {(x,y) € R2 : y = 2x 4- I}, abbre-
viations such as R for the set of reals being allowed as their definitions
are reducible to finite sequences of letters from the theory's alphabet.)
Different formal expressions may represent the same 'intuitive' mathe-
matical object;however in most everyday situations our formal systems
will not be strong enough to decide, given two expressions,whether or
not they represent the same object, even if they are strong enough to
prove all usual mathematical results.
To be more specific, we suppose that our theories are formalized
withina first-order classical predicate calculus with equality. (It is also
convenient to suppose that our formal theories T include Russell's i
symbol; in that case the extended theory is a conservative extensionof
the theory without that particular variable-binding term operator.)
We follow the notation of (da Costa, 1991a) with a few changes that
are explicitly indicated; in particular v will denote the set of natural
numbers, Z is the set of integers, and R are the real numbers. Let
T be a first-order axiomatic theory that contains formalized arithmetic
N and such that T is strong enough to include the concept of set and
classical elementary analysis. (We can simply take T = ZFC, where
ZFC is Zermelo-Fraenkel set theory with the axiom of choice.) If Lt

is the formal language of T, we suppose that we can form within T

a recursive coding for Lt so that it becomes a set Lt G T of formal
expressions in T. Objects in T will be denoted by lower case italic
letters, such as x, y, z, f, g, a, Predicates in T will be noted P,
Q, The use of Greek letters and more particular notational features
(such as p, q for polynomial functional symbols) will be clearfrom the
context. (Predicates are open one-variable formulae in the language of
From time totime we will play with thedistinction between an object
and the expression in Lt that represents it. If x, y are objects in an
intended interpretation of the theory T,theyare in generalnotedby term-
expressions £, £ that belong to the formal language Lt of T. In general
there is no 1-1 correspondence between objects and expressions; thus
we may have different expressions for the same functions: 'cos 7r' and
'0' are both expressions for the constant function 0. If £ is a set in an
intended interpretation of T, we noteby \x] a set of expressions for the
elements in x. We allow the following abuse of language: predicates P
sometimes apply to objects in T and sometimes apply to expressionsin
Lt(P(OY> meaning will be clear from context.
In particular we notice that since our theory T includes formalized
arithmetic N, we will sometimes need the distinction between a partial
recursive function and the algorithm that computes it. For any compu-
tation 4>(x), (f>(x) Imeans that the computation converges (stops and
produces an output), while (f>(x) |means that the computation diverges
(enters a never-endingloop).
We emphasize that proofs in T are algorithmically defined ways of
handling the objects of Lt', for the conceptof algorithm see (da Costa,
1991a; Rogers,1967).


Suppes predicates (or Bourbaki species of structures) were first used

by Suppesin the 'fifties as a way of directly axiomatizing any mathe-
matically-based theory. Suppes'smain contention is that"to axiomatize
a theory is to define a set-theoretic predicate." A formal treatment was
then givenby daCostaand Chuaquiin1988,andimmediate applications
ensued. (For details, applications and references see (Bourbaki, 1957,
1968; daCosta 1988, 1990b, 1992a; Suppes, 1967, 1988). Everything
here is supposed to happen within our arithmetically consistent theory

Structures and Predicates

A mathematical structure w is a finite orderedcollection of sets (which
may be particularized to relations and functions) of finite rank over the
union of the ranges of two finite sequences of sets, xj,xi, ,xm and
> 0. If weare doing ourconstructions
y\ ,yii " " 2/n» where m > 0and n

within ZFC, w is thus a ZFC set. The xs and the ys are called thebase
setsofw; the xs are the principalbase sets, while the ysare the auxiliary
base sets.
The auxiliary base sets can be seen as previously defined structures,
while the principal base sets are 'bare' sets; for example, if we are
describing a real vector space, the set of vectors is the only principal
base set, while the set of scalars, R, is the auxiliary base set.
A Suppespredicate or a speciesof structuresin the senseofBourbaki
isa formula of set theory whose only free variables are those explicitly

P(W, Xi,X 2,
"" " , m,2/1,2/2,---, 2/n)-

P defines it; as a mathematical structure on the principal base sets

... ...
x\, xm,with the auxiliary base sets y\ , ,yn,subject to restrictions
imposed on w by the axioms we want our objects to obey. As the
principal sets x\, ...vary over a class of sets in the set-theoretical
universe,we get the structures of species P, or P-structures.
The Suppes predicate is a conjunction of two parts: one specifies
the set-theoretic process of construction of the P-structures, while the
other imposes conditions that must be satisfiedby the P-structures. This
second piece contains the axioms for the species of structures P.
We write the Suppes predicate for a general w as follows:

W(q) <-> 3xi3x2...3xmP(iy,xi,X2,...,x m,yi,...,2/n).

The auxiliary sets are seen as parameters in the definition of w. All of

everyday standard 'professionals' mathematics canbe formalized along

Deduced andDerivedStructures
Given a structure wof species P(w, x\,... ,x ,2/1 ,... ,y ), let
m n 21,
... , z ,be p(p> 0) sets of finite rank over the union of ranges of the

""" ,%mi y\i - - " yn,
also let v\ , ...,v q (q
> 0) be q arbitrary sets. If the Suppes predicate
defines w* as a structure on the principal base sets z\, ...
with the v\, auxiliary base sets, we say that the structure w* of species P* is
deducedfrom the structure w of species P.
We can obtain new structures out of (sets of) already defined struc-
tures by the means of two basic procedures:
1 With the help of set-theoretic operations, such as Cartesian prod-
ucts, passages to the quotient, and the above-described operation
of deduction of structures;
2. Through the imposition of new axioms to already existing set-
theoretic structures.
Therefore we can introduce the notion of derived structure. When
we define anew structure w from a set s ofother structures with the help
of the two procedures described above, we say that w is derived from
the structures s. The Suppes predicate of w can be expressedin terms
of the Suppespredicates of the elements of s. The conceptof deduction
of structures is a particular case of derivation of structures.
The set s is the set of ground structures for w.
Finally, let w and w1 be twostructures of species P and P', respec-
tively. We suppose that P and P' differ only in connection with their
sets of axioms, but that the conjunction of the axioms of P' implies
each axiom of P, with quantifiers restricted to sets of finite rank over
theunion of the ranges of the base sets for w. If that is the case, we say
that the P'-structure is richer than the P-structure (or that P' is richer
than P). For instance, the species of commutative groups is richer than
the species of groups.
The Q'-structure g' is then derived from the Q-structure g if Q' is
richer than Q, or Q' canbe obtained from Q in the way we have already
described above. The above ideas can also be extended to the concept
ofpartial structures introduced by (da Costa, 1990a).
TheAxiomatics ofEmpirical Theories
As a first approximation we see empirical theories as triples
where (i) Mis a Suppes-Bourbaki species of mathematical structures;
(ii) A is the theory's 'domain of definition',and (iii) p gives the 'inter-
pretation rules' or 'characteristic examples' that relate Mand A. We
can be more specific about (ii) and (iii), however, as we did elsewhere
(da Costa, 1992c); in any case we consider A to be a set-theoretic
construct. (In that case p in general contains nonrecursive aspects (da
Costa, 1991a, to appear, 1992a).)

An Example: SuppesPredicatesfor Classical FieldTheories

in Physics
We follow the usual mathematical notation in this subsection. In par-
ticular, Suppes predicates are written in a more familiar but essentially
equivalent way. Therefore some symbols willhave a different meaning
thanin the remainingportions of the paper.
The species of structures of essentially all physical theories can
be formulated as particular dynamical systems derived .from the P =
(X,G,fi), where X is a topological space, G is a topological group,
and p is a measure on a set of finite rank over X UG. Thus we can say
that the mathematical structures of physics arise out of the geometryof
a topological space X: physical objects are those that exhibit invari-
ance properties with respect to the action of G and the main species
of structures in 'classical' theories can be obtained out of two objects,
a differentiable finite-dimensional real Hausdorff manifold M and a
finite-dimensional Lie group G.

DEFINITION 2.1. The species of structures of a classical physical

theory is givenby the 9-tuple
E= (M,G,P,T,A,I,G,B,V<p = l)
which is thus described:

1. The GroundStructures. (M,G), where Mis a finite-dimensional

real differentiable manifold andGisa finite-dimensional Lie group.

2. The Intermediate Sets. A fixed principal fiber bundle P(M,G)

over M with G as its fiber plus several associated tensor and
exterior bundles.
3. TheDerivedField Spaces. Potential space A, field space Tand the
current or source space X. A, T andTare spaces (in general,man-
ifolds) of cross-sections of thebundles that appear as intermediate
sets in our construction.
4. AxiomaticRestrictions on theFields. The dynamical rule Vy? = t,
the relation ip = d(a)a between a field <p c T and its potential
a c A, together with the corresponding boundary conditions B.
Here d(a) denotes a covariant exterior derivative with respect to
the connection form a, and V a covariant Dirac-like operator.
5. The Symmetry Group. Q C Diff(M) ® Q', where Diff(M) is
the group of diffeomorphisms of M and Q' the group of gauge
transformations of the principal bundle P.
6. The Space of Physically Distinguishable Fields. If /C is one of the
T, A or Z field manifolds, then the space of physically distinct
fields is


(In more sophisticated analyses we must replace our concept of theory

for a more refined one. Actually in the theory of science we proceed
as in the practice of science itself by the means of better and better
approximations. However for the goalsof the present paperour concept
of empirical theory is enough.)
We show elsewhere (da Costa, 1992a) that what one understands
as the classical portion of physics fits easily into the previous scheme.
We discussindetail two examples,Maxwellian theory and Hamiltonian

Maxwell's Electromagnetic Theory

Let M = R4, with its standard differentiable structure. Let us endow
M with the Cartesian coordination induced from its product structure,
and let 7/ = diag(— l,+l,+l,+l) be the symmetric constant metric
Minkowskian tensor on M. If the F^v {x) are components of a differ-
entiable covariant2-tensor field on M, p, v = 0, 1,2, 3, thenMaxwell's
equations are:


d^p+ dpFay + dvFpu = 0

Thecontravariant vector fieldwhose components are givenby the set
of four smooth functions j^(x) onMis the current thatserves as source
for Maxwell's field F^. (We allow piecewise differentiable functions
to account for shock-wave-like solutions.)
Itisknown thatMaxwell's equations are equivalent to the Dirac-like

V(/? = L,

= (1/2)^7^,
- Jul ,

V = yp dp,

(where the {7^ : p — 0, 1,2, 3} are the Dirac gamma matrices with
respect to 77). Those equation systems are to be understood together
with boundary conditions that specify a particular field tensorF^ 'out
of the source jv (Doria, 1977).
The symmetry group of theMaxwell field equations is theLorentz-
Poincare group that acts onMinkowski space Mand in an induced way
onobjects defined over M. However,since weare interestedin complex
solutions for the Maxwell system, we must find a reasonable way of
introducingcomplex objects inour formulation. One may formalize the
Maxwellian system as a gauge field. We sketch the usual formulation:
again we start from M = (R4,rj), and construct the trivial circlebundle
P — M x Sl over M, since Maxwell's field is the gauge field of the
circle group 51(usually writtenin that respect as U( 1)). We form the set
£ of bundles associated to P whose fibers are finite-dimensional vector
spaces. The set of physical fields in our theory is obtained out of some

of the bundles in S: the set of electromagnetic field tensors is a set of

cross-sections of thebundle F = A 2 0 s 1(M) of all 2-forms
on M, where s is the group's Lie algebra. To be more precise, the set
of all electromagnetic fields is T C Ck (F), if we are dealing with Ck
cross-sections (actually a submanifold in the usual Ck topology due to
the closure condition dP = 0).
Finally we have two group actions on T: the first oneis theLorentz-
Poincare action L which is part of the action of diffeomorphisms of
M; then we have the (here trivial) action of the group Q1 of gauge
transformations of P when acting on the field manifold T. As it is
well known,its action is not trivial in thenon-Abelian case. Anyway,
it always has a nontrivial action on the space A ofall gauge potentials
for the fields in T Therefore we take as our symmetry group Q the
product L ® Q' of the (allowed) symmetries of Mand the symmetries
of the principal bundle P.For mathematical details see (Doria, 1981).
We must also add the spaces A of potentials and of currents, X,
as structures derived from M and Sl Both spaces have the same
underlying topological structure; they differ in the way the group Q' of
gauge transformations acts upon them. We obtain I = A1 l(M) and
® s
A= 1 = C k(I). Notice that XjQ' = X while A.
Therefore we can say that the 9-tuple

{M,SI,P,T,A,g,X,B,V<p = i)

where Mis aMinkowski space, and B is a set of boundary conditions

for our field equations V</? = t, represents the species of mathematical
structures of aMaxwellian electromagnetic field, where P, T and Q are
derived from Mand Sl TheDirac-like equation

V<£> = L
should be seen as an axiomatic restriction on our objects; the boundary
conditions B are (i) a set of derived species of structures from Mand
S\ since, as we are dealing with Cauchy conditions, we must specify
a local or global spacelike hypersurface C in M to which (ii) we add
sentencesof the form Vx G Cf(x) = /o(x), where /o is a set of (fixed)
functions and the / are adequate restrictions of the field functions and
equations to C.
Hamiltonian mechanics is the dynamics of the 'Hamiltonian fluid'
(Arnold, 1976b). Our groundspecies of structures are a2n-dimensional
real smooth manifold, and the real symplectic group Sp(2n,R). Phase
spaces in Hamiltonian mechanics are symplectic manifolds: even-
dimensionalmanifolds like Mendowed with a symplectic form, that is,
a nondegenerateclosed 2-form Q, on M. The imposition of that form
can be seen as the choice of a reduction of the linear bundle L(M)
to a fixed principal bundle P(M,Sp(2n,R)); however given one such
reduction it does not automatically follow that the induced 2-form on
Mis a closed form.
All other objects are constructed in about the same way as in the
preceding example. However, we must show that we still have here a
Dirac-like equation as the dynamical axiomfor the species of structures
of mechanics. Hamilton's equations are

— —dh,
where %x denotes theinterior product with respect to the vector field X
over M, andhis theHamiltonian function. That equationis (locally, at
least) equivalent to:

LxSl = 0,


d{ixQ) = 0,

where Lx is theLie derivative with respect to X. Thecondition dip = 0,

with (f = ix&>, is the degenerate Dirac-like equation for Hamiltonian
mechanics. We do not get a full Dirac-like operator V d because
M, seen as a symplectic manifold, does not havea canonical metrical
structure, so that we cannotdefine (through the Hodgedual) a canonical
divergence 6 dual to d. The group that acts on M with its symplectic
form is the group of canonical transformations; it is a subgroup of the
group of diffeomorphisms of M so that symplectic forms are mapped
ontosymplectic forms under a canonical transformation. We can take as
'potential space' the space of all Hamiltonians on M(which is a rather
simple function space),andas 'field space' the space ofall 'Hamiltonian
fields' of the form ixsl-

The constructionof Suppespredicates for gravitation theory(general

relativity), classical gauge fields, Kaluza-Klein unified field theories,
and Dirac's electron theory seen as a classical field theory can be found
in (da Costa, 1992a). We notice that Dirac-like dynamical equations
havebeenobtained for all those (Doria, 1975, 1986). Themathematical
background is in (Cho, 1975;Dell, 1979; Kobayashi, 1963).


Wenow review previous material andobtainnew results on theundecid-

ability and incompleteness of classical analysis; the chief new results
are the construction of several intractable problems and undecidable
sentencesin T which cannot be reduced (in T) to arithmetical problems.

DEFINITION 3.1. Tis arithmetically consistent if and only if the stan-

dardmodel Nfor N is a model for the arithmetic sentencesof T. ■
Now let \V~\ be the algebra of polynomial expressions on a finite
number of unknowns over the integers Z; we identify \V\ with the
set of expressions for Diophantine polynomials in T. Let \£~\ be the
set of expressions for real elementary functions on a finite number of
unknowns, while [JF] is the set of expressions for real-valued elemen-
tary functions on a single variable (da Costa, 1991a).
... ...
Givena polynomial expressionp(x\, ,x m,yi, ,2/ n),letrm(xi,
, xm) be the function that effectively codes m-tuples of natural
numbers (xi, ..., xm) by a single natural number (Rogers, 1967,
p. 63). Letr =rm((...». We abbreviate p(xi, ... ,xm,2/i, ...
,y„) =
Let us inductively construct out of a polynomial qm {x\ ,...,x
n) an
n given through following
R -defined and R-valued function aqm the

Initial Step. Suppose that we are given the expressions qm as below:

"If qm(x\ , ...,x ) =c, where cis a constant, then we put aq
n m
"If qm{x\ ,"" " ,xn) = xi, then aqm =x]+ 2.
Induction Step. We suppose that qm is given as indicated. We then
obtain the corresponding aqm as follows:
"If qm =Sm ± tm,then aqm = asm + atfm .
"If qm = smtm, then agm = aSmOLtm .
We then write ki(m, x\, ...,x ) = adiq
n m (x\ , ...,x ), where <%
Suppose now that weisolate some of the variables in our polynomi-
als as parameters. Then

DEFINITION 3.2. The map a : \V] -+ \£~\, givenby:

p(xi,...,x m,2/i,...,2/n) "->" ap(xi,...,x m,2/i,...,2/n)
= (n+ 1)V(xi,...,x ,2/i,...,2/n) +
+ XXsin27TXi)kj(xi,..., m,2/1, "" " ,2/n),


isRichardson 'sFirst Map.

COROLLARY3.3. Given a polynomial expressionpm £ Lt, there is

an algorithm that allows us to obtain an expression apm GLt for the
image ofpm under Richardson'sFirst Map.
Proof. Immediate, from the definition of a.

We now define:
h(x) = xsinx,
g(x) — xsinx3 .
Given a set ofreal variables x\, . xn,we define the following maps:
= h(x),
x2 = hog(x),
= hogog(x),
= ho go ... og(x),
(whereg is composed n
— 2 times), and
xn = gogo ... og(x).

Here g is composed n times.

Given a polynomial expressionpm(x\ , ...,x ) 6 \V~\ ,we define:

DEFINITION 3.4. The maps i' : -+ \T] and i" : \V] -+ \T],
(1) pm(xi,...,x n) i-> i\prn{x\,...,x n)]{x)
= apm(h(x),hog(x),..
gogo ...og(x)),
where a is Richardson's First Map; and
(2) Pm(xi,...,X n) t"[pm(xu... ,X„)](x)
= t /[pm(xi,...,X„)](x) - ,
areRichardson 'sSecondMap of the first (i1) and second {i") kinds. ■
COROLLARY 3.5. Given a polynomial expression pm G Lt, there
is an algorithm that allows us to obtain expressions t!p m £Lt and
i"Pm €Lt for the images ofpm under Richardson 'sSecond Map.
Proof. Immediate, from the definition of i' and i" .
We assert:

PROPOSITION3.6 (Richardson's Functor). Letpm(x u x2, ,xn) = ...

0be afamily of expressionsfor Diophantine equationsparametrizedby

the positive integer m in an arithmetically consistent theory T. Then
there is an algorithmic procedure a : \V~\ * \£~\ such that out of
pm €V we can obtain an expression

Jm \X\, X 2, .."
,Xn) — Q!Pm \Xli X 2,
, ,
"" " Xn)

fm €£, such that fm =0ifand only iffm <1ifand only ifthere are
positive integers x\,x 2 , xn such thatpm(x\ , ,xn) =0. ... —
Moreover, thereare algorithmic proceduresi',i" :P * T such that
we can obtain out of an expressionpm two other expressionsfor one-
variable functions, gm(x) = i'pm(x\, .) and hm(x) = i"pm(x\, .) ..
such that there are positive integers x\, with pm(x\ , .) =0ifand
only ifgm{x) = 0andhm(x) < 1, for all real-valued x.

Proof See (da Costa, 1991a).

PROPOSITION3.7. IfT is arithmetically consistent, and ifweadd the

absolute value function \x\to \T] andcloseit to obtainanextended set
of expressions \T*~\, we have:
1 (Undecidability) We canalgorithmically constructin Ta denumer-
able family of expressionsfor real-valued,positive-definite func-
tions k m(x) > 0 so that there is no general algorithm to decide
whether one has, for all real x, km(x) =0.
2. (Incompleteness) For a model Msuch that T becomes arithmeti-
cally consistent, there is an expression for a real-valued function
k(x) such thatM |= Vx <E Rfc(x) = 0 while TV- Vx € Rfc(x) = 0
andTY3x G Rfc(x) 0.
Proof. See (da Costa, 1991a).

If km (as in Proposition 3.7) results out of pm,we write km = Xpm -

Equality Is Undecidable in Lt
COROLLARY 3.8. IfTis arithmetically consistent then for an arbi-
trary real-defined and real-valued function f there is an expression
f G Lr such thatM f= £ = /, whileTY^ = f andT V- -.(£ = /).
Proof. Put £ = / + k(x), for fc(x) as in Proposition 3.7.
The Halting Function andExpressionsfor CompleteDegrees in the
Arithmetical Hierarchy
Now let Mn(q) be the Turing machine of index n that acts upon the
natural number q (Rogers, 1967). Let 9(n,q) be the halting function
for Mn(q), that is, 0(n,q) =1if and only if Mn{q) stops over q, and
0(n,q) = 0 ifand only if Mn(q) does not stop over q.
We need a definition and a lemma.

DEFINITION3.9. For the following real-defined andreal-valued func-


(i) m={-*:*<2:

fl,a; > 0,
/ \

-10,< 0.

x — >
v J x-y=
y { f x yy,' x
y ~ 0,
[0, x—2/ 0.
( +1, x > 0
(4) a{x) = < 0, x=0
[ -1, x<o

LEMMA 3.10. If T is (arithmetically) consistent, then each of the

following operations generates the others within T:
(1) +,x,|
(2) +, X,7/(...).

(3) +,x, (...-...).

(4) +,X,<7(...).
Proof. Immediate.

Let pn,q(xi,X2, ,x n) be a universal polynomial (Jones, 1982).
Since \T*~\ has an expression for |x| (informally one might have
|x| = +Vr),ithas an expression for the sign function a(x). Therefore
we can algorithmically build within the language of analysis (where we
can express quotients and integrations) an expression for the halting
function 6{n,q):

PROPOSITION 3.11 (The Halting Function). IfT is arithmetically

consistent, then:

o(n,q) = a(Gnjq),

a -

~ 2
Cn.i(x>ce x dx
1 + Cn,,(i)
— ,...,Xr ).
Proof See (da Costa,1991a).

There follows:

PROPOSITION 3.12. IfT is arithmetically consistent then we can

explicitly and algorithmically construct in Lt an expression for the
characteristic functionofa subset of uj of degree 0".

Remark 3.13. That expression depends on recursive functions defined

on uj andon elementary real-defined andreal-valued functions plus the
absolute value function, a quotient and an integration, as in the case of
the 9 function givenby Proposition 3.11. ■

Proof. We could simply use Theorem9-IIin (Rogers, 1967, p. 132).

However,for the sake of clarity we give adetailed, albeit informal proof.
Actually the degree of the set described by the characteristic function
whose expression weare going to obtain will dependon the fixed oracle
set A; soour construction is a general one.
Let A C uj be a fixedinfinite subset of the integers:

DEFINITION 3.14. The jump of A is written A'; A1 = {x : <j>£{x) |},

where 4>^ is the A-partial recursive algorithm of index x. ■
1. An oracle Turing machine (f>^ with oracle A can be visualized
as a two-tape machine where tape 1 is the usual computational
tape, while tape 2 contains a listing of A. When the machine
enters the oracle state so, it searches tape 2 for an answer to a
question of theform 'is w G AT Onlyfinitely many such questions
are asked during a converging computation; we can separate the
positive and negative answers into two disjoint finite sets DU(A)
and D*(A) with (respectively) the positive and negative answers
to those questions; notice that Dv C A, while D* Cuj— A. We
can view these sets as orderedk- and A;*-pies; v and v are recursive
codings for them (Rogers, 1967). The DU(A) and D*(A) sets
can be coded as follows: only finitely many elements of A are
queried during an actual convergingcomputation with input y; if

k' is the highestinteger queriedduring one such computation, and

if dA C ca is an initial segment of the characteristic function ca,
we take as a standby for D and D* the initial segment dA where
the length l(dA ) =k'+ 1.
We can effectively list all oracle machines with respect to a fixed
A, so that, givena particular machine we can computeits index (or
Godel number) x, and given x we can recover the corresponding
2 Given an A-partial recursive function <ft£, we form the oracle
Turing machine that computes it. We then do the computation
4>x(y) = z tnat outputs z. The initial segment dy A is obtained
during the computation.
3. The oracle machine is equivalent to an ordinary two-tape Turing
machine that takes as input (y, dy^)', 2/ ls written on tape 1 while
dVy A is written on tape 2. When thisnew machine enters state so it
proceeds as the oracle machine. (For an ordinary computation, no
converging computation enters so, and dViA is empty.)
4. The two-tape Turingmachine canbemade equivalent to a one-tape
machine, where some adequatecodingplaces on the single tape all
the information about {y, dy^)- When this thirdmachine enters so
it scans dy> A-
5. We can finally use the standard map r that codes n-ples 1-1 onto
uj andadd to the preceding machine a Turing machine that decodes

the single natural number r((y, dy,A)) into its components before
proceeding to the computation.
Let w be the index for that last machine; we noteit <j> w
If x is the index for 4>£, we note w = p{x), where pis the effective
1-1 procedure described above that maps indices for oracle machines
into indices for Turingmachines. Therefore,

<f>£(y) = 4>p(x)((y,d ,A))'


Now let us note the universal polynomial p(n, q,X\, ...,x ). We

can define the jump of A as follows:

A1= {p(z) : 3xi,...,xn € v

p(p(z),(z,dZjA ),x v ...,xn) = o}.
With the help of the A map defined following Proposition 3.7, wecan
now form a function modelled after the 9 function in Proposition 3.11;
it is the desired characteristic function:
CO/(x) =0(p(x),(x,4 O/». j

(Actually we have proved more; we have obtained

ca'(x) = 9{p(x),{x,d )), X)A

with reference to an arbitrary A C uj.)

We write 9^2\x) =cr(x).

We recall (Rogers, 1967):

DEFINITION 3.15. The complete Turingdegrees, 0, 0', , o^\...

p < uj, are Turing equivalence classes generated by the sets 0, o', 0",
Now let 0(n) be thenth complete Turing degrees in the arithmetical
hierarchy. Let r{n, q) = mbe thepairing function in recursive function
theory (Rogers, 1967). For 9(m) = 9(r(n,q)), we have:

COROLLARY3.16 (Complete Degrees). IfT is arithmetically consis-

tent, for all p £ uj the expressions 9p(m) explicitly constructed below
representcharacteristic functions in the complete degrees0^p\
Proof. From Proposition 3.12,
'0(°) = c (m) =O,
< 0O)(m) = cy(m) = 0(m),
0(n) (m),

for ca as in Proposition 3.12.

Incompleteness Theorems
We now state and prove several incompleteness results about N andits
extension T; they will be needed when we consider our main examples.
'— '
Werecall that is a primitive recursive operation on uj.

The starting point is the following consequence of a well-known


PROPOSITION 3.17. IfTis arithmetically consistent, then we can

algorithmically construct a polynomial expression g(xi, ,xn) over
... ...
Z such that M(= Vxi, ,xn G ujq(x\ , ,xn) >0, but
TV- Vxi,...,x„ G wg(xi,... ,xn) > 0
TV- 3xi,"" " ,xn G ujq(x\ , ...,x ) =0.

Proof. Let ( G Itbe an undecidable sentence obtained for T with

the help of Godel's diagonalization; let n^ be its Godelnumber andlet
rriT be the Godel codingof proof techniquesinT (of the Turing machine
that enumerates all the theorems of T). For an universal polynomial
p(m, q, x\,..., xn) we have:
g(xi,...,x n) = (p(mT,n^xv ... ,x )) .

COROLLARY 3.18. IfH is arithmetically consistent then we can find

within it apolynomialp as in Proposition 3.17 ■
We can also state and prove a weaker versionof Proposition 3.17:

PROPOSITION3.19. IfTis arithmetically consistent, there is a poly-

nomial expression over Z.p(x\ , ,xn) such that Mf= Vxj, ,xn G ...
ujp(xi, ,xn) > 0, while
TF Vxi,...,x„ G wp(xi,...,x n) > 0
TP 3xi," " " ,xn G ujp(x\, ... ,x ) =0.

Proof. See (Davis, 1982): if p(m,x\, ..., xn), m = r(q,r),

is a universal polynomial (r is Cantor's pairing function (Rogers,
1967), then {m : 3xi ... ..
G ujp(m,x\, .) =0} is recursively enu-
merable but not recursive. Therefore there is an mo such that Vxi
...uj(p(m 0,xi,...)) 2 > 0. ■
Predicates or properties in T are representedby formulae with one
free variable in Lt.

DEFINITION 3.20. A predicate P in Lt is nontrivial if there are term-

expressions (,CG LT such that T h P(£) and T h -iP(C). ■


PROPOSITION 3.21. If H is arithmetically consistent and if P is

nontrivial then there is a term-expression ( G Ln such that N |= P(C)
while N V P(() and N V -.P(C).
... -
Prao/ Put (=£ 4- r(xi, ,xn)v, for r= 1 erg2,gas in Propo-
sition 3.17 (or asp in 3.19). ■

Remark 3.22. Therefore every nontrivial arithmetical property P in

theories from arithmetic upwards turns out to be undecidable. We can
generalize that result to encompass other theories T that include arith-
metic; see below. ■
We now give alternative proofs for well-known results about the
arithmetical hierarchy that will lead to other incompleteness results:

DEFINITION 3.23. The sentences £, £ G Lt are demonstrably equiv-

alent if and only if T h £ <-» (. ■

DEFINITION 3.24. The sentence £ G Lt is arithmetically expressible

if and only if there is an arithmetic sentence £ such that Th^<->(. ■


PROPOSITION3.25. IfT is arithmetically consistent, then for every

m G uj there is a sentence £ G T such thatM |= £ while for no k < n
is there aEfc sentence in N demonstrably equivalent to £.
Proof. The usual proof for N is given by Rogers (Rogers, 1967,
p. 321). However, we give here a slightly modified argument that
imitates Proposition 3.19. First notice that
0("+" ={x : V)}
is recursively enumerable but not recursive in 0( m). Therefore 0(m+U
is not recursively enumerable in 0( m), but contains a proper 0(m)-
recursively enumerable set. Let us take a closer look at those sets.
We first need a lemma: form the theory T^m+l^ whose axioms are
those for T plus a denumerably infinite set of statements of the form
'n0 G 0(m)', 'ni G 0(m)', ,that describe 0<m). Then,

LEMMA 3.26. IfT^n+l^ is arithmetically consistent, then 4>x (x) J.
ifand only if
T(m+l) |_ 3xU ...,Xn eUJ
p(p(z),(z,dy0(m) ),xi,...,x n) =0
Proof. Similar to the proofin thenonrelativized case; see (Machtey,
1979, pp. 126ff). ■
Therefore wehave that the oracle machines (f>x (x) [ ifand only if
T(m+l) |_ 3xU ...,Xn eUJ
p(p(z), (Z, dy>o(m)), xi, n ) =0.

However, since 0(m+1) is not recursively enumerable in 0(m) then

there will be an index mo(0 (m^) = (p(z),(z, d 0 (m> )) such that

M (= Vxi,...,xn[p(m0,xi,...,xn)]2 > 0,
+ -
while it cannotbe provednor disproved within J'(m 1) it is therefore
demonstrably equivalent to a nm+i assertion. ■
(m) (m) 2
Now let g(mo(0 ), xi..)= p(rao(0 ), xv " -)) ■be asin Pro-
position 3.25. Then:

COROLLARY 3.27. IfTisarithmetically consistent, thenfor:

/?(m+l ) = a(C(mo(0 (n))),
G(mo(0("))) = [°C(rno(^),x)e-° 2

C(mo(0(n)),x) = A g (mo(0(n)),xi,...,x r),

T {n) y. _,^(m+l)
M |= /3(m+l) = 0 but for alln < m + 1, T^ V flm+^
= 0 and


COROLLARY3.28. IfTis arithmetically consistentand ifLT contains

expressionsfor the 9^ functions as givenin Proposition 3.16, then for
any nontrivialpredicate P in N there isaC, G Lt such that theassertion
P(C) is T-demonstrably equivalent to and T-arithmetically expressible
as a nm+i assertion, but not equivalent to and expressible as any
assertion with a lower rank in the arithmetic hierarchy.
Proof. As in the proof of Proposition 3.21, we write:
C = S + [1
- ff(p(m (0m),x ,
o 1 ...,x ))>,

where p{. .) is as in Proposition 3.25.

Remark 3.29. Rogers discusses the rank within the arithmetical hierar-
chy of well-known open mathematical problems (Rogers, 1967,p.322),
suchas Fermat's Conjecture - whichin its usual formulation is demon-
strablyequivalent toa ni problem,or Riemann'sHypothesis,alsostated
as a111 problem. Rogersconjectures that our mathematical imagination
cannot handle more than four or five alternations of quantifiers. How-
ever the precedingresult shows that any arithmetical nontrivial property
within Tcan give rise to intractable problems of arbitrarily high rank.
We obviously need the extensionT D N, since otherwise we would
not be able to find an expression for the characteristic function of a set

with a high rank in the arithmetical hierarchy within our formal lan-
guage. ■
An extension of the preceding result is:
COROLLARY 3.30. IfT is arithmetically consistent then, for any
nontrivial property P there is a £ G Lt such that the assertion P(()
is arithmetically expressible, M |= P(Q but it is only demonstrably
equivalent to a Tln+\ assertion andnot to a lower one in the hierarchy.
Proof. Put
where oneuses Corollary 3.27.

Undecidable Sentences OutsideArithmetic


0^ = {(x,t/) :xG0(y) },
for x,y G uj.


#H(m) = C0 (w)(m),
where c^{u)(m) is obtained as in Proposition 3.12.


_ (oHy_

COROLLARY3.34. O^1 ) is the degree o/0( w+l).

COROLLARY 3.35. 9^ u+l\m) is the characteristic function of a

nonarithmetic subset of uj of degree0^u+l\ ■
COROLLARY 3.36. IfTis arithmetically consistent, thenfor:

/^+1) = (7(G(mo(0(v;) )),

m («M«))=
Tc(m0(»(")),x)e-' 1

il+ C(ra„(oH),z)^
C(mo (0^) ),x) = A,(m0(0^) ),xi,...,x r ),

M(= /?^+1) = 0but T V /?^+1) =OandTY -,(p(v+l) = 0). ■

PROPOSITION 3.37 (Nonarithmetic Incompleteness). IfT is arith-

metically consistent then given any nontrivialproperty P:
(1) There is a family of expressions (m GLt such that there is no
general algorithm to check, for every m G uj, whether or not
(2) There is an expression( G Lt such that M |= P(C) while T V
(3) Neither £ m nor £ are arithmetically expressible.
Proof. We take:
(1) Cm = x9^+l \m) + (1 - 9^+l\m))y.
(2) ( = x + yP^+l\
(3) Neither 9^UJ+l\m) nor /3 W+l) are arithmetically expressible. ■
Remark 3.38. We have thus produced out of every nontrivial predi-
cate in T intractable problems that cannot be reduced to arithmetical
problems. Actually there are infinitely many such problems for every
ordinal a, as we ascend the set of infinite ordinals in T. Also, the
general nonarithmetic undecidable statement P(() has been obtained
without the help of any kind of forcing construction. ■

AllNontrivial Properties Are Undecidable

We motivate our nextintractability results with aconcreteexample from
Hamiltonian mechanics.
Suppose that we have axiomatized Hamiltonian mechanics over a
fixed phase space M with the help of a Suppes predicate within a first-
order theory T (which we can take to be ZFC) as done in Section 2.
We then have a predicate H(£) in Lt that asserts, 'the expression£ is
a Hamiltonian function. We can enumerateall other predicates P& in
Lt, k anatural number,and wecan also enumerateall the theorems in
T. We start such an enumeration, and select theorems of T which have
the form:
(2) For fc, Zj, i/j,G Lr,T hPfc(&) and T h -Pfc fe).
Out of that we list allnontrivial predicatesPfc that apply to Hamilto-
nian functions.
We have proved:

PROPOSITION3.39. IfTis arithmetically consistent, we can obtain

a recursive enumeration ofpairs of expressions £ 2i, &i+h ia natural
number, that representdifferentfunctions and such that:
(1) Both T h Hfoi) andTY H(&i+l);
(2) T h PifoO o/k/T h -P;(6i+i)/
where the P{ arenontrivialpredicates(relative to theHamiltonians over
M) in T that range over Hamiltonian functions. ■

Remark 3.40. The previous result allows us to obtain, out of an enu-

meration of the theorems in T and of all predicates Pi in the language
of T, an enumeration of different expressions £2*. £2^+l* ia natural
number,for different Hamiltonians that satisfy (and do not satisfy) each
predicate Pi. ■

We can state:

PROPOSITION3.41. IfTis arithmetically consistent, then there is a

countable undecidablefamily Cm of expressionsfor Hamiltonians in the
language ofT such that there is no generalalgorithm to decide, for any
nontrivial predicate Pk relative to the Hamiltonians over M, whether
that expressionsatisfies (or does not satisfy) Pk in T.
Moreover, there is no problem in the arithmetic hierarchy that can
be made equivalent to the (algorithmetically unsolvable) decision pro-
ceduresfor Cm-
Sketch of the proof. We proceed in a stepwise manner. Let £1,
... ,be an infinite countable sequence of mutually independent Godel
sentences in T. Let m\, ... -
, be the corresponding Godel numbers.
Form the 0{ = 6(mi). Put: e = 1 9j,c) = 9j, all j G uj.
Let mrn be a variable that ranges over all 2n binary sequences of
length n. Code those by ordered n-tuples of O's and l's and establisha
map / between those n-bit binary sequences andall n-factor products
... '
£f e2 £"" so that in the 7-th position (0)1 ■-» ej(e|). Given m,rn,
the associated product is /(rn). Given a specific model for T, for a
prescribed length, all such sequences equal 0but for a single one, that
equals 1.
Order the predicates Pi, P2, ;we write that T h P»(C2x-i) while
Now list all finite binary sequences and select from those an infi-
nite set of mutually incompatible sequences such as 1, 01, 001, 0001,
... ..
Note the incompatible set {t\,t2, .}; if Tj is any sequence, note
Tj>,Tjii, ., its extensions. (A general extension is noted r^.)
The expression we require is:

C = Ci[(l/2)/(ri) + (1/2)2 £ */(rr ) +...] +

+ C2[(l/2)/(r2) + (1/2)2 ]T 7(r2' + ..J +
(s2* denotes sumover all extensions of equallengthin the factors c.) ■

Remark 3.42. Notice that, as each 9^ k \m) is either 0 or 1, C will equal

a single one among the Q, despite the fact that itis an expression with a
countably infinite recursively defined number of symbols. Also, since
the actual value of C can only be determined if we solve an intractable
problem in the arithmetic hierarchy, and since it contains (possibly
repeated) representative expressions ofall properties in T, we will only
be able to check for any property if we first solve a complete problem
of degree > o'.

