Logical reflections on the semantic approach

April 26, 2013

Abstract
Patrick Suppes is one of the first and most eminent defenders of
the semantic approach to scientific theories. As is well known, the
main feature of the semantic approach is supposed to be the shift from
the focus on languages and sentences to models. According to this
view, it is models that play the most important role in representing
reality. Even though this is taken to be the “standard” approach to the
problem nowadays, there is little consensus on what models are, and,
consequently, on what scientific theories are. A great variety of propos-
als have been developed in the literature, and Suppes has emphasized
that, in this case, models should be understood as set-theoretical struc-
tures. This is a major step towards converting the semantic approach
into a rigorous position. However, even if we choose that route, some
fine logical points remain to be discussed. In this paper we propose to
reflect on the role of the underlying mathematical basis employed in
the formulation of the semantic approach in the Suppesian sense. Once
models are taken to be set-theoretical structures, we should not expect
that the underlying set theory should go unnoticed. We analyze the
influence of the underlying set theory employed in the construction of
the models and argue that philosophers of science would benefit from
that kind of foundational study. Some common slogans of the seman-
tic approach, such as “a theory is a class of models”, “in philosophy
of science we should employ mathematics and not metamathematics”
and “a theory is language free”, will be examined and their ‘truth’ will
be properly assessed.
Key-words: models; set-theoretical structures; semantic approach;
scientific theories.

1
1 Introduction
It is widely held nowadays that the semantic view successfully replaced the
syntactic view on what concerns the philosophical discussion on the nature
of scientific theories and has become the current orthodoxy. As philosophical
legend has it, the syntactic view suffered from a broad range of difficulties,
spanning from some intricacies due to some features of the project to exter-
nal (well-founded) criticism. The locus of the main opposition between the
semantic and the syntactic approach, as it is usually put, concerns the use
of language in the formulation and identification of a scientific theory: while
the syntactic view focused on formal axiomatizations, the semantic view fo-
cused on models; that is, while the first approach (it is said) took theories to
be linguistic entities, the latter (it is said) considered them as non-linguistic
entities.
As the tale usually goes, the syntactic conception of scientific theories
identified a theory with its linguistic axiomatic formulation. In order to
present a theory, then, one has to present both a logical and a non-logical
vocabulary, to state the rules of sentence formation, to choose among the
sentences some of them to play the role of axioms, and finally, to go ahead and
derive theorems. The relationship between such a theory and the empirical
data is provided by a set of correspondence rules that attributes meaning (at
least) to some of the theoretical terms [18, chap.4, E].
In some sketches of this view, philosophers have held that the logical
vocabulary and the logical axioms comprise only what is presently known
as first-order logic, so that the deductive apparatus would comprise only
that logic.1 By its very linguistic nature, the identification of a theory with
a linguistic object already raises some questions which may have troubled
the proponents of the syntactic view; here we shall recall only one, which
is directly related to the identification mentioned. The trouble comes from
the fact that by employing distinct sets of primitive non-logical vocabulary
we may present alternative axiomatizations of what one would intuitively
call “the same theory”. In this case, despite the fact that we are getting
different axiomatizations of “the same subject”, identification of a theory
with its formulation would amount to counting the distinct axiomatizations
as distinct theories, for they may differ concerning the primitive assumed
1
Although Suppe acknowledges that in what he calls “the final version” of the approach
the first-order languages could be enlarged by modal operators [18, p.50].

2
terms and postulates.
On the other hand, it is generally told that the semantic view steered
away from language and linguistic formulations and took models as playing
the major role in the presentation of a scientific theory. To present a theory,
one must somehow describe a class of models, the models of the theory (of
course, that is plainly circular if taken as a precise characterization, but let us
just leave that for the moment). A quick look in the philosophical literature
on models is enough to grant that the word model is far from univocal, and to
be sure one must specify first what is meant by that word. In the beginning
of the model theoretic approach in the 1960s, the precursors of the view
relied on models in the same sense as the newly born Tarskian model theory:
models are taken to be mathematical entities, set theoretical structures of
the same kind as used in logic. Being non-linguistic,2 characterizing a theory
in terms of models, it seems, the semantic approach would not suffer from the
same difficulties as the ones the syntactic approach had to face. To mention
one of the advantages of the semantic approach over the syntactic approach,
it is said to be much closer to actual scientific practice, since scientists are
model-builders by nature, while axiomatizing and proving theorems would
be usually not their concern at all.
The opposition between both views, as we have presented it, is, we agree,
rather sketchy and schematic (and we do not agree with everything the tradi-
tion attributes to the syntactic view). However, it is enough for us to present
the main tension created by the growing self-awareness inside the semantic
tradition itself that some use of language is not wholly indispensable for this
approach. That is, one must use the resources of some linguistic formulation
of a theory in semantic tradition too, and that reliance on linguistic formu-
lations may be more than a mere auxiliary step to go to the required class of
models. To put the point again, given that a theory is to be considered as a
class of models, how do we select the relevant models? Just point to a set of
sentences in some language which may have the required structures as mod-
els, that is, sentences forming an axiom system for that class of structures.
That is clear from the very Tarskian tradition being followed by the seman-
tic view. However, as some have already pointed out (see Chakravartty [2],
Muller [16], and Thomson-Jones [26]), that feature makes the alleged major
distinctions between the semantic view and the syntactic view look rather
2
In terms, for A. Church, in his classical book [5, §09] has already called our attention
to the fact that “there is a sense in which semantics can be reduced to syntax”.