Moreover, since every arithmetical statement has a finite, bound-

ed degree in the hierarchy, nontrivial properties of C cannot be made
equivalent to anything in the hierarchy.
Finally, when we (informally, butin an actual computation) 'open
up' the expression for C, weobtain an infinite formal sum whose terms
are products of iterated integrals (the expressions for the 9^); such
expressions are certainly uncommon in classical mechanics, but they
are the standard staple of quantum and particle physicists. So, nothing
out of the mainstream here. ■

SCHOLIUM 3.43. The lowest degree in (m may be arbitrarily high.

Proof. For any p>o wecan substitute 9p+k for 9k in the expression
for Cm-

COROLLARY 3.44. The decision problem within an arithmetically

consistentTfor Cm cannotbe made equivalent to a decision problem in
the arithmetic hierarchy.
Proof. From Remark 3.42.

Remark 3.45. Rogers gives the proofby a diagonal argument (Rogers,

1967, Section 14.8, Theorem XIII) of the following assertion: if the
axiomatization for ZFC is consistent then there is a sentence in ZFC
that cannotbe made equivalent to an arithmetic sentence with the tools
of set theory. As in most proofs by diagonalization the counterexample
obtainedis a legitimate assertioninZFCbut itas meaningless as Godel's
original undecidable sentencein arithmetic. However, out of the previ-
ousresult (Proposition3.41) wecan immediately obtain a nonarithmetic
sentence in set theory. ■

Sincethe set of theorems of T (supposed [arithmetically] consistent)

is a creative set (Rogers, 1967), its complement is productive. There-
fore, we can add the axiom 'There isno solution foxpo{x\ , ,xn) — 0
over thenaturalnumbers' to anew polynomial p\ (x\ , ,xr ) such that
the Diophantine equation p\ = 0 has no solution in M, but such that
again we cannot prove that fact from T. Again, due to productiveness
we obtain a third polynomial, p2,whichleads to a thirdundecidable sen-
tence and to another extended theory T" We thus generatea sequence
of unsolvable Diophantine equations po —o,p\ =0, , pj =0,...
... , which lead to undecidable sentences in the theories T(°) =T,
j. ..
, rp(i)
j. ...
That construction is equivalent to obtaining a recursively enumer-
able set out of a productive set (Rogers, 1967); no such construction
willexhaust the productive set,and severalalternative recursively enu-
merable subsets of the productive set can be found.
We now use those pj as follows: if A is the map defined following
Proposition 3.7, we form the sequence ki(x) = Xpi, iG uj,and obtain:

PROPOSITION 3.46. IfT is arithmetically consistent, thenM |= /% =

0, all i, where the j3i are given by:

Pi = v(Gi),


However, T^ V Pi = 0, all k < i.

Proof. See (da Costa, 1991a).

PROPOSITION3.47. IfT is arithmetically consistent, then for every
recursively enumerable extension of the axioms ofT there is a Hamil-
tonian C all whose nontrivial properties cannot be proved within that
Proof. The expression we require is:
C = Ci[(l/2)/(tI) + (1/2) £*/(t,.) + ---] +

+ C2[(l/2)/(r2) + (l/2)2 £7(T2' + ...] +

(J2* denotes sum over all extensions of equal length in the factors c.)
Here the #'s in the r's are substituted for /3's. ■

These exampleslead to the immediate proofofa generalundecidabil-

ity and incompleteness theorem that deals with those systems for which

no nontrivial property can be algorithmetically decided or proved. Let

Q{x, ai,a 2 , " ,an) be a Suppes predicate on the fixed parameters a\,
Suppose given an enumeration of the predicates Pk in T. Again
we suppose that:
(2) For &,&,** j, cLT,TY Pfc(&) and T h -Pfc fe).
(3) Out of that we list all nontrivial predicates Pk that apply to Q-
defined objects.
Out of that we obtain:

PROPOSITION3.48. IfT is arithmetically consistent, we can obtain

a recursive enumeration ofpairs of expressions 2i, * a natural
number, that represent different objects andsuch that:
(1) Both T h Q(i2i ) andTY Q(&i+i).
(2) T h P*fe) «^T h -Pifei+i),
where the Pi are nontrivial predicates (relative to Q) in T that range
over Q-objects. ■


PROPOSITION3.49. IfTis arithmetically consistent then:

Undecidability. There is a countable family (m of expressionsfor
Q-objects in T such that there is no generalalgorithm to decide,
for any nontrivial Q-property Pk in T whether that expression
satisfies (or does not satisfy) Pk.
Incompleteness. There is a Q-object for which none of the non-
trivial Q-properties can be provedwithin T. ■
So, incompleteness of the nastiest kindis to be expectedeverywhere
in the axiomatized sciences. For density theorems related to those
intractable problems see (da Costa, 1993a).


We discusshere two problems that havebeen recently handled with the

help of the present methods. Arnolds 1974 questions on polynomial
dynamical systems and the incompleteness of the theory of finite Nash
games. Full details will appear in references (da Costa, 1992d, 1992e).

PolynomialDynamicalSystems Are Undecidable andIncomplete

In the 1974 AMS Symposium on the Hilbert Problems, V. I.Arnold
suggested the following questions for the list of 'Problems of Present
Day Mathematics' that was compiled at that symposium:
Is the stability problemfor stationarypointsalgorithmically decid-
able? The well-known Lyapunov theorem solves the problem in
the absence of eigenvalues with zero real parts. In more compli-
cated cases, where the stability depends on higher order terms in
the Taylor series, there existsno algebraic criterion.

Let a vector field be given bypolynomials of a fixed degree, with

rationalcoefficients. Does analgorithm existallowing us to decide
whether the stationary point is stable?

A similar problem: Doesthereexistan algorithm to decide whether

a planepolynomial vector field has a limit cycle?
See(Arnold, 1976a). We notice that Arnoldsays nothing aboutbound-
ary values,despite the fact that the system's behavior may vary wildly
as a function of the initial vales; he also imposes no conditions on the
geometry of the manifolds where those dynamical systems are to be
fond. Therefore we may offer as a counterexample a specific poly-
nomial dynamical system over Z with fixed initial values such that no
algorithm in the sense of Arnold exists for the requiredproperties.
LetPbeanontrivial predicate (relative to the adequate vector fields)
in T. We assert:

PROPOSITION4.1. If Tis arithmetically consistent, then:
(1) There is an expressionfor a vectorfield v over Rn in T such that
M\='v is-> a smooth vectorfield on Rn, n> 2 with property P ',
and T V 'v is a smooth vectorfield on Rn with property P
(2) There is an expression for a vector field v in T such that the
sentence 'v is a smooth vector field on Rn, n> 2 with property
P'is T-arithmetically expressible as a nn+i sentence, m>\,but
such thatfor no m it willbe equivalent to a Tim sentence.

(3) There is an expression for a vector field v in T such that the'

sentence 'v is a smooth vectorfield on Rn,n> 2 with property P
cannot be taken to be T-demonstrably equivalent to any sentence
in the arithmetical hierarchy.
(4) We can explicitly find an expression for a polynomial vector field
s over7. on Rm,mfixed, such that, for a given P,M \= P(s), but
TV P(s)andTY^P(s).
HereMis amodel where T is arithmetically consistent.
Proof. The first three assertions are immediately proved with the
methods presented in (da Costa, 1991a) and expanded in the previous
section of this paper. Actually they generalize our previous results on
dynamical systems (da Costa, 1991a).
The last assertion is the negative solution (in an obvious sense) to a
problem related to Arnolds problem; for details on Arnolds see (da
Costa, 1992e). We make two remarks:
" We write x = (xi,...,Xj), where j = dim x.
" We notice that if the polynomialp(x) is not identically zero, then
given the expressionbelow for a smooth elementary function,

(1) «(x, y) =(j+ 1)V(x) + £ 0&4


if rG R, then for s(x,y) = u(x,y) —r, R27 s _1 (0,0) is open
- .
in R27 ,where vis taken as areal-valued function on R2j (See on
v aboveDefinition 3.2.)
The first lemma we require is:

LEMMA 4.2. IfTis arithmetically consistent, then:

(2) tt(x, y) =(j + l)4b2(x) +J2 yM(*)l

together with

(3) yi
— sin7TXi, wi = cos7TXj,
is the unique solution of the following polynomial dynamical system
with coefficients in Z over R3j+l endowed with the usual Cartesian
coordinates and the correspondingEuclidean metric tensor:
dxi dv
dt dxi
for i=l,2, 3,...,j,plus the boundary condition u(0,0) =p(0) (with
an obvious abuse oflanguage);
dyi = idu

(5) W Z '
Id dx~-
dwi i dv
dt dxi
whereagain i= 1, ...,j, withboundary conditions 7/i(0) =0, Wi(0) —
1. Andfinally,
" t=°
with boundary condition z(0) = 7r.
Proof. Equations (4) are immediately integrated (since they are
gradientequations) to Equation (2). Equation(4.2) trivially implies that
z = 7r, and therefore (also from Equations (4)) we have:
—dyi = irwi,
dwi = -*yi,
(no sumon i). That systemhas the solutions yi = sin7TXi,Wi = cos7TXf.
Since the du/dxi are polynomials over Z whichare not identically zero,
the lemma is proved. ■
We need a secondlemma: here we allow v to range over R — {I}.
Therefore for each xi andfor each yi at most a countably infinitenumber
of points are to be deletedfrom thecorrespondingdomains, those points

that are solutions of the (fixed) equation v 1 = 0 over the reals. We

then state:

LEMMA 4.3. IfTisarithmetically consistent, then

(8) v = l-v
is the unique solution to the polynomial equations on R — {I},

where c=—l, v G (— oo, 1) and c=o, v G (1, +oo); and

Tt= c *
dTj dv

together with the conditions v(0) =1

— ti(0), v G (— oo, 1), and v= 0
/oru G (I,+co).
Proof. Immediate.

We conclude the proof with still another lemma: let S, T be poly-

nomial vector fields over Z on a A;-dimensional real smooth manifold
M such that T h P(S) and T h + vT). Let us form the vector
= =
fields S (x, 0) and T (0,y). Then:

LEMMA 4.4. IfTis arithmetically consistent, then thereis an expres-

sion for the polynomial vectorfield over Z of the form:
(11) $= £ + 7jT
such thatM \= $ = E, but T V $ / E. (AgainM is a model where T
is arithmetically consistent.)
Proof. It suffices to take the polynomial p(x) asin Propositions 3.17
or 3.19, and get v as in Equations (2) and (3), and finally obtain v out
ofit. ■

This lemma concludes our proof.

Let M'be the manifold where the expression for $ is defined. Then
SCHOLIUM 4.5. // Tis arithmetically consistent, then T h 'The
dimensions ofMl,m<2k + 3(j + 1) '.

SCHOLIUM 4.6. IfT is arithmetically consistent, then M |= The
dimension ofMl,m = k', while
TV 'The dimension ofM',m = k\
TV -> 'Thedimension ofM',m = k '.

So we might end up dealing with a planar vector field despite the

fact that T will not recognize it!

COROLLARY 4.7. IfT is arithmetically consistent, then it has an

expression(in M)for aplanarpolynomial autonomousvectorfield with
coefficients in Z such that we cannotprove (from the axioms ofT) that
thatfield is planar. ■

With the help of Proposition 3.37 we can state:

COROLLARY 4.8. If T is arithmetically consistent, then it has in

M an expression for a planar polynomial autonomousvector field x
with coefficients in Z such that the assertion 'x is planar' cannot be
made demonstrably equivalent within Tto any arithmeticalassertion. ■

Other consequencesof the previous results are easily stated andproved.

Remark 4.9. Our example ofa polynomial vector field withundecidable

properties has either a high apparent dimension or an apparently very
high degree. If we take k = 2 and the original Diophantine universal
polynomial p with 11variables, m= 46 and the degree is ridiculously
high. For degree 16 we will have that m = 250. A discussion on
the possibility of reducing those values can be found in (da Costa,
1992e). ■

Incompleteness of FiniteNash Games

The results given here are presented in full detail in (da Costa, 1992d).
It is intuitively obvious that, for a finite Nash game, we can algorith-
mically check for equilibria in it. However, things turnout to be much
more delicate when we work within a formalization for the theory of
finite Nash games. We start from:

DEFINITION 4.10. A noncooperativegame is given by the yon Neu-

mann triple T = (N,Si,Ui), with i= 1,2,...,N, where N is the
number of players, Si is the strategy set of player % and Ui is the real-
valued utility function Ui : Yli Si -+ R. ■


PROPOSITION4.11. IfTis arithmetically consistent, then there is a

noncooperativegame T where each strategyset Si is finite but such that
we cannot compute its Nash equilibria.
Proof. Let V and T" be two different games with the same number
of players but with different strategy sets and different equilibria. If
P = /3(w+l ) in Corollary 3.36, wecan form the utility functions:
fJUi .
Therefore the game T = (N,Si,Ui) does not have a decidable set of
equilibria. ■
SCHOLIUM 4.12. Determining the equilibrium set ofT is a nonarith-
metic problem. ■
We cantake T barely beyond formalized arithmetic N: let N* be the
theory whoseaxioms are N plus a compatible version of the separation
axiom. Therefore we can give an 'implicit' definition for sets of num-
bers. We put Ui : Y\iSi —+ uj. We now use q as in Proposition 3.21
and form the utility functions Ui = u\ + qu" The remainder of the
argumentgoes as usual. Thus:
PROPOSITION4.13. 7/N* is arithmetically consistent,
' then there is a
game f in N* such thatthe assertion 'u{ = u\ is undecidable m N*. ■


In 1987 the authors started a research program whose goals were

twofold: we wished to axiomatize as much as possible of physics in
order to search for physically (i.e. empirically) meaningful undecidable
sentences and physically meaningful unsolvable problems within our
formalizations. The axiomatization program was fulfilled in part with
our construction of Suppes predicates for classical physics out of an
unified framework that reminds one of ideas and tools from category
theory (da Costa, 1992a). The incompleteness portion of the program
was certainly much more difficult to pursue: we had two candidates
for undecidable problems in physics,and we hoped thatthose problems
might lead us to thedesired incompletenessproof. Thecandidates were
classification schemes in generalrelativity andHirsch's problem on the
existence of algorithms to check for chaos in dynamical systems. (See
the references in (da Costa, 1990b, 1991a).
At first we believed that forcing models might provide the indepen-
dence proofs we were looking for (da Costa, 1990b, 1992b), since forc-
ing is an obvious source of so many mathematically relevant undecid-
able statements in ZFC. Therefore we explored our Suppes-formalized
version of generalrelativity and the corresponding sets of noncompact
space-times, but the results on incompleteness were meager at best.
Then, due to a suggestion by Suppes, we turned to Richardson's 1968
undecidability results in analysis. It was immediately apparent to us
that those results entailed a full-fledged incompleteness theorem, which
weextended, after some failed efforts, to the explicit construction of the
halting function for Turing machines within classical elementary real
Wethen started to turnoutseveral undecidability andincompleteness
resultsin dynamical systems andrelated questions,as wellas counterex-
amples to the existenceof algorithms of the kind that Hirsch was asking
for in chaos theory. Again Suppes pointed out to us that there was
some sort of a very generalundecidability and incompleteness theorem
at work here; the first version of that theoremis Proposition 3.29 in (da
Costa, 1991a).

That theorem is so general (it reminds one of Rice's theorem in

computer science (Machtey, 1979; Rogers, 1967), from which it can
be derived) that for a moment we felt it might announce some kind
of imminent disaster, as every nontrivial property in the language of
classicalanalysis canbemadeintoan undecidability andincompleteness
theorem. So, in generalnothing interesting can be decided and very
little can be proved. (This point was also emphasizedby Stewart in his
commenton our results (Stewart, 1991).) At the same time we received
notice that our results were at first met with disbelief by researchers in
dynamical systems, precisely due to that very generalincompleteness
statement;it was then argued thatthoseresults were correct but 'strange'
due to some undetected conceptual flaw in the current view on the
foundations ofmathematics.
Yet our assertions have a very clear, letus say, practical meaning:
if what we mean by 'proof in mathematics is algorithmic proof, that
is to say, something that can be simulated by a Turing machine, then
very difficult problems are to be expectedeverywhereat the very heart
of everyday mathematical activity; more and more innocent-looking
questions are to be found intractable (as we now see in chaos theory)
and (to repeat an example that we haveoffered before) systems willbe
formulated that have a tangled, chaotic appearance when approximated
ona computer screen,but such that no proof of their chaotic properties
will beoffered within reasonable axiomatic systems such as ZFC.
Godel incompleteness isnooutlandishphenomenon; itisanessential
part of the way we conceive mathematics.
What can we make out of that? We do not think that there is any
essential flaw in thepresent-day view aboutfoundational concepts;how-
ever we think that our incompleteness theorems point out very clearly
where the problem lies. Nothing will be gained by adding 'stronger'
and 'stronger' axioms to our current axiomatizations. But we certainly
must strengthen our current concept of mathematical proof. Turing-
computable proofs are not enough, for Church-Turing computation is
not enough. We must look beyond.


It certainly is a pleasure to dedicate this work to Pat Suppes on the

occasion of his 70th anniversary. We deem that dedication to be espe-
cially adequate sinceour workon the undecidability and incompleteness
of axiomatized theories has been inspired from its very beginnings by
many fruitful intuitions and suggestions of his.
We also wish to acknowledge the constant cooperation andinterest
of our coworkers: J.A. de Barros, A. F.Furtado do Amaral, D. Krause,
andM.Tsuji. The present ideas werealsodiscussed withM. Corrada,R.
Chuaqui, and D. Mundici,to whom we owe friendly remarks andcriti-
cisms. A first presentation of our incompleteness results was carefully
read by M. Hirsch,to whom weare indebted for severalcorrections and
The present paper was completed while the second author held a
visiting professorship at the University of Sao Paulo's Institute for
Advanced Studies under aFAPESP scholarship program. The authors
also acknowledge support from the CNPq(Brazil) scholarship program
in philosophy and wish to thank the ResearchCenter on Mathematical
Theories of Communication in Rio de Janeiro (CETMC-UFRJ) for the
use of its computer facilities.

N.C. A. da Costa,
Research Group on Logic and Foundations,
Institute for Advanced Studies,
University of Sao Paulo,
05655-010 SaoPaulo SP, Brazil
F.A. Doria,
Research Center on Mathematical Theories of Communication,
School of Communications,
Federal University at Rio de Janeiro,
22295-900 Rio RJ, Brazil


Arnold, V I.: 1976a, 'Problems of Present Day Mathematics', XVII (Dynamical

Systems and Differential Equations), Proc. Symp. Pure Math., 28, 59.
Arnold, V I.: 1976b, Les Methodes Mathematiques de la Mecanique Classique,
Moscow: Mir.
Atiyah,M. F: 1979, Geometry of Yang-Mills Fields,Pisa: LezioneFermiane.
Bourbaki, N: 1957, Theoriedcs Ensembles, Paris:Hermann.
Bourbaki, N.: 1968, Theory of Sets, Boston: Hermann and Addison-Wesley.
Dalla Chiara,M. L. andToraldodi Francia, G.: 1981, Le TeorieFisiche, Boringhieri.
Cho, Y.M.: 1975, J. Math. Phys., 18, 2029-2035.
da Costa, N. C. A. and Chuaqui, R.: 1988,Erkenntnis, 29, 95-112.

da Costa, N. C. A. and French, S.: 1990a, 'The Model-Theoretic Approach in the

Philosophy of Science', Philosophy ofScience, 57, 248-265.
da Costa, N. C. A., Doria, F. A., andde Barros, J. A.: 1990b, Int. J. Theor. Phys., 29,
da Costa, N. C. A. and Doria,F. A.: 1991a, Int. J. Theor. Phys., 30, 1041-1073.
da Costa, N. C. A. and Doria, F. A.: 1991b, Found. Phys. Utters, 4, 363-373.
da Costa,N. C. A. and Doria,F A.: 1992a, 'Suppes Predicatesfor Classical Physics',
to appearin: A. Ibarra et al. (Eds.), The Space ofMathematics, De Gruyter.
da Costa, N. C. A. and Doria, F. A.: 1992b, 'Structures, Suppes Predicates and
Boolean-ValuedModels in Physics', to appear in: P. Bystrov and J. Hintikka
(Eds.), Festschrift inHonor ofV. I. Smirnov onHis 60th Birthday.
da Costa, N. C. A. and Doria,F. A.: 1992c, 'Jaskowski'sLogic', preprint CETMC-10.
da Costa, N. C. A., Doria, F. A., and Tsuji, M.: 1992d, 'The Incompleteness of Finite
NoncooperativeGames withNash Equilibria',preprint CETMC-17.
da Costa, N. C. A., Doria, F A., Baeta Segundo, J. A., and Krause, D.: 1992e, 'On
Arnolds 1974 Hilbert Symposium Problems',inpreparation.
da Costa, N. C. A., Doria, F. A., and Furtado do Amaral, A. F: 1993a, Int. J. Theor.
Phys. 32,2187-2206.
da Costa,N. C. A., Doria,F. A.,Furtadodo Amaral, A. F, andde Barros, J. A.: 1994a,
'Two Questions on the Geometry of Gauge Fields',Found. Phys., to appear.
da Costa,N. C. A. andDoria, F A.: to appear, Metamathematicsof Physics.
Davis, M.: 1982, 'Hilbert's Tenth Problem Is Unsolvable', in: Computability and
Unsoh'ability, 2ndedition, Dover.
Dell,J. and Smolin, J.: 1979, Commun. Math. Phys., 65, 197-212
Doria, F. A.: 1975, Let.Nuovo Cim.,14, 480^182.
Doria, F. A.: 1977, J.Math. Phys., 18, 564-571.
Doria, FA.: 1981, Commun.Math. Phys., 79, 435-456.
Doria, F A., Abrahao, S. M., andFurtadodo Amaral, A. F: 1986, Progr. Theor. Phys.,
75, 1440-1446.
Jones, J. P.: 1982, J. Symbol. Logic, 47, 549-561.
Kobayashi,S.andNomizu, X.: 1963/67, Foundationsof Differential Geometry, Vols.I
and11, New York: Wiley.
Krantz, D. H., Luce, R. D., Suppes, P., and Tverski, A.: 1971, The Foundations of
Measurement, Vol. I,New York: AcademicPress.
Kreisel,G.: 1976, 'A Notion of Mechanistic Theory',in: P. Suppes (Ed.), Logic and
Probability in QuantumMechanics, Dordrecht: D. Reidel.
Lewis, A. A.: 1991, 'On Turing Degrees of Walrasian Models and a GeneralImpossi-
bility Result in the Theory of Decision Making', preprint, University ofCalifornia
at Irvine, Schoolof Social Sciences.
Lewis, A. A. and Inagaki, V: 1991, 'On the Effective Content of Theories', preprint
University of California at Irvine, Schoolof Social Sciences.
Machtey, M.and Young, P.: 1979, An Introduction to theGeneral Theory of Algorithms,
Amsterdam: North-Holland.
Narens, L. andLuce,R. D.: 1986, Psych. Bull, 99, 166-170.
Richardson, D.: 1968, J. Symbol. Logic, 33, 514-520.
Rogers, H., Jr.: 1967, Theory of Recursive Functions and Effective Computability,
Scarpellini,B.: 1963, Zeitschriftfur Math. Logik v. Grundl. derMath.,9, 265-289.
Stegmuller, W.: 1970, Theorie undErfahrung,Vol. I, Berlin: Springer.
Stegmiiller, W.: 1973, Theorie undErfahrung, Vol. 11, Berlin: Springer.
Stegmuller,W.: 1979, The Structuralist View of Theories: A Possible Analogue of the
BourbakiProgramme in Physical Sciences,Berlin: Springer.
Sternberg, S.: 1964, Lectures onDifferential Geometry, Prentice-Hall.
Stewart, I.: 1991, Nature, 352, 664-665.
Suppes, P.: 1967, Set-TheoreticalStructures in Science, mimeo., Stanford University.
Suppes, P.: 1988, Scientific Structures andTheirRepresentation,preliminary version,
Stanford University.
Taubes, C. H.: 1987, J. Diff. Geometry,25, 363^112.


For over half a century since Godel's famous results, the incompleteness
of any axiomatization of classical analysis has been known, but in the
past several decades other results have stressed how far we are from
having anything like a real algorithmic approach to any significant
class of mathematical problems. One example is the realization that
Tarski's decision procedure for elementary algebra is not feasible, in
the technical sense that the computations grow exponentially in the
length n of any formula whose truth is to be decided. So even positive
results on decision procedures themselves do not guarantee practical
applicability. Itused tobe thought ofas partof the folklore,but ofcourse
not in any sense proved, that 'most' problems in mechanics that were
not too complicated to formulate wouldhave relatively straightforward
solutions. Therecent discoveryofchaotic systems inallsorts of domains
has shownhow the problems that fill the textbooks of mechanics are a
carefully selected group. That such difficulties were lurking about has
really beenknown since theintensivework on the three-body problem in
the nineteenth century and the culminating negative results ofPoincare.
One way to put it is that any undergraduate in physics can derive the
differential equations governing familiar cases of the restricted three-
body problem- therestricted problem is when the massof the third body
is negligible and therefore does not influence the regular motion of the
other two bodies. But the problem of finding mathematical solutions of
the differential equations is whollyunmanageable for most cases.
Da Costa and Doria have embarked on a program, as indicated by
the many additional references in their paper, to show how widespread
the presence of unsolvable problems is in physics andother sciences.

That difficulties in alldirections can be found with alittle effort was

anticipated in some sense by Richardson's 1968 results on undecidable
problems involving elementary functions of a real variable, but what
Newton andChico have doneis extend thiskindofresult much further.
It has probably been the feeling of most nonlogicians working in these
areas that results on incompleteness or undecidability were really not
going to stand in the way of standard mathematical progress. These
foundational results were quirky ones on the edges and could not have
anything to do with anything lying in the heartland of analysis and
mathematical physics. Newton and Chico are certainly in the process
of showing how thisis not the case.
I have learned a lot from both Newton and Chico over the years,
especially through their livelyparticipation in my seminars atStanford. I
especially benefittedfrom long conversations withChicoDoria acouple
of yearsago whenhe spenta year atStanford. Infact those conversations
about thefoundations of physics,and recent discussions with Acacio de
Barros, his former student from Brazil, have completely rekindled my
interest in working on the foundations of quantum mechanics.
The kind of results that Newton and Chico have obtained havealso
been a motivation for the work that Rolando Chuaqui and I have been
doing on constructive foundations of infinitesimal analysis, especially
aimed at physics. Ourobjective hasbeen to give afree-variable positive
logic formulation of infinitesimal analysis which is finitarily consistent
and which is still strong enough to prove, in somewhat modified form,
many of the standard theorems that underlie constructive methods in
theoretical physics. Itis a feature of theoretical physics that it is mainly
elaborate computations, either of a symbolic or numerical kind from a
mathematical standpoint. Little is ever done in theoretical physics as
suchabout provingexistence theorems. Here, of course,I am drawing a
distinctionbetween theoretical andmathematical physics, a distinction
thatis now relatively wellestablished. Of course, necessarily such weak
systems as Rolando and I have been working on cannot do everything
that one wants. We do think that muchcanbe done anditis important to
understand where the boundary exists. The important methodological
point is that by using infinitesimals, wecan give a free-variable formu-
lation of such standard theorems ofcalculus as themean value theorem
or Green's theorem. Ishould mention that these theorems are proved
in approximate form, that is, equality is replaced by an equivalence
relation that means there is a difference that is infinitesimal.

I do not mean to suggest that the system that Rolando and I have
been working on gets around all the problems uncoveredby Newton and
Chico. What it does show is that weak constructive systems, demon-
strably consistent,are sufficient for a great deal of the work.


Chuaqui,Rolando and Suppes,Patrick: to appear, 'Free-Variable AxiomaticFounda-

tionsofInfinitesimalAnalysis: A Fragment withFinitary Consistency Proof.
Suppes, Patrick andChuaqui, Rolando: inpress, 'A FinitarilyConsistentFree-Variable
PositiveFragment of InfinitesimalAnalysis',Proceedingsofthe 9thLatinAmerican
Symposium onMathematicalLogic,held at Bahia Blanca,Argentina, August 1992.


ABSTRACT. This paper sketches an account of scientific explanation based on a

semantic or model theoretic conception of scientific theories derivativefromthe work
of Suppesandhis collaborators[8, 10, 14] and aBayesianaccount ofscientific reasoning
suggested by Rosenkrantz [9], HowsonandUrbach [7] and Earman [3] amongothers.
The model theoretic conception of scientific theories has been described in detail by
Sneed [12], and Balzer, Moulines and Sneed [I]. A somewhat similar view the
semantic view has been developed by vanFraassen [15], Suppe [13] andothers. The
discussion in this paper employs the conceptual apparatusdevelopedin [I]. However,
most of whatis said can beapplied to amore general 'semantic' conceptionoftheories.
The applicationofthis conceptionofscientific theories toscientific explanationhasbeen
developedfromsomewhatdifferentperspectives byForge [4,5] andSintonen [11]. This
paper generalizes (and somewhatsimplifies)the workof Forge, indicates how it applies
to so-called 'functional explanations', and connects it with the Bayesian account of
scientific reasoning to provide an account of howcompetingscientific explanationsare
(should be) evaluated.


On the model theoretic view, the simplest kind of scientific theory

(theory element in the vocabulary of [1]) is an ordered pair T (K,I)

consisting of a conceptual core X and a range of intended applications
I. Very roughly, X defines a set of possible situations or ways things
could be which we call
The theory isused to say something -make a claim -about itsintended
applications I. The claim is simply that I is one of those situations
characterized by K. Formally, this claim is just that
(1) c Content^).
Note that formally Content(if) is a set of sets (of something). So
that it makes sense to say that a set / is a member of Content^).
Roughly, we should think of Ias the totality of potential data the theory

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 195-216.
© 1994 Kluwer Academic Publishers. Printed in theNetherlands.

is supposed to account for. Individual members of / are pieces of that

data individual applications of the theory. Content(if) characterizes
entire configurations of data not just single individual applications.
In cases of genuine scientific interest, it appears that I will be an
intentionally described, 'open-ended' class - for example the set of all
particle collisions - the set of all state-level societies. In these cases,
though we may have goodreasons tobelieveI to be finite, we willnever
be certain that we haveexamined all its members. Thus,however sure
we may beof the truthof (1), there exists the possibility of discovering
additional members of / which would cause us to revise our epistemic
attitude toward (1). Indeed,I think one might persuasively argue that a
necessary condition on 'real' science is that / be 'open-ended' in this
It is perhaps worth noting that (1) may be recast into the form of a
'general statement':

(1') then X is a Content^).

For all X, if X is an I
Inthis form theclaimof a theory looks very muchlikethe 'general laws'
that play a major role in traditional accounts of scientific explanation.
Roughly, whatmakes claims like (1), (1') 'interesting' is that they point
out some feature that all members of Iboth those we have examined
and those we have not allegedly have. The major differences between
the account of explanation I willoffer and the traditional account is that
my account makes explicit use of the set-theoretic properties of things
that are (in) / and things that are (in) Content(if). The 'features' that
members of / allegedly have are 'structural' features most naturally
described by the language of set theory.
Again, in cases of genuine scientific interest, Content(i^) has the
formal property that, whenever V is in Content(i^), so is every non-
void subset of Y. Intuitively, this means that inductive evidence for
(1) may be provided by examining subsets of / and discovering them
to be in Content(i^). Indeed, one, overly simple picture of 'theory
confirmation' is that we just examine more andmore members of /.If
the set of those we have examinedcontinues to be in Content(if ), we
become more andmore confident that (1) is true.
Itis convenient to view this simple picture of theory confirmation as
a 'research program' carried out over time. Thus, at any time t during
the course of this program there will be some subset Ft of / that is
'believed with good reason' (by members of the scientific community
pursuing the research program) to be in Content(if). We may call Ft
the set of 'firm' applications of Tat time t Firm applications that are
in fact in Content(if) we may call 'successful' applications. On the
customary account of 'knowledge', theusers of Tknow that successful
applications arein Content(if). (Theseideas are elaborated indetail in
[I], Ch. V.) One may provide an account of explanationbasedeither on
'firm' or 'successful' applications. The difference is just in the epistemic
strength of the notion of explanation. To make matters simple, I will
opt for 'successful' and an epistemically strongnotion of explanation.
So far, very little has been said about the explicitly model theoretic
nature of this conception of scientific theories. This becomes apparent
only when we characterize X in more detail. Intuitively, X consists of
a vocabulary or conceptual apparatus characteristic of the theory (Mp)
from which a non-theoretical (Mpp ) may be distinguished, some laws
(M) formulated in the full vocabulary and some constraints (C) which
limit the ways in which theoretical concepts may be applied across
different applications of the theory. Thus Xis an ordered 4-tuple:
(2) X = (Mp,Mpp,M,C).
What makes this conception 'model theoretic' is that each of the ele-
ments of X may be viewed as a class ofmodels set theoretic structures.
These model classes may sometimes be characterized by sentences in
first-order logic. But, for theories of genuine scientific interest, char-
acterizing them in the language of set theory by defining a set theoret-
ic predicate is almost always easier and sometimes the only practical
means. The insight that classes of models characterized in this way
could serve to illuminate parts of interestingempirical science is due to
Suppes and his collaborators [8, 10, 14].
These model classes hang together in the following way. Non-
theoretical structures members of Mpp -are obtainable frommembers
of Mp simply by lopping off their theoretical components. Formally,
there isa 'forgetful functor' that maps Mp ontoMpp Mis just a subset
of Mp It represents scientific laws in that this subset is acharacteristic
way in which the 'values' of the components of single members of
Mp are related. C is a subset of the power set of Mp (Pot(Mpp )). It
represents characteristic waysin which the values of the components in
different members of Mp arerelated.
Content(iiT) consists of all the sets N of non-theoretical structures
(subsets of Mpp) that canbe filled out with theoretical components in a

way so that: (i) each individualmember ofAf is filled out to a memberof

M(satisfies the laws);and (ii) the set of theoretical structures produced
by filling out every member of N is a member of C (satisfies the
The set of intended applications / is a set of non-theoretical struc-
tures described in the manner sketched above. Thus the claims of the
theory (1) amounts to saying that / (and every subset of/) can be filled
out with theoretical components in some way that satisfies these condi-
tions. That this rendition of the empirical claimof theories like classical
particle mechanics,classical equilibrium thermodynamics andothers is
plausible has been arguedin detailin [12, I]. These arguments will not
be repeated here.
Somewhat more complicated kinds of scientific theories can be rep-
resentedas 'nets' of theories of the sort describedhere (theory elements)
and an appropriate concept or 'content' can be defined for these nets.
Sintonen [11] hasexploited thismore generalconceptof scientific theory
to explicate some contextual, pragmatic aspects of the concept of sci-
entific explanation. This macroscopic, net-level account of explanation
effectively presupposes a microscopic, theory-element-level account
which is not explicitly provided. This paper complements the work of
Sintonen by providing the microscopic account it presupposes.


These ideas may be applied to scientific explanation in the following

way. Veryroughly, the occurrence ofXis explainedby showingit tobe a
specific part of somethingelse thatregularly occursin a well-understood
way. For example, the occurrence of long-neck, brown bottles in the
recycling bin in front of my house is explainedby showing it to be a
part of a regularlyoccurring pattern of beverageconsumption andrefuse
disposal characteristic of my household.
A bit more precisely, to explain some phenomenon X is to show
that X is a specific aspect or part of some largerphenomenon sharing
some interesting features with (generally occurring in the same ways
as) a wider class of phenomena. To recall Hempel's classic example,
we explain the bursting of my pipes lastDecember 28 bynoting that my
pipes are part of a thermodynamic system and thermodynamic systems
behave in certain general ways described in detail by a physical theory
like classical equilibrium thermodynamics. We explain the entry of
the 8-ball into the corner pocket by noting the context in which this
event occurred - a collision sequence involving the 8-ball, the cue-
ball and a third ball. Further, this sequence of collisions has some
general properties characterized by classical collision mechanics. We
explain the presence of turquoise fragments on part of the surface of
a grooved stone by noting features of the context in which the stone
was found. It was in a stone enclosure on a river bank about a mile
from a turquoise outcropping. We conclude the stone was part of a
certain kind of 'technological system' - perhaps a 'mineral extractive
system. Further, weare able tosay something in generalabout 'mineral
extractive systems'. That is, we have a 'theory' about them.
Though these rough 'explanation sketches' do not make it explicit, it
is worth noting the somewhat obviousfact that the intuitive plausibility
of these explanations depends on identifying the phenomenon to be
explained with a specific kind of part of the larger system. Though it
is true, it is not enough to say the turquoise fragments were just any
part of a mineral extractive system. We must specifically identify them
as debris resulting from the use of a specific kind of tool used to break
turquoisebearing rock.
This rough, intuitive gloss of 'explanation' does not sound too differ-
ent (Ihope) from more traditional accounts. Surely, it should not. My
intention is to capture the same intuitiveidea as thetraditional accounts.
The difference between my account and the traditional accounts lies in
the way these initial intuitions are made precise. Note too that these
examples are prima facie mute on the traditionally controversial issue
of whether (a description of) the explained phenomenon is logically
entailed by the explanation.
What all these exampleshavein commonis thata phenomenon Xis
brought within the scope of some plausible theory. The epistemic value
of explanationis that itallows us to view X as an aspect of 'the familiar'
- things we think we understand. To begin to make this account more
precise, we might say roughly this. To explain some phenomenon X is
to show that X'is' (may be viewed as) a part of some known,successful
applications of a theory that can be taken seriously. On this account,
providing an explanation for X can be analyzed into three parts:
(A) Showing X is a specific part of some A;
(B) Showing A are successful applications of some theory T;
(C) Showing T canbe taken seriously.

Let us postpone saying what it means to say a theory can be taken

seriously. Iwill return to this aspect of explanation in Sections 4and5
below. Rather, I
will focus on (A) and (B) explaining what it means
to say that X is brought within the scope of theory T.
The phenomenon X is explainedby bringing it within the scope of
theory T = (X,I). This simply means that we identify X as a specific
part of some intended applications A of T that are in fact successful
applications of T. Thus we explain why the 8-ball went in the corner
pocket by describing a collision of which its path to the corner pocket
was a part and noting that this collision is a successful application of
classical collision mechanics i.e. it conserves momentum without
requiring us to assign masses to the billiard balls that are incompatible
with data about other collisions in which they have been observed. In
this example X is shown to be a part of a single successful application.
This is perhaps the typical case. However, the suggested formulation
also allows for the possibility that X is shown to be a part of a set of
successful applications. For example, X might be data about the role
- -
of a single particle the cue ball in several different collisions.
Note that 'part' here is to be understood to include 'improper part.
Thus, showing that X is (all of) some successful intended application
counts as explaining X. In particular, showing that X is an intended
application that can be added to the set of unsuccessful applications (at
time t) withoutthe resulting set falling outside of Content(X) counts as
explaining X. This appears to be the case considered by Sintonen [11]
for theory nets. This kind of explanation is 'progress' of the research
program associated with the theory. But thereare more mundane kinds
of explanation -perhaps those involving technological applications of
the theory that are not cases of 'progress' in any interesting way. The
account offered here includes both cases.
Very roughly wehave something like this:

(3) T = (X,I) explains X iff X is a specific part of a set of some

successfulintended applications of T.
Intuitively, a 'part of an application is just something that canbe filled
out to a (full) application. For example, it might be a specification of
the velocities of a subset of the particles involved in acollision; it might
be the specification of the velocities of all the particles before, but not
after, the collision. We need to talk about parts of sets of structures and
this may be done in roughly the following way.
(4) X is part of A iffeach member of X is part of exactly one member
of A and each member of A contains exactly onemember of X.
Formally, a 'part of an individual set-theoretic structure is just a
substructure. A set-theoretic structure haslots of different substructures
- different parts of the structure. An explanation of X must specify
which specific substructure is identified with X. For present purpose,
I will use the notation 'part-i' to denote a specific kind of substructure
of intended applications of T. In doing this, lam ignoring certain
technical questions about how substructures of intended applications
are specified. Thus, a slightly more precise version of (3) is:
(3') T = {X,I) explains X iff there are some isuch that X is part-i of
a set of some successful applications of T.
Wenow need to be a littlemore explicit about the pragmatic aspects
of explanation. Successful applications of a theory (X,I) at time t are
simply intended applications that are known by the users of T toat time
t to be in Content(i^). Thus:
(5) A is a set of successful applications of theory T = (X,I) at time
(i) A CI;
(ii) A 6 Content^);
(iii) at time t users of T have good reason to believe that A c
Thus, we have the following accountor definition of 'explanation':
(6) (X, /, A) is an explanation for X at time t iff:
(i) (X,I)is a theory element;
(ii) A CI;
(iii) A 6 Content^);
(iv) at time t users of T have good reason to believe that A €
Content (X).
(v) Xis part-i of A.
This accountof explanationis not free fromknown problems. These
turn on the concept of 'part' that is appropriate here. Roughly, some
'parts' of structures seem to be inappropriate candidates for explanation
according to this model. For example, one might take a 'part' of a
classical collision mechanical structure to be simply a subset of the set
of particles appearingin the structure. It seems counterintuitive to say

that we 'explain' this set of particles by embedding it in a model for

classical collision mechanics. The obvious solution to this problem is
to define 'part' in sucha way that subsets ofnon-basic sets in structures
must appear in 'parts' of these structures.
Another problem appears in connection with theories (like classical
collision mechanics) whichhave aconcept of 'temporal order. Suppose
we 'explain' a given structure byidentifying it as that part of a classical
collisionmechanical structure consisting of (some of) the particles plus
their initial velocities. One might contend that doing this explains
nothing about 'why the particles have these velocities'. In contrast,
identifying parts of the same given structure with the final velocities
of some other models of collision mechanics does seem to provide an
intuitively satisfactory explanation.
Roughly, the same data may be explained by classical collision
mechanics in several ways. Some of these ways appear to be intu-
itively satisfactory. Some do not. The source of this difference clearly
lies in our intuitions about temporal order. In the intuitively unsatis-
factory explanations we appear to be explaining the given structure by
showing 'what happens next' rather than 'what happenedbefore. We
might avoid this by placing additional restrictions on the temporal order
of 'parts' appearing in explanations. Those who think that 'causes' are
an essential feature of explanation would endorse this move. Thought
this might be a feasible move for theories with a concept of temporal
order,itis hard to see what (ifanything) might correspond to this for the-
ories (like equilibrium thermodynamics) without a concept of temporal
Perhaps the accountof explanation offered here should be viewed as
characterizing arather generalframework for explanation whichcan, in
some cases,be augmented withfurther conditions to characterize 'causal
explanations'. However, it should be noted that this augmentation
will apparently usually involve adding something that is not naturally
regarded as a part of the conceptual apparatus of the theory providing
the explanation. That is, 'causality' even when we can make sense of
it is usually not going to appear explicitly in the conceptual apparatus
that is capturedby the set-theoretic axiomatization.
Having said what Ican about known difficulties, Iwill explore
implications of this account of explanation in two directions. First, I
willconsider thetraditional question about the symmetry of explanation
and prediction in relation to the question of functional explanations.
Second, Iwill consider how comparative evaluations of explanations
are made.


Traditional accounts of explanatioirhave tended to view an explanation

as a kind of argument. Rough, the conclusion of the argument is that
which is explained while the premises are providedby the theory that
provides the explanation, together with some 'collateral information'
or 'initial conditions'. How can we compare the present account of
explanation with this 'deductive model' of explanation?
The most obvious approach to comparison is trying to convert the
various pieces of a model theoretic explanation into linguistic entities
that at least couldbe pieces of a deductive argument. For thetheoretical
partof the explanation,the obvious candidate is simply the claim of the
theory T = (X,I) that

(1) c Content^).
It is also reasonably clear that A the set of intended applications
used in the model theoretic explanation plays a role' corresponding
roughly to that of the collateral information in the deductive model.
A bit more precisely, what remains of the set-theoretic structures in A
- —
when the explanandum Xis deleted this A{ }X' is the collateral

information. More intuitively, A is the total context within which we

locate X for purposes of bringing T to bear on it. A{—}X is that part
of the context which was not initially given, but rather added (perhaps
discovered) in the process of providing the explanation for X.
Thus, to provide additional premises for a deductive argument we

need to assert that the collateral information A{ }X is,in fact, part of
an intended application of T. This we may do with two propositions:
(7) AC I
(8) A{-}XispartofA.
With (8) we simply claim that the additional structure A{—}X which
we have added to X is a part of the total context. This is, of course,

trivially true. Non-trivialis the claim (15) that the total context A is an
intended application of T.
What is now the appropriate conclusion to this argument? I suggest
that we would like to be able to conclude from (1), (7), and (8) that:
(9) X is part of A.
That is, we would like to conclude from the truth of the theory's claim;
the fact that the theory applies to the total context A and the fact that
the added context A{—}X is a part of A that X is a part of A.
Thus,I suggest that themost natural way to convert amodel theoretic
explanationinto a deductive argument yields the following argument:

(1) / <E Content^). [T is true.]

(7) ACL [T applies to A.]

(8) [collateralinformation]

(9) X is partof A [explanandum]

The problem with this argument form is thatit is not generally valid.
Whether it is valid depends on the nature of the theory T used in the
explanation and upon the nature of the application A of the theory.
The reason is roughly this. For applications A in Content(K), parts
of A having the structure characteristic of yl{-}X will not always be
associated with unique parts of A having the structure characteristic
of the part X. Thus simply knowing how the part is
instantiated does not generally tellushow theX-like part is instantiated.
Tomake this clearer, consider the special casein which X contains
no theoretical concepts and trivialconstraints. That is,

Mpp = Mp; C = Pot(Mp).

Inthis case,

Content^) = Pot(M)

and nothing islost by simply forgetting about multiple applications and


Content(if) = M.
Suppose that theexplanandum £is a substructure of some mp inMp .
That is, xis obtained from mp bydeleting someparts of mp Intuitively,
we may think of systematically going through all of Mp removing from
eachmember parts isomorphic to those we removed from mp to obtain
x. Call the set of structures we obtain in this way '5X [MP]\ Providing
a precise definition of '5X[MP]' is somewhat tedious so I omit it here.
Intuitively, SX [MP] is the set of all x-type data structures. The structure
x is just one 'value' or 'instantiation' of this datatype.
To explain x is simply to add the remaining structure required to
make it an a in Mp which is also a model for T, i.e. a in M. This
remaining structure is a{-}x. As before, we may considerSa{_} x [Mp]
- intuitively the set of all a{ — }x-type data structures.
Now, starting with

a{-}x in Sa{ _ }x [Mp]

we may generally add the remaining x-type data to produce a model
for T in a variety of ways. That is the emendations of a{ —}x in M
form a 'blob' rather than a point (see Figure 1). In turn, this blob of
emendations to modelsfor T projects downinto Sx [Mp] as ablob,rather
thana point. In this blob are all the possible ways of filling out a{ —}x
with x-type data that yields amodel for T.

Thus,knowing only that a{ }x is a part ofa model for T, the most
wecan infer about the value of the x-type part of this model is that it
is somewhere in the blob projected into Sx [Mp] In some special cases,
this blob will be a singleton. But generally it will not be. Whether or

not there is a unique way of making a{ }x into a model for T depends

on the specific nature of T and a{ }x.
Thus structural explanations are, in general, what Hempel called
'partial explanations' ([6], pp. 16-17). They allow us to deduce some
features of the explanandum. But, they do not, in general, allow us to
deduce everything we know about it - not even everything we know
about it that is expressible in the vocabulary of the explaining theo-
ry. What the structural account of explanation allows us to see (which
is, at least, not readily apparent in linguistic accounts) is that partial

Fig. 1.

explanations occur in the 'hard' sciences as well as the 'soft. Phys-

ical theory offers partial explanations that are, in their logical form,
no different than the functional explanations appearing in biology and
cultural anthropology. The only difference is this. In some cases, at
least, physical theories can provide 'complete' explanations. It might
be the case that the explaining theories used in sciences like biology
and anthropology are simply so weak that the functional explanations
they provide can neverbe more than partial explanations. Indeed, one
accountof functional explanation suggests that this is true [16]. Iftheo-
ries supporting functional explanations are 'optimization models' with
multiple local optima, then it is intuitively clear that the explanations
will always be partial.

The obvious way to apply a Bayesian accountof scientific reasoning to a

model theoretic conception of scientific theory is to consider conditional
probabilities of the following form:

(10) P(I c Content^) |E)

where Eis the 'total evidence' available. E might simply be a conjunc-
tion of things like

Ii c Content(if)
not (Ij c Content(K))

P(li C I) ± 0.
But there is no reason at this point to rule out otherkinds ofevidence.


It is common practice among empirical scientists (andothers) to debate

the relative merits of different explanations offered for the same phe-
nomenon. Do we best explain an apparent rise in temperature in a
palladium cell by the presenceof 'cold fusion' or by 'faulty instrumen-
tation. Dowebest explain the construction of large masonry structures,
roads, etc. centered around Chaco Canyon, NMin the 10th— 11th Cen-
tury AD by some variant of 'indigenous development' or some variant
of 'foreign influence. At least a half dozen alternative explanations
for this phenomenonare evaluated comparatively in a recent discussion
([l l],p.391-402). Supposing that the alternative explanations in these
examples couldbe mashed into the format suggestedabove, would this
illuminate the discussion of the alternatives?
Given that we have two distinct putative explanations for X, how
should we compare them? Intuitively, two considerations appear to
be relevant: (1) how good is the theory involved in the explanation?;
(2) how probable isit thatthe theory involved really does explain XI On

a Bayesian account of scientific reasoning this suggests that something

(11) P(Ic Content^) & (X,J, A) is an explanation for X |E)
could be taken as a 'figure of merit' for explanations. That is, alterna-
tive explanations of X, (X, /, A) and {Xl,Kl,I',A'), should be evaluated
according to the value they give for (11). In fact,it appears that in many
cases of interest this expression can be considerably simplified.
First, it appears that in many cases of interest, the conjuncts in (11)
will be probabilistic independent so that (11) reduces to:
(12) P(I G Content^) |E) x P
({X,I,A) is an explanation for X \ E).
Clearly this independence assumption will not generally be true. In
some cases that (X, /, A) explains X may provide evidence that / €
Content(if ), that is
P(IE Content(X) {X,I,A) explains X & E)
> P(I c Content(X) |E).
Intuitively, this is because
AC Iand A c Content^)
provide evidence that

i" € Content^).
However, in the case of theories with lots of evidence of this sort, the
contribution of the evidence from specific putative explanationmay be
negligible. So theproduct form may be acceptable as an approximation.
This would most likely be the case when the theories were wellestab-
lished and supportedby substantial amounts of evidence apart from the
explanatory contextin question.
The product form (12) has the nice intuitive feature of clearly sepa-
ratingjudgements about themerits of the theory invokedin explanation
from judgements about whether the theory explains the phenomenonin
question. Focusing on thelatter judgement,some additional simplifica-
tion is possible. Considering(6) again, one might most naturally view
(6-i), (6— ii), and (6— iii) as purely formal conditions so the probability
of each (and the conjunction) will always be either 1 or 0. We could
complicate things here by considering non-trivial probabilities for the
formal conditions, as for example when we are uncertain about our
mathematical or calculational results. But this complication would add
little to the present discussion. Thus in the case the formal conditions
are satisfied (P(6-i) & 6— ii) & 6— iii)) = 1), the probability that we have
an explanation at all is just the probability that the empirical condition
(6-iv) is met. Thus, the probability that (X,/, A) explains X is:

(13) P(Xpartof,4 \E).

Thus far wehave avoidedexplicit discussionofthe formal properties
of the explanandum X. In some cases X may be (described as) a
set-theoretic structure in a fragment of the vocabulary of the theory
T = (X,I) used in the explanation. In this case, whether X is a part
of A may be a purely formal question. In the more general case X
will be described insome vocabulary disjoint from that of T and it will
remain a 'contingent' question whether this description of X has been
appropriately 'translated' into the vocabulary of T.
Under these simplifying assumptions, the goodness of the theory is
measured by the probability (given the available evidence E) of the
empirical claims of the theory, i.e.

P(I <E Content^) |E).

The probability that the theoryexplains X is:
P(Xpartof A),
and the 'figure of merit' for explanations is simply the product of these
probabilities. Thus:

(14) If (X, I,A) and (X', /', A') are putative explanations for X then:

(X, /, A) is better than (K\ I',A')

P(I £ Content^) | E) x P(X part of A \ E)
> P(l' G Content^') | E) x P(X part of A' \ E)

Thus far we have assumed that the theories involved in alternative

explanations are totally different. That is, both their vocabulary and
laws are different. This is not always so. In some cases of interest, the
alternative theories will share all or part of their vocabulary. The most

interesting of these cases are thosein which the theories in question are
'competing' theories about the same data -thatis,when Mpp Mpp and
1= I. In this case, the conditional probability of the theories' claims
(10) do not always provide a good indication of the relative merit of
the theories. The reason is roughly this: relatively weak theories may
have claims with high probabilities. But, intuitively, stronger theories
whoseclaims havelower probabilities might be preferable. How should
theories be compared in these cases? Andonce theories are compared,
how should explanations based on them be compared?
Fromamodel theoretic point of view,the contentof theories provides
an indication of their relative strength. Clearly,
(15) T = {X,I) is stronger than T' = (X',I)

Content^) C Content^')
whether the 'if in (15) should be 'if and only if is not entirely clear.
More explicitly, it is not evident how we should compare the strength
of theories with identical classes of non-theoretical structures whose
contents intersect partly or fail to intersect at all. For present purposes,
it appears that little is lost by simply taking (15) to be necessary,as well
as a sufficient, condition. We may, somewhat arbitrarily, say T and T'
are equally strong (equipotent) in allother cases. Thus:
(16-i) T = (X,I) is stronger than T' = (X',I)

Content^) C Content^')
(16— ii) Tis equipotent with T
T is stronger than T' and T' is not stronger than T.
Having a way of comparing the strength of theories allows us to address
the question of how probability trades off against strength in evaluating
theories. Isuggest that the trade off is made lexicographically with
probability dominating strength. That is,strictly more probable theories
are alwayspreferred to strictly less probable. Only in thecase of equally
probable theories does strength play role. Then, stronger theories are
preferred to weaker. Thus:
(17) T = (X,I) is better than T' = (X',I)
P(I G Content^) |E) > P(I G Content(A") | E)

P(I G Content^) | E) = P(I G Content^') |E)

T is stronger than T'
This seems plausible for the following reason. Numerous weak the-
ories with high probability are easy to construct. But most such the-
ories receive short-lived attention because it is also relatively easy to
strengthen some of the theories without substantially diminishing their
probability. The theories we consider seriously are those which retain
their probability under strengthening. Having a 'better than' ordering
for theories allows us to readdress the question of a 'better than' ordering
for explanations. Here the question is this. How does the probability
that the theory provides an explanation trade off against the quality of
the theory in evaluating the quality of the explanation? Again, I suggest
a lexicographic criterion with the quality of the theory dominating.
(18) If (X, /, A) and {K\ I,A'} are explanations for X then:

{X, I,A) better than (X',I,A')

{X,I) strictly better than (K1,1)

(K,I) equivalent (K\l)


P(Xpa.rtofA) > P(X part of A').
This 'better than' ordering for explanations clearly emphasizes thequal-
ity of the theory rather than the probability that the theory explains. Is
this plausible? The main argument for its plausibility is this. The pur-
pose of explainingXis to show that Xcanbe integrated intoour existing
body of knowledge - without substantially revising that body. Explana-
tion is reducing theprima facie unfamiliar to the familiar. Among other
things, this makes us more comfortable with the already -
familiar. We
feel comfortable that existing knowledge is adequate weneed not take
the trouble to learn something new. But, some things are more familiar
thanothers. Integrating Xinto theframework ofa moreprobable theory
is just 'more satisfying.

Department ofHumanities andSocial Sciences,

Colorado School ofMines,
Golden, CO 80401, U.S.A.


Iam indebted to Mr. Michael Pierce for calling my attention to these problems and
discussing them with me.


1. Balzer, W., Moulines, C. U., andSneed, J. D.: 1987, An Architectonicfor Science:

The StructuralistProgram, Dordrecht: D.Reidel.
2. Balzer, W., Moulines, C. U., and Sneed, J. D.: 1986, 'The Structure of Empirical
Science Localand Global', in:R. B.Marcus, G. J. W. Dorn, andP. Weingartner
(Eds.), Amsterdam: North-Holland,pp. 291-306.
3. Earman, J.: 1992, Bayes or Bust, Cambridge, MA: BradfordBooks, to appear.
4. Forge, J.: 1980, 'The Structure of PhysicalExplanation',Philosophy of Science,
47, 203-226.
5. Forge, J.: 1985, 'TheoreticalExplanationin Physical Science',Erkenntnis, 23,
6. Hempel,C. G.: 1962, 'Explanationin Science and History',in: R. G. Colodny
(Ed.), Frontiers of Science and Philosophy, Pittsburgh: University of Pittsburgh
Press, pp. 9-33.
7. Howson,C. andUrbach,P.: 1989, Scientific Reasoning: The BayesianApproach,
La Salle: Open Court.
8. McKinsey,J. C.C, Sugar, A. C,and Suppes,P.C: 1953, 'AxiomaticFoundations
of ClassicalParticleMechanics', Journalof RationalMechanicsandAnalysis,2,
9. Rosenkrantz, R. D.: 1977', Inference, MethodandDecision: Towarda Bayesian
Philosophy ofScience, Dordrecht, D. Reidel.
10. Rubin, H. and Suppes, P. C: 1954, 'Transformations of a System of Relativistic
ParticleMechanics',Pacific Journal ofMathematics,4, 563-601.
11. Sintonen, M.: 1989, 'Explanation:In Searchofthe Rationale',in: W. Salmonand
P. Kitcher (Eds.), Scientific Explanation:Minnesotastudies, Vol. 13, Minneapolis:
University of Minnesota Press.
12. Sneed,J. D.: 1979, The LogicalStructure of MathematicalPhysics, (2nd edition),
Dordrecht: D. Reidel.
13. Suppe,F: 1989, The Semantic Conception of Theories and Scientific Realism,
Urbana: University of Illinois Press.
14. Suppes, P. C: 1959, Introductionto Logic, New York: VanNostrand.
15. vanFraassen,B. C: 1980, The Scientific Image, Oxford: ClarendonPress.
16. vanParijs,P.: 1981, EvolutionaryExplanationin theSocialSciences: AnEmerg-
ing Paradigm, Totowa: RowmanandLittlefield.
17. Vivian,R. G: 1990, The Chacoan Prehistory of theSanJuan Basin, New York:
Academic Press.


I am sympathetic to Joe Sneed's general structural viewpoint as well as

his Bayesian approach. The analysis of structure, the characterization
of scientific structures, and the analysis ofthe natureof the structures, as
wellas a strong Bayesian bent, havebeen long-standing features of my
own work in the philosophy of science. On the other hand,as the years
goby, I find myself increasingly skeptical of very generalphilosophical
accounts of explanation or causality or even of the viability of very
generalBayesian methods. My view of sciencehas moved increasingly
from that of a foundationalist to the viewpoint that the conceptual con-
tent of science is best analyzed in terms of a diverse set of methods for
solving a wide variety of problems. Schemes of great generality about
scientific explanation, such as those of Sneed, can perhaps be useful
in providing some kind of general framework, but I think they miss
the main part of what we intuitively want from good scientific expla-
nations. Moreover, our sense of satisfaction with good explanations
is particularistic in nature and not in any natural way subsumed under
the general schemes proposedby Joe. Above all,I am skeptical that a

general notion of part, as he uses it, can do the work he would like to
have it do.
For those raised in traditional philosophy or even traditional phi-
losophy of science, with the search for generality and universality of
conceptual schemes so dominant, it is not easy to accept or even be
sympathetic with a view that is skeptical of the success of any of the
general schemes aimed atproviding a traditional philosophy of science.
The picture of scientific activity Iincreasingly favor myself is clos-
er to that of apprenticeship than to the propositional organization of
knowledge. Perhaps only in mathematics do we have the generality of
structure and generality ofresult that would satisfy the hearts as well as
theminds of philosophers. Even physics, the most sophisticated of the
empirical sciences, has only a pretense at great generality. Individual
problems must be tackled by individual methods. The methods that
areused to attack a particular problem dependupon the experienceand
insight of the investigator, not upon the sharp and exact codification
of theory and its range of application. It is only a myth engendered
by philosophers - even in the past to some extent by myself - that
the deductive organization of physics in nice set-theoretical form is an
achievable goal. A look at the chaos in the current literature in any
part of physics is enough to quickly dispel that illusion. This does not
mean that set-theoretical work cannot be done, it is just that its severe
limitations must be recognized.
There are many ways of expanding upon the views Ihave just
expressed. A central one is the current realization of how few even
simple physical systemscan be thoroughly understood,in the sense that
detailed and precise predictions about the behavior of the system can
be successfully verified. In a way, this is a lesson that was ready for
understanding already in the nineteenth century in terms just alone of
the massive andunsuccessfuleffort to master the three-bodyproblem in
classical mechanics. But the modern developments of chaos havemade
the facts much more salient and very much increased the awareness of
how difficult it is to make successful predictions. But once the rarity of
systems or structures whose behavior can be successfully predicted is
recognized then much of the older talk about general schemes of expla-
nation seems unsatisfactory for dealing with the rich details of actual
science. This is not the place for a full-scale exposition of my views,
but let me try to give one or two examples.
Iwant to contrast our attitude toward various cases of failure of
prediction in classical mechanics. Let us begin with a case already
mentioned, that of the three-bodyproblem. Our inability to solve the
differential equations, even in principle, in a satisfactory analytical
form has been recognized for over 100 years. On the other hand, the
derivation of the differential equations governing the motion of three
bodies acted uponby the force of gravitation alone is an easy exercise.
There is a very general belief that the differential equations accurately
reflect many real situations to an extraordinarily highdegreeof accuracy.
There is no need to say 'with complete accuracy' for no real systems
are sufficiently isolated, just to mention one central reason. What is
important is thebelief that the equations hold to a very high degree of
accuracy. But they are unmanageable from the standpoint of solutions.
A closely related example is that of a single body sliding or rolling
down an inclined surface of variable degrees of roughness. We can
write a differential equation including friction that also, because of the
complicated nature of the surface, we cannot solve except numerically.
But in this case the status of the equation is quite different. We do not
think that it is possible to write down, to the same degree of precision
at all,a differential equation governing the complicated physical phe-
nomena that reflect theinteraction between the surface and the moving
body. Because our feelings of definiteness about these two cases seem
so strong, a theory of structural explanation should provide a very com-
pelling account of their difference. But to provide this requires entering
into the details in a way that is not a feature of the current work.
Let us continue the same line of examples. A favorite one of many,
also used by Sneed, is the way in which we accept the explanation of
the motion of billiard balls on a table based upon the mechanical laws
of collision. However, if we consider a somewhat more complicated
billiard ball - 1have in mind thekind studied by Ornstein et al, Sinai,
and others we enter an entirely new realm. The most striking theo-
rems are that when the obstacles onour new and wonderfully different
billiard table are convex in their shape then we cannot distinguish no
matter how many observations we take between the motion of the
billiard balls following the usual laws of mechanical collision,and the
motion of theball being described by a probabilistic Markov process. It
is one of the great insights of modern mechanics, whose philosophical
importance I have tried to stress in several publications over the years,
that the separation between deterministic mechanical systems and ran-

dom probabilistic ones is not at all what it was once thought to be.
Structuralist views of this kind of example are as yet missing from the
The examplesjust cited are in a way too mathematical in character.
The analysis ofa physicalproblem we hope to solve, think wecan solve
and are willing to tackle,is more open-ended,less honed down to a few
sets of variables than thekinds of examples I have just given. Here the
apprenticeship of the physicist seems to me of the greatestimportance.
Toknow what to countand what to discard as unimportant in analyzing
the givenphysical situation is not something wherein no endof training
inmathematics andin the solution of differential equations willbe much
of help. As wemove from applications of theory to anew situation and
the organization of experiments, what I have to say is even more true.
Moreover,in complicatedhigh-technology experiments involvinglarge
numbers ofindividuals itis fair to say that noone commands any longer
all the details of the experiments. It is not just apprenticeship, but
a collection of mature apprenticeships, that are required to organize,
prepare and execute the experiments planned. Individual papers with
more than 100 authors are now not uncommon in high-energy physics,
yet the obviously social natureof these experimentsandthe complicated
sets of skillsrequired toexecutethemhaveas yetreceivedlittle attention
in the analysis of physics by philosophers of science. In making these
remarks about the collective enterprise of runningexperiments in high-
energyphysics, I am not interested as such in the sociology of science
but more in the bewildering variety of patches of theory and patches
of skills required to put the whole thing together. Itis that it would be
wonderful to have a detailed structural analysis of. Perhaps Sneed and
his energetic collaborators canbe persuaded to enlightenus allon some
of these detailed matters.




The type of theory of measurementwediscussis the representational one

that is treated in great detail in the three volume treatise Foundations of
Measurement.l Weattempt tomake the present treatment self-contained,
including definitions of major concepts, but these volumes may be
consulted for greater detail.
Our aim is to discuss two classes of, probably difficult, issues that
needclarification. The first type, covered in Section 3, has to do with
principled arguments to justify the representationalperspective andpos-
siblemodifications of it to deal with error and with limiting properties
of passing from finite, through denumerable, to continuum represen-
tations. These issues are, at least partially, philosophical in character.
The second type, covered in Sections 5 and 6, is more technical and
focuses primarily on the issues of relating the useful, but abstract, ideas
(to be defined below) of scale type, Archimedeaness, and Dedekind
completeness to observable properties of qualitative structures.
Many of the problems we describe are primarily conceptual andtheir
resolution probablyrequires somethingof an intellectual break-through.
Once that is achieved,however, some may well prove to be technically
easy to solve. Others that we cite are well defined within existing
mathematical ideas, and some of these appear to be difficult.


2.1 Historical Background
Mathematicians and scientists, among them Helmholtz (1887), Holder
(1901) and yon Neumann (in yon Neumann and Morgenstern, 1947),

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 219-249.
© 1994 Kluwer Academic Publishers. Printed in theNetherlands.

clearly had the idea of axiomatizing ordered qualitative structures as

possible models of empirical attributes and in the latter two cases of
establishing, as mathematical theorems, theexistenceand uniquenessof
numerical representations. The power and generality of this approach,
however, was not widelyappreciatedby empirical scientists or philoso-
phers of science during the first half of the 20th century, as evidenced
by the discussions of measurement in, for example, Bridgman (1922,
1931), Campbell (1920, 1928), Cohen and Nagel (1934), and Ellis
(1966). More than anyone else, Suppes brought to the attention of
non-mathematicians this axiomatic style of studying the measurement
ofattributes. Particularly important wereScottand Suppes (1958),Sup-
pes (1951), and Suppes and Zinnes (1963). The latter was especially
influential among mathematical psychologists, andit was the forerunner
of the Foundations ofMeasurement.
For over 30 years, representation and uniqueness theorems2 were the
mode of research. New ordered mathematical structures that appeared
to haverelevance to the measurementofcertain attributes were isolated,
and two theorems were established: the existence of a representation
into or onto some prescribed numerical structure, and the uniqueness
of that representation in the sense of formulating the class of trans-
formations relating equally good representations into or onto the same
numerical structure. This has come to be called the Representational
Theory of Measurement, abbreviatedRTM.
Manykey concepts andmethodsof RTM, asitis now formulated,had
analogues in 19thcentury geometry,particularly in (i) the assignment of
coordinate systems to qualitative geometries(e.g.,Hilbert, 1899), (ii) the
idea that invariants of structure preserving mappings had geometrical
significance(Klein, 1893), and (iii) the idea that the geometric structure
of space is the same at each point, i.e., that space is homogeneous
(Helmholtz, 1868;Lie, 1886).

2.2. The Underlying Principles

The formal part of RTMcan be reduced to five main ideas:
1 A qualitative situation is specifiedby a (usually ordered)relational
structure X consisting ofa domain X,finitely many relations on
X andfinitely many special elements of X.

These relations,subsets,andelements are called the primitives of X.

Measurement axiomsare then stated in terms of primitives of X. These
axioms areintended to be true statements about X for some empirical
identification andareintended to capture important empirical properties
of X, usually ones that prove useful in constructing measurements of
its domain, X.
2. The representationaltheory requires that the primitives of X can
be given an empirical identification.
In particular, if R is an n-ary primitive relation on X, then it is
required that the truth or falsity of R(x\, ...,xn), for any particular
choice of the n-tuples X{, be empirically decidable.
3. As much as possible, measurement axioms, statedin terms of the
empirically identifiedprimitives, shouldbe empirically testable.
Next,a numerically-based,representing structure Mis selected, and
for each structure X that satisfies themeasurementaxioms,the setS(X)
of homomorphisms of X into M isconsidered.
4. Measurement of X is said to takeplace ifand only ifthe following
two theorems canbe shown:

(i) (Existence Theorem). <S( X) is non-empty for each X that

satisfies the measurementaxioms.
(ii) (Uniqueness Theorem). An explicit description is provided
about how the elements of S(X) relate to one another. In
practice this descriptionusually consists ofspecifying a group
offunctions G such thatfor each <f> in S{X)
S{X) = {9*<t> :^isinC},

where * denotesfunction composition.

A concept of meaningfulness similar to that proposed by Stevens
(1946) is used for judging the empirical significance of quantitative
statements: an n-ary relation Son the domain of the numerically based
structure J\f is said to be meaningful about X if and only if for all
iri, ,xn andall </> and ip in S{X),
S[(f)(xi),...,(f)(xn)] iff S[if>(xi),...,il>(x n)].
FollowingSuppes and Zinnes (1963)

5. RTMidentifies empirical significance with meaningfulness.

The concept of meaningfulness with respect to S(X) is easily extend-
ed to numerical statements involving measurements of elements of the
domain: meaningful statements are those whose truth value is unaffect-
ed by the particular representationinS(X) used to measure X.
It should be noted that the concepts of empirical significance and
meaningfulness as embodied in RTM are independent of truth; both
true and false statements canbe meaningful andempirically significant.
Although the following two features are not part of the formal the-
ory, they hold in many important measurement situations with infinite
6. Using the measurement axioms, sequences of equal spaced ele-
- -
ments standard sequences are constructed.

These standard sequences are then used to establish constructively

the set of homomorphisms of X into Af. Toconstruct a specific one, one
... ...
proceeds as follows: a first standard sequence s\, ,s\, is chosen
to produce the first level of approximation 0 of a homomorphism (j).
l ,where <
For each element xin X, find the n such that s\ <x < s n+l
denotes the primitive of X that orders X. Set
<j>i(x) = ra/1.
A secondlevelof approximation iscarried out bychoosing the sequence
5\,..., 52,..,s 2,...,where for each positive integer i,s\{ = s}. The second
approximation to </> is obtained by setting
(f)2 (x) = k/2,

wherekis such that s\ <x < s|+1 Continuing in this way,a sequence
of approximations (f)m is constructed, andfor many measurementaxiom
systems it canbe shown that as m— > 00, the limit of </>m exists and is
the desired homomorphism 0 of X into Af.
Beginning with a different standard sequence will usually yield a
different homomorphism.
The accuracy of measurement, inpractice, iscontrolled by how many
terms in this approximation are used.
The ancient geometerEudoxesused standard sequencesin much the
same way as in RTM.They also played a critical role in the approaches
taken by Helmholtz (1887) and Holder (1901). Later Luce and Tukey
(1964) used related kinds of standard sequences to obtain their repre-
sentation and uniqueness results for additive conjoint measurement.3
Krantz (1964) and, in a moreuseful way, Holman (1971) recognized
that the standard sequence approach of Luce and Tukey closely related
to the algebraic structure describedby Holder for physical measurement
-an Archimedean,4 totally ordered group. Basically, Krantz and Hol-
man defined an operation on one component that captured completely
the trade-off structure between the components. Under the axioms, the
operation was shown to be associative and so the problem wasreduced
to an application of Holder's theory.
Mostof FMI is devoted to recastingvarious measurementsituations,
often using the ideas inherent in Krantz andHolman's approach, so that
Archimedean ordered groups (or large semigroup portions of them)
come into play. It has been remarked, with some justice, that the
theory of measurement is largely an application of Holder's theorem,
i.e.,of standard sequences in the guise of Archimedean ordered groups.
Moreover, as we shall see in Section 4.2, this continues to be the case
even for non-additive structures.
7. The representational theory offers an abstract theory of the kinds
of well-behaved scales that one encountersin science.
This theory is based on a concept of homogeneity that abstractly is
much like the geometric one mentioned above and provides qualita-
tive and empirical criteria for empirical structures to be homogeneous.
(Issues involving homogeneity are discussed more fully in Sections 4
and 5.)
Until the 1980s, the major emphasis of RTM has been on produc-
ing existence and uniqueness theorems for various kinds of empirical
situations. Scott and Suppes (1958) remarked:
A primary aim of measurement is to provide a means of convenient computation.
Practical control or prediction of empirical phenomenarequires that unified, widely
applicablemethods of analyzingthe importantrelationshipsbetweenthe phenomenabe
developed.Imbedding the discoveredrelationsin variousnumerical relational systems
is the most important such unifying methodthathas yet beenfound (pp. 116-117).


Although mostcritics accept thatRTM, viaits axiomatizations of impor-

tant empirical situations and the various representation and uniqueness
theorems, is a major contribution to our understanding of measurement
in science, they criticize it as a theory of measurement. The criticisms
are of four major types: (1) questions about the definition of measure-
ment in terms of scales of homomorphisms, (2) objections to the lack of
a theory of error for RTM, (3) concerns about exactly what invariance
concepts have to do with the concept of meaningfulness, and (4) doubts
about the heavy use of infinite structures and continuum mathematics
when, after all, the universe is, according to contemporary physics,
composed of only finitely many particles. A number of these concerns
are expressed by Kyburg (1984) and in papers in Savage and Ehrlich

3.1. Homomorphism Definition ofMeasurement

The criticisms of the homomorphism definition of measurementare of
two types: those arguing that the definition is toonarrow in the sense that
it does not accountfor certainkinds of measurement; and those arguing
that itis toobroadin that itpermits toomany kinds of measurement. The
former usually contain elements of the following criticism of Adams
It seems to me that in characterizing measurement as the assignment of numbers to
objects according to rule, the proponentsof the representational theory have fastened
on something whichis undoubtedly of great importance in modernscience, but which
is not by any means an essential feature of measurement. What is important is that
the real numbers provide a very sophisticatedand convenient conceptual framework
whichcan be employed indescribing the results of making measurements: but, what
can be conveniently described with numbers can be less conveniently described in
other ways, and these alternative descriptions no less 'give the measure' of a thing
than do the numericaldescriptions. Note, too, that the ancientGreeks didnot have
our concepts of rational, muchless realnumbers, yet it seems absurd to say that they
could not measure because they did not assign numbers to objects. In sum, Iwould
say that the employment of numbers in describing the results of measurement is not
essentially differentfromtheir employmentinothernumerical descriptions,andthatthis
employment is neither a necessary nor a sufficient conditionfor making or describing
measurements (pp. 129-130).
Even granting the assumption that measurementnecessarilyinvolves assigningnum-
bers, it seems to me to be far from true thatinmaking these assignments it is always the
case that mathematical operations andrelations are made to correspond to or represent
empirical relationsand operations. The situationis worse with most ofthe widely
used measures in the behavioral sciences, like I.Q.s and aptitude test scores. It may
be claimed, of course, that these are not really measurements at all, but to justify this
[claim], some argument wouldhave to be given,unless the theory of representational
measurement is not to degenerateinto a mere definition (I.Q.s are not measurements
because they do not establishnumerical representations of empirical operations and
relations)(p. 130).

Niederee (1987, 1992) and others criticize RTMas being toobroad.

He notes that requiring a structure to be numerically representable into
structures based on the real numbers places little restriction on the
structure beyond cardinality and even that has not been justified. This
kind of restriction is too liberal because
... it does not involveany concept of measurement whatsoever. Indeed measurement
theorists wouldhardly be prepared to accept [it] as a sufficient
criterionfor a structure
to be called representablein terms of fundamental measurement. What seems to be
lacking hereis an analysis of whatit should mean that a 'number' expresses an ideal
valueof measurement (Niederee, 1992, p. 245).
The above criticisms of RTM suggest the following problem:

PROBLEM 1 Justify in a philosophicallyprincipledfashion RTM(or
a largepart of it) as a generaltheory of measurement without severely
restricting itspositive uses.5

A few comments are in order. First, theories of ability measure-

ment alluded to by Adams, although based on quantitative data, are not
in principle ruled out by RTM. Empirical structures can be based on
numerical information just as long as those numbers arise from empiri-
cal means - whichis the case for ability measurement. What is lacking
from the RTM perspective are axiomatic theories for the various kinds
of ability measurement. We do not see any in principle impediments to
the development of such axiomatic theories, although in practice none
has been devisedand it does not appear to be easy to do so.
Second, either directly or indirectly standard sequences are used to
establish scalesinalmostall6 of the major results ofFM.Thus, since the
process of measurement through standard sequencesis usually taken as
paradigmatic of 'measurement processes', almost all of the represen-
tational results of FM are valid not only from the RTM viewpoint but
from a number of different perspectives about what measurement is.

Third, some attempts at solving Problem 1 already exist, including

Michell (1990) and Niederee (1987, 1992). The present authors are not
yet persuaded that their attempts are useful beyond the additive cases.
In particular, as we shall seein Sections 4and 5, the greatestprogress to
date in understanding non-additive structures involves mapping certain
general types of structures into a particular subgroup of their automor-
phisms, showing that this subgroup is Archimedean ordered, and then
using Holder's theorem to map these into the multiplicative positive
real numbers. Because we do not know, in general, direct structural
descriptions of individual automorphisms, itis not obvious thatthe pro-
cedure proposed by Niederee for constructing a suitable representation
will capture the known results. Considerably more work is needed
to establish how these approaches lead to these general measurement
Fourth, much of FM is concerned with issues that are not limited
to measurement per se, but are closely related to measurement. The
most prominent of these is the testing of scientific theories through the
use of RTM. An example of this is subjective expectedutility (SEU)
theory, which holds that a subject's preference ordering over lotteries
is explainable as if the person is trying to maximize a numerical SEU,
whichis computedin theobvious wayusinga simultaneously construct-
ed subjective probability over the family ofuncertain eventsand a utility
over the set of consequences. One way to test this scientific theory is
to formulate it as a quantitative model, gather data, and test the model
using standard statistical methods to determine the degree to which the
model accounts for most of the variance in the data. This approach often
involves approximate construction of standard sequences. A different
way is to formulate the scientific theory as a qualitative, axiomatic
measurement model that through RTM is equivalent to the quantita-
tive SEU, and then test the axioms' qualitative theory through direct
empirical observations of samples of the stimuli.
Three additional examples of theories that are closely related to
measurement issues are the axioms for distributive triples, which can
be used to construct the space of physical quantities (see Section 6.1);
a variety of psychophysical models involving more than one sensory
attribute (one of which is described in Section 6.2); and a theory of
certainty equivalents to uncertain monetary gambles (Section 6.3).

3.2. ErrorinMeasurement Theory

The second class of criticisms concern error andhow it shouldbe treated
vis-a-vis measurement models. While 'error' considerations for mea-
surement give rise to many different kinds of problems, we cite three
treatments of error that are intimately connected with key concepts of
RTM. The first is based on the idea that probabilities underlie the obser-
vations that are made. Although some progress has been made on this
general approach(Falmagne, 1979, 1980; Falmagne and Iverson,1979;
Iversonand Falmagne, 1985;Michell, 1986), little effective use has yet
beenmade of it to test measurementaxiomatizations in detail.

PROBLEM 2. Specify aprobabilistic version of measurement theory

and the related statistical methods for evaluating whether or not a data
set supportsor refutes specific measurementaxioms.

One approach, taken by a number of people,is to assume the data base

consists of probabilities of binary or more complex choices, with the
major question being the conditions on these probabilities that corre-
spond to the existenceof an underlyingrandom variable representation
in which the largest value observed determines the choice. A survey of
such work is found in Chapter 17 of FM2. One of the most interesting
recent contributions is Heyer and Niederee (1992) who study when a
probabilistic structure, e.g., of choices,can be considered to arise from
a probability distribution over a family of conventional measurement
structures that all satisfy the same axioms.
From a fundamental measurementperspective, this approach is not
fully satisfactory because it assumes as primitive a numerical structure
of probabilities and thus places the description of randomness at a
numerical, rather than qualitative, level. One would like to formulate
the qualitative primitives soas simultaneously to capture at aqualitative
level both the structural and the random qualities of the situation.
The following is a reasonably concrete instance of what we have in
mind. Consider the typical extensivesituation whereboth judgments of
order can be made and entities can be concatenated to form new ones
exhibiting the same attribute. One would like axioms that lead to a
random variable representation that specifies the random variables with
respect both to their structuralrelations witheachother and to the nature
of their distributions. For example, if © denotes concatenation within

the extensive structure, it would be interesting to arrive at qualitative

axioms sufficient to guarantee that the representing family canbe taken
to be the gamma family. Mathematically, the problem undoubtedly
entails finding a functional equation characterization of the gamma
family which, to ourknowledge,has never been given. More generally,
one would expect the distribution of therandom variable associated to
x © y to be the convolution of the distributions of those associated to x
and to y. With such a representation, the expectedvalue would behave
as a traditional measurementrepresentation, in particular

x®y =z iff E[4>{x)}+E[qb(y)]=E[(t)(z)].

Thus, we state the problem as:

PROBLEM 3. Extend the qualitativeprimitives ofRTMin sucha way

that the objects of the domain are represented by random variables
(instead ofby numbers = constant random variables).

Theideathatrandomness canbe capturedqualitatively within arelation-

al structure is not totally idle because, in a certain sense, that is exactly
what has been done in theories of subjective utility theory (Savage,
1954, andmuch subsequentliterature). In such theories,choices among
uncertain alternatives are used to infer arandom variable -utility and
probability distributions over families of events underlying the uncer-
tain alternatives. The problem is to doit in contexts more analogous to
extensive and conjoint measurement, not preference among uncertain
Progress on Problem 3 hasbeen made for the special case of finite,
ordinal empirical structures, i.e., there is just one primitive, an order-
ing relation (Cohen and Falmagne, 1990). Suppes and Zanotti (1992)
have followed a different tack in attempting to axiomatize qualitative
moment information and to use that to characterize a random variable
Alternative approaches to random variable ideas of error may be
appropriate for extensions of RTM.In particular, applications of Bool-
ean-valued andother multi-valued logics and fuzzy logics seem poten-
tially interesting.
PROBLEM4. Extendthe RTMapproach to include axioms formulated
in terms ofmulti-valued logics.

Efforts in this direction shouldbe carried out soas to produce either new
kinds ofresults -not just translations ofknown random variable repre-
sentation results - or new insights into measurement through concepts
not available in the standard approaches to random variables. Some
progress hasbeen made by Heyer andNiederee (1989), but much more
research on the topic is needed.7

3.3. Meaningfulness
The meaningfulness part of RTMhas not received as much attention as
the existence and uniqueness parts and so it is less fully developed. In
particular, as with the definition of a measurement scale, the criterion
for meaningfulness invokedin RTMhas not been adequately justified.
Although FM3 describes methods for linking qualitative correlates to
meaningful quantitative relations, no justification is provided for why
these qualitative correlates are indeed empirical.
Narens (1988) showed that these qualitative correlates are definable
in terms of the primitives through a very powerfulhigher-orderlogical
language that includes individual constant symbols for purely mathe-
matical entities. Because empirical definitions requireonly much weak-
er logical languages, Narens' results establish that the qualitative corre-
lates of non-meaningful qualitative relations cannot be defined empir-
ically in terms of the primitives of the empirical structure. Thus, the
correlates of non-meaningful relations are non-empirical with respect
to the primitives. These results can also be used to show that there
exist qualitative correlates of meaningful quantitative relations that are
necessarilynon-empirical. The conclusion to be drawn from thisis that
theRTM conceptof meaningfulnessgives a necessary but not sufficient
condition for empirical significance. It should be remarked that most
of the applications of the meaningfulness concept, such as dimensional
analysis (Bridgman, 1922, 1931; Luce, 1971, 1978), use it only as a
necessary condition for empirical significance, i.e.,use itas acondition
for eliminating from consideration non-meaningfulrelationships.

PROBLEM 5. Amend RTM's concept ofmeaningfulness so that it cap-

tures in amoreappropriatefashion the concept ofempirical significance
(with respect to the qualitative structure). Obviously, a major part of
thisproblem is giving acoherent formulationof 'empirical significance

It may well be the case that the solutions to Problems 1 and 5 are
intimately connected.

3.4. Continuum Representations

Although there are someresults onfinite measurementstructures, some
of which are quiteusefulin applications, andondenumerable structures,
the general consensus is that the strongest, most elegant results are
about structures that map onto a continuum. Examples were cited in
Section 2.1 and additional ones are given below.
The results about structures on finite sets are of two distinct types.
One in essence axiomatizes a finite standard sequence, which leads
in the usual way to an integer representation and a simple uniqueness
theorem. Theother establishes a set of inequalities that must be satisfied
and the existence of a numerical representation is established but we
usually are unable to give a compact anduseful characterization of its
Assuming a universe of finitely many particles, which many believe
is implied by current physical knowledge, why does one ever need
to look at denumerable let alone continuum results; yet it is these,
especially the latter, that seem to be of greatest import for scientific
measurement. The problem divides naturally into two phases: from
finite to denumerable and from denumerable to continuous. One can
envisage some sort of theory concerning a nesting of finite systems
that leads, asymptotically, to a denumerable structure from which the
finite systems can be thought of as samples. But it is well known that
the step from denumerable systems, even the rational numbers, to the
continuum is delicate. For example, the results on scale type to be
described in Section 4become vastly more complex in the denumerable
case (Cameron, 1989). Considerable research exists on the Dedekind
completions8 of certain classes of structures (seeNarens, 1985; Ch. 19
of FM3), but this has not yet been given an adequatejustification for
using continuum models.
PROBLEM 6. Provide aprincipled account of why it is scientifically
useful to replace finite structures by continuum ones. In particular,
it is important to make clear just what limiting processes give rise to
Dedekind completeness or Archimedeaness in the continuum,and also
what gives rise to the properties ofhomogeneity andfinite uniqueness
thatare discussed next.

This problem may well be closed related to Problems 1 and5.


Coexisting with the representational approach has been another theme
which, beginning in 1981, began to flourish as a major alternative.
During the earlier debates over the existenceof psychological measures
of any sort, as distinctfrom physical ones, Stevens (1946, 1951) placed
great emphasis on the uniqueness of representations. His list of scale
- -
types ordinal, interval, ratio, and absolute is famous, and most
empirical examples fell within it. But not all. For example, Narens and
Luce (1976) showed that one could simply drop the associativity axiom
from classical extensive measurement and still show the existence of
an (inherently non-additive) numerical representation. But they failed
to provide a satisfactory description of its uniqueness beyond the fact
that specifying a single point rendered it unique. The reason for their
failure did not become apparent until the work of Cohen and Narens
(1979) in which the uniquenessproblem was first treated as essentially
equivalent to understandingthe group of symmetries (=automorphisms)
of the structure, i.e., isomorphisms of the structure onto itself. They
showed that for non-associative concatenation structures the group of
automorphisms is surprisingly simple: an Archimedean ordered group.
Holder (1901) had completely characterized such structures byshowing
that each is isomorphic to some subgroup of the multiplicative positive
real numbers (MPRN). Ratio scales onto the positive reals are those
cases where the automorphism group is isomorphic to theentireMPRN

4.1. Homogeneity, Finite Uniqueness, andScale Type

The results just described were a precursor to Narens' (1981a, b) pro-
posed classification of the automorphism groups of ordered relational

structures in terms of two major properties, called the degree of homo-

geneity and degree of uniqueness. The automorphism group is saidto be
M-point homogeneousif for any two ordered sequences of M distinct
elements, thereexistsan automorphism that takes the first sequenceinto
the second. That group is said to be N-point unique if whenever two
automorphisms agree at Ndistinct points, they are necessarily identical.
For an M-point homogeneous structure with more than Mpoints, it is
easy to see that M < mm N. The scale type of the structure is the
ordered pair ofnumbers (max M, mm N). Ratio scale structures are of
type (1,1); interval scales ones,of type (2, 2).
The question posedby Narens was: what scale types are possible?
A number of general, but still partial, answers are known for (simply)
orderedrelational structures in which the primitive relations are of finite

4.2. A Recipefor Constructing Numerical Representations

The major results rest on properties of a subset of the automorphisms
called translations: the identity map together with all automorphisms
that do not have any fixed points. Three major questions about them
(i) Dothe translationsform amathematical group? The only real prob-
lem is in showing that they are closedunder function composition
which is equivalent to showing that they are 1-point unique,
(ii) Under the ordering of the translations naturally induced from the
order of the given structure, 9 are the translations Archimedean? 10
(iii) Are the translations 1-point homogeneous?
Once these facts are established,one can construct a numerical repre-
sentation in whichthe translations appear as multiplication by constants
(i.e., the similarity group) as follows: using homogeneity, map the
structure isomorphically onto the group of translations;using Holder's
theorem, map the translation group, and so the structure, into MPRN
(Alper, 1987; Luce,1986, 1987). Thatleaves unanswered the question
about the rest of the automorphisms. A simple answer is known when
the underlying ordered relational structure is order dense:11 the auto-
morphisms are a subgroup of the power group x— ► rxs, which means
that mm N < 2 (see FM3, Theorem 20.7, Corollary 2).
4.3. Some Structures with Translations Satisfying the Recipe
So a key problem is to try to understand what structural properties
give rise to these three properties of translations. So far, no one has
derived any necessary structural properties from these three properties
of translations. All that is known are certain sufficient conditions. We
cite two partial results.
First, if the ordered relational structure is Dedekind complete and
order dense, then (i) implies (ii) (see Theorem 20.6, FM3). Second,
and this is the most generalresult known at present,if the ordered rela-
tional structure canbe mapped onto the real numbers, is homogeneous
(maxM > 1),and is finitely unique (mm N< oo), then the threecon-
ditions are satisfied and the possible scale types are (1, 1),(1, 2) and (2,
2), with the first beinga ratio scale, the third an interval scale, and the
(1, 2) case falling between the two. An example of the (1, 2) case is
the group of real transformations x— > k nx +s, where k>o is fixed,
n ranges over all integers, and s ranges over all real numbers. Narens
(1981b)proved the part of this result for which max M = mm N, and
Alper (1987), using a very different approach, proved theresult without
that restriction. Alper's methods first made very clear the importance of
the three properties of the translations. Surprisingly, the most difficult
property to establish is that the translations form a group.

4.4. Structural Equivalents to the KeyProperties of Translations

Given the results just mentioned, it is clear that we still have much to
learnabout the conditions under which the translations form a homoge-
neous,Archimedean-ordered group.

PROBLEM 7. What structural properties are implied by each of the

translation properties group, Archimedean, and homogeneity sep-
arately or together? As none are currently known, a simpler problem
may be: what structural properties are sufficientfor each of the three
properties of translations, either separately or jointly ? Ideally, one
would like tofinddistinct structuralfeatures that correspondseparately
to eachproperty,although that may wellprove infeasible.

For any homogeneous and finitely-unique ordered structure having a

binary, monotonic operation, not necessarily on a continuum,one can

prove directly that minN < 2 (Luce and Narens, 1985). Little is
known in general about the automorphism groups in this case, except,
as was noted above, when the operation is also positive, solvable, and
Archimedean, one can prove, without assuming homogeneity, that the
automorphism group is Archimedean ordered. We do not know of
comparable results in the remaining homogeneous case in which the
operation is necessarily idempotent.12 In particular, we do not know
of conditions that result in the translations forming an Archimedean
ordered group.

PROBLEM 8. What algebraic properties on homogeneous, finitely

unique relational structures are just sufficient' to prove mm A^ < 2?
The twoknown results shouldbe special cases. When mmN < 2 can
be proved, what can be saidabout the set of translations?

(It strikesus as unlikely that areasonable set of necessary and sufficient

conditions willbe found, and so the criteria for 'just sufficient' should
be considered to be flexible.)


5.1. Combining Scale Type andStructural Conditions

Once theresults on scale types began to be discovered,the possibility of
applying them to specific measurementproblems began to be explored.
Basically the strategyhas been to combine the homogeneity and finite
uniqueness conditions with more specific structural assumptions and to
characterize in greaterdetail theclasses of structures that canarise. One
example of this is found in Luce and Narens (1985) in which homo-
geneous, finitely unique concatenation structures on the continuum are
described quite fully. Closely related is the fact that for the class of
concatenation structures that are positive, solvable, Archimedean, and
Dedekind complete, if the ordered automorphism group is order dense,
thenit isalso homogeneous. A third example is continuous semiorders;
they havea (1,oo) group of automorphisms with a subgroup of transla-
tions that is homogeneous andcan be ordered so thatit is Archimedean.
We do not know of other cases where the numerical representation of
a structural axiomatization has been studied by proving that its transla-
tions form a homogeneous, Archimedean ordered group.

PROBLEM 9. Explore the possibility of using the recipe described

above to construct numerical representations of particular ordered
structures that do not fall under the scope of the theorem of Alper
and Narens mentioned in Section 4.3.

5.2. Structures That Are Not Finitely Unique

A structure may fail either homogeneity or finite uniqueness; the two
failures are very different.
Ordinally scalable structures are not finitely unique; indeed,it takes
at least an order dense subset of the domain to fix an automorphism.
Moreover, they are also M-pointhomogeneous for each finite M. We
say they are of scale type (oo,oo). Roberts and Rosenbaum (1985)
established thatfor an ordered relational structure that is M-pointhomo-
geneous, if Mdoes not exceed the cardinality of the domain andif the
order of each defining relation of the structure is not greater than M,
then the automorphism group of that structure is identical to that of just
the ordered domain. Thus,if these conditions are met and the ordered
domain is isomorphic to the ordered real numbers, then the structure is
of scale type (00, oo). Droste (1987) has characterized in aconvenient
form the automorphism groups of such structures.
Thisleaves the (M,oo) cases,about which relatively little is known.
Althoughit hasbeen shown thatcontinuous semiorders are examples of
(1,oo) structures (Narens, 1994), weknow of no explicit measurement
structures with 1 < M < 00. Thus, many questions remain unan-

PROBLEM 10. What canbe saidabout structures and their automor-

phism groups of scalar type (M,oo) with 1 < M < oo?

5.3. Structures That Are Finitely Unique ButNot Homogeneous

For generalfinitely unique,non-homogeneous structures on the contin-
uum, Alper (1987) provideda description of the possible automorphism
groups, but so far his classification has not been used successfully to
describe the corresponding structures. Luce (1992a) followed a far

more restricted tack, but one that seems highly relevant to some mea-
surement applications. He defined a point in a structure to be singular
if it remains fixed (or invariant) under every automorphism of the struc-
ture. The concept ofa translationis generalizedtobe either the identity
or any automorphism whose only fixed points are singular ones. Clearly,
if the structure is finitely unique, then ithas only finitely many singular
points and soitis meaningful to speak of such astructure as being trans-
lational homogeneousbetween adjacent singularpoints. For a class of
structures that he calls generalizedconcatenation structures whichhave
a monotonic 13 n-ary operation,he gave a fairly complete description of
the possibilities when the structure isboth finitely unique andtranslation
homogeneous between adjacent singular points. What makes matters
simple is that such structures can have at most three singular points:
a maximum, a minimum, and an interior one. An example of such a
structure is the multiplicative, positive real numbers augmented by 0
and oo with translations x— ► xr, where r> 0. The singular points are
the two extremeones, 0 and 00, andone interior one,1 .
If, in addition, the group of translations commute,14 thenan interior
singular point, call it c, acts like a generalized zero15 in the follow-
... ...
ing sense: if F denotes the operation, then F(e, ,c, Xi, c, ,c) ==
6i(xi), where oneither sideof c the function 6i agrees with a translation
of the structure. Moreover, if any singular point is a generalized zero,
then any other singular point, c', acts like an infinity in the sense that
if c' is an argument of the function, then the value of Fis c! Finally,
for structures on a continuum, Alper's results can be used to derive a
numerical representation in which translations on each side of the inte-
rior singularity are multiplication by a constant, the two constantsbeing
simply related by a power relation. As we shall seein Section 6.3, these
and relatedresults have been applied effectively in devising a theory of
certainty equivalents for gambles (Luce, 1992b).

PROBLEM 11. What structures are there that are usefulfor applied
measurement, beyond those based on a general operation, for which
homogeneity fails at selected points? And what structures, although
failing homogeneity more globally, still have a fairly rich automor-
phism group, such as x— > k nx, where k>o is a fixed constant and n
rangesover the integers?
Important non-homogeneous structures lie outside this framework.
The most notable examples are qualitative probability structures. They
are non-homogeneous not only because of their extreme points the
null anduniversal events butbecause two events that are qualitatively
equally probable need not exhibit the same relational patterns to other
events. Also, the only automorphisms of such a structure are ones, like
the identity, that take an event into an equallyprobable one. Obviously,
the previous tacks we have taken in structures that are more or less
homogeneous are completely useless in such contexts. Yet, clearly the
probability case exhibits a great deal of regularity.

PROBLEM 12. What kind of useful classification can be given for

structures that exhibit a great deal of regularity, such as is seen in
probability structures, but that only have automorphisms a for which
a(x) is equallyprobable to x?


6.1. Product-of-Powers Compatibility in Bounded Cases

One feature of physical measurement is the existence of two distinct
kinds of attributes: those having an internal structure, like mass, time,
length,and charge,and those having a trade-off structure between com-
ponents, like energy, momentum, density, etc. A major feature of clas-
sical physical measures is that the conjoint ones can be represented as
products of powers of the extensiveones. For example, kinetic energy is
given by mv2 and densityby m/V. This aspect of the representation
is reflected in the fact that theunits of the conjoint measures are always
products of powers of theunits of extensivemeasures. How this arises
from measurement considerations is discussed in Ch. 10 of FMI and
again,much more generally, in parts ofChs. 20and 22of FM3.
The most general result to date involves two major properties: a
qualitative concept of how a structure on one component of a solvable
conjoint structure distributes16 in the latter structure (Definition 20.6,
FM3), and the assumption that the component structure is such that its
translations form a homogeneous Archimedean ordered group. These
two properties force the conjoint structure to have a multiplicative rep-
resentation involving a power of the representation of the component
attribute (Theorem 20.7,FM3).

This model of interrelated measures covers much of classical mea-

surement and makes clear exactly which generalization of extensive
- -
structures basically any ratio scale structure can be added to the
physical structure without disrupting the product-of-powers feature.
Unfortunately, it fails to cover everything of importance. The most
notable exceptionis relativistic velocity, which is bounded from above
by the speed of light.
As is well known,physicists have elected to keep the multiplicative
conjoint relation s = vt among distance, velocity, and time, and to use
a non-additive and bounded representation of the associative concate-
nation of velocities, namely, u(&v= (v + v)/{\ + uv/c2 ), where c
denotes the maximum velocity, that of light. (In principle, they could
have adopted an additive representation of ©, but at the very consider-
able expenseof foregoing the simple relation s = vt). Itis important to
realize that © is not reallyan operation on the velocity component of the
definingconjoint structure. The reason is that v and v © v are velocities
in a single frame of reference whereas v is in a different frame, one in
which the distances and times are modified relative to the first frame.
Thus,from theperspective ofrelativistic measurement,itis inappropri-
ate to think of © as an operation on the conjoint structure. But even
if one does, it fails to be distributive in the conjoint structure. Within
the context of a binary operation, distribution comes to: if ut = u't'
and vt = v't', then (v © v)t = (u1 © v')t' It is easy to verify that this
fails. More deeply, when distribution holds, the automorphism struc-
tures of the conjoint and component structures agree in the sense that
every translation of the component structure is that component's con-
tribution to an automorphism of the conjoint structure, and conversely.
This is not true in the case of the velocity component. Yet, at the same
time multiplication by a positive constant r, whichis an automorphism
of the conjoint structure, simply maps one velocity representation into
another, with c becoming re. This lack of a connection between the
automorphism groups of the two structures and yet their tight product
of powers representation is not understood in terms of measurement
This problem is important not only to complete our understanding
of how relativistic velocity ties into the classical measures, but also to
allow the introduction of other bounded measures. The most obvious
of these is probability which, in computingexpectations, has a product
relation with other measures, as in (subjective) expectedutility theory
(Savage, 1954; yon Neumann and Morgenstern,1947). Aside from the
boundedness, the two cases are very different. Other examples may
arise in the behavioral sciences wheremeasures of subjective intensity
are almost surely bestmodelled as bounded from above.

PROBLEM 13. What links the bounded component structures of a

conjoint structure and the conjoint structure itself so as, again, to lead
to a product-of-powers representation? Possibly one should initially
assume the components are homogeneous between bounds, as in the
velocity case, but ultimately that restriction must be removed to deal
with the probability case.

6.2. Compatibility ofPsychological and Physical Theory

Matching is a psychological procedure in which a subject 'matches' a
stimulus in one intensity domain,such as brightness, to a given stimulus
in another intensity domain, such as loudness. Formally, there are two
physical structures Xand S, withdomains X and S, thateachhaveratio
scale representations; classically, they are simply extensive structures
but the theory applies to any with a ratio scale representation. The psy-
chologicaldata may be summarized as follows: for each ann X, there is
some s in S such that, according to the subject, s matches x,which we
symbolize as xMs. In this situation, there are two physical relational
structures and a purely psychological connecting relation Mbetween
them. Empirically, to a first approximation at least, the matching rela-
tion can be described as a power relation between the two physical
ratio-scale measures. The question is to what does this correspond.
Luce (1990) suggesteda principle of compatibility between the two
domains that may be described as follows. Toeach translation r of X,
thereis a corresponding translation aT of S such that for all x 6 X and
all s £ S:
xMs if and only if r(x)MaT (s).
This is easily shown to be equivalent to a power relation between the
ratio scale representations, and a number of other relationships are
explored. The multiplicative constant of the relation has, of course,
dimensions that dependupon the exponent involved.
Clearly, the application just described is highly special to a particu-
lar situation, but the general idea of asking which psychological laws

relating physical variables are in fact compatible with the translation

structure of these variables is a far more general principle. The fail-
ure of such compatibility means non-physical variables are required to
describe the phenomenon.

PROBLEM14. Is there any deepscientific or philosophicalgrounding

for the supposition thatpsychological laws should be compatible with
the automorphism structures of the physical variables that they relate?
Are there other examples in which application of this principle can be
illustrated? And are there examples whereit clearly is violated?

6.3. Applications ofResults about Structures with Singular Points

Todate just onenew application(beyond relativistic velocity) has been
made in the measurementliterature of the results about structures with
general operations that are finitely unique, have singular points, andare
translation homogeneous between adjacent singular points (see Sec-
tion 5.3). It concerns certainty equivalents to uncertain monetary
alternatives.17 For a fixed event partition, a certainty equivalent can
be viewed as a monotonic function of money arguments the pay-offs
associated with the several subevents into an amount of money that is
indifferent totheuncertain alternative. A sharp distinction ismaintained
between gains and losses, making 0 an interior singular point. Assum-
ing the structure is homogeneous oneither side of 0 and finitely unique,
which has been typical of utility theories, and definingutility to be the
isomorphism that represents the translations as multiplication by pos-
itive constants yields a linear weighted average utility representation.
The fact that 0must be a generalizedzero is very important in construct-
ing the weighting functions ofrank- and sign-dependent character. We
do not go into any of the details,for the only point in mentioning ithere
is as evidence that such generalresults do have applications.

PROBLEM 15. Given that we gain some results about structures with
variousforms of non-homogeneity (see Problems11 and12) including
ones withinterior singularpoints, are there applications that up to now
have been overlookedbecausepreviously we didnot know how to deal
with suchsituations?

The representational theory of measurement,despite its positive contri-

butions,has been subjected to attack on a number of fronts. Onebasic
issue is how to formulate clearly what one means by measurement in
such a way that the representing structure is derived from the axioms
of the qualitative structure. Of considerable interest are structures that
do not map into the real number system, but rather into families of
random variables or into structures basedon multi-valued logics(Prob-
lems 1-5). Further, even in the case of numerical representations, one
can wonder why anything as idealized as the continuum is relevant to
science (Problem 6). The remaining problems are all considerably less
philosophical and involve somewhat complex issues in the theory of
scale types. We know a lotmore about scale type than we did 12 years
ago, but much about structures remains shrouded;hard work and new
ideas are probably needed to gain a deeper understanding. Some of
these problems are listed as 7-15.


This work has been supported in part by National Science Foundation

grant SES-8921494 to the University of California, Irvine. We wish to
thank Jean-Claude Falmagne, Dieter Heyer,Reinhardt Niederee,and A.
A. J. Marley for detailed and helpful comments on earlier drafts of this

Institute for Mathematical Behavioral Sciences,

University of California,
Irvine, CA 92717, U.S.A.


1Krantz,Luce, Suppes,
and Tversky (1971) willbereferred to as FMI;Suppes,Krantz,
Luce, and Tversky (1989) as FM2; and Luce, Krantz, Suppes, and Tversky (1990) as
Some of the subtletiesinvolved in formulating uniqueness results are dealt withby
Roberts andFranke (1976).
A conjoint structure is a weakorderingof a Cartesianproduct;it is additiveif it admits
an additiverepresentationover its components.

Archimedeaness simply means that any bounded standard sequence is finite. Put
another way, no positive element is infinitesimal relative to another element of the
structure. In practice, the impact of Archimedeanessis to permit homomorphisms into
the realnumbers rather than ordered extensions of the realnumbers such as the non-
standard reals.
5 Among
the positiveuses weinclude the testingof mathematically formulatedtheories
relating several variables,although we recognize that othersmay not want to takesuch
a consideration into account in finding a solution to this problem. Subjective expected
utility,mentionedlater,is one example;three more are given inSection6.
6 The primary exceptions are purely ordinal cases including variants such as interval
ordersand semiorders.
7 See Heyer and
Mausfeld (1987) for a discussion of some conceptual problems con-
nected withBoolean-valued approaches.
8 A structure is Dedekind complete if every subset of elements thatis bounded from
above has a least upperbound in the domain. A structure that can be embeddedin a
Dedekind complete one of the same algebraic formis said to have a Dedekindcomple-
For automorphismsa,(3, a <' (3 if andonlyifforeveryelement x of X, a(x) P(x).
Forany two translationsone of whichis greater than the identity, finitely many appli-
cationsof the positive one willexceedthe other.
If x < y, there exists z in X such that x < z < y.
Homogeneityimplies, for allelements x, either weak positivity, x o x > x, idem-
potence, x o x
~ x, or weak negativity, x o x < x. The first and third are formally
identical if > is replacedby <.
13 The definitionofmonotonicis theusualoneexcept that somecareis neededindealing
with extreme points,if such exist.
This is true if theycan berepresentedin the multiplicativerealnumbers. The assump-
tionmay be redundant,butLuce failed to deriveit fromthe other assumptions.

The term 'zero'is appropriatewhen one thinks of homomorphisms to the realnum-
bers where0 is the interior singularpointand thetranslations are x » rx.
Let C = (A x X, >) be the conjoint structure withA a relational structure on A
whoseorder is thatinduced fromC. For fixed x, y € X, let a be the functiondefined
by allsolutions to (a,x) ~ (a(a),y). Then, A distributes in C if every such a is an
automorphism of A.
17 This is closely related to the prospect theory of Kahneman andTversky (1979) and to
such extensionsof it as Luce andFishburn (1991) and Tversky andKahneman(1992).


Adams, E.: 1966, 'On the Nature andPurposeofMeasurement', Synthese, 16, 125-169.
Alper, T.M.: 1987, 'A Classificationof All Order-Preserving HomeomorphismGroups
of the RealsThat Satisfy Finite Uniqueness',JournalofMathematicalPsychology,
31, 135-154.
Bridgman, P.: 1922, 1931, Dimensional Analysis, New Haven, CT: Yale University
Cameron, P.J.: 1989, 'Groups ofOrder-Automorphisms oftheßationals withPrescribed
Scale Type', Journalof MathematicalPsychology, 33, 163-171.
Campbell, N. R.: 1920/1957, Physics: The Elements, Cambridge: Cambridge Univer-
sity Press. Reprinted as Foundations of Science: The Philosophy of Theory and
Experiment, New York: Dover, 1957.
Campbell, N. R.: 1928, AnAccount of the Principles ofMeasurementand Calculation,
London: Longmans, Green.
Cohen, M.andFalmagne,J.-C: 1990, 'RandomUtilityRepresentationofBinary Choice
Probabilities: A NewClass of Conditions', Journal ofMathematicalPsychology,
34, 88-94.
Cohen, M. and Narens, L.: 1979, 'FundamentalUnit Structures: A Theory of Ratio
Scalability', Journal ofMathematicalPsychology, 20, 193-232.
Cohen, M.R. and Nagel,E.: 1934, An Introduction to Logic and Scientific Method,
New York: Harcourt, Brace.
Droste, M.: 1987, 'Ordinal Scales in the Theory of Measurement', Journal ofMathe-
maticalPsychology, 31, 60-81.
Ellis, B.: 1966, Basic Concepts of Measurements, London: Cambridge University
Falmagne,J.-C: 1979, 'On a Class of ProbabilisticConjoint Measurement Models:
Some Diagnostic Properties',JournalofMathematicalPsychology, 19, 73-88.
Falmagne,J.-C: 1980, 'AProbabilisticTheory ofExtensiveMeasurement', Philosophy
of Science, 47, 277-296.
Falmagne,J.-C.andIverson,G: 1979, 'Conjoint WeberLawsand Additivity',Journal
ofMathematicalPsychology, 86, 25^43.
Helmholtz,H. yon: 1868, 'Über die Thatsachen, dieder Geometriczu Grunde liegen',
GottingenNachrichten, 9, 193-221.
Helmholtz, H. yon: 1887, 'Zahlen und Messen erkenntnis-theoretisch betrachtet',
Philosophische Aufsatze EdwardSeller gewidmet, Leipzig.Reprinted in Gesam-
melte Abhandl, Vol. 3, 1895, pp. 356-391. English translation by C. L. Bryan:
1930, Counting and Measuring, Princeton,NJ: VanNostrand.
Heyer, D. andMausfeld,R.: 1987, 'On Errors, ProbabilisticMeasurement andBoolean
ValuedLogic', Methodika,1, 113-138.
Heyer, D. and Niederee, R.: 1989, 'Elements of a Model-TheoreticFramework for
ProbabilisticMeasurement', in: E.E.Roskam (Ed.), MathematicalPsychology in
Progress, Berlin: Springer, pp. 99-112.
Heyer, D. and Niederee, R.: 1992, 'Generalizingthe Concept of Binary Choice Sys-
tems Induced byRankings: One Way of ProbabilizingDeterministicMeasurement
Structures', MathematicalSocial Sciences, 23, 31-44.
Hilbert, D.: 1899, Grundlagen derGeometric(Bth cd., with revisions and supplements
by P. Bernays, 1956), Stuttgart: Teubner.
Holder, O.: 1901, 'Die Axiomeder Quantitat und dieLehre vomMass',Berichte über
die Verhandlungen der Koniglich Sachsischen Gesellschaft der Wissenschaften zu
Leipzig, Mathematisch-Physische Klasse,53, 1-64.
Holman,E.: 1971, 'A Note on AdditiveConjoint Measurement', JournalofMathemat-
icalPsychology, 8, 489-494.
Iverson,G. and Falmagne, J.-C: 1985, 'StatisticalIssues inMeasurement',Mathemat-
icalSocial Sciences, 10, 131-153.

Kahneman, D. and Tversky, A.: 1979, 'Prospect Theory: An Analysis of Decision

under Risk',Econometrica, 47, 263-291.
Klein, F: 1872 (this was the oralErlangen address; transcribedinto English, 1893), 'A
ComparativeReview of RecentResearches inGeometry',Bulletinof theNew York
MathematicalSociety, 2, 215-249.
Krantz, D. H.: 1964, 'ConjointMeasurement: The Luce-Tukey Axiomatizationand
SomeExtensions', Journalof MathematicalPsychology, 1, 248-277.
Krantz, D. H., Luce, R. D., Suppes, P., and Tversky, A.: 1971, Foundationsof Mea-
surement, Vol. I, NewYork: AcademicPress.
Kyburg,H. E., Jr.: 1984, Theory andMeasurement,Cambridge:CambridgeUniversity
Lie, S.: 1886, 'Bemerkungen zuHelmholtz'Arbeitiiber dieThatsachen, diederGeome-
tric zuGrunde liegen',Berichteüber die Verhandlungender Koniglich Sachsischen
Gesellschaft der Wissenschaften zu Leipzig, Mathematisch-Physische Klasse,38,
Luce,R. D.: 1971, 'Similar Systems andDimensionallyInvariant Laws', Philosophy
of Science, 38, 157-169.
Luce, R. D.: 1978, 'DimensionallyInvariantNumericalLaws Correspond to Meaning-
ful QualitativeRelations', Philosophy ofScience, 45, 1-16.
Luce, R. D.: 1986, 'Uniqueness andHomogeneity of OrderedRelational Structures',
Journal ofMathematicalPsychology, 30, 391-415.
Luce, R. D.: 1987, 'Measurement Structures with ArchimedeanOrdered Translation
Groups', Order,4, 165-189.
Luce, R. D.: 1990, 'On the Possible Psychophysical Laws Revisited: Remarks on
Cross-ModalMatching',PsychologicalReview, 91, 66-77.
Luce, R. D.: 1992a, 'Singular Points in GeneralizedConcatenation Structures That
Otherwise AreHomogeneous', MathematicalSocialScience 24, 79-103.
Luce, R. D.: 1992b, 'A Theory of Certainty Equivalents forUncertain Alternatives',
JournalofBehavioralDecision Making 5, 201-216.
Luce, R. D. and Fishburn, PC: 1991, 'Rank- and Sign-Dependent Linear Utility
Models forFiniteFirst-Order Gambles',Journal of Risk and Uncertainty, 4, 29-
Luce,R. D., Krantz,D. H., Suppes, P., and Tversky, A.: 1990, Foundations of Mea-
surement, Vol. 111, New York: Academic Press.
Luce, R. D. and Narens, L.: 1985, 'Classification of Concatenation Measurement
Structures According to Scale Type', Journal of Mathematical Psychology, 29,
Luce, R. D. and Tukey, J. W.: 1964, 'SimultaneousConjoint Measurement: A New
Type of FundamentalMeasurement',JournalofMathematicalPsychology, 1, 1-27.
Michell, J.: 1986, 'MeasurementScales andStatistics: A Clashof Paradigms',Psycho-
logical Bulletin,100, 398^107.
Michell, J.: 1990, AnIntroduction to the Logic of Psychological Measurement, Hills-
dale,NJ: Erlbaum Associates.
Narens, L.: 1981a, 'A General Theory of Ratio Scalability withRemarks about the
Measurement-Theoretic Concept of Meaningfulness', Theory andDecisions, 13,
Narens, L.: 1981b, 'On theScales ofMeasurement',JournalofMathematicalPsychol-
ogy,24, 249-275.
Narens,L.: 1985, AbstractMeasurement Theory, Cambridge, MA: MIT Press.
Narens, L.: 1988, 'Meaningfulness and the Erlanger Program of Felix Klein',
Mathematiques Informatique et Sciences Humaines, 101, 61-72.
Narens, L.: 1994, 'The Measurement Theory of Continuous Threshold Structures',
Journal ofMathematicalPsychology, in press.
Narens,L. and Luce, R. D.: 1976, 'The Algebraof Measurement', Journalof Pure and
AppliedAlgebra, 8, 197-233.
Niederee,R.: 1987, 'OntheReference toReal Numbers inFundamentalMeasurement:
A Model-Theoretic Approach',in: E.E. Roskam and R. Suck (Eds.),Progress in
MathematicalPsychology, Vol. 1, pp. 3-23, New York, Amsterdam:Elsevier.
Niederde, R.: 1992, 'WhatDo NumbersMeasure? A New Approach to Fundamental
Measurement', MathematicalSocial Sciences, 24, 237-276.
Roberts,F. S. andFranke, C. H.: 1976, 'Onthe Theory ofUniquenessinMeasurement',
Journalof MathematicalPsychology, 14, 211-218.
Roberts,F. S. andRosenbaum, Z.: 1985, 'SomeResults on Automorphisms of Ordered
Relational Systemsand the Theory of Scale Type inMeasurement',in: V Alavi et
al.(Eds.), GraphTheoryandIts Applicationsto Algorithms andComputer Sciences,
New York: Wiley, pp. 659-669.
Savage, C. W. andEhrlich, P. (Eds.): 1992, PhilosophicalandFoundationalIssues in
Measurement Theory, Hillsdale, NJ:ErlbaumAssociates.
Savage,L.J.: 1954, TheFoundationsofStatistics, New York: Wiley.
Scott, D. and Suppes, P.: 1958, 'Foundational Aspects of Theories of Measurement',
JournalofSymbolic Logic, 23, 113-128.
Stevens,S. S.: 1946, 'Onthe Theory ofScalesofMeasurement',Science, 103, 677-680.
Stevens,S.S.: 1951, 'Mathematics,MeasurementandPsychophysics', inS. S. Stevens
(Ed.), Handbook ofExperimentalPsychology, New York: Wiley, pp. \-A9.
Suppes,P.: 1951, 'A Set of Independent Axioms forExtensive Quantities', Portugal.
Math., 10, 163-172.
Suppes, P., Krantz, D. H., Luce, R. D., and Tversky, A.: 1989, Foundationsof Mea-
surement, Vol. 11, New York: AcademicPress.
Suppes, P. and Zanotti, M.: 1992, 'Qualitative Axioms for Random-Variable Repre-
sentationsof Extensive Quantities', in: C. W. SavageandP. Ehrlich (Eds.), Philo-
sophicalandFoundationalIssues inMeasurement Theory, Hillsdale,NJ: Erlbaum
Associates, pp. 39-52.
Suppes, P. and Zinnes, J. L.: 1963, 'BasicMeasurement Theory',in:R. D. Luce, R. R.
Bush, andE.Galanter (Eds.), Handbook ofMathematicalPsychology, Vol. 1, New
York: Wiley, pp. 1-76.
Tversky, A. and Kahneman, D.: 1992, 'Advances in Prospect Theory: Cumulative
Representationof Uncertainty', JournalofRiskand Uncertainty 5, 297-323.
yon Neumann,J.andMorgenstern, O: 1947, Theory GamesandEconomicBehavior,
Princeton: Princeton University Press, secondedition.


Luce and Narens givea substantiallist of problems that are central to the
representationaltheory of measurement, a topic on whichDuncan and I
have worked together with other colleagues for more than a quarter of a
century. Duncan andLouis bring out nicely the range of philosophical
and scientific problems still to be faced in the theory of measurement.
Unfortunately, most of these problems as well as the ones that have
been solved in the past have not attracted the interest of philosophers
of science in the way one might have thought they would. Ithas turned
out that in spite of the philosophical roots of much of the work in the
theory of measurement, the current developmentshave mainly beendue
to scientists and mathematicians.
The problems of measurementin thebehavioral and social sciences
present foundational and conceptual issues of considerable subtlety
which now have a large literature, especially those surrounding the
measurementof subjective probability and utility. Without question the
problems formulated by Luce andNarens alldeserve attention, although
of course some are of more generalinterest than others. It isareflection
of the generalnature of the theory of measurement that my own list of
problems would overlap but still be rather different from that presented
by them. Sayingsomething about my own list is notmeant to be a crit-
icism of theirs, but is a way of emphasizing the range of philosophical
issues still open in the theory of measurement. The two large topics that
I would organize problems around and that are not directly mentioned
by Luce and Narens are geometricalproblems of measurement - what
are alsocalled multidimensional scalingproblems -and secondly,com-

Geometrical Problems. As Luce and Narens point out in various

places briefly and casually, there is a long history of representation
and uniqueness problems in geometry. There is a substantial review
of classical work in Foundations of Measurement, Vol. IIbut there are
ways in whichthe modern representationaltheory of measurementcalls
for new developments in geometrical representation theory. The most
important direction,in my own judgment, is to develop representation
theories like those of measurementthat are embeddings and not isomor-
phisms to standard analytic representations. Typical examples would
be sufficient, and where possible, necessary and sufficient conditions to
embed a bounded fragment of Euclidean geometryinaEuclidean space
of the same dimension and with the usual results on uniqueness for
that embedding. Another kind of example of interest in current physics
is a qualitative axiomatization of a discrete lattice of points for spe-
cial relativity to provide a qualitative geometrical framework for lattice
computations andother discrete geometricalreasoning characteristic of
much modern physics.
Ina similar vein, but with undoubtedly a somewhat different concep-
tual apparatus, we couldmuch benefit from deeperqualitative analysis
of the geometrical structures that arise in multidimensional scaling.
The closerelation between topology and measurementin many con-
texts is perhaps the part of geometrymost neglectedin theFoundations
of Measurement, but deliberately so for reasons given in Volume I.
Topologyprovides a general set of concepts for formulating axioms on
qualitative structures of a different sort thanhavebeen mainly exploited
in the theory of measurement thus far, although there has been a cer-
tain body of work ineconomics using topological conditions rather than
algebraicones in the theory of utility. On theother hand, problems of the
relationbetween topology and measurementin the theory of perception
haveas yet been little explored.
Some of the most interesting recent work of Luce and Narens has
been on characterizing the automorphism group of a suitable measure-
ment structure as an Archimedean ordered group. There is a similar
but deeper and more far-ranging set of problems in geometry on the
connection between various transformation groups for geometry and
the characterization of these transformation groups as themselves being
instances of manifolds. For example, the group ofrotations ofEuclidean
three-space is a nonsingular surface with a system of local coordinates
providedby familiar Euler angles. Exploitation of thiskind ofrelation-
ship has scarcely begun as yet in the theory of measurement.

Computation and the Continuum. Another important aspect of recent

work by Duncan andLouis in the theory of measurementis thatof scale
type. Here the work has depended almost exclusively on assuming
the measurementstructure is isomorphic to the continuum ofreal num-
bers. Theseresults naturally raise philosophical questions that call for
a deeper analysis of why the continuum is important in the theory of
measurement. Initially we think of measurement as a very finitistic
constructive procedure used in almost all domains of science to assign

numerical quantities to various empirical properties. It would be sur-

prising if highly nonconstructive aspects of the entire continuum of
real numbers really did play some essential philosophical role in our
conception of measurement. In saying this, of course, I am expressing
philosophicaldisagreement with Luce and Narens, a disagreement that
we have discussed on various occasions.
As I have stated in comments onother articles,I look upon classical
analysis and continuum mathematics as beingmainly computationally
important. The differential and integral calculus is just that, a calculus,
not a foundational view of how the universe is really put together. I as
much as Luce and Narens, perhaps even more so, cherish infinitesimals
and what they can do for providing efficient methods of computation,
especially in physics andengineering. I do not for a momentnecessarily
believe that infinitesimals are really out there in the real world. I would
be quite prepared to accept the fact that space is ultimately discrete
and we cannot go below a certain smallest measurement of length or
of other quantities. This would not for a moment shake my confidence
in the importance of infinitesimal methods in science which have been
used so successfully for over two centuries, but it is rather to insist
that they provide wonderfully efficient methods of computation, not
a fundamental view of the world. Ilike very much the derivation
of standard diffusion equations from taking the limit of very discrete
random walks. I am quite happy to look upon the limit operation as an
ideal one abstracted from thereal detail of particles and the spaces in
which they operate.
It is this kind of philosophical view of mine that has led me to be a
much stronger supporter of finite structures andfinitistic or constructive
measurementprocedures than are Duncan andLouis. That debate will
continue and probably no endis in sight. That they remain unconstruct-
ed continuum advocates is clear from their statement of problems at
the beginning of their paper. But I found puzzling their statement in
Section 3.4 that we couldhave "doubts about the heavy useof infinite
structures and continuum mathematics when, after all, the universe is,
according to contemporary physics, composed of only finitely many
particles." This is not the real problem of using the continuum. We
could very well require the continuum if there were only three particles
but they were moving along continuous paths. The realcommitment is
to there beingonly discrete space anddiscrete properties or,if you wish,
only a finite number of spatial points, at least in any bounded region,
anda finite numberof values of anyproperty of anyparticle. Thiskind
of constructivism is clearly very far removed from a large number of
their Problems.
In raising and pursuing once again this dialogue with them, I am
not suggesting that Ihave any strong commitment that the way Iam
suggesting is the only way to proceed. In fact, it may well turn out
that the line of attack they have taken will be more fruitful than a
more constructive finitistic approach. Actually, from a philosophical
standpoint I now favor a view that has as yetnot been developed very
far technically. Thisis theview that in the framework of current physical
theory we cannot empirically determine whether space is continuous or
discrete, and possibly a decision that was empirically supported could
only be for discreteness. In any case, given current physics, the choice
is transcendental,i.e., beyond experience. Consequentlyit isof interest
to develop theories whose fundamental concepts are invariant under
appropriatemappings from the infinite to the finite.
As a simple example, let (X,d) and (Y,d') be two metric spaces
with X being a bounded infinite set and V a finite set. Then (X,d) is
e-homomorphic to (V, d 1) if and only if there is a mapping / from Xto
V such that Vx, y in X
\x-y\<€ iff \f(x)-f(y)\<e.
More to the point,rather than this simple abstract example, are themeth-
ods currently used to approximate continuous processesby the discrete
ones used in computer simulations, or the classical approximation of
the discrete by the continuous,as in approximating the binomial by the
normal distribution. Studies of invariance and meaningfulness in these
contexts would in all likelihood be conceptually enlightening not only
in terms of theories of measurement.



ABSTRACT. This paperstudies themeaningfulness (invariance) of the ordinalcom-

parison f(a) > f(b),madeusing a scaleof measurement /. Specifically,it investigates
scales / whicharise fromhomomorphismsinto general orderrelationalsystems, which
are theunions of m-ary relations induced on the set of real numbers by rankings of
the integers in {1,2,...,m}. The results generalize earlier work in the literature,
whichsets out to make a systematic analysis of conditionsunder which the assertion
/(a) > f(b) is meaningful.


Inthe representationaltheory of measurement,aspioneeredin the papers

by Scott and Suppes (1958) and Suppes and Zinnes (1963) and in the
book by Krantz,Luce,Suppes,and Tversky(1971), oneconsiders scales
of measurementas arising from homomorphisms of relational systems.
We will be specifically concerned with homomorphisms into a very
generalkindofrelational system which wecall a generalorder relational
system. Inmeasurement theory, one studies admissible transformations
taking one homomorphism into another and calls a statement using
scales of measurement meaningful if it is invariant under admissible
transformations. In this paper we study the meaningfulness of the
ordinalcomparison f(a) > f(b). A systematic study of this comparison
for homomorphisms into generalorder relational systems was begunby
Roberts (1984) and continued in Harvey and Roberts (1989). The
purpose of this paper is to generalize the results in those two papers to
cover a much wider variety of situations.
Inthis section we present some of the backgroundmaterial on mea-
surement theory needed for the rest of this paper. In Section 2 we
introduce the general order relations that will be a primary topic of
interest here. We then study the meaningfulness ofordinal comparisons
with homomorphisms into general order relational systems. Section 3
starts with meaningfulness of equality comparisons f(a) = f(b). Sec-
tion 4 begins the study of f(a) > f(b). Section 5 shows how to reduce

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 251-274.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.

the problem in many situations to homomorphisms into special general

order relations called strict and Section 6handles ordinal comparisons
for homomorphisms into strict general order relational systems.
We shall follow the notation and terminology of Roberts (1979).
Suppose A andB are sets andR and S are ra-ary relations on A and B,
respectively. We shall be interested in the systems (A, R) and (B,S),
and sometimes we shall just refer to these systems as m-ary relations.
A homomorphism from (A,R) to (B,S) is a function / from A into B
so that for all a\, ,am from A (notnecessarily distinct),
(1) (ai,...,am) G R<r+ (/(ai), ...,/(a m )) G5.

We shall sometimes interchange the terms scale and homomorphism.
We shall also speak of the representation (A,R) ► (B,S) and the
problem of finding a homomorphism from (A,R) into (B,S) as the

representationproblem. Anadmissible transformation of / isa function
<p : f(A) ► B so that cp o f is again a homomorphism. A statement
involving scales of measurement is called meaningful if its truth or
falsity is unchanged if admissible transformations are applied to all
of the scales or homomorphisms in the statement. This concept goes
back to Suppes (1959) and to Suppes and Zinnes (1963). In some
cases of relations (A,R) and (B,S), not every homomorphism from
(A,R) to (jB,S) can beobtainedfrom every other homomorphismby an
admissible transformation. In this case, Roberts andFranke (1976) call
the representation problem (A,R) into (B,S) irregular and introduce
the following more general concept: a statement involving scales of
measurement is called meaningful if its truth or falsity is unchanged
if all of the scales or homomorphisms in the statement are replaced
by other scales or homomorphisms. This definition is reasonably well
accepted, at least as a necessary condition for 'meaningfulness',and we
shall adopt it here. However, thereare special situations whereit might
not be appropriate. See Falmagne and Narens (1983) for a discussion
and see Roberts (1985, 1989, 1994) and Luce, Krantz, Suppes, and
Tversky (1990) for surveys and general results about the concept of
Suppose / is a homomorphism from (A,R) into (E, S), where E
is the set of real numbers. Many assertions can be made using the
scale /. We shall be specifically interested in the simple assertion that
f(a) > f(b). When is it meaningful? We say that / is of ordinal scale
type if the representation (A,R) into (E,S) is regular (not irregular)
and the collection of admissible transformations of / is exactly the
collection of order preservingmaps from f(A) into E, transformations
(p so that

> X2 ip(xi) > <p(x2).

If / isofordinal scale type, then theassertion f(a) > f(b) is meaningful

for all a, b in A.


In this paper, the primary object of investigation will be a homomor-

phism from an m-ary relation (^4, R) into a general order relation on
the reals. To define general order relations and explain why they are
important, we firstintroduce some preliminary notions.
— ...
Let m be a fixed integer greater than 1andletM {1, ,m}. By
a ranking of M we will mean a linear ordering of Mwith ties allowed.
(Roberts (1979) calls this a weak order.) For example, if m = 4, we will
write 3, 2-A, 1 to represent that ranking of M which ranks 3 highest, 2
and 4 tied for next, and 1last. If a particular ranking of M happens to
have no ties, we will refer to it as a strict ranking of M. The height of
a ranking is thenumber of different levels it contains. For example the
height of the ranking, 3, 2-4, 1 is equal to 3, which we will frequently
denote as ht(3, 2-4, 1) = 3. Note that if n is a ranking of M then
1 < ht(7r) < m and that ht(7r) = m if and only if ir is a strict ranking.
At the other extreme,the ranking 1-2-. -m, whose height is equal to
1, will be called the tie ranking of Mand designatedas c.
If 7r is a ranking of M, wecan use it to define an m-ary relation on
the reals, Tn,as follows:
(xi,...,x m ) GTT iff
(xi > Xj <->■ iis ranked ahead of jin 7r).
For example, if tt is the ranking 3, 2-4, 1, then we have that (14, 23, 87,
23) G Tx while (94, 18, 37, 18) g Tv .
We now define an m-ary relation Son the reals to be a generalorder
relation or GOR if

s=7vlur,r2 u...ur7rp

for rankings tt\,7r2, ir
p .
of M,for some p > 1 We will let 11(5) =
n = {7ri,...,7TP}. If all 7r G n happen to be strict rankings, we will
say that S is a strict generalorderrelation or SGOR.For example, with
m = 2, some common binary relations on the reals qualify as GORs.
For instance, Tj)2 is >, T2)i is <, Ti_ 2 is =,ThU Ti_ 2 is >, Th
2 2U T2)l
is Ti_ 2 U T2)i is <. Of these, >,<, and are SGORs. With'm =3,
UT3)2 i yields theusual 'betweenness' relation onthereals. Thus,
many of therelations of interest in measurement theory arise as GORs.
The GORs arose from the concept of m-point homogeneity which
plays a central role in the theory of scale type and meaningfulness. A
relation S onE is called m-point homogeneousif whene-ver

x\ >x 2> ... >x m and y\> y2 > ...> ym

for real numbers X{ and yi, there is an automorphism <p of (E, S) (one-
to-one homomorphism from (E, S) onto (E, S)) so that <p(xi) = yi,
i= 1, 2, ,m. The literature of m-point homogeneity is summarized
in Luce, Krantz, Suppes and Tversky(1990), Roberts (1989), and Luce
and Narens (1986, 1987). Roberts (1984) showed that if S is m-ary,
then (E, S) is m-point homogeneous ifand only if S is 0 or S isa GOR.
He then pointed out that in the systematic study of the meaningfulness
of ordinal comparisons f(a) > f(b) for homomorphisms /, it was
natural to study thecase where those homomorphisms were into m-point
homogeneous m-ary relations or GORs, and he initiated that study for
the case where S = T*. Harvey andRoberts (1989) studied the same
problem for general GORs and in particular resolved all cases where
p = 2and m *= 2,3. Themain purpose of this paperis topresent results
about this problem in a much more general setting. Homomorphisms
into general order relational systems, systems (E, 5i, ,5&) where
each Si is a GOR,are also ofinterest. However, their study will be left
to a later time.
Inthis paper, we shallbeinterestedin homomorphisms fromrelations
(A,R) whichare m-ary andfor whichR 0 andR Am,where Al7l is
the Cartesian product of A with itself m times; we shallcall such (A,R)
nontrivial. A homomorphism from anontrivial m-ary relation (A,R)
into an m-ary GOR (E,S) willbe called an m-ary GORhomomorphism.
If (E, S) is an SGOR, / will be called an m-ary SGOR homomorphism.

Before studying the meaningfulness of the ordinal comparison f(a) >

f(b), we consider thatof the equality comparison /(a) = /(&). We first
prove a lemma which we will use throughout.

. ...
LEMMA 1 Suppose ir is a ranking of { 1,2, ,m], fis a function
from A into E, |/(A)| > ht(7r) > 2, and a,b are elements of —A with
f(a) f(b). Then there are a\% ,am in A with as =a,at b for
some 5, t € {1, ,m} such that
\{au ...,am}\ =ht(7r)
(/(a1 ),...,/(am))GT7r .
Proof Assume without loss of generality that f(a) > f(b). There
are b u ..., b ht{n) so that (*) /(&i) > ... > f(b ht{7r)). If /(a) is
any f(bi), replace bi by a. If not, let ibe the largest index such that
f{bi) > f(a) or, if there is no such index,let ibe 1. Replace bi by a.
We still have (*). Similarly, we can replace some bj by b and preserve
(*). A complication arises if bj is a. Then replace bj+\ by b, unless
j = ht(ir). In that case, drop b\, lower every index by one and let
b = b ht^y Again, (*) holds. Lastly, permute the bi appropriately and,
at components where 7r declares a tie, throw in a repeat of the corre-
sponding f(bj). This defines a\, ...,a m with the desired properties. ■

THEOREM1. Suppose f : (A,R) -► (E,S) is an m-ary GOR homo-

morphism. Then the statement f(a) = f(b) is meaningful for all
a,b G A.
Proof. Assume there are a, b G A for which /(a) = f(b) is not
meaningful. Then there are homomorphisms g and h from (A, R) into
(E, S) so that g(a) = g(b) and h(a) h(b). Assume without loss
of generality that h(a) > h(b). The analysis now breaks down into
two distinct cases, depending on whether or not the tie ranking c =
1-2-. -m appears in n.
Case 1. c G" 11. If we let k = min{ht(7r) : it G n}, then we
have k > 2. It is important to note that \h(A)\ > k. Otherwise,
for any ai, ,am Gi we would have that (h(a\), ,h(am)) £ ...

S, which would in turn imply (because h is a homomorphism) that

(0i, ,am ) G" R. This contradicts our assumption that R 0 (since
(A,R) is nontrivial). Hence, \h(A)\ > k. By Lemma 1, since \h(A)\ >
k> 2, there must be a\, ,am in A with as =a,at=bso that

(2) \{a1,...,am }\ = k

and (h(a\), ,h(am)) G5. Since /iis a homomorphism, this implies
that (ai,...,am) G i?, which, because gis a homomorphism, in
turn implies that (g(a\), ,g(am)) GS. By definition of k, G =
{g(ai), ,g(am)} has cardinality at least k. However, by (2), it has
cardinality at most k. Hence, it has cardinality k and, in particular,
g{a) 7^ g{b). This is a contradiction.
Case 2. c G n. The proof proceeds much as in Case 1,but now
we let k = min{ht(7r) : it G" n}. Note that k exists because IIdoes
not contain all possible rankings of M. Otherwise, for any a\, ,am
in A, we would have (h(ai), ,h(am )) G S, which would in turn
imply that (ai,...,am ) GR. This would contradict our assump-
tion that R/ Am (since (^4, R) is nontrivial). Since we are assum-
ing that c G n, we know that k > 2. It is important to note that
\h(A)\ > k. Otherwise, for any ai,...,am GA, we would have
... ...
(/i(ai), ,/i(am)) G5,so (ai, ,am) G#, and we wouldbe forced
to conclude that R = A
Because \h(A)\ >k> 2, Lemma 1
implies that there are a\, ...
,am in A with as = a, at = bso that
(2) holds and (h(ai), ,h(am )) 0 5. Since h is a homomorphism,
this implies that [a\, ,am) G" J? and so, since pis a homomorphism,
(p(ai), ,g(am)) &S. Since everym-tuple onEis insome T^ on M,
by definition of k, G = {<?(ai), ,j(am)} has cardinality at least k.
However, by (2), it has cardinality at most k. Hence, it has cardinality
k and, in particular, g(a) g(b). This is again a contradiction. ■

The meaningfulness of the equality statement /(a) = f(b) plays a

fundamental role in the analysis of the measurement process. Specif-
ically, Roberts and Franke (1976) prove that a homomorphism from a
relation (A,R) to a relation (B,S) is regular if and only if the state-
ment f(a) == f(b) is meaningful for all a, b G A. Hence, we have the
following Corollary.
COROLLARY 1.1. Iff : (A,R) -> (E, S) is an m-ary GOR homo-
morphism, then
(a)/ is regular;
(b)f is ofthe ordinalscale type ifandonly ifthe statement f(a) > f(b)
is meaningful for all a, 6 G A.
Proof Part (a) follows since by Theorem 1, f(a) = f(b)>is mean-
ingful for all a, b G A. To prove part (b), suppose that /(a) f(b) is
meaningful for all a,b € A. Thenfor anyhomomorphism g : (A,R) >

(E,S) and any a, b G A,
(3) f(a) > f{b) ~ g{a) > g(b).
Let (p be an admissible transformation of / andlet g = <p o /. Then by
(3), for any a,b G A,
(4) f(a)>f(b)~<pof(a)>tpof(b).
It follows that (/? is order-preserving on f(A). Conversely, suppose </?
is any order-preserving map from f(A) into E Then clearly ip o / is
again a homomorphism from (A,i?) into the GOR (E, S). Hence, <p is
admissible. We conclude that / has ordinal scale type. Next, suppose
that / has ordinal scale type andconsider any homomorphism g = p°f.
Then (4) holds and hence (3) follows so f(a) > f(b) is meaningful. ■


In the previous section, we showed that measurement into a GOR is

of the ordinal scale type if and only if f(a) > f(b) is meaningful for
all a,b G A. Ourmain result in this section, Lemma 2, shows exactly
when meaningfulness fails, and thereby enables us to state conditions
that insure measurementinto aGOR be of the ordinal scale type.
There is one special case which is not covered by the general char-

acterization that is to appear inLemma 2. We deal with the special case
first. Let us say that a homomorphism / : (A,R) ► (E, S) is singular
(ai,...,am) eR-> f{a\) = ... = f(a m ).
The following theorem exhausts the theory of singular homomorphisms.
Note that part (c) generalizes Theorem 4of Roberts (1984).

THEOREM 2. If f : (A,R) -+ (E, S) is a singular m-ary GOR

homomorphism, then
(a) Everyhomomorphism from (A,R) into (E, S) is singular.
(b) The tie ranking c = 1-2-. -m is in 11.
(c) IfUcontains any ranking other than c, then

(5) \f(A)\ < min{ht(7v) : TV eU and TT^e}

(d) |g(A)| > 1for any homomorphism g : (A,R) — ► (E, S).
(c) f(a) > f(b) is meaningless for some a,b G A.
Proof (a) Let qbe any other homomorphism. If (a\, ,am ) G...
.R, we have f(a\) = = f(am), which implies (by Theorem 1) that
g{a\) = ... = g(a0, ).
... ...
(b) Sincei? 7^ there are ai, ,am GA. such that (ai, ,am ) G
#. Then (/(ai ),..., /(am )) G S. Since /(ai) = ... = /(a ),
m we are
forced to conclude that c G n.
(c) Suppose that 7r has minimum height of all rankings in IIdifferent
from c. If the inequality (5) is false, then \f(A)\ > ht(7r) > 2and so by
... ...
Lemma 1 thereare ai, ,am G A such that (f(a\), ,f(am)) G T^.
But then (ai, ,am) Gi?even thoughit is not true that f(a\ ) = ... =
f(am). This contradicts the fact that /is singular.
(d) If \g{A)\ =1, then, by part (b), R = Am, which is contrary to
(c) By part (d), we know that \f(A)\ > 1. Pick c,d G A such that

/(c) > f{d). We now define a mapping g : A » E as follows: for all
a G A, let
r /(c) if f(a) = /(d)
/(d) if /(a) = /(c)
g(a) =
( /(a) otherwise
Note that <j(c) < g(d). Thus,it suffices to show that g is a homomor-
phism. Suppose that (a\, ,am) G i?. Then /(ai) = ... = f(a ).
By the definition of g, g{a\) = = ...
g(am). Hence, by part (b),
(g{ai),...,g(am)) G5. Conversely,suppose that (#(ai), ,g(am )) G
S. Suppose one of the /(a*) is /(c). Define 6j, j = 1, ,m, so that
/(M = gfaj)- Thus,(/(6i),...,/(6m )) G 5 and hence/ (61) = = ...
/(6m). Then for all j,g{a3 ) = f(bj) = f{bz ) = g(ai ) = f(d). Thus,
for all j,/^) = /(c). Itfollows by part (b) that (/(a1 ),..., /(am)) G
S and hence that (ai, ,am) GR. The proof is similar if one of the
f{ai) is /(d). If none of the /(a*) is /(c) or /(d), then (/(a^) = /(%")
.. ...
for all j. Thus, (/(ai),. ,/(am)) = (a(ai), ,#(am )) G S and so
(0'i,... ,am ) G .R. ■
If we stay away from singular homomorphisms, we have the following

LEMMA 2. Suppose f : (A,R) > (E, 5) isanonsingularm-ary GOR
homomorphism. Then the following are equivalent:
(a) /(a) > /(&) w meaningless for some a, 6 G A.

(b) 77z£re w a homomorphism g : (A,i?) > (E, S) and there are
a\ , ,am G A such that

(6) (/(ai),...,/(am))eTT. and (g(ai)t ...,g(am )) G TV,

for some ni, 7Tj G n, 7Tj 7^ ttj.

Proof. From the definition of singularity, we know that there must be
6i,...,6m GA such that (6b ..., 6m) G i?and |{/(&i),...,/(&m)}| >
1. Since /is a homomorphism, (f(b\),...,f{b'm)) G S and so
(/(&i),..., /(6m)) € 2>. for some Gn. Note that
ht(7Tz ) = |{/(61),...,/(6m)}|>l,
and in particular \f{A)\ > ht(7Tj). Now assume (a), and in particular

assume that there is another homomorphism g : (A,R) ► (E, S) so
that for some a, b G A,
(7) f(a) > fib) ~ g(a) > gib)
fails. By Theorem 1, we know that f(a) / f{b). By Lemma 1, there
are a\, ,am G A with as =a,at = b for some s, t G {1, ,m} ...
.. ...
and (/(ai),...,/(am)) € TT Hence, (/(ai), ,/(am)) G 5 and
(ai, ,am ) G i?. Since gis also a homomorphism, this implies that
.. ..
(g(ai), g(a m)) G S and, in particular, that (g(ai), g{a m )) G
Ttt for some -kj G n. Moreover, since (7) fails, 7r; 7^ ttj. This
establishes part (b).
Next,weassume (b). Since 7^ 7Tj, theremust be r,s G {1, ,m} ...
which are ranked differently by 7r; and ttj. Hence, (7) is violated with

a= ar and b = as Part (a) follows.

Iffor some pair of m-ary GOR homomorphisms / and g from (A,R)

to (E, S) and ai, ... ,am in A, (6) holds, we will refer to 7r7; and -kj as
co-rankings. Although the condition in part (b) is difficult to verify
directly, it can be used in the following way: if it can be shown that
it is impossible to have a iti and -kj which are co-rankings, then we
are guaranteed that f(a) > f(b) is meaningful for all a, b G A. As a
simple example, consider the following, which is essentially the same
as Theorem3 ofRoberts (1984).

THEOREM3. Suppose f : (A,R) ► (E,S) is a nonsingular m-ary
GOR homomorphism and \U\ =1. Then f(a) > f(b) is meaningfulfor
all a, b G A.
Proof. Since n contains only a single ranking, it cannot contain
co-rankings. ■

The next theoremuses the same approach and subsumes Theorems 4

and 7 of Harvey and Roberts (1989). Suppose is an equivalence
relation on {1, ,m}. We shall call this the signature of a ranking
7r if i jexactly when iand j are tied in tt. Thus, for example, the
rankings iti = 3-4,5, 1-2and -kj = 1-2,?>-A, 5 have the same signature,
while the rankings 7T; = 3, 2-1 and -Kj = 3, 2, 1do not.

LEMMA 3. Ifthereis an m-ary GOR homomorphism from (A,R) into

(E, S), then co-rankings have the same signature.
Proof Let 7r; and ttj be co-rankings. Since / and g are homomor-
phisms, Theorem 1tells us that forall r, s G {!,».. ,m},/(ar ) = f(as )
if and only ifg(ar ) = g(as ). Hence, 7^ and 7Tj have the same signature.

THEOREM 4. Suppose f : (A,R) * (E,S) is a nonsingular m-ary
GOR homomorphism andno two distinct iri and ttj in Hhave the same
signature. Then f(a) > f(b) is meaningfulfor all a,b G A.
Proof By Lemma 3, there areno co-rankings. Hence, theconclusion
follows by Lemma 2. ■
Lemma 3 can actually be used to yield an even stronger result. We
begin by partitioning n according to signature. Let n be partitioned
into disjoint, nonempty sets Hi,...,n„ where all 7Tj, itj G Ilk have the
same signature and whenever tt, and -Kj have the same signature, they
are in the same n^. For all iG {1, ,n}, let
Sz = U{2V : 7T G Ui}.
For example,suppose that
(8) S — Ti,2-4,3-5 U25,2,1-3-4 U22,5,1-3-4 UT2_4,3_5,i
U 22_5,i,3,4 U T2_5,i,4,3 U 22-5,3,4,1
= T^UT„
2 UT^U2V4 U2V5 U2V6 U2V7
Then there are three different signatures here, a partition of n is
lii= {7Ti,7r4 }, n2 = {7r2,7r3 }, n3 = {7r5,7r 6,7r 7},
(9) S\ = 21,2-4,3-5 U 22-4,3-5,1 >
(10) 52 = 25,2,1-3-4 U 22,5,1-3-4 >
(11) 53 = 22_5,1,3,4 U T2-5,1,4,3 U T2_5,3,4,l .
The partition of n together with the homomorphism / induces a
correspondingpartition {R{, ,R^} of R, where

(12) (0i,...,dm) €R{ ~ (/(ai),. "" ,f(am)) G SZ.5 .


(Note that R{ could be empty even if all the n^ are nonempty). By

(12), /is a homomorphism from (A, RJ) into (E, Si ) for alli. We shall
show in the next theorem that R( is the same for all homomorphisms /
from (A,R) into (E,S). The next theorem also shows how the question
of the meaningfulness of /(a) > f(b) canbe decomposedaccording to
signature. It says, essentially, that f(a) > f(b) is meaningful for all

a, b G A in the representation (A,R) > (E, S) if it is meaningful for
all a, b G A in some one of the representations (A, Rj) —> (E,Sj ).

THEOREM 5. Suppose f : (A,i?) -> (E,5) w on m-ary GORhomo-

morphism. Then

(a) Ifg : {A, R) * (E,5) is another homomorphism, then g is also
a homomorphism from (A, ) into (E, Si ) for alii G { 1, ,n}
RJ ...
and, moreover, R^ — Rf.
(b) Suppose there is j G {1, ,n} so that whenever h\, h2 are
homomorphisms from {A,R?) into (E, Sj), thenfor all a,b G A,

(13) hx (a) > hx {b)

~ h (a) >h (b).
2 2

Then f{a) > f(b) is meaningfulfor all a,b G A.

Proof, (a) Suppose (0i, ,am) G R{, some i. Then (f(a\), , ...
/(flm)) G sj, so (/(ai),...,/(am)) G2V for some tt G n,and(/(ai),
"" " , f(a>m)) G5. It follows that (a\, ...
,am) G -R, so {g{a\), , ...
g{am)) G 5 and therefore, for some j, (g{a\), ,g{am)) G2V for...
some 7r' Gnj and (<?(ai), ,g(am )) G <Sj. But then, by Lemma 3,
7r and 7r' must have the same signature. Thus, by the definition of the
partition, imust equal jand we conclude that (g(a\), ,g{am )) G Si....
Conversely, suppose that (g(a\), ,g(am)) G S». By reversing the
direction of the previous argument, one shows that (a\, ,am) G ...
Rj for some j, and in particular that (#(ai), ,g(am)) G T^/ and
(/(ai), ,/(am)) G Ttt for 7r G n^- and tt' G n^. Again, by using
Lemma 3, one shows that i— j. Hence, (ai, ,am ) G ...
Thus, we i?f .
have shown that g is a homomorphism from (A,R() into (E, Si). By
this conclusion and definition ofRf wehave
(ai,...,am) GR\ <-► (tf(ai ),...,tf(am)) GSi
<-► (ai,...,am) G i?f,
so R{ = Rf.
(b) Let jbe as in the theorem, andlet g be any homomorphism from
(A, R) into (E,5). By (12) and part (a), both / and g are homomor-
phisms from (A, Rj) into (E, Sj ). Thus, by the hypothesis of part (b),
(13) holds for all a, b G A with h\ = f and h2 =#. This implies that
/(a) > f(b) is meaningful for all a,b c A. ■

Leti?7 denote the common R{j — Rj- of Theorem5.

COROLLARY 5.1. Suppose f : (A,R) -> (E,5) is an m-ary GOR
homomorphism. Then f(a) > f(b) is meaningful for all a,b E A

iffor some j andfor some homomorphism g : (A,Rj) > (E,Sj),
g(a) > gib) is meaningfulfor all a, b G A.

It should be noted that the converse of Corollary 5.1 is false. Let

A = {x,y,z}, R = {(x,x,y J z),(x,y,y,z),(y,y,x,z),ix,z,z,y)}
and 5 = =lTni where tti = 1-2, 3, 4, tt2 = 3, 1-2, 4, tt 3 = 1, 2-3, 4,
and 7T4 =1, 4, 2-3. Then Si = 2V, U 2V2 ,52 = 2V 3 U 2V4 ,f(x) = 100,
f(y) = 50, f(z) = 0is a homomorphism from (A,R) into (E, 5), R\ =
{(x,x,y,z),(y,y,x,z)}, and R2 = {(x,y,y,z),(x,z,z,y)}. Now
any other homomorphism g from (A, R) into (E, 5) has g{x) > g(y) >
g(z), so the statement f(a) > f{b) is meaningful for all a, 6 G A If
h(x) = 50, h(y) = 100, /i(z) = 0 and k(x) = 100, k(y) = 0, fc(z) = 50,
then / and h are homomorphisms from (A,R\ ) into (E, Si ) and / and
k are homomorphisms from (A,R2 ) into (E, S2). Since f(x) > f(y)
while h(x) < h{y), the statement g(x) > g(y) is meaningless for all
homomorphisms g from (A,R\) into (E, S\ ). Similarly, since f(y) >
f(z) and k(y) < k(z), the statement g(y) > g(z) is meaningless for all
homomorphisms g from (A,R2 ) into (E, S2).
Next, we note that Theorem 4 is an immediate consequence of The-
orems 3 and 5 because,if distinct -Xi and ttj in ncannot have the same
signature, then each consists of a single ranking.
More importantly, Theorem5 andits Corollary suggest the following
approach to the meaningfulnessquestion. When faced with an arbitrary
(E,S), a good first step is to partition n by signature, and carry out
the analysis on each of the (E, Si) separately. If meaningfulness of
f{a) > f(b) for all a,6 G A can be demonstrated for any one of
these for homomorphisms from {A,Ri) into (E, Si), then it is also
demonstrated for the original GOR. What is especially attractive about
this approach is that each Si uses only rankings of the same signature,
and can essentially be treated as an SGOR. For instance, the GOR Si
of (9) can be handled by studying
(14) S?=rw UT2>3,i,
which is arrived at by merging the second and fourth positions and the
third and fifth positions in the original Sj. Similarly, the GOR S2 of
(10) can be handled by studying
(15) S2 = T3,2,iUT2,3,i

and the GOR 53 of (11) by studying

(16) S3* = 22,1,3,4 U22,1,4,3 U22,3,4,1 .

We discuss thisidea in detail in the next section.


Suppose ~ is an equivalence relation on {1,...,m} and there are

q equivalence classes. Consider a set A. Given ai,...,am GA,
define (ai, ,am)* by dropping all components but the first com-
ponent from each equivalence class of subscripts. For instance, if
m = 9 and has equivalence classes 1-4, 2-3-6, 5, 7-9, and 8,
... ...
then (ai, +,ag)* is (ai,a2,a5,a-],aB). Given a\, ,aq GA, define
... ...
(ai, ,aq ) byinsertingcopies of theelements a\, ,aq in the prop-
er places to obtain an ra-tuple with signature ~. For example, with
~ +
as above, (b\,b2,hMM) is {hMMMMMMMM). The
... ... +
notations (ai, ,am)* and (ai, ,aq ) are ambiguous since they
dependupon the equivalencerelation ~. However, in what follows,
will always be clear.
If (C, S) is any m-ary relation, define S* on C by

(17) 5* = {(ci,...,cm)* :(ci,...,cm)GS}.

For example, if S3 is defined by (11) and ~is the common signature of
the 71-j's defining S3, then S3 of (17) is givenby (16). In general, as in
this example, if (E, S) isaGOR in which all itj have the same signature
~, then (E,S* ) is an SGOR. (The exception to this statement is when
q = 1, i.e., the common signature of all the ttj declares all elements
ofMtied. In this case, (E, S* ) is a1-ary relation,and weare not calling
1-ary relations SGORs or even GORs.) If (D,Q) is any g-ary relation,
q+ = {(<*!,...,<y+ (dv ...,dq )eQ}.
Note that if ai =aj whenever i~ j,and in particular if [a\ , ...,a m )G
2V for 7r of signature ~,then
(18) (ai,...,am)* = (ai 1Q"m)-
Thus,if (C, S) is anyra-ary GORin whichall -Xi have the same signature
~, then S* =S. Also, we always have
(19) (au ...,ag ) = (ai,...,a9 ).
THEOREM 6. Suppose S is a GOR in which each -Xi has the same
signature with q > 1 equivalence classes and f is an m-ary GOR
homomorphism from (A, R) into (E, S). Then ifß* andS* are defined
from ~, thefollowing hold:
(a) Any homomorphism from (A, 22) into (E, S) is also a homomor-
phism from (A,R*) into (E, S*).
(b) Any homomorphism from (A,R*) into (E, S* ) w also a homomor-
phism from (A,22) into (E, S).
(c) /(a) > f(b) is meaningful for all a,b G A in the representation

(A, 22) > (E,S) ifand only iff(a) > f{b) is meaningful for all

a,b € Am the representation (A,R*) ► (E, S* ).
Proof, (a) Suppose that g is a homomorphism from (A, 2?) into
(E, S). Let ai, , a 9 be elements of+A. Suppose that (g{a\), , ...
g{aq )) G S*. Then (g(ai), ..., g(a q )) G S since S*+ =S. Since
g is a homomorphism from (A,2?) into (E,S), we conclude that (ai,
... , aq)+ G22 and hence (ai, , aq) G 22* by (19). Conversely,
... +
... ..
suppose that (a\, ,aq ) G R*. We first show that (0i, .,aq ) GR.
... ...
Since (a\, , aq ) G R*, there must be (&i, ,bm ) GR so that (&i,
... ... ...
, &m )* = (ai, ,aq). Then (/(&i), , f(bm)) G S since /is
a homomorphism from (A, R) into (E,S), and so we must have that
/(&i) = /(&j) whenever t ** jbecause Sisa GORin which each+7Ti has
signature ~. Hence, it must also be true that (/(ai), ,f(aq )) G S
... + ...
and then (ai, ,aq) G i?, as desired. Now, (g(a\), ,g(aq )) G S
because g is+a homomorphism from (A, i?) into (E, S). Hence, (#(ai),
... , g(aq)) * G S*. By (19), (^(ai), , g(aq)) G S*. Hence, we
have shown that g is a homomorphism from (A, 22*) into (E, S* ).
(b) Suppose that g is a homomorphism from (A, jR*) into (E, S*).
Suppose that (<xi, ...
, am ) G -R. Then by definition of 22*, (ai, ,
am)* G 22*, and since gis a homomorphism from (A, 22*) into (E, s* ),
(0(ai),...,0(am))* G S*. Then(£(ai),...,£(a m))*+ G S*+ S,
the equality by our earlier observation. To prove that + ig(a\), ...,
g(a>m)) GS, it suffices to show that (g(ai), ..., m ))* = (g(a\),
■ ■" i g(cbm))- This follows by (18) if we can show that g{ai) = #(%")
whenever i~j Note that since /isa homomorphism from (A, 22) into

— ...
(E, S) and since {a\, , am) G 22, we have (/(ai), ,/(am)) G S
~ ...
and so f{ai) f{af) whenever i j. By part (a), / is a homomor-
phism from (A, 22*) into (E, S* ). Applying Theorem Ito / and g, we
conclude that g(ai) = g{aj) whenever i /, as needed. (To apply
Theorem 1, we need to show that R* / 0 and 22* .
Aq The former
follows since 22 / 0. To see the latter, note that since q > 1, there is an
iso that 1 idoes not hold. Pick ias small as possible. Then {a\, , ...
am) G22 implies a\ ai and so (ai,ai, ,a\) is not in 22*.)
Tocomplete theproof of part (b),suppose that {g{a\),...,g(am)) G
S. Then (g(cn), ..., g(am))* GS* and so (a\ , ,am)* G 22* since g
is a homomorphism from (A,22*) into (E, S* ). Since {a\ , ,am )* G ...
... ...
22*, there is {bu ,6m) G22 such that (&i, ,bm)* = (ai,... ,am)*.
Note that (/(&i), ,f(bm)) G S since / is a homomorphism from
(A, 22) into (E,S). We next show that
(20) (/(&i),...,/(6m )) = (/(ai),...,/(am)).
To see why,note that since (bi, ...,b m )* = (ai, ...,am )*, we have
(21) (/(61);...,/(6m)r = (/(0,),...,/(om))*.

(22) f(bi) = f(bj) whenever i~ j

because (/(6i),..., /(6m)) G Sand
(23) g(ca) = g(o>j) whenever i~j
because (^(ai), ,g(am )) GS by hypothesis. We can now apply
Theorem 1 to g and /. (By part (a) and the hypothesis of part (b), both
are homomorphisms from (A, 22*) into (E, S*); andas above, 22* / 0
and 22* / Aq.) By Theorem 1, wecan conclude from (23) that
(24) f(di) = f(aj) whenever i~ j.
Now (21), (22), and (24) yield (20), as desired. Since (/(&i), ...,
f(bm)) GS, we have (/(ai), "" " , f(o>m)) GS. Since /is a homo-
morphism from (A, 22) into (E, S), weconclude that (a\ , ,am) G 22,
as required. We conclude that # is a homomorphism from (A, 22) into
(c) This follows trivially from parts (a) and (b).
Let us apply Theorem 6 to the example (8). From (9) and (14),
we note that we can reduce the question of the meaningfulness of
f{a) > f(b) for a homomorphism / from (A,22i) into (E, Si) =
(E,Ti,2 _4,3-5 U 22_4,3-5,i ) to the same question for / treated as a
homomorphism from (A,22*) into (E,S*) = (E,Ti,2,3 UTyj). A
similar conclusion holds for S2 and S3 .

In this section we examine the meaningfulnessof the statement f{a) >

f{b) in the special case where Sis an SGOR, i.e.,consists of only strict
A strict ranking is nothing more than a permutation of the elements
of M. We can think of such a ranking as a mapping which assigns to
each preference level aparticular position in the ra-tuple. For example,
the strict ranking -k = 3, 1, 4, 2 can be thought of as the permutation
7T : M -* M for which tt(1)
= 3, vr(2) = 1, tt(3) = 4, 7r(4) = 2. The
interrelationships between therankings will also play acriticalrole. For
each, i, j G {1,. ,p}, where p = \U\, we define a^ :M — ► Mby
° = 7Tj, i.e.,

Oij(k) = 7ij(7r. \k)).

For example, if 7Ti = 2, 4, 1, 3 and 7Tj = 4, 3,1,2, then
7Ti(l)=2, 7T,(2)=4, 7Ti(3) =l, (4)=3,

=4, 7^(2) =3, =!, =2,

and Gij is givenby
o-ij(l) =1, aij(2) =4, <Tij(3) =2, o-y(4) = 3
Note that, in general, it is always true that a^ is the inverse of Oji and
that on is the identity map. For what is forthcoming, we will find it
useful to define Sj to be the following set of permutations:
Si = {(Tij : j = !,...,£>}.
Inparticular, note that each £i contains an, the identity map. Note also

(25) (xi,...,x m) 6 2V, <-> (^-"(i),.-.,^-!^))cTnj

Tosee why (25) holds,let (yu ,ym)= {xa-i{l),..., (m) ). We
wish to show the following: for all k,
(26) vis kthin 7Ti and x v is fcth highest among x\, ...,x m

holds ifand only if

(27) vis kth in 7Tj and ?/w is A;th highest among yi,...,ym

holds. Note that o~j = Oji. Suppose that (26) holds. Thus, 7Ti(A;) =
v. Suppose v = 7Tj(k). Then cr~j l(v) = crji('u) = Oji(7Tj(k)) =
7Ti(7r~ (7Tj(A;))) = 7Ti(A;) = v, y v is x v and is kth highest among
2/b " "" > 2/mI (27) follows. Conversely,suppose (27) and suppose v
TTi(k). Then, again as above, cr~j iv) =v,and so y v is xv and xv is kth
highest among x\, ,xm\ (26) follows. Hence, wehave (25).
Itis straightforward to seethat ifaisa permutation of { 1, ,m] and...
for some (x v ...,xm) G T^, wehave (vi(i)v,V(m)i € 2V,-,
then cr must be cty.
The next few results demonstrate the importance of the &ij and the
Sj in the analysis of SGORs.

LEMMA 4. Iff: (A,R) »" (E,S) isan m-ary SGOR homomorphism,
7Ti w m 11/ S, a«J (f{a\), ,f(am)) G T^, ?/ze«
a G Si <-► (a -i(!),...,a -i( )) G 22.
a CT m

Proof. If a G £j, then there is j G {1,... ,p} such that 7Tj GII for
S and cr= c^-. But then, by (25),
(/(aa-i(l))'--''/( a^-i(m)))

Hence, (/(ao-i(i)), "" " an^ since / is a homomor-
phism, (aa-i(1) ... aa-ifm)) G 22.
, ,
Conversely,if cr 0 Ei, then for all j G {1, ,p},by theobservation
right after the verification of (25), we must have

(/(a<r-'(l))> " "> /(V'(m))) 0 2Vr

Hence, if(aa-l{l)), ..., /(aa-i(m))) G" S and (c^-i^, ...,a a-wm) )
0 22. ■
The nexttwo results echoLemma 3 and Theorem 4.

LEMMA 5. If S is an SGOR and 7Ti and ttj are co-rankings, then

Zji =
Proof. Let / and g be two m-ary SGOR homomorphisms from
(A, 22) into (E,S) satisfying (6). Then by two applications ofLemma 4,
we have
oG Sz <-► (aa-i(1), ...,(V-i(
m )) G22 <-+ aG Sj
THEOREM 7. Suppose f : (A, 22) -> (E, S) is an m-ary SGOR
homomorphism. If there is an iso that Si Sj for all j i, then
f(a) f(b) is meaningful for all a,b G A.
Proo/ Let g be any other homomorphism from (A,22) into (E, S)
and a, 6 be in A. By Theorem 1, we need only consider the situation
where f(a) f(b) and g(a) g(b). Since 22 0, there is (&i, , ...
b m) G 22. Hence, (/(&i), .., .
f(bm)) G 2V fc for some k. Since 7rfc
is a strict ranking, we conclude that |/(A)| > m ='ht(7rjt) = ht(7Ti).
Since m > 1throughout this paper,Lemma 1 implies that there are a\,
... , am in Aso that as =a,at = b and (f(a\), ..., f(am)) G TXi .
Hence, (ai, ...
,am) G 22, so ig{ax ), ..., g(am )) G S and (g(a\), ...,
g(a<m))— G 2V. for some j. Since Si Sj for z/ j, Lemma 5 implies
thati j. Thus,
fia) > fib) «-> g{a) > gib),
as desired.

As an application of this result, consider the special case of Theorem 6

of Harvey andRoberts (1989) where S = 2V, U 2V2 with tti = 1, 2, 3
and 7T2 = 2, 3, 1. Then Sj = {crn,cr^}, where a\\ is the identity map
and (Ti2 is givenby
<712(1)=2, (712(2) =3, (3) =1,

and S2 {(72i, 022}, where is the identity map and

021 is given by
0-21(1) =3, (7
2l(2) =l, (7
2i(3) =2.

Since Si S2, weconclude that if /is an m-ary GOR homomorphism

from (A, 22) into (E, S), then /(a) > /(&) is meaningful forall a,b G A.
As a second application of Theorem 7, suppose that S is givenby
(8) and / is an m-ary GOR homomorphism from (A, 22) into (E,S).
If Si is given by (9) and is the common signature for rankings for
Si, then S* is givenby(14). By definition, / isa homomorphism from
(A, 22i) = (A, 22{) into (E, Sj"). It is an m-ary GOR homomorphism.
To see why, note that since it is an m-ary GOR homomorphism from
(A, 22) into (E,S), 22 Am, and so R{ C R Am If 22i =0, .
then by (9), |/(A)| < 3. Hence, by (8), R = 0. Since all m in
Hi have q > 1 equivalence classes, Theorem 6 implies that / is a
homomorphism from (A, 22|) into (E, Sjf ). By (14) and the discussion
in the previous paragraph, /(a) > fib) is meaningful for all a, b G A
for the representation (A,R*) * (E, S* ) and thus, by Theorem 6, also
for the representation (A, 22i ) » (E, Si ). Note by way of contrast that
if S2 is givenby (10), then S| is givenby (15). For SJ, Si = S2 so
Theorem7 does not apply. In fact, by Theorem5 of Harvey andRoberts
(1989), forall m-ary GOR homomorphisms g from iA,R%) into (E, SJ ),
theconclusion gia) > gib) is meaningless for some a, b G A so long as
\giA)\ > 2 and |#(A)| < 00. However, we still have meaningfulness
for m-ary GOR homomorphisms from (A,22) into (E, S) for S of (8)
since Corollary 5.1 tells us that we need to find only one Si for which
we have meaningfulness and Si has this property.
The binary relation SJ of Equation (15) gives us a counterexample
to the converse of Theorem 7. As noted above, Sj = S2. We can
easily find a set A and binary relation 22 on A such that 22 0, 22 7^
Am, and there is a homomorphism / from (A, 22) into (E, SJ) with
/(A) unbounded. (Thus, in particular, |/(A)| = 00 and Theorem 5
of Harvey and Roberts (1989) does not apply.) Now given a, b in A
with /(a) > fib), there is a c in A such that /(c) > /(a) > fib).
Hence, (/(&),/(a), /(c)) G r3)2)i g SJ, so (&,a,c) G 22. But then,
for any homomorphism g from (A, 22) into (E, SJ), (#(£>),#(&), #(c)) G
SJ =>23,2,1 U 22,3,1, which implies that gia) > gib). We conclude that
fia) f(b) is meaningful for all a, b in A.


In this paper, we have studied the meaningfulness of the ordinal com-

parison fia) > fib), for all a,b c A, for m-ary GOR homomorphisms

/ : (A, 22) —♥ (E, S). We have given necessary and sufficient condi-
tions for meaningfulness in many situations, though in a few situations
wehave just given sufficient conditions.
Tosummarize the situation,Theorem 2 settlesall the situations where
/ a singular homomorphism. Theorem 4 settles all the situations
where / is nonsingular and all the 7Ti have different signatures.
In Theorem 5 (Corollary 5.1), we consider the case of / with some
signatures allowed to be the same and partition n into sets of rankings
of the same signature and, correspondingly, let Sj be the union of TlTi
for the rankings 7Ti of the jth signature. We study meaningfulness
of ordinal comparisons for homomorphisms into (E, Sj ) and note that
meaningfulness in one of these cases implies meaningfulness for the
whole case. However, the converse of this statement is false, and
exactly how to handle the situation when meaningfulness fails in all of
the homomorphisms into (E, Sj ) is still an open question. The question
of determining meaningfulness in each suchcase (E, Sj ), i.e., where all
the rankings in S = Sj have the same signature, is handled by reducing
to SGORs.
Theorem 6 gives a reduction to the SGOR situation. Theorem 7
gives a sufficient condition for meaningfulness if the homomorphism
is into an SGOR. However, the converse of this Theorem is false. We
are left with the problem of handling meaningfulness if there is no ifor
which Si Sj for all j i.
As we have noted, the m-ary GORs correspond to m-point homoge-
neous m-ary relations. The theory of the meaningfulness of the ordinal
comparisons /(a) > fib) remains to be systematically developed for
homomorphisms into other kinds of relations or relational systems.
Some special cases of interest with which to start wouldbe homomor-
phisms into m-point homogeneous n-ary relations where m n; and
homomorphisms into relational systems (E, 5i, ,S&) where each Si
is an rrii-ary GOR.


This paper is dedicated to Patrick Suppes, on the occasion of his 70th

birthday and hisretirement from Stanford University.


FredRoberts acknowledges the support of the National Science Foun-

dation under grant number IRI-89-02125 to Rutgers University. Both
authors thank Denise Sakai, Shaoji Xv, and Wenan Zang for their help-
ful comments.

Fred S. Roberts,
DepartmentofMathematics andCenter for Operations Research,
Rutgers University,
New Brunswick,NJ 08903, U.S.A.
Zangwill Samuel Rosenbaum,
Department ofMathematics and Computer Science,
Wilkes University,
Wilkes-Barre, PA 18766, U.S.A.


Falmagne, J.-C. and Narens, L.: 1983, 'Scales and Meaningfulness of Quantitative
Laws', Synthese, 55, 287-325.
Harvey, L. H. and Roberts,F. S.: 1989, 'On the Theory of Meaningfulness of Ordinal
Comparisons inMeasurement ll',Annals N.Y. Acad ofSci., 555, 220-229.
Krantz, D. H., Luce, R. D., Suppes, P, and Tversky, A.: 1971, Foundations ofMea-
surement, Vol. I,New York: AcademicPress.
Luce, R. D., Krantz,D. H., Suppes, P., and Tversky, A.: 1990, Foundations of Mea-
surement, Vol. 111, New York: AcademicPress.
Luce,R. D. and Narens, L.: 1986, 'Measurement: The Theory of Numerical Assign-
ments', Psychol. Bull.,99, 166-180.
Luce, R. D. and Narens, L.: 1987, 'MeasurementScales on the Continuum', Science
236, 1527-1532.
Roberts,F S.: 1979, Measurement Theory, withApplications toDecisionmaking, Utility,
and theSocial Sciences, Reading,MA: Addison-Wesley.
Roberts, F. S.: 1984, 'On the Theory of Meaningfulness of Ordinal Comparisons in
Measurement', Measurement, 2, 35-38.
Roberts,F S.: 1985, 'Applicationsof the Theory of Meaningfulness to Psychology', J.
Math. Psychol., 29, 311-332.
Roberts, F S.: 1989, 'Meaningless Statements, MatchingExperiments, and Colored
Digraphs (Applicationsof GraphTheory andCombinatorics to the Theory of Mea-
surement), in: F S. Roberts (Ed.), Applications of Combinatorics and Graph
Theory in the Biological andSocial Sciences, IMA Volumes inMathematics and
Its Applications, Vol. 17, New York: Springer-Verlag, pp. 277-294.
Roberts, F S.: 1994, 'Limitationson Conclusions Using Scales of Measurement', in:
A. Barnett, S. M. Pollock, andM.H. Rothkopf (Eds.), Operations Researchand
Public Systems, Amsterdam: Elsevier,pp. 621-671, inpress.
Roberts,F. S. and Franke, C. H.: 1976, 'On theTheory ofUniquenessin Measurement',
J. Math. Psychol., 14, 211-218.
Scott,D. and Suppes,P.: 1958, 'FoundationalAspects ofTheories of Measurement', J.
SymbolicLogic, 23, 113-128.
Suppes, P.: 1959, 'Measurement, Empirical Meaningfulness and Three-Valued Log-
ic', in: C. W. Churchman and P. Ratoosh (Eds.), Measurement: Definitions and
Theories, New York: Wiley, pp. 129-143.
Suppes,P. and Zinnes, J.: 1963, 'Basic Measurement Theory', in: R. D. Luce, R. R.
Bush, and E. Galanter (Eds.), Handbookof MathematicalPsychology, Vol. 1, New
York: Wiley, pp. 1-76.


The study of the meaningfulness of a statement when various measure-

ments are referred to in the statementhas receivedconsiderable attention
in various articles in these three volumes. Several of the outstanding
problems formulated by Luce and Narens, for example, are concerned
with meaningfulness. The article by Roberts andRosenbaumis a con-
tinuation of a seriesof papers by Fredandhis colleaguesthatprovide one
of the most thorough studies of meaningfulness to be found anywhere,
namely, of the meaningfulness ofordinal comparisons as representedby
anumerical function. This means thatthe papers are concerned with the
meaningfulness of the ordinal numerical comparison /(a) > fib). In
the early days of meaningfulness theory, it wasassumed that the mean-
ingfulness ofordinal comparisons wasa straightforward andcompletely
elementary, indeed, almost trivial, exercise. Some thirty years later it
is clear that complexities liebeneath the surface evenin what appear to
be the simplest cases of meaningfulness.
The analysis ofRoberts andRosenbaum on these matters is thorough
indeed. Ionly regret that they did not amplify their detailed discussion
withadditional details ofactual scientific examples wherecontroversies
about meaningfulness have arisen,whichis certainly thecase forordinal
comparisons subject to statistical analysis.
It wouldalsobeinteresting to see how their results work for theaffine
order relation of betweenness which they mention briefly at one point,
and also for the projective order relation of separation which has its
representation theorem formulated ordinarily in terms of the projective

concept of the crossratio of thenumerical representationof four points.

The familiar condition being that distinct points a, b, c, d stand in the
separationrelation ab Scd if and only if the cross ratioof a, b, c, d with
respect to a numerical representation / is negative. (The cross ratio is
f(a) ~ /(c) fjb) - fjd)
fia) -fid) fib) -fie)
where we have to extend thenumerical operations on the real numbers
to include infinity -the details are given in Section13.2 ofFoundations
of Measurement, Vol. II.)
Still another direction of interest mentioned in my comments on
Luce and Narens, is the introduction of topological ideas where the
requirement is made that the order relation be continuous, the kind of
assumption usedcritically in someof Debreu's (1954) well-known work
on ordinal utility. Havinginmind application to theutility of commodity
bundles in a space ofann-dimensional Euclidean space,Debreu proves
a theorem of a more general character, i.e., for a topologicalspace that is
connected, separable and for which there isa complete order relation of
preference. An assumption of continuity is oftenused inessential ways
ineconomics. It would be useful to enlarge the study of meaningfulness
to include preservation of continuity.
Theselast remarks are in the spirit of remarks I havemade to several
of the measurement papers, to the effect that we needmore extensions
of concepts from the theory of measurementto the rich array of exam-
ples and applications of the same general ideas that may be found in


Debreu,G.: 1954, 'Representationof a Preference Ordering by a Numerical Function',

in: Thrall,Coombs, andDavis (Eds.), DecisionProcesses, New York: John Wiley
& Sons.
Suppes, P., Krantz,D.,Luce,D., and Tversky,A.: 1989, Foundationsof Measurement,
Volume11, San Diego: AcademicPress.



ABSTRACT. The firstpart ofthis essay deals with the generalproblem of identifying
empiricaltheories; an identity criterionfor theories as nets ofmodelclassesis proposed
whichmaybe seenas an expansionof Suppesianideas onthisissue. In the secondpart,
the criterionproposed is applied to the particular case of combinatorialmeasurement
theory. This theory is represented as a tree-like net of eighteen so-called model-
elements. Some of them are only implicitly (or not at all) containedin the literature
on measurement, especially in Foundations of Measurement. The corresponding set-
theoretical predicates are precisely defined for each case and some new results are
proven. A particular innovationof the presentreconstructionof measurement theory is
the systematic treatment ofso-callednull and limit elements.

Philosophy of science owes Pat Suppes a number of far-reaching in-

sights. One of them concerns the identity of empirical theories, another
the foundations of measurement. The two are interrelated (though not
in an immediately obvious way) and havebeen very influential on sub-
sequent work. Suppes' views on the identity of theories have led to
a family of approaches that have come to be known as 'the semantic
view of theories'; more specifically, it has inspired a program in the
formal reconstruction of empirical theories usually called 'the struc-
turalist view. Now, there may be some quarrel over the question of
whether structuralism belongs to the semantic approach or not. In our
own view, this is a rather uninteresting quarrel about words. There
are certainly some significant differences between the ideas about the
nature of empirical theories propoundedby,say, Bas vanFraassen and
Frederick Suppe on the one hand, and those of Joe Sneed and Wolf-
gang Stegmuller on the other. But, in our opinion, the differences are
overarched by some convergent views on the identity of a theory at a
very fundamental level -one might say: at thelevel of 'ontology'. This
shared metatheory of empirical theories may bestbe summarized by the
following tenets:

P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher, Vol. 2, 275-299.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.

(1) An empirical theory is not a linguistic entity;more concretely, it is

not just a set of axiomatic sentences ora conjunction of them. Itis
rather an abstract, mathematical sort of entity.
(2) The most basic component of a theory's identity is a class of
structures, andmore specifically aclass ofmodels inTarski's sense.
(3) Typically, this class of models is not a singleton, not evenaclass of
isomorphic structures. Categoricity is never to be found in devel-
oped empirical theories, nor is it a desideratum. The desideratum
for empirical theories is rather to have a genuinely heterogenous
class of empirical applications.
(4) The most convenient way to pick out the class ofmodels essentially
characterizing a theory'sidentity is by means of a set-theoretical
predicate, i.e. by defining a 'second-order' predicate in terms of
naive set theory.
Now this set of tenets, shared by all philosophers of science who call
themselves semanticists, or structuralists,or, perhaps, something else,
is the common gift we all have gotten from Pat Suppes. In a more or
less explicit way we found them expoundedin his writings on general
philosophy of science and,particularly, in the book he has (alas!) never
published,Set-Theoretical Structures in Science (Suppes, 1970, Chs. 1
and 2).
In his characteristic way, Suppes has not gone on to philosophize
much around his general view on theories but has preferred instead to
implement his approach in anumber ofreconstructions of concrete cas-
es,ranging from relativity theory andclassical mechanics to psychology
andeconomics. Andhere we find the connection with the second great
gift we have received from him in philosophy of science, namely a
huge program in the construction and reconstruction of measurement
theories in many different disciplines, but especially in the social sci-
ences. This program has culminated in the three volumes of the Summa
metrica he has set up with Krantz, Luce, and Tversky (Krantz et al,
1971-1990, in the following abbreviated as FM for 'Foundations of
Measurement'). Put in a nutshell, the connection between the two
strands of Suppes' philosophy of science consists in explicating differ-
ent kinds of measurement as different classes of qualitative structures
as models characterized by set-theoretical predicates. The axiomatic
conditions defining these predicates have to be satisfied by observation-
al structures which consist of certain qualitative objects, relations, and
operations. This proposal to explicate measurementis what Suppes and
his collaborators call the 'axiomatic-representational approach'; they
characterize its essentials as follows:
The most pervasive abstraction in measurement theory consists in formalizingbasic
observations as a relational structure, that is, a set with some primitive relations and
operations. This abstractionarises from consideringthe natureof empirical,qualitative
operations (FM, Vol. 3, p. 201).

As in physical theories, we should not expectcategoricity for the class

of models determining the identity of a measurement theory:
[The] theory of measurement avoids categorical systems of axioms. The reason is ...
apparent: diverse empirical applications are intended, and it would be surprising if
different applications led to isomorphic models of the theory. The same attitude is
strikingly evident in physics Unless the object of physics is to create the one true
modelof the entire universe, categoricity of a physical theory is a defect, not a virtue
(FM, Vol. 3, pp. 247-248).

Consequently, according to the Suppesian view of science, there is

no essential difference between the discipline of measurement and the
discipline of physics, or any other empirical science for that matter.
Bothconsist of theories in the sense ofabstract entities that may bestbe
identified through a class of models characterized by a set-theoretical
predicate; and since each theory has a plurality of genuinely different
intended applications, categoricity is not to be supposed to be among
the characteristic features of such theories.
In the rest of this essay, we shall dwell upon these two strands of
Suppes'philosophy of science: what exactly is the identity criterion for
an empirical theory and what consequencesdoes it have for therecon-
struction of the discipline of measurement? We will do this from the
standpoint of the structuralist program which was started by Sneed and
Stegmuller in the early 'seventies. As pointed out above, structuralism
has at least some of its roots in the Suppesian approach. This does not
mean that there is complete convergence of views about the nature of
theories between Suppes and the structuralists. There are some signif-
icant divergences over the semantics of empirical theories. They were
already discussedinanother Festschrift essay some yearsago(Moulines
and Sneed, 1979, as well as Suppes, 1979). Itis not the purposeof this
essay to go back to this discussion. We shallrather concentrateon those
aspects of a theory's identity which structuralists think are essential and
which,in our view, are already implicit in Suppes' approach, so that he
may agree on them. More concretely, we try to enlarge Suppes' concept
of an empirical theory in a way which, we think, makes the structural

pictureof a theory more adequate. And weillustrate this enrichednotion

in the case of a particular measurement theory.
Suppes apparently identifies a given theory with a class of (model-
theoretical) structures. This class is defined by a set-theoretical pred-
icate. And the conditions or 'axioms' which, in turn, define this set-
theoretical predicate all have the same logical and epistemological sta-
We now wish to modify thispicture in the following way. A theory is
not to be identified with a class of structures but rather with ahierarchi-
cally organizedarray of classes of structures. Each class of structures
we call a 'model-element' and the whole array we call a 'theory-net.
That is, a giventheory(usually) is ahierarchical netofmodel-elements.
The reason to undertake thismodification is mainly this: allaxioms
of a given theory are equal, but some are more equal than others.
That is, they are allaxiomatic in the sense of being neither definitions
nor theorems, but they have a different status they play different
methodological and epistemological roles. We distinguish at least two

(a) basic axioms;

(b) special laws.
Basic axioms are those thatshould be satisfiedin any conceivable appli-
cation of the theory. They are common to all cases. This category
should, in turn, be subdivided into two subcategories: what we call
characterizations and fundamental laws. The firstare purelyconceptual
or mathematical in nature and determine the theory's conceptual frame.
The second are empirical laws of very general scope. But we need not
go into the details of this subdivision in the present context. Suffice it to
know that both characterizations and fundamental laws are common to
all applications of the theory, however heterogeneous they might be -
but that, for that very reason, they alone usually arescarcely informative.
To get more concrete informationabout the systems which are supposed
to bemodels of the theory,you have to require, in addition, speciallaws.
These special laws are not common to all cases of application of the
theory but are put up according to thekind of application we envisage.
The more specialized they are, the more empirical information they
provide, but their application in the theory presupposes the validity of
the basic axioms (i.e. of the characterizations and fundamental laws).
If we call Mq[T] the class of structures satisfying the basic axioms,this

o o o
o o o o
o o o o


means that the class of structures Mi[T] satisfying any of these special
laws fulfils the condition: Mi[T] C Mo[T]. The specializations may
go in different directions,i.e. they may be branched out. Therefore, a
theory's identity is not just givenby the class of structures determined
by thebasic axiomsbut by an open array ofclasses and subclasses. The
picture of the structure of a theory T we get accordingly is that of a
ramified tree with a basic common element - for example, something
of the sort shown in Figure 1 where Mj C Mfj for all i, j,and the
superindex indicates thelevel of specialization, so that, for any i> 1,
and for any j, there is a ft such that iMj C Mkl l).
Now the assertion that any theory can be brought into this form is a
theoretical claim about theoretical science, or if you wish,a metatheo-
retical claim. Such a claim, as in any other theoretical endeavour,has
to be not only made as precise as possible but also checked against the
empirical data. Now, the data here is somewhat different from what
we usually think of when we speak about the empirical data of a the-
ory. It does not consist of natural items like particles, gases, genes or
mental states; itrather consists ofcultural products, namely the theories
themselves that deal with particles, gases, and so on. These prod-
ucts are typically found in standard scientific literature like textbooks,
encyclopedias, and so forth. Leaving this ontological difference aside,

however, the methodological situation is completely analogous to the

one weknow from the methodology ofchecking any theoryby means of
data. This means that we have to find out whether the theory's intended
instances really fit the theoretical model. The particular procedure by
means of which the philosopher of science checks his/her general the-
ory of science against the data is the logical reconstruction of concrete
cases of scientific theories. This procedure only works if we are ready
to accept acertain degree of abstraction andidealization with respect to
the standard presentations of theories by working scientists. But this is
no specific problem of the methodology of philosophy of science: any
theoretization of data - be it in physics or metaphysics only works
to some degree if we accept abstractions and idealizations of the 'raw
material. Up to now,the generalmetatheoretical claim to the effect that
any empirical theory may be represented as a tree-like theory-net has
been checked in about thirty case studies from different disciplines
from physics throughbiology and psychology to economics. The degree
of abstraction and idealization that had to be assumed to consider the
claim positively confirmed in all these cases has varied but it hasnever
been intolerably high at least, it has not been significantly higher than
the degree of abstraction andidealization we encounter when checking
thebest theories of the social sciences. Inthis sense, it may be said that
the claim that theories are tree-like theory-nets has beenconfirmed for a
wide spectrum of scientific theories without much 'forcing' of the 'raw
On the other hand, the claim in question has been confronted with
another sort of methodologicalproblem, which seems to be specific to
the metatheory of science, or at least does not appear to be so acute
in other theoretical sciences. This problem has to do with the extreme
variance of identity criteria for theobjects of study in ourcase. Whereas
in usual empirical science there are intuitive criteria of identity for
the objects investigated which are reasonably precise and commonly
accepted,it is just a fact that the presystematic intuitions about what an
empirical theory has got to look like are extraordinarily vague. This,
of course, poses a problem when trying to check our central claim.
For example, if you think that the thing called 'classical mechanics'
is just one theory (as a superficial glance at physics textbooks may
suggest), then you willfind that thereis no reasonable way to represent
the structure of this thing as a tree-like theory-net. So, the claim would
appear to be falsified in an important case. However, the situation
changes radically if you do not consider that 'classical mechanics' is
the name of one theory, but rather an amorphous denomination for a
bunch of different theories like Newtonian particle mechanics, rigid
body mechanics, classical hydrodynamics, anda few more. Then, you
canfind quite nice and natural representations ofthese 'smaller'items as
tree-like theory-nets. A similar situation emerges when analyzing that
thing called 'phenomenological thermodynamics' and in many other
cases. The things representable as tree-like theory-nets are just there
but it may not be a completely trivial task to detect and differentiate
The representation of given theories as tree-like theory-nets has pri-
marily a metatheoretical value: it allows us to identify each theory
precisely, to analyze its formal structure in detail, and also to investi-
gate the nature of possible intertheoretical relationships of allkinds as
formal links between two or more graphs of this sort. But in the case
of young, still not well-established disciplines, this representationmay,
in addition to the epistemological role, also play a methodological one.
We mean by this that it may help the working scientist herself and
not only the philosopher of science in providing her with some clues
about possible developments of her theory she might not have thought
of before. Since one of thebasic ideas of the theory-netrepresentation
consists in viewing a theory as a sort of 'pyramid' withdifferent precise-
ly defined levels of generality, it follows that if we notice somelacunae
in the hierarchy of levels of existingtheory-elements, we may wonder
whether we might not fill theselacunae with new elements having some
possible use in actual research. In other words, the completion of a
theory-net in a natural way may have not only an aesthetic but alsoa
quite practical value.
We now come to the promised illustration of the foregoing ideas
by means of an example from the discipline of measurement. The
purpose of this reconstruction is not only to provide an illustration of
our generalmetatheory, but also to make acase for two more particular
claims suggestedabove. First, in the case of 'measurement theory', as
in so many other cases, we have a misnomer: there is no such thing as
the one and only theory of measurement, but rather there are several
(though not many) different theories of measurement (if we accept the
identity criterion proposed above). We reconstruct the historically and
methodologically most conspicuous one: the theory of combinatorial
measurement. The second thing to note is that, by representing this

theory as a tree-like theory-net, we can say exactly at what place in the

pyramid the usual expositions of this theory seem to have overlooked
some potentialities that might be empirically relevant. These lacunae
may be filled in a rather easy and natural way and the corresponding
representation theorems may be obtained.
To abbreviate, let us call the theory of combinatorial measurement
TCM'. We claim that TCMis a theory-net with a basic model-element
and seventeen specializations so far detected. Some of them appear to
have beenoverlooked by the authors of FM.
TCMneeds three basic concepts: a domain of objects, an ordering
relation, and an operation of combination. That is, themodels of this
theory, whichmake up the conceptual framework for all its specializa-
tions, are structures of the sort (D, £, o). Thisis one of the reasons why
TCMis a theory different from other measurement theories (according
to our identity criterion): itsmodels have adifferent structure byhaving
a component (the combination operation) the othermodels do not have.
Theother reason (which is related to the first one) is that this theory has
abasic axiom (a weak monotonicity condition) whichis not common to
all theories of measurementbut iscommon to allkinds of combinatorial
TCM's basic class of models, Mq [TCM], is givenby the following
basic axioms.2

DEFINITION1. x G MO°[TCM] iff there are D, £, o such that:

(2) £C D x D is strongly connected and transitive
(3) o is a function from D x D into D
~ — ~
~ —
(4) (a) (a b ► a o c b o c)
(b) (a 6 ► c o a c o b)3
Condition (3) of Definition 1 is more restrictive than is actually needed
for a thoroughreconstruction of TCM.The present formulation requires
that the operation o be closed, i.e. that combinations be defined on all
pairs of objects. However, we know that, in some cases, combinations
might be empirically impossible (for example, because the resulting
object would be just too 'big', or would explode, or whatever). We
could weaken Definition l-(3) so that the domain of obe only a subset
of D x D, andthe rest of thereconstruction would retain a similar form.
However, the formulation of definitions and theorems would become
more cumbersome. For lack of space, we choose the simplified (more
restrictive) formulation. Froma metatheoretical point of view,nothing
essentialis thereby lost.
For the sake of abbreviation, let us call any model of TCMa 'com-
binatorial structure' and write 'C(x)' instead of x G Mq[TCM]; that is,
'Cis the fundamental set-theoretical predicate ofour theory.
From Definition 1an obvious corollary follows:4

THEOREM1. a~b&c~d— >aoc~bod&coa~dob.

Proof. From a b and Definition l-(4b) we get a o c bo c.
~ ~
From c d and Definition l-(4b) we get b o c b o d. Therefore,by
~ ~
Definition l-(2), a o c b o d. Analogously for c o a d o b.

The first level of specializations of the basic element makes a fun-

damental distinction. It distinguishes between the way the ordering of
objects behaves after combination. This differentiation is fundamental
in the sense thatall other interesting properties of specific kinds of com-
binatorial measurement, and in particular the form the representation
theorem adopts, will depend on it. When combining two objects a and
b, it may be that the resulting object iseither 'greater' or 'lesser' than a
and b but that, in any case, it preserves the same ordering with respect
to both; or else, on the contrary, the resulting compound ao b may be
sometimes greater than a but lesser than b, or conversely.
For the first possible situation let us choose the label 'external com-
binatorial metric' iEC). The second situation we call acase of 'internal
combinatorial metric' (IC). We shalldefine themjn amoment,but before
that, let us make a remark ona further property externalcombinatorial
metrics may have. Ininternal measurement, combination is necessarily
idempotent. Not so in external measurement. But,here, there might be
some specialobjects for whicha particular kind of equivalence is valid.
A sort of object / let us call them 'limits' - might appear, for which
it is the case that, for any other object aof the domain, a o I= I; when
combining velocities, the speed of light might come to our minds as
an example of this situation. On the other hand, there might also be a
kind of objects n -call them 'null' - for which it is true that ao n = a
for any a; when combining chemical substances and considering their
therapeutic effects, a placebo might be a case in point here. The pos-
sibility of having both limits and nulls is responsible for some formal

complications in the reconstruction of the theory. But since there are

goodempirical as well as formal reasons not to exclude these possibili-
ties apriori, we will have to consider them in our definition of external
combinatorial metrics. In internal measurement, they have no role to
play. Letus introduce first the correspondingauxiliary definitions: 5

DEFINITION 2. Let x = (D, £, o) with C(x).

(a) Nia) iffao6~6oa~6
(b) L(a) iffao&~6oa~a

Our limits are reminiscent of the notion of an 'essential maximal

element' as introduced in FM (Vol. 1, Ch. 3, Sec.7),though they do not
fully coincide.6 At any rate,nulls andlimits haveanumber of properties
we intuitively expectfrom them. First, nulls are equivalentonly to nulls
while limits are equivalent only to limits, and all nulls are mutually
equivalent while the same goes for limits.

THEOREM 2. (1) N(a) & a b -> N(b)
(2) L(a) & a b -> L(b)
(3) Nia) & Nib) -* a b
(4) L(a) & Lib) -» a b
Proof. (1) Assume Nia) and a b. Then,by Definition l-(4), for
~ ~ ~
any c, a o c b o c. Since Nia), a o c c. Thus, for any c, c 6 o c.
Therefore, N(b). The proof of (2) is entirely analogous. (3) From
~ ~
Nia) it follows that a o b b and from Af (6) it follows that a o b a.
Therefore a b. Similarly for (4).

Secondly, nulls and limits have the mutual relationship we should

expect: if there are at least two non-equivalent objects in the domain,
then no null may be equivalent to a limit.

THEOREM 3. 3a, b(a * b) -> ->(N(c) & L(c))

Proof. Take some a, b with a *> b. Suppose there were a c with N(c)
and Lie). Then, by iV(c), aoc~a&6oc~6;and, by L(c), a o c c
and 6 o c c. Therefore, a~aoc~6oc~6,i.e. a b (!). ~
For the rest of this essay, we shall just assume that the premise of The-
orem 2 holds. Further, null objects allow for a stronger monotonicity
condition with respect to o:

THEOREM4. N(a) -* (ao& £ aoc *-* b y c) & (fcoa y coa <-► b y c).
Proof. Assuming Nia), we get 6~ao6^aoc~c, i.e. by
Definition l-(2), 6 >3 c. Similarly for the rest of the theorem.

Wenow introduce the first specialization of TCM which is, in turn, the
basic model-element of one of the two main 'branches' of the combi-
natorial 'tree.


DEFINITION 3. EC{x) iff there are D, fc, o such that x = {D, fc, o)
and Cix) and
(1) a o b £ a, b V a, 6 £ a o b
(2) a o 6 a -* £(a) V JV(6)

In external combinatorial measurement, nulls and limits have the ex-

tremal properties we intuitively expect from them: when combination is
alwaysincreasing, i.e., when the combination of two objects is always
greater thanboth of them,nulls are minimal andlimits maximal; when
combination is alwaysdecreasing,nulls are maximalandlimits minimal.
Wehave said that the characteristic feature of this kind of measure-
ment is that the result of the combination of a and b is either greater
than a and b or lesser than both. But this does not imply that, for all
pairs (a,b), it will always be greater or always.lesser thanboth. Itcould
be the case that it is sometimes greater and sometimes lesser than both
a and b. If the latter is the case, we speak of a cyclic metric. If the
former is the case, i.e. if ao b is greater than a and b for all (a,b) or
less than a and b for all (a,b), we speak of a linear metric. Moreover,
all linear metrics share a strict monotonicity condition. Linear metrics
may, in turn, be subdivided into the positive case (where the result of
combination is always greater than the components) and the negative
case (where the result is always lesser than the components). The most
conspicuous subcaseof the positive linear metric is in turnthe extensive

one, which historically was the first one to be studied and which has
found very wide application in physics.


DEFINITION 4. LECix) iff there are D, y,o such that x = {D, fc, o)
and ECix) and
(1) Va, b(a o b y a, b) V Va, 6(a,b fc a o b)
(2) (a) iayb*-+ao C yboc)
(b) ->L(c) a ycob)

For the next specialization (positive metrics), besides positivity, a natu-

ralcondition shouldbe imposed: the well-known condition of solvabil-


DEFINITION 5: PLECix) iff there are D, £, o such that x = (D, £, o)

and LECix) and
(1) aob>za,b
(2)ayb^ 3c(-^N{c) & a y b o c)

The most important specialization on this branch of the tree are the
extensive metrics, i.e. combinatorial metrics with additive represen-
tation. They are those metrics everybody thinks of when considering
positive linear combinatorial measurement, though they are possibly
not the only specialization of this kind relevant for empirical science.
In order to get this important specialization with the correspondingrep-
resentation theorem, we need to add the requirements of commutativity,
associativity, and archimedianity to the conditions already set forth.


DEFINITION 6. EPLECix) iff there are £>, £, ° such that x = {D, fc,
o) and PLECix) and
(1) a o b b o a1
(2) (a o b) o c a o (6 o c)
(3) Any sequence (a;);eN of combinable objects for which it is true
that: ->N(ai) & Va^i >2-*a{ = a<_i o aj) & 36(-iL(6) &
Vai(6 >- Oi)), is finite.

Theseextensive metrics are essentially those representedin FM (Vol. 1,

Ch. 3, Definition 3), but enriched with nulls and limits; the following
representation theorem can be proved:

THEOREM5. Ifx = (D, £ o) with EPLECix) then:

(1) There is f :D ► [0, oo] such that
(a) a y b <- fia) > fib)
(b)/(ao6) = /(a) + /(6)
(2) Any function f satisfying (1) is such thatfor any a € D:
(a) Nia) «- fia) = 0
(b) L(a) «-> /(a) = oo
R ,such that, for any
(3) //"/ arcd /' satisfy
(1) then there is an r £
aeD: fia) r " /(a).
Proof The proof is rather easy, though somewhat lengthy. We can
only sketch it here.
To prove Theorem5-(l), the essentialideais to show that, if we take
all nulls and limits out of D, then the resulting structure (£)', y' o') is
an extensive metric in the usual sense and the representation theorem
can be proved for itin the way it is shown,for example, in FM (Vol. 1).
That is, this 'reduced' structure has arepresentation function /* with the
required properties. If we then construct / = /* U {(a,0) : a G D &
N(a)} U {(a,oo) : a € D & £(a)}, it is easily seen that / is a function
satisfying the ordering and the additivity conditions.
To prove Theorem 5-(2), first assume /(a) =0; for any b e D, we
thenhave /(a) + fib) = fib), i.e. /(a o6) = fib); therefore aob ~b,
i.e. Nia); thus: /(a) = 0 -+ Nia). Now, assume /(a) = 00,
for any b, we then have /(a) + f(b) = fia), i.e. /(a o 6) = /(a);
therefore ao b ~ —
a, i.e. L(a); thus: /(a) = oo > L(a). On the
other hand, assuming ./V(a) take any b with ->L(6) (such a b exists by
Theorem 3); wehave just seen thatthis implies/(6) oo; JV (a),in turn,

implies ao b ~ b and thus /(a o b) = fib); i.e. /(a) + /(&) = f{b);

since /(o) 7^ 00, it must be /(a) = 0; thus N(a) ► /(a) = 0.
Assuming Lia), take any 6 with -»JV(6); we have seen that this implies
/(&) 7^ 0; Lia), in turn, implies ao b ~a, i.e. /(a 06) = /(a),
i.e. /(a) + /(&) = /(a); since f(b) =^ 0, it must be /(a) = 00; thus:
Lia) -+ /(a) = 0.
Finally, Theorem 5-(3) follows easily from the fact that, if / and /'
are two functions satisfying Theorem 5-(l), then f'(a) = r " /(a) for
any r € M, if a is a null or a limit (by Theorem5-(2)); and for the rest of
objects whichare neither nulls nor limits, justapply the usual argument
as given in FM (Vol. 1).


Thoughextensivepositive measurementwith its corresponding additive

representation is the most outstanding case of combinatorial measure-
ment, it isbut one of themainbranches of TCM's tree. You get another
branch - which seems to have been ignored in the literature if you
consider negative metrics within the linear case. They correspond to
- -
those properties which decrease instead of increasing after combina-
tion. A physicalexample wouldbe theparallel combination ofelectrical
resistances, where the total resistance of the system decreases as more
resistors are combined.
As may be gathered from the proof of the representation theorem
below, negative metrics may be seen, from a purely formal point of
view, just as trivial 'rewriting' of the positive ones. However, whatis
not so trivialis the fact that we really get the conversecase of the posi-
tivemetrics by just assuming that combinationmakes thecorresponding
property decrease. Also, it is interesting tonote that (as the example of
electrical resistance shows) one and the same qualitative ordering may
'behave' in different ways depending on the kind of combination we
use. It would not be natural to take one ordering for one combination
and another for the other one. Negative metrics are therefore concep-
tually necessary to provide a complete account of different empirical
Ina way completely analogous to thepositive case, we first introduce
negative combinatorial measurementiNLEC) andits extensive version
{ENLEC) as a specialization.
DEFINITION7. NLEC{x) iff thereare D,fc, o such that x = (D, fc, o)
and LECix) and
(1) a,by ao b
{2)ayb-+ 3c{^Nic) & a o c £ 6)

DEFINITION8. ENLECix) ifthere are D, y,o suchthat a; = (D, £, o)

and NLEC{x) and
(1) a o 6 6 o a
(2) (a o 6) o c a o (6 o c)

(3) Any sequence (a2 )ieN of combinable objects for which it is true
&Vai{i > 2 -» a{ a;_i o ai) & 36(^L(6) &
Maiiai y b)),is finite.
Now, extensivity in the case of negative metrics doesnot entail additivity
in the numerical representation on the positive real numbers. The
representation theorem we get in this case is rather different. To state
it, let us first introduce a numerical operation, call it '*', such that, for
any r,s G [0, oo]:
(a) r *s = -£&, if r,s G E+
(b) r * 5 = r,if s = oo
(c) r * s = s, if r = oo
(d) r * 5 = 0, if r = 0 or 5 = 0
The representation theorem thenruns as follows

THEOREM 6. Ifx = {D, y,o) with ENLEC{x), then:

(1) There is f : D — > [0, oo] so //*atf
(a) a y b «-» f{a) > f{b)
{b) f{aob) = f{a)*f{b)
(2) Any function f satisfying (1) is such thatfor any a G D:
(a) Nia) <-> f{a) = oo
(b) Lia) <- f{a) = 0
(3) Same as Theorem 5-(3).

Proof (sketch). The first thing to notice is therather trivial fact that,
if (D, y,o) isa negative metric, then (D, o) (where is theconverse
of y) is a positive one. This entails,in turn, that (by Theorem 5) there
willbe a homomorphism from (£>, ;<, °) into ([O, oo], > +). Then, take
the isomorphism
as defined by the condition:
Fir) = 1/r.
Itis easily seen that the functional composition / o Fis itself a homo-
morphism from (D, o) into ([O, oo], <, *), which,in turn, entails that
itis also a homomorphism from (D, y,o) into ([O, oo], >, *).
The composition / oFis thus a homomorphism from {D,£,o) into
([O, oo], <, *), whichis equivalent to saying that itis itself a homomor-
phism from (D, y,o) into ([O, oo], >, *).


A third important branch of TCMcontains the measurement of cyclic

magnitudes {CEQ. A portion of thisbranchhas already been studied in
FM (Vol. 1, Ch. 3), though within a different setting. Limits make no
sensehere but there may be null objects. Monotonicity takes a special

DEFINITION 9. CEC{x) if there are D, y, o such that x = (D, y,o)

and EC{x) and
(1) 3a, 6(a o b y a, 6) & 3a, 6(a,b y a o b)
(2) a^^c^aocV&oc^c)

Thereare two kinds of cyclic metrics: inferior and superior ones. The
former {ECEC) are characterized by the fact that null objects are those
that correspond to no cycle at all (minimal objects), whereas, in the
latter (SCEC), the null objectscorrespond to acomplete cycle (maximal
objects). Correspondingly, we shall have two different set-theoretical
DEFINITION10. ICECix) iff thereare D, y, o such that z = (£>,£ o)
and CECix) and
r\j r\j r^*j rsj rsj

DEFINITION11 SCECix) iff thereare D, y,o suchthat x = (Z>, £, o)
and CEC{x) and

Ininferior cyclic metrics, nulls are minimal, while they are maximal in
superior cyclic metrics.
If we add therequirements of commutativity, associativity, andarchi-
medianity to the conditions of Definition 10, respectively Definition 11,
we get the extensive version of inferior cyclic metrics, respectively
superior cyclic metrics.

DEFINITION12. EICECix) iff there are D, y, o such that x = (D, y,

o) and ICECix) and
(1) a o b b o a
(2) (a o 6) o c a o (6 o c)
— —
(3) If for any integer n,na = a when n \, and na = in \)a o a
when n > 1, then: a y b —+ 3n(a y na & nb y b).

DEFINITION13. ESCEC{x) iff there are D, >% o such that x = (D, y,

o) and SCECix) and
(1) Same as Definition 11-(l).
(2) Same as Definition 11-(2).

(3) Under the same assumption as in Definition 11-(3), a y b >
3n(a y na &nb y b)

Both kinds ofcyclic metrics allow fora 'natural' representation theorem.

The range of the representation function is an interval [0, r) for EICEC
and an interval (0,r] for ESCEC - in the first case it assigns 0, in the
second r, to the null objects. There is a sort of additivity herebut it is
'modulo r\ To express it
+ + formally, weneed some auxiliary definitions:
let rG R and h,k £R U {o}. Then

(a) nr {h) is the greatestinteger such that n " r < h

(b) h+ ~ k =df (h +k)— r " nr {h +k) (where the index li* stands
for 'inferior').
(c) h+fk = h+^k,ifh+ (where V stands for 'superior');
otherwise it is = r.

THEOREM7. If x = (D, y,o) with EICECix), then,for any r G K+,

f/zere is <z unique f : D -^> [o,r) so that:
(2)/(ao6) = /(a) + i/(6).
(3) Nia) <- f{ a ) = 0
Proof. The proof of Theorem 7-(l) and (2) is completely analogous
to the one to be found in FM (Vol. 1, Ch. 3, Theorem 2), except for
the fact that here, for any a, fia) r. (This is actually the mark
of inferior cyclic metrics.) Indeed, suppose it were /(a) = r. Then,
for any b, fia o b) = fib) + j f{a) = f{b) + j /(&) (by definition of
'+;:'). This entails N(a). Sincenulls are minimal here, b y a, therefore
fib) > /(a)- Since, by construction of /, r > fib), the assumption
r = /(a) would imply /(a) > /(&), thus /(a) = f{b) for any 6, i.e.
a 6 for any 6, which violates the general assumption of Theorem 3.

As for Theorem 7-(3), the >-part follows from the fact that, for any b,
ao b
~ -
b, thus fia) + l fib) = /(&), but by definition of '+£', this
is true only if f{a) = 0or it is a multiple of r,but since /(a) G [0, r),
it must be /(a) = 0. Now, take any a with /(a) = 0. This entails
f(b) + j f\a) = fib) for any b. Therefore f(b o a) = /(&) for any 6,
i.e. a o 6 6 for any 6, thus N{a).

THEOREM8. If x = (£>,£, o) vWtf* ESCECix), then for any r G M+,

//zerg is a unique f :D — ■> (0,r] so //za/:

(1) a t b «-> fia) > fib)

(2)/(ao6) = /(fl) + J/(6)
(3) Nia) «-» fia) = r
Proof (sketch). First, it has to be shown that ESCEC{{D,y,o))
iff EICECi(D,3, o)). Assume ESCECi(D, £, o». It is well known
that < is alsoa weak order, o is the same operation, and commutativity
and associativity are still satisfied. As for monotonicity, a b iff by a
iff {bocyaocyc or cybocyaoc or aocycyboc) iff
icy aoc^boc or aoc^boc^c or boc-<c^,aoc). As for
archimedianity, assume a-< b, which means b y a. By Definition
13-(3), 3n{b ynb &na y a), which means: 3n(a -< na &nb ■< b),
which is Definition 12-(3). By a completely analogous argument we
get ESCECi(D, y, o» from EICEC{{D, 3,o)).
Now, assume ESCEC((D,y,o)). It follows EICEC{(D,£,o)).
This implies, by Theorem 7, that there is a representation function
/' from D into [0, r). We now define a function / as follows:
f{a) = r — fia). Clearly, / is a function from D into (0, r). Fur-
* a iff /'(&) > f'(a) iff r f'(a) > r /'(&)
thermore, a >z b iff b - -
iff /(a) > fib), thus we get Theorem 8-(l). Using the numerical
that, -
for any s, t G [0, r), r (s + M) = (r - s) + J (r - £),
- -
we obtain: /(a ob) = r /'(a o b) = r if{a) + J /'(&)) =
- +
(r /'(a)) + (r-- f'(b)) /(a) J /(&), thus we get Theorem 8-
J =
(2). Theorem 8-(3) is immediate.


TCM'slast 'bigbranch' arises from what we have called 'internalcom-

binatorial metrics (/C)'. In this case, if a and b are combinable, then
a^3 ao^^3^' or^^ ao^~ a Nulls and limits make no sense in
thesemetrics. Indeed, in internal metrics, if two objects are equivalent,
then the property of internality obviously entails that they will also be
equivalent to their combination, (i.e. idempotence), so that equivalent
objects would always play the role of nulls and limits simultaneously
with respect to each other. ConditionDefinition 14-(2) does not follow
from internality but is the 'converse' of idempotence.

DEFINITION 14. IC{x) iff there are D, y, o such that x = (D, y, o)

and C(x) and
(1) a>zaobybVbyaobya
(2) ao6~a^a~6

A particular kind of internal metrics has already been studied in the

literature, although under a different label. They are those requiring,

in addition to internality (actually: idempotence),monotonicity and so-

called 'bisymmetry' (seeDefinition 15— (2)). We may call these internal
metrics 'proportional' {PIC), since their numerical representation is
such that we get one and the same proportionality factor for all objects,
as we shall see below.

DEFINITION 15. PlC{x) iff there are D, y,o such that x = (D, £, o)
and IC{x) and
(1) ayb<r->aocyboc
(2) (a o b) o {c o d) (a o c) o b o d)

Proportional internal metrics,in turn, may be subdividedinto two cases

to allow for a 'nice' numerical representation. In one case, the com-
bination operation is necessarily closed and the domain of objects may
be infinite {CPIC); in the other case, the operation is not closed but the
domain is assumed to Definite in any application {FPIC). We may fur-
ther specialize both kinds into the case where o is commutative, which
allows for a still 'nicer' numerical representation.

DEFINITION16. CPIC{x) iff thereare D,£, o such that x = {D,y,o)

and PIC(x) and
(1) (a) c o b y a y d o b — ► 3e(e o b a)
(b)6oc>-a>-6od— » 3e(6 o c a) ~
(2) Any sequence (a^)i€N such that, first, there are c, d with cy dand
~ ~
{ai oc ai+ \ odV coai doa^+i), and second, thereis a b such
that b y ai for any ai, is finite.

The next theorem is the representation theorem for thiskindof metrics.

A similar theorem hasbeen provenin FM. But the representation func-
tion there is somewhat more complicated. It can be shown, however,
that our representation is derivable from theirs (seeProof'below).

THEOREM9. Ifx = (D, £, o) with CPIC{x) then:

(1) There is f: D — >" Eand an r G (0,1) so that
(a) a y b <- f{a) > f{b)
(b)/(ao6) = (l-r)-/(a) + r-/(6)
(2) Iff and f satisfy (1) then thereare s GM and tG X such that
{a) f{a) = s-f{a) + t
(b) r' = r.
Proof. InFM (Vol. 1,Ch. 6,pp. 295-296) itis proved that,under the
assumption of Theorem 9, without internality there are functions / and
/' satisfying (la) and (2a), and there are real numbers k, I, m with k,
I> 0 such that, for any a, 6 G D: {A) f{aob) = k- f{a) + l- f{b) +m,
and so that,for any other /', k' = k and /' = /.
Now, we only need to prove that (lb) follows from (A)plus internal-
ity (i.e. idempotence). Assume (A). We know that any object a has an
equivalent b (atleast itself). So, take any a,b with a b. Itfollows from
~ ~
Definition 14-(l)thatao& a b. Therefore, / {aob) = f{a) = f{b).
By(A) and f{aob) = f{a), we get k f{a) + /"/(&) +m = f{a). But

since f{a) = f{b), we havek " f{a) + / " f{a) + m = f{a). Since this

is true for any a, it must be m = 0. Therefore {k + l l)-f{a) = 0, i.e.
k + l-\ =o,i.e. k = \-l. By (A), /(ao6) = (1 -/) -/(a) + /"/(&).
This is Theorem 9-(lb). Theorem 9-(2b) is immediate.

It is easily seen that, by adding commutativity to the conditions of

CPIC, we would get a very simple representation, which we might call
'bisection' {BCPIQ:

fin « M =
/( a ) + /W
/(a°o) 5 '

Consider now the case of proportional internal metrics over a finite


DEFINITIONI7. FPIC{x) iff thereare D,£, o such that x = (£>,£, °)

and PlC{x) andDis finite.

Since D is finite, it is moreover well ordered by the linear ordering

generatedby y on the ~-equivalence classes. This allows for anatural
definition of 'distance' between any two objects a, b as the number
of 'intermediate' objects separating a and b in the ordering. Clearly,
there couldbe various ways of metricizing this sort of distance,thereby
getting different metrical representations. We shall not get into an



Fig. 2.

exploration of these possibilities here. Suffice it to note that, if we add

commutativity to the conditions of FPIC, we obtain a bisection again
{BFPFC), as in Theorem 9-(lb). A corresponding theorem, which has
already been proven in FM (Vol. 1, p. 297), could easily be translated
into our own terms here.


If we use the abbreviations of the set-theoretical predicates defined so

far to indicate each model-element of TCM, we may represent the form
and content of this theory as a tree-like theory-net as shownin Figure 2.

C. Ulises Moulines,
Institutfur Philosophic, Logik
und Wissenschaftstheorie,
Munich, Germany
Jose A.Diez,
Departamento de Filosofia,
Universidadde Barcelona,
Tarragona, Spain

We owe to Professor R. Duncan Luce some helpful remarks on a previous version of
this paper. They contributed to an improvedpresentation.
2 To
simplifythe formalization we agree that open formulae are to be interpreted as
universallyquantifiedover the domainof objects. We make explicituse oftheuniversal
quantifier only in somecases whereconfusion might arise.
3 ~
In (4a) (and correspondinglyin (4b)) we do not require that a o c b o c implies
~ 6. The reason is that the possible existenceof so-called 'limit objects' (see below)
would violate this requirement.
Fromnowon, we implicitly assume in each theorem that the domainof objects is the
domain of a modelof TCM.
For the sake of perspicuity, we have slightlysimplifiedthe followingdefinitions. A
thorough treatment ofthesenotions wouldrequire some formalcomplicationsbut this
wouldaddnothing essential to the argument.
When thispaper was already finished, Professor Luce kindly sent usa copy of a recent
research report of his, where he introduces the notions of 'infinite point' and 'zero
point', which are quite similar to our 'limits' and 'nulls' (see Luce, 1990, especially
p. 8).
Actually,it can be proved that commutativity follows from the rest of conditions of
EPLEC, under the assumption that o is closed. However, for the sake of perspicuity
and to allowfor an immediategeneralizationto thecase whereo isnot closed, we state
commutativity explicitly as a conditionforEPLEC.

Krantz, D., Luce, R. D., Suppes, P., and Tversky, A.: 1971-1990, Foundations of
Measurement, Vol. 1(1971), Vol. 2 (1989), Vol. 3 (1990), San Diego: Academic
Luce, R. D.: 1990, 'GeneralizedConcatenationStructures That Are TranslationHomo-
geneous betweenSingular Points', MathematicalBehavioral Sciences, Technical
Report Series, Irvine: University of California.
Moulines, C. U. and Sneed, J. D.: 1979, 'Patrick Suppes' Philosophyof Physics', in:
R. J. Bogdan(Ed.), Patrick Suppes, Dordrecht: D. Reidel, pp.59-91.
Suppes, P.: 1970, Set-TheoreticalStructures in Science, Typescript, Stanford: Stanford
Suppes,P.: 1979, 'Replies to Moulines and Sneed', in: R. J. Bogdan (Ed.), Patrick
Suppes, Dordrecht: D. Reidel,pp. 207-211.


Ican hardly disagree with the general things that Moulines and Diez
say at the beginning of their essay about my approach to the philosophy

of science. Considering other parts of the structuralist program with

which Ulises has been associated closely with Stegmuller and Sneed, I
was surprised to find no discussions of a specific kind about empirical
interpretations, butrather a quite detailed characterization of different
structures of measurement.I find the latter enterprise, the one engaged
in here, very much easier to understand and to appraise. I have liked
the attempt by the structuralist to characterize the subset of models that
are empirical but I have not as yet been overwhelmed by the detailed
arguments. On the other hand, the kindof fine-grainedanalysis givenby
Moulines andDiez of special axioms in the general theory of extensive
measurementleads to a very refined classification anda tree-like graph
for classesof structures ordered by set-theoretical inclusion. Atthe most
basic level the structures satisfy the axioms of the theory of combina-
torial measurement. Characterization of this class of structures and the
resulting many different specializations remind me of the Bourbakian
enterprise of identifying mother-structures that are basic in mathemat-
ics. In alimited way this is how I would understand the Moulines and
Diez enterprise.
On the other hand, I do have one skeptical query for them. Unlike
Bourbaki, they claim a strong empirical component for their theorizing
and remark, in a vein that I much approve of, that the identification
of their hierarchical net of model elements, to use their terminology,
is an empirical enterprise concerned with the data in just the way any
empirical theory is concerned with data. In their case the data are
generatedby theories of measurement that have been produced. What
Ido not find is the thorough canvass that one might expect to back
up the empirical side of the claim. Essentially everything that is said
is sensible and all the structures identified make sense, but there is no
real support for the claim that the analysis hasbeen exhaustive or that
it corresponds in any detailed way to what is present in the very large
literature of measurement.
Interesting examples that lie outside their tree structure are the mid-
point algebras introduced by Wanda Szmielew (1983) in affine geom-
etry. These midpoint algebras are closely related to the bisymmetric
structures introduced in Definition 16 by Moulines and Diez. But it is
natural in geometry to first introduce them without order, and this is
the case for Szmielew. In fact, more generally there are various rea-
sons for introducing structures without order which may still lead to
measurementresults. I am not claiming that there are not goodreasons
to exclude midpoint algebras without ordering in the analysis givenby
Moulines andDiez. What I find missing is the empirical pursuit of the
many different alternative structures lying on the edges of the standard
topics they address. How are we to make the argumentone way or the
other for exclusion or inclusion of a given class of somewhat aberrant
structures? From a philosophical standpoint this is the aspect of their
methodology that is not clear to me, if weare to regard the enterprise as
empirical in character.


Szmielew, W.: 1983, From Affine to Euclidean Geometry: An Axiomatic Approach,

Dordrecht: Reidel.

Abrahao, S.M. 190 Descartes, R. 46, 51,53, 60

Adams, E.W. 129-131 passim, 138, 145, Detles, R. 21
224, 242 Dewey 98
Aharanov,Y. 21 Diez, J. 275-299
Albert, D. 19,20,21,25,28 Div,B. 54
Alper, T.M. 232,233, 235, 242 Doria, F.A. 151-193
Amaral, do, A.F. Furtado 189, 190 Dorling79
Archimedes 59 Droste, M.235, 243
Aristotle 55,59 Duhem 37
Arnold, V.I. 152, 161, 181, 182, 190 Durr, D. 21
Aspect, I. 12
Atiyah, M.F. 189 Earman, J. 74, 94, 95, 212
Eddington, A.S.79, 95
Balzer, W. 144, 145, 212 Ehlers, J. 91,95
Barros, de,J. Acacio 26, 28, 189, 190 Ehrlich, P. 224
Bell, J. 12, 19,20,21,32,35,38 Einstein, A. 11, 21, 29, 31, 36, 38, 45, 77,
Bertrand 112 83,93
Beth, E. 144, 145 Ellis, B. 220,243
Bethe, H. 41 Eudoxus of Knidos 59
Binder, K. 117 Euler, L. 60, 120
Blaquiere, A. 54
Bogdan, R.J. 19,21 Falmagne, J.-C. 228, 241, 243, 252,272
Bohm, D. 11, 12,21,32 Feld 79
Bourbaki, N. 154, 189, 298 Fermat 60
Bridgman, P. 220, 229, 242 Fermi 52
Broglie, de, L. 45 Feynman, R. 40, 42, 43, 54, 55
Burgess, J. 66, 95 Field, H. 66, 95
Fine, A. 29^13
Cameron, PJ. 230, 243 Fishburn, PC. 244
Campbell, N.R. 220, 243 Forge, J. 212
Carnap, R. 125, 143, 145 Fourier 43, 47, 48 f., 54,56
Catton, P. 77, 95 Fraassen, van,B. 20, 22, 115,118, 126, 144
Chiara, Dalla, M.L. 189 146, 213, 275
Cho, Y. 162, 189 Francia, di, G. Toraldo 189
Chuaqui, R. 154, 189, 189, 192, 193 Franke, CH. 241,245, 252, 256, 273
Church, A. 119 Fresnel46, 47, 51
Cohen,M. 220,228, 231,243 Friedman, M. 77, 83, 95
Cohen-Tannoudji, C. 54
Comte, A. 47 Galileo 59, 60, 61
Corrada, M.189 Gleason, A.M.12, 19, 21
Costa, da, N.C.A. 151-193 Godel, K. 191
Cuche, V 54 Goldstein, S. 21
Gridgeman, N. 113, 117
Daneri, A. 21 Griinbaum, A. 80,145, 145
Davis, M. 170, 190
Debreu, G. 274 Harvey, L.H. 251, 254, 260, 270, 272
Dell, J. 162, 190 Heerman, D.W. 117


Heisenberg, W. 52, 56 Luce, R.D. 59, 95, 97, 190, 219-249, 251
Helmholtz, yon, H. 102,219, 220, 222, 243 252, 254, 272, 274,276, 297
Hempel, C.G. 198, 205,212
Heyer, D.227, 229, 241,243 Machtey, M. 172, 188, 190
Hilbert, D. 243 Malebranche53
Hiley, D. 21 Manders, K. 74, 95
Hirsch, M. 189 Margenau, H. 54
Holder, O. 102, 219, 222,223, 232, 243, 223 Marley, A.A.J. 241
Holman, E. 243 Maudlin,T. 19, 21
Horwich, P. 72,95 Mausfeld, R. 243
Howson, C. 213 Maxwell, J.C. 42, 48
Hughes, R.I.G. 20, 21 McClelland, J. 118
Hume, D.47 McKinsey, J.C. 213
Humphreys, P. 103 Metropolis,N. 111,117
Huygens, C.46,51,53, 55, 60 Michell, J. 226, 244
Minkowski 77
Inagaki, V 190 Moivre, de, A. 102, 121
Iverson, G. 243 Moler, N. 70, 95
Morgenstern, O. 219, 239, 245
Jones, J.P 190 Moulines, CU. 63, 65, 96, 144, 145, 212
Kahneman, D. 242, 244, 245 Mundici, D. 189
Kallen, G. 41,43 Mundy, B. 59-102
Kant, I. 27,47, 51,52 Murphy, G.M. 54
Kanter, J.J. 54
Keller, E. 117 Nagel, E. 220
Kelvin, Lord60 Narens, L. 190, 219-249, 252, 272
Kepler, J. 60 Naylor, T.H. 105, 117
Klein, F. 68, 244 Nelson, E. 19,21,25,28
Kobayashi, S. 162, 190 Neumann, yon,J., 7, 8, 12, 19, 22, 31 32.
Kochen, S. 19,21,32,38 39,118,219,239,245
Kolmogorov, A.N. 119 Newton, I. 41,46, 60, 120
Krantz, D. 59,95, 97, 190, 223, 244, 251, Niederee, R. 225,226, 227, 229, 241, 243
252, 254, 272, 274,276, 297 245
Krause, D. 189 Nomizu, K.190
Kreisel, G. 152, 190 Norton, J. 94, 95
Kuhn, T.S. 79, 95,143, 145
Kyburg, H.E., Jr. 224, 244 Parijs, van, P. 213
Pasch, M.102
Lagrange 45,60 Pauli, W. 41
Lakatos, I. 143, 145 Pearce, D.145
Laloe,F. 54 Pearle, P. 6, 21
Landau, L.D. 41,43 Peirce, C.S. 98, 131
Laplace 60, 61, 112, 121 Peres, A. 29, 32,35, 37,38, 39
Laudan, L. 143, 145 Pierce, M.212
Leibniz, yon, G.F.W. 53, 60 Pirani,FA. 91,95
Leighton, R.H. 54,55 Planck, M.53
Leverrier 138 Podolsky,B. 11,21,38
Lewis, A.A. 151, 190 Poincare, de, H. 191
Lie,S. 220, 244 Popper,K. 30, 39
Lifshitz, E.M.41,43 Posidonius, Stoic 46
Loewer, B. 3-28 Prosperi, G.M. 21
Longer,A. 21 Pythagoras 49

Rayleigh 120 comments on Fine 39

Redhead, M. 20, 22 comments on Humphreys 118
Reichenbach, H. 79, 80, 96 comments on Loewer 22
Rice 188 comments onLuce and Narens 246
Richardson, D. 152, 187, 190, 192 comments on Moulines and Diez 297
Rimini 20 comments on Mundy 97
Robb, A.A. 78,79, 80, 96 comments on Roberts and Rosenbaum
Roberts,F.S. 235, 241, 245,251-274, 251 273
Rogers, H., Jr. 154, 162, 165, 167, 169, 170, comments on Sneed213
172, 173, 178, 179, 188, 190 comments on Vuillemin 55
Rohrlich, F. 105, 117 comments on Wojcicki 146
Rosen, N. 11,21,38 Suszko, R. 144, 145
Rosenbaum, Z.S. 235, 245,251-274 Szmielew, W. 298, 299
Rosenbluth, M. 117
Rosenbluth, A.117 Tarski, A. 119, 126, 129,144, 146, 276
Rosenkrantz, R.D. 213 Taubes, CH. 191
Rubin, H. 213 Teller, P. 32, 38, 77, 81, 82, 97, 117
Rumelhart, D. 118 Tsuji, M. 189,190
Russell, B. 153 Tukey, J.W. 222, 223,244
Turing, A. 119
Sakai, D. 272 Tversky, A. 59, 95, 97, 190, 242, 244, 245,
Sands, M.H. 54, 55 251,252,254,272,274,276, 297
Savage, C.W. 224, 228, 239, 245
Scarpellini, B. 152, 191 Ulam, S. 118
Schild, A. 91,95 Urbach, P. 213
Schrodinger, E. 29—43 passim
Schwinger, J. 41 Vivian, R.G. 213
Scott, D. 220, 223, 245, 251, 272 Voirol, V. 54
Sintonen, M. 198, 200, 213 Vuillemin, J. 45-57
Smolin, J. 190
Sneed, J. 63, 65, 96, 129-131passim, 138, Weber, P. 20
144, 145, 195-216, 275, 277, 297 Whitehead, A.N.74
Solomon, W. 77
Wigner, E.P. 19, 22
Sparrow, C. 119 ff. Winnie, J.A. 80, 97
Specker,E.P. 19,21,32,38 Wittgenstein, yon, L. 48
Spinoza 53 Wojcicki, R. 125-149
Stegmuller, W. 131, 191,275 Woodhouse, N. 91,97
Sternberg, S. 191 Woodward, J. 144, 146
Stevens, S.S. 68. 96,231
Stewart, L. 188, 190 Xv, Shaoji 272

Sugar, A.C. 213 Young, P. 46, 51,190

Suppe, F. 213, 275
Suppes, P. 12, 19, 28, 29, 37, 95, 96, 145, Zangi N 2i
151, 154, 187, 188, 189, 197, 220, Zanotti, M.102, 228, 245
223, 228, 245,251, 252, 254, 272, Zeemail) EC 78> 80> 92> 97
274, 275, 276 Zinnes, JL. 220, 245, 251,252, 272
comments on da Costa andDoria 191

affine manifold93 complexity, computation and119

algebra, causality productof47 computability 114
mid-point 298 insufficiency of189
annealing, simulated 116 computation, complexity and119
application, firm, successful 197 continuum and 247
intended 138 connectionism 116
Archimedean group 223 conservation rule 46
argument,explanationas 203-207 consistency, arithmetical 162, 172
Arrow-Debreu equilibrium 151 contextuality, quantum17
atom, Bohr model 140 continuum, computation and 247
collision 46 representation 230f.
representationof nature 45 correspondence, failure of30
average,quantum probability as 23 rules 125
axiom, basic 278 counting argument75, 77
necessary 81
structural 77 data, Suppes model of 137
axiomatic-representationalapproach 277 decision procedure, Tarski 191
axiomatics, empirical theories 157 Dedekindcompletion 230
definition, coordinating 125
Bohm theory, dfficulties of 23 degree, comlete 169
Boolean logic 8 in arithmetical hierarchy 165
Born rule 3 f., 5-8,10, 14, 18,41 determinism, transcendental 26f.
contradiction with other quantumlaws 7 diagonalization 178
Bourbaki species of structures 154 ff. differentiable manifold93
Brownian motion 25 structure 91, 92, 93
Buffon needle 112 Diophantineequation 151, 164, 179
Buffon-Laplace method113 Dirac set 159
dynamical system, polynomial,
calculus, differential andintegral 60 undecidability of 181 f.
cat, quantum state of11 dynamics, Kantian 51
category 86
Cauchy condition 160 eigenfunction 48 f.
causal structure 80 eigenstate-eigenvalue rule 5, 6
causality, product ofalgebra 47 contradiction with other quantumlaws 7
causation 79 eigenvalue 48 f.
chaos 191 embedding 71
chaos theory 119 f.,187, 188 empricism 105
Churchthesis 114 epistemic interest 196
claim, empirical 131 EPR paradox 11, 12
Co-ranking 260 Schrodinger version 29-43
collapsepostulate 7 equality comparison, meaningfulness 255 ff.
empirical testing of 10 undecidability of165
collision, atomic 46 equivalence, demonstrable 171
combinatorial measurement theory 282 essential maximal element 284
communication theory 50 ether, luminiferous 47
comparison,ordinal, meaningfulness of existence theorem 221
251-274 experiment,hierarchy of models for 148


experimentation, numerical 103-121 theory 25

theoretical model105 hierarchy, arithmetical, rank in 173
explanation as argument 203-207 Hirsch problem 187
evaluation of 207-212 hole argument 94
functional 206 homogeneity 254
partial 205 finite uniqueness,scale type and 231 f.
pragmatic aspects 201 homomorphism 252
scientific 198-203 into SGOR, meaningfulness of ordinal
sketch 199 comparison 267-270
structural 195-216 Hopfieldnet 116
expressibility, arithmetical 171 Huygens equation 26
hypothesis, nearly true 132
factual interpretation 142
Fermat conjecture 151, 173 idealization 71
ferromagnetism 106ff. immaterialism 72
Feynmanpropagator 24 impact, laws of 46
field magnitude 48 incommensurability 79
classical 45-57 incompleteness, central to mathematics 188
field theory,classical, Suppes predicates for finite Nash Game 186f.
157 nastiestkind of 181
theory, XIX century 51 non-arithmetical 175
finite uniqueness,homogeneity and, polynomial dynamical systems 181 ff.
structures 235 ff. theorem 169-174
structures without 235 undecidability and 162-181
formalization, semantic 65 infinite regress 137
set theoretical 64 instrument, scientific 105
syntactic 65 integration, Monte Carlo methodand
functor, forgetful 197 108-112
intended application doctrine 129-131
game, finite, incompleteness of theory 152 interpretation, factual 142
gauge field 159 into representation71
general order relation 253 intrinsic axiom 92
strict 254 statement 90
system 257-264 invariance 65
geometry, analytical 60 Ising model 110, 111
axiomatic 67-71 ferromagnetism 106 ff.
differential, representationtheory 81-90
Euclidean 134 jump 167
representation, quantity and 59-102
representation theorems in 98 Kaluza-Klein unified field theory 162
space-time, axiomatic 77-80, 90-94, 101 Kochen-Specker theorem 4, 12, 17, 38
Gleason theorem 4
Gbdel diagonalization 170 language, formal, undecidability in 153
number 168 informal, unlimitability of149
gravitation, Newtonian 134 representational 87
group, totally ordered Archimedean 223 large numbers, strong law of 109
law, Newton's second 131
halting function 165, 166, 187 special 278
Hamiltonian mechanics 161 Lie algebra 160
heat, diffusion of 47 light, mechanical theory 47
Heisenberg model, ferromagnetism 106 ff. wave, Huygens 51
hidden variable, algebraic constraint 32-35 logic, propositional,quantum, classical 9
quantumtheory, Bohm 12-19 Lorentztransformation 78

Lorentz-Poincare action 160 linear combinatorial, positive 286

group 159 extensive 286 ff.
Lorenz equation 119 f. negative 288-290
Lyapunov theorem 181 Metropolis algorithm 105, 109, 111, 112,
Mach principle 93 Minkowski space 159, 160
magnitude, physical, quantum definition of model class, theory as 195-216
45 element, hierarchical net 278
manifold, topological 84 finite 76
map, Richardson first 163 hierarchy of, for experiments 148
second 164 mathematical 103
materialism 71-77 optimization 206
mathematics, applied vs. pure 128 semantic 127
pure, empirical theory vs. 130 semantical vs. theoretical 139-142, 146
theoretical science vs. 147 set theoretical, Suppes approach
reality and113 126-129
Maxwell equations 26 theoretical, of empiricalphenomenon
mean value theorem 108 ff. 133-139
meaningfulness 65, 229 f. theory and 125-149
condition 62 theory based 140
generalized theory 87-90, 100 Monte Carlo estimator 113
group 63, 82 method 104
ordinal comparison 251-274 integration and 108-112
syntactic criterion 88
theory of68 Nash equilibrium 152
measurement, fundamental 64 Nash game 181
homomorphismdefinition 224 finite, incompleteness 186f.
problem,quantum 18 net, hierarchical 278
Bohm solution 14 theory as275-299
orthodox resolution 7, 11 no-go theorem 35, 36ff.
quantum6 numerical unease 119
representationaltheory 59,61-64,
219-249, 251 observable 105, 114
ancestry 98 contextual 17
history 219 f. quantum, Bohm theory 16
principles 220f. onto representation 71
problems 224 ff. operator, noncommuting 50
structure, finite 230 theory 56
theory, combinatorial, net 275-299 oracle Turing machine 167
error in 227 ff. order relation, general 253
mechanics, analytical 60 temporal 202
classical 66 ordinal comparison, meaningfulness
Hamiltonian 161 257-264
Newtonian134 for homomorphisms267-270
non-Markovian stochastic 37, 40 scale 252
Mercury, movement of,theory 133 ff.
method, quantitative physical 60 pairing function 170
methodology, physical, XVIII century paradigm 128, 143
consensus 60, 98 particle interaction, causal thinking on 41
metric, cyclic 290-293 phenomenon, empirical, theoretical model of
external combinatorial 285 f. 133-139
internal 293-296 photonpropagator 41
linear cobinatorial 286 properties 26

scattering 42 reality, mathematics and113

physical law, intrinsic formulation 66 problem 6
numerical 67 quantum, Bohm solution 14
quantity 64 reasoning, scientific, Bayesian 208
formal theory 94 reduction 264
theory, intrinsic formalization 64-67 relation, intrinsic 61
physics, abstract vs. concrete 53 meaningful 221
finitism in 99 relational structure, primitives of221
psychological theory and 239 f. system,general order, ordinal
pi, value of, empirical estimation 112 f. comparison in 251-274
point, singular, structure with 240 homomorphismof251
stationary, stability problem 181 relationism 71-77, 94
possible world, Bohmian 14 embedding 75, 77
orthodox quantum 8 materialist 75
predicate,nontrivial 171 second order 73 f.
Suppes 155 relationist theory 73
variable degree 76 relativity 60, 78, 79
probability, classical, in Bohm quantum general, classification schemes in 187
theory 15 principle 83
law, quantum 13 representationcondition 62, 63
quantum,average 23 numerical 61
interpretationof 4 construction 232
theory and 3-28 problem252
problem, analytically intractable 115 quantity, geometry and 59-102
unsolvable, construction of151-193 theorem 62, 68
product ofpowers 237 Klein-Stevens approach 62
rule 32,35 representationalcondition 70, 90
proof,algorithmic 188 proposition87
proposition,Bohmian 14 research program 143, 196
representaton 87 tradition 143
psychology, physical theory and239 f. Rice theorem 188
Richardson functor 164
quantity, physical 64 Riemann hypothesis 173
formal theory 59 ff., 94 Riemannian geometry 92
representation, geometry and 59-102
quantumelectrodynamics 41 sampling, simple 110
entanglement 30 scale type,applications 237-240
logic, nonstandard 8 f. combination with structural conditions
state 5 234 f.
collapse 8 homogeneity, finite uniqueness and
theory 60 231f.
Bohm 12-19 scaling 64,67
probability in 15 scattering, photon-photon 42
locality of 12 theory 41
nonrelativistic 5 Schrodinger's cat 5, 6
orthodox, inexplicitness of 10 ff. equation 5, 6, 8, 13, 20, 23, 24
probability and 3-28 contradiction with other quantum
laws 7
random walk 112 science, axiomatic 151-193
quantum,25 f. computational 103
ranking 253 theoretical, pure mathematics vs. 147
rationality, Bayesian 207 semantics, modern126
realism, classical 5 sequence, standard 222

set, finite, structure 230 net,

treelike 278-282passim
theory, Suppes approach to 126-129 physical, intrinsic formalization 99
ZF 151-193 passim psychological andphysical 239 f.
significance, empirical 222 semantic view 275
situation, qualitative 220 theoretical model and 125-149
space-time geometry, axiomatic 77-80 thought experiment, Bohm11
spatialism 71-77 topological manifold 91
spectrum rule 35 transcendental determinism 26 f.
square rule 31 transformation, admissible 252
state, initial 18 translation, structural equivalent to 233 f.
statement, meaningful 252 structure 233
structuralist view 275 trigonometry 70
structure, classification by scale type truth 126
231-234 approximate 131-132
deduced and derived 156 partial or relative132
ground 156 Turing degree 169
mathematical 155 machine 165, 168, 170, 188
partial 156
set theoretic, part of 201 uncertainty relation 1 8
sumrule 31 undecidability, extra-arithmetical 174-181
Suppes predicates 151-193 incompleteness and 162-181
symmetry group 63, 82 nontrivialproperties 176
syntactic transform 88 polynomial dynamical systems 181 ff.
system, economic, incompleteness of Suppes predicates and 151
Hamiltonian model151 uniqueness, finite, homogeneity, scale type
systems theory, dynamical 151 and 231f.
generalized theory 83
tensor analysis 89 group 63, 82, 84
testability,emprical 221 problem 85
theory application vs. formation 140 theorem 62, 68, 77, 221
as net 275-299 unsolvability, algorithmic 177
better thanordering 211
competition between 210 vector analysis 48 ff.
definition 127 field, plane polynomial, limit cycle 181
empirical, axiomatics of 157 quantity 69
identity 276 velocity law, quantum13
pure math vs. 130
formal core of 130 wave, analysis of concept 53
formation 199 equation,boundary constraint 48
Maxwell electromagnetic 158-160 temporal form50
modelbased 140 wave-particle dualism 52
model class 195-198 world view 134
natureof 120
Volume 1: Probability andProbabilistic Causality

Paul Humphreys /Introduction xiii


Karl Popperand David Miller / Some Contributions to

Formal Theory of Probability / Comments by Patrick Suppes 3
Peter J. Hammond / Elementary Non-Archimedean
Representations of Probability for Decision Theory and
Games / Comments by Patrick Suppes 25
Rolando Chuaqui / Random Sequencesand Hypotheses
Tests / Comments by Patrick Suppes 63
Isaac Levi/ ChangingProbability Judgements /
Comments by Patrick Suppes 87
TerrenceL.Fine / Upper andLower Probability /
Comments by Patrick Suppes 109
Philippe Mongin / Some Connections between
Epistemic Logic and the Theory ofNonadditive Probability /
Comments by Patrick Suppes 135
Wolfgang Spohn / On the Properties of Conditional
Independence / Comments by Patrick Suppes 173
Zoltan Domotor / Qualitative Probabilities Revisited /
Comments by Patrick Suppes 197
Jean-Claude Falmagne / The Monks' Vote:
A Dialogue on Unidimensional Probabilistic Geometry /
Comments by Patrick Suppes 239

312 table of contentsTO VOLUME 1


Paul W Holland / Probabilistic Causation Without

Probability / Comments by Patrick Suppes 257
I.J.Good / Causal Tendency,Necessitivity and
Sufficientivity: AnUpdated Review / Comments by
Patrick Suppes 293
Ernest W. Adams / Practical CausalGeneralizations /
Comments by Patrick Suppes 317
Clark Glymour,Peter Spirtes,andRichard Schemes /
InPlace of Regression / Comments by Patrick Suppes 339
D.Costantini / Testing Probabilistic Causality /
Comments by Patrick Suppes 367
Paolo Legrenziand Maria Sonino / Psychologists
Aspects of Suppes'sDefinition of Causality /
Comments by Patrick Suppes 381
Name Index 401
Subject Index 407
Table of Contents to Volumes 2and 3 415
Volume 3: Language,Logic, and Psychology


DagfinnFOLLESDAL / Patrick Suppes' Contribution to the

Philosophy of Language / Comments by Patrick Suppes 3
Michael Bottner / OpenProblems in Relational
Grammar / Comments by Patrick Suppes 19
William C. Purdy / A Variable-Free Logic for Anaphora/
Comments by Patrick Suppes 41
J.Moravcsik /Is Snow White? / Comments by Patrick
Suppes 71
Paul Weingartner / CanThereBe Reasons for Putting
Limitations on Classical Logic? / Comments by Patrick Suppes 89
Jaakko Hintikka and IlpoHalonen / Quantum Logic
as a Logic onIdentification / Comments by Patrick Suppes 125
Maria Luisa Dalla Chiara andRoberto Giuntini /
Logic and Probability in Quantum Mechanics / Comments by
Patrick Suppes 147

W. K.Estes / From Stimulus-Sampling to Array-Similarity

Theory / Comments by Patrick Suppes 171
Raimo Tuomela and GabrielSandu /Action as
Seeing to Itthat SomethingIs the Case /
Comments by Patrick Suppes 193
Colleen Crangle / Command Satisfaction and the
Acquisition of Habits / Comments by Patrick Suppes 223



Maria Carla Galavotti / Some Observations on Patrick

Suppes' Philosophy of Science / Comments by Patrick Suppes 245


Patrick Suppes / Postscript 273

Bibliography of Patrick Suppes 275
Name Index 325
Subject Index 329
Tableof Contents to Volumes 1and 2 333

106. K. Kosik, Dialectics of the Concrete. A Study on Problems of Man and World.
[BostonStudies in the Philosophy of Science, Vol. LII] 1976
ISBN 90-277-0761-8; Pb 90-277-0764-2
107. N. Goodman, The Structure of Appearance. 3rd cd. with an Introduction by G.
Hellman. [BostonStudies in the Philosophy of Science, Vol. LIII] 1977
ISBN 90-277-0773-1; Pb 90-277-0774-X
108. K. Ajdukiewicz, The Scientific World-Perspective and Other Essays, 1931-1963.
TranslatedfromPolish. Edited and with an Introductionby J. Giedymin.1978
ISBN 90-277-0527-5
109. R. L. Causey,Unity ofScience. 1977 ISBN 90-277-0779-0
110. R. E. Grandy, AdvancedLogicfor Applications. 1977 ISBN 90-277-0781-2
111. R. P. McArthur, TenseLogic. 1976 ISBN 90-277-0697-2
112. L. Lindahl, Position and Change. A Study in Law and Logic. Translated from
Swedishby P.Needham. 1977 ISBN 90-277-0787-1
113. R. Tuomela,Dispositions. 1978 ISBN 90-277-0810-X
114. H. A. Simon, Models of Discovery and Other Topics in the Methods of Science.
[BostonStudies in the Philosophy of Science, Vol. LIV] 1977
ISBN 90-277-0812-6; Pb 90-277-0858-4
115. R. D. Rosenkrantz, Inference, Method and Decision. Towards a Bayesian
Philosophy of Science. 1977 ISBN 90-277-0817-7; Pb 90-277-0818-5
116. R. Tuomela, Human Action and Its Explanation. A Study on the Philosophical
Foundations of Psychology. 1977 ISBN 90-277-0824-X
117. M. Lazerowitz, The Language of Philosophy. Freud and Wittgenstein. [Boston
Studies in the Philosophy of Science, Vol. LV] 1977
ISBN 90-277-0826-6; Pb 90-277-0862-2
118. Not published
1 19. J. Pelc (cd.),Semiotics inPoland, 1894-1969. TranslatedfromPolish. 1979
ISBN 90-277-0811-8
120. I.Porn, Action Theory and SocialScience. Some FormalModels.1977
ISBN 90-277-0846-0
121. J. Margolis,Persons andMind.The Prospects of Nonreductive Materialism.[Boston
Studies in the Philosophy of Science, Vol. LVII] 1977
ISBN 90-277-0854-1; Pb 90-277-0863-0
122. J. Hintikka, I. Niiniluoto, and E. Saarinen (eds.), Essays on Mathematicaland
PhilosophicalLogic. 1979 ISBN 90-277-0879-7
123. T.A. F.Kuipers, Studies in Inductive ProbabilityandRationalExpectation. 1978
ISBN 90-277-0882-7
124. E. Saarinen, R. Hilpinen,I.Niiniluoto and M. P. Hintikka (eds.), Essays inHonour
ofJaakko Hintikka on the OccasionofHis 50th Birthday. 1979
ISBN 90-277-0916-5
125. G Radnitzky andG. Andersson (eds.), ProgressandRationalityinScience. [Boston
Studies in the Philosophy of Science, Vol. LVIII] 1978
ISBN 90-277-0921-1; Pb 90-277-0922-X
126. P. Mittelstaedt,Quantum Logic. 1978 ISBN 90-277-0925-4
127. K. A. Bowen,Model Theory for ModalLogic. Kripke Models for Modal Predicate
Calculi.1979 ISBN 90-277-0929-7
128. H. A. Bursen, Dismantling the Memory Machine. A PhilosophicalInvestigation of
Machine Theories of Memory.1978 ISBN 90-277-0933-5

129. M. W. Wartofsky, Models. Representation and the Scientific Understanding.

[BostonStudies in the Philosophy of Science, Vol. XLVIII] 1979
ISBN 90-277-0736-7; Pb 90-277-0947-5
130. D. Ihde, Technics andPraxis. A Philosophy of Technology. [Boston Studies in the
Philosophy of Science, Vol. XXIV] 1979 ISBN 90-277-0953-X; Pb 90-277-0954-8
131. J. J. Wiatr (cd.), Polish Essays in the Methodology of theSocial Sciences. [Boston
Studies in the Philosophy of Science, Vol.XXIX] 1979
ISBN 90-277-0723-5; Pb 90-277-0956-4
132. W. C. Salmon (cd.), Hans Reichenbach: LogicalEmpiricist. 1979
ISBN 90-277-0958-0
133. P. Bieri, R.-P. Horstmann and L. Kriiger (eds.), Transcendental Arguments in
Science. Essays inEpistemology.1979 ISBN 90-277-0963-7; Pb 90-277-0964-5
134. M. Markovic andG. Petrovic (eds.), Praxis. Yugoslav Essays in the Philosophy and
Methodology of the Social Sciences. [Boston Studies in the Philosophy of Science,
Vol. XXXVI] 1979 ISBN 90-277-0727-8;Pb 90-277-0968-8
135. R. Wojcicki, Topics in the Formal Methodologyof Empirical Sciences. Translated
from Polish. 1979 ISBN9O-277-1004-X
136. G. Radnitzky andG. Andersson (eds.), The Structure andDevelopment of Science.
[Boston Studies in the Philosophy of Science, Vol. LIX] 1979
ISBN 90-277-0994-7; Pb90-277-0995-5
137. J. C. Webb, Mechanism, Mentalism and Metamathematics. An Essay on Finitism.
1980 ISBN 90-277-1046-5
138. D. F. GustafsonandB.L. Tapscott (eds.),Body, Mind andMethod. Essays inHonor
of Virgil C. Aldrich. 1979 ISBN 90-277-1013-9
139. L. Nowak, The Structure ofIdealization. Towards a SystematicInterpretation of the
MarxianIdea of Science. 1980 ISBN 90-277-1014-7
140. C. Perelman, The New Rhetoric and the Humanities. Essays on Rhetoric and Its
Applications. Translated from French and German. With an Introduction by H.
Zyskind. 1979 ISBN 90-277-1018-X; Pb 90-277-1019-8
141. W. Rabinowicz, Universalizability. A Study in Morals andMetaphysics. 1979
ISBN 90-277-1020-2
142. C. Perelman, Justice, Law and Argument. Essays on Moral and Legal Reasoning.
TranslatedfromFrench and German. With an Introduction by H.J.Berman. 1980
ISBN 90-277-1089-9; Pb 90-277-1090-2
143. S. Kanger and S. Ohman (eds.), Philosophy and Grammar. Papers on the Occasion
of the QuincentennialofUppsalaUniversity. 1981 ISBN90-277-1091-0
144. T. Pawlowski, Concept Formation in the Humanities and theSocialSciences. 1980
145. J. Hintikka, D. Gruender and E. Agazzi (eds.), Theory Change, Ancient Axiomatics
and Galileo's Methodology. Proceedings ofthe 1978 Pisa Conference on the History
and Philosophy of Science, Volume 1.1981 ISBN 90-277-1126-7
146. J. Hintikka, D. Gruender and E. Agazzi (eds.), Probabilistic Thinking, Ther-
modynamics, and the Interaction of the History and Philosophy of Science.
Proceedings of the 1978 Pisa Conference on the History and Philosophy of Science,
Volumell. 1981 ISBN 90-277-1127-5
147. U. Mdnnich (cd.), Aspects ofPhilosophicalLogic. SomeLogicalForays intoCentral
Notions of Linguistics and Philosophy. 1981 ISBN 90-277-1201-8
148. D.M. Gabbay,SemanticalInvestigations inHeyting's Intuitionistic Logic. 1981
ISBN 90-277-1202-6

149. E. Agazzi (cd.), ModernLogic- A Survey. Historical, Philosophical,and Mathemati-

cal Aspects ofModernLogic and Its Applications.1981 ISBN 90-277-1137-2
150. A. F. Parker-Rhodes, The Theory of Indistinguishables. A Search for Explanatory
Principles below theLevelof Physics. 1981 ISBN 90-277-1214-X
151. J. C. Pitt, Pictures, Images, and ConceptualChange. An Analysis ofWilfridSellars'
Philosophy of Science.1981 ISBN 90-277-1276-X; Pb 90-277-1277-8
152. R. Hilpinen (cd.), New Studies in Deontic Logic. Norms, Actions, and the Founda-
tions ofEthics. 1981 ISBN 90-277-1278-6; Pb90-277-1346-4
153. C. Dilworth, Scientific Progress. A Study Concerning the Nature of the Relation
between Successive ScientificTheories. 2nd, rev. and augmented cd., 1986. 3rd rev.
cd., 1994
ISBN0-7923-2487-0; Pallas Pb0-7923-2488-9
154. D. Woodruff Smith and R. Mclntyre, Husserl and Intentionality. A Study of Mind,
Meaning, and Language. 1982 ISBN 90-277-1392-8; Pb 90-277-1730-3
155. R. J. Nelson, The Logic ofMind.2nd. cd., 1989
ISBN 90-277-2819-4; Pb90-277-2822-4
156. J.F. A. K. van Benthem, The Logic of Time. A Model-TheoreticInvestigationinto
the Varietiesof TemporalOntology, and TemporalDiscourse. 1983; 2nd cd., 1991
ISBN 0-7923-1081-0
157. R. Swinburne(cd.), Space, Timeand Causality. 1983 ISBN 90-277-1437-1
158. E. T. Jaynes, Papers on Probability, Statistics andStatistical Physics. Ed. by R. D.
Rozenkrantz. 1983 ISBN 90-277-1448-7; Pb (1989) 0-7923-0213-3
159. T. Chapman, Time: A PhilosophicalAnalysis. 1982 ISBN 90-277-1465-7
160. E.N. Zalta, Abstract Objects. An Introduction to Axiomatic Metaphysics. 1983
ISBN 90-277-1474-6
161. S. HardingandM.B. Hintikka (eds.), Discovering Reality. FeministPerspectives on
Epistemology,Metaphysics,Methodology,and Philosophy of Science. 1983
ISBN 90-277-1496-7; Pb 90-277-1538-6
162. M.A. Stewart(cd.), Law, Morality andRights. 1983 ISBN 90-277-1519-X
163. D. Mayr andG. Sussmann (eds.),Space, Time, andMechanics. Basic Structures of a
Physical Theory. 1983 ISBN 90-277-1525-4
164. D. Gabbay and F. Guenthner (eds.), Handbook of Philosophical Logic. Vol. I:
Elements of ClassicalLogic. 1983 ISBN 90-277-1542-4
165. D. Gabbay and F. Guenthner (eds.), Handbook of Philosophical Logic. Vol. II:
Extensionsof Classical Logic. 1984 ISBN 90-277-1604-8
166. D. Gabbay and F. Guenthner (eds.), Handbook of Philosophical Logic. Vol. Ill:
Alternative to Classical Logic. 1986 ISBN 90-277-1605-6
167. D. Gabbay and F. Guenthner (eds.), Handbook of PhilosophicalLogic. Vol. IV:
Topics in the Philosophy of Language. 1989 ISBN 90-277-1606-4
168. A. J. I. Jones, Communication and Meaning. An Essay in Applied Modal Logic.
1983 ISBN 90-277-1543-2
169. M. Fitting, Proof Methodsfor Modaland Intuitionistic Logics. 1983
ISBN 90-277-1573-4
170. J.Margolis, Culture andCulturalEntities. Towarda New Unity ofScience. 1984
ISBN 90-277-1574-2
171. R. Tuomela, A Theory ofSocialAction. 1984 ISBN 90-277-1703-6
172. J. J. E. Gracia, E. Rabossi, E. Villanueva and M. Dascal (eds.), Philosophical
Analysis inLatin America. 1984 ISBN 90-277- 1749-4
173. P. Ziff, Epistemic Analysis. A Coherence Theory of Knowledge. 1984

ISBN 90-277-1751-7
174. P. Ziff, Antiaesthetics. An Appreciation oftheCow with theSubtile Nose.1984
ISBN 90-277-1773-7
175. W. Balzer, D. A. Pearce, and H.-J. Schmidt (eds.),Reduction in Science. Structure,
Examples, PhilosophicalProblems.1984 ISBN 90-277-1811-3
176. A. Peczenik, L. Lindahl and B. van Roermund (eds.), Theory of Legal Science.
Proceedings of the Conference on Legal Theory andPhilosophy of Science (Lund,
Sweden,December 1983). 1984 ISBN 90-277-1834-2
177. I.Niiniluoto,IsScience Progressive? 1984 ISBN90-277-1835-0
178. B. K. Matilal and J. L. Shaw (eds.), Analytical Philosophy in Comparative
Perspective. Exploratory Essays in Current Theories and Classical Indian Theories
of Meaning and Reference.1985 ISBN 90-277-1870-9
179. P.Kroes, Time:ItsStructure andRole inPhysical Theories.1985
ISBN 90-277-1894-6
180. ].H.Fetzer,Sociobiology and Epistemology. 1985
ISBN 90-277-2005-3; Pb 90-277-2006-1
181. L. Haaparanta and J. Hintikka (eds.), Frege Synthesized. Essays on the Philosophical
andFoundationalWork ofGottlobFrege. 1986 ISBN 90-277-2126-2
182. M. Detlefsen,Hubert's Program. AnEssay onMathematicalInstrumentalism.1986
ISBN 90-277-2151-3
183. J.L. Goldenand J. J.Pilotta (eds.), Practical Reasoning inHuman Affairs. Studies
in Honor of Chaim Perelman. 1986 ISBN 90-277-2255-2
184. H. Zandvoort, Models ofScientific Developmentand the Case of Nuclear Magnetic
Resonance. 1986 ISBN 90-277-2351-6
185. I.Niiniluoto, Truthlikeness.1987 ISBN 90-277-2354-0
186. W. Balzer, C. U. Moulines and J. D. Sneed, An Architectonic for Science. The
Structuralist Program. 1987 ISBN 90-277-2403-2
187. D. Pearce, Roads to Commensur-ability. 1987 ISBN 90-277-2414-8
188. L. M. Vaina (cd.),Matters ofIntelligence.ConceptualStructures inCognitive Neuro-
science. 1987 ISBN 90-277-2460-1
189. H. Siegel, Relativism Refuted. A Critique of Contemporary Epistemological
Relativism.1987 ISBN90-277-2469-5
190. W. Callebaut and R. Pinxten, Evolutionary Epistemology. A Multiparadigm
Program,with a CompleteEvolutionaryEpistemologyBibliograph.1987
ISBN 90-277-2582-9
191. J. Kmita,ProblemsinHistoricalEpistemology. 1988 ISBN 90-277-2199-8
192. J. H. Fetzer (cd.), Probabilityand Causality.Essays inHonor of Wesley C. Salmon,
with an Annotated Bibliography. 1988 ISBN 90-277-2607-8; Pb1-5560-8052-2
193. A. Donovan, L. Laudan and R. Laudan (eds.), Scrutinizing Science. Empirical
Studiesof Scientific Change. 1988 ISBN 90-277-2608-6
194. H.R. Otto andJ.A.Tuedio (eds.),PerspectivesonMind. 1988
195. D. Batens and J.P. van Bendegem (eds.), Theory and Experiment. Recent Insights
and New Perspectiveson TheirRelation.1988 ISBN 90-277-2645-0
196. J. Osterberg, Selfand Others. A Study of Ethical Egoism. 1988
ISBN 90-277-2648-5
197. D.H. Helman (cd.), Analogical Reasoning. Perspectives of Artificial Intelligence,
Cognitive Science, and Philosophy. 1988 ISBN 90-277-2711-2

198. J. Wolenski, Logic and Philosophy in theLvov-WarsawSchool. 1989

199. R. Wojcicki, Theory of Logical Calculi.Basic Theory of ConsequenceOperations.
1988 ISBN 90-277-2785-6
200. J. Hintikkaand M.B. Hintikka, The Logic of Epistemologyand the Epistemology of
Logic. SelectedEssays. 1989 ISBN 0-7923-0040-8; Pb 0-7923-0041-6
201. E. Agazzi (cd.),Probability in theSciences. 1988 ISBN 90-277-2808-9
202. M.Meyer-(ed.), FromMetaphysics to Rhetoric. 1989 ISBN 90-277-2814-3
203. R.L. Tieszen, Mathematical Intuition. Phenomenology and Mathematical
Knowledge. 1989 ISBN 0-7923-0131-5
204. A. Melnick, Space, Time, and Thought inKant. 1989 ISBN 0-7923-0135-8
205. D.W. Smith, The Circle of Acquaintance. Perception, Consciousness, andEmpathy.
1989 ISBN 0-7923-0252-4
206. M.H. Salmon (cd.), The Philosophy of Logical Mechanism. Essays in Honor of
Arthur W. Burks. With his Responses, and with a Bibliography of Burk's Work.
1990 ISBN 0-7923-0325-3
207. M. Kusch, Language as Calculus vs. Language as UniversalMedium. A Study in
Husserl, Heidegger, andGadamer.1989 ISBN 0-7923-0333-4
208. T.C. Meyering, Historical Roots of Cognitive Science. The Rise of a Cognitive
Theory of Perception from Antiquity to theNineteenthCentury. 1989
209. P. Kosso, Observability andObservationinPhysical Science. 1989
210. J. Kmita,Essays on the Theory ofScientific Cognition. 1990 ISBN 0-7923-0441-1
211. W. Sieg (cd.),Acting and Reflecting. The Interdisciplinary Turn inPhilosophy. 1990
ISBN 0-7923-0512-4
212. J. Karpiriski, Causality inSociological Research. 1990 ISBN 0-7923-0546-9
213. H.A. Lewis(cd.), PeterGeach:PhilosophicalEncounters. 1991
ISBN 0-7923-0823-9
214. M. Ter Hark, Beyond the Inner and the Outer. Wittgenstein's Philosophy of
Psychology. 1990 ISBN 0-7923-0850-6
215. M. Gosselin, Nominalism and Contemporary Nominalism. Ontological and
EpistemologicalImplications of the Work of W.V.O. Quine and of N. Goodman.
1990 ISBN 0-7923-0904-9
216. J.H. Fetzer, D. Shatz and G. Schlesinger (eds.), Definitions and Definability.
PhilosophicalPerspectives. 1991 ISBN 0-7923-1046-2
217. E. Agazzi and A. Cordero (eds.), Philosophy and the Origin and Evolution of the
Universe. 1991 ISBN 0-7923-1322-4
218. M. Kusch, Foucault's Strata and Fields. An Investigation into Archaeological and
Genealogical ScienceStudies. 1991 ISBN 0-7923-1462-X
219. C.J. Posy, Kant's Philosophy ofMathematics.Modern Essays. 1992
ISBN 0-7923-1495-6
220. G. Van de Vijver, New Perspectives on Cybernetics. Self-Organization,Autonomy
andConnectionism. 1992 ISBN 0-7923-1519-7
221 J.C. Nyiri, Traditionand Individuality. Essays. 1992 ISBN 0-7923-1566-9
222. R. Howell, Kant's TranscendentalDeduction. An Analysis of Main Themes in His
CriticalPhilosophy. 1992 ISBN 0-7923-1571-5

223. A. Garciade la Sienra, The Logical Foundations of the Marxian Theory of Value.
1992 ISBN 0-7923-1778-5
224. D.S. Shwayder, Statement and Referent. An Inquiry into the Foundations of Our
Conceptual Order.1992 ISBN 0-7923-1803-X
225. M. Rosen, Problemsof the HegelianDialectic.Dialectic Reconstructed as a Logic
of HumanReality. 1993 ISBN0-7923-2047-6
226. P. Suppes,ModelsandMethods in the Philosophyof Science: SelectedEssays. 1993
227. R. M. Dancy (cd.), Kant and Critique: New Essays inHonor of W. H. Werkmeister.
1993 ISBN0-7923-2244-4
228. J. Woleriski (cd.), PhilosophicalLogic in Poland.1993 ISBN0-7923-2293-2
229. M.De Rijke (cd.), DiamondsandDefaults. Studies inPure and AppliedIntensional
Logic. 1993 ISBN0-7923-2342-4
230. B.K. Matilal and A. Chakrabarti (eds.), Knowing from Words. Western and Indian
PhilosophicalAnalysis of Understanding and Testimony.1994
ISBN 0-7923-2345-9
231. S.A. Kleiner, The Logic of Discovery. A Theory of the Rationality of Scientific
Research. 1993 ISBN 0-7923-2371-8
232. R. Festa, Optimum Inductive Methods. A Study in Inductive Probability, Bayesian
Statistics, and Verisimilitude.1993 ISBN 0-7923-2460-9
233. P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher. Vol. 1: Probability and
ProbabilisticCausality. 1994
ISBN 0-7923-2552-4; Set Vols.1-3 ISBN0-7923-2554-0
234. P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher. Vol. 2: Philosophy of
Physics, Theory Structure, andMeasurement Theory. 1994
ISBN 0-7923-2553-2; Set Vols.1-3 ISBN0-7923-2554-0
235. P. Humphreys (cd.), Patrick Suppes: Scientific Philosopher. Vol. 3: Language,
Logic,andPsychology. 1994
ISBN 0-7923-2862-0; Set Vols. 1-3 ISBN 0-7923-2554-0
236. D. Prawitz and D. Westerstahl (eds.),Logic and Philosophy of Science in Uppsala.
Papers from the 9thInternationalCongress of Logic,Methodology,andPhilosophy
of Science. 1994 ISBN 0-7923-2702-0
237. L. Haaparanta (cd.), Mind, Meaning andMathematics. Essays on the Philosophical
Viewsof Husserl and Frege. 1994 ISBN 0-7923-2703-9
238. J. Hintikka (cd.): Aspects ofMetaphor. 1994 ISBN 0-7923-2786-1
239. B. McGuinness and G. Oliveri (eds.), The Philosophy of Michael Dummett. With
Replies fromMichaelDummett.1994 ISBN0-7923-2804-3
240. D. Jamieson (cd.),Language, Mind, and Art. Essays in Appreciation and Analysis,
InHonor ofPaul Ziff. 1994 ISBN0-7923-2810-8
241. G. Preyer, F. Siebelt and A. Ulfig (eds.), Language, Mindand Epistemology. On
DonaldDavidson's Philosophy. 1994 ISBN0-7923-28 11-6
242. P. Ehrlich (cd.), Real Numbers, Generalizations of the Reals, and Theories of
Continua. 1994 ISBN0-7923-2689-X
243. G Debrock and M. Hulswit (eds.): Living Doubt. Essays concerning the epistemol-
ogy of Charles SandersPeirce. 1994 ISBN 0-7923-2898-1

Previous volumes are still available.