3
exaggerated. Maybe some of the features of the syntactic view that were
seen as deplorable by its critics’ eyes may plague also the semantic view.
However, the repudiation of the syntactic view, as we mentioned before,
was not only related to the essential use of a linguistic formulation, but also
with some of the internal difficulties associated with the project. Here, ‘the
project’ means the Logical Positivists’ view of science that was coupled with
the syntactic approach from its very inception. Their desire to attribute
empirical significance to science according to their verificationists tenets re-
quired that the specific vocabulary of a theory should be divided in two parts:
the observational vocabulary and the theoretical vocabulary. Observational
vocabulary had empirical significance through a somehow direct correlation
with experience, while theoretical vocabulary had a more difficult and indi-
rect association. Anyhow, this distinction and the necessity to link theoret-
ical terms with empirical observation led the Positivists to a whole bunch
of concepts that were, indeed, very far away from actual scientific practice.
Probably it is their required distinction between theoretic and observational
terms and their verificationist theory of meaning applied to the interpreta-
tion of scientific theories that led the syntactic view to its demise, not the use
of a linguistic formulation of a theory. Of course, that does not mean that
the syntactic view without this division in vocabulary and its consequences
is a better alternative for philosophers of science, but rather that the main
difference between the semantic and the syntactic approaches lies not in the
use of some language in presenting the theory, but in how theory relates to
experience. That is, it is not necessary that by employing language in some
way the semantic view should be directly identified with the syntactic view:
there is a specific approach to the relation between theory and experience
which accounts for their main feats and differences.
In this paper we focus on the relationship of the semantic view with
language, hoping to enlighten this relationship. We shall first argue that
there is a historical confusion which the pioneering works of Suppes some-
how avoided, and which should be reconsidered on a reasonable formulation
of the semantic approach. Also, after having freed ourselves of that confu-
sion we shall present two alternative ways of formulating a scientific theory,
both pursued by the characterization of a set-theoretical predicate. The first
one, which we call axiomatization in Suppes’ style, writes the axioms in the
language of set theory (i.e some set theory) properly,3 while the second one,
3
Other frameworks could be used instead; it should be noted that Carnap [3] made use

4
which we call da Costa-Chuaqui style, advances the axioms in a specific for-
mal language (here we follow the distinction advanced in Krause, Arenhart
and Moraes [14]). Both approaches have their merits, both, as we shall see,
are in complete agreement with the semantic tradition as we formulate it, and
more than that, they both serve to distinct purposes for the philosopher of
science. We finish by discussing the role of formal methods in the philosophy
of science. Our view is that rigorous discussions backed by logical analysis
should be provided for where available; increase in clarity and avoidance of
pitfalls may help us gaining a deeper understanding of the semantic approach
and, consequently, of the nature of scientific theories.

2 Suppes against the syntactic approach

To begin by making clear how did some philosophers arrive at the idea that
the semantic approach requires that we avoid any use of language, let us
begin by recalling how Suppes himself saw the opposition between the new
semantic approach and the syntactic approach. We believe that by recalling
some of Suppes’ arguments against the syntactic tradition we may have a
better idea of the underlying motivations of the semantic approach and of
the problems it was supposed to solve. This may serve our purposes in
establishing that it is not the case that any kind of use of language in the
semantic approach should be seen as a return to the deceased syntactic view;
language, properly employed, is a part of the semantic approach. Indeed,
the way language is employed in the semantic approach may have important
consequences for us to understand its scope and its limitations.
We first recall that the syntactic approach is identified by Suppes him-
self as including axiomatization within a system of first-order logic as one of
its main features (which he calls standard formalization, or intrinsic char-
acterization [24, p.5]). The problem, as he puts it, is whether a theory can
be axiomatized this way and whether such an axiomatization is useful for
gaining philosophical insights into the nature of the scientific enterprise.
Of course, the main trouble with such a framing of the approach comes
from the fact that the whole of the mathematics employed by the theory being
axiomatized will need to be developed from scratch too. The presentation of a
of higher-order logic, contrary to the standard view on the syntactic approach. Also, we
don’t know yet the advances in this field by considering category theory, but this possibility
should be investigated too.

5
highly mathematical theory like General Relativity will have a mathematical
introduction which will take more space and work to be stated in the required
formal framework than the whole specific aspects of the theory proper. So, it
seems, introductory work will require so much effort that the main goal, the
study of the nature of scientific theories will be lost in the details involved
in the axiomatization process. Furthermore, that way of proceeding is so
far away from actual scientific practice that one cannot hope to get a better
understanding of current science by following this approach.
To overcome that kind of difficulty, Suppes suggests that it would rather
be more appropriate to have an axiomatization in his sense, namely, an ex-
trinsic characterization, which defines the theory’s class of models directly
inside some informal set theory. Since all of the required mathematics is
supposed to be already available, all we must care about is to furnish some
axioms to capture the intended set of models. So, the shift to models al-
lows us to deal only with the required structures that may be employed to
characterize the theory. Furthermore, in defending the semantic approach
against the syntactic one, Suppe [19, p. S104] claims that this is exactly
what we do in practice, so that the semantic approach is in tune with scien-
tific practice. And he is right: in usual mathematical practice, for instance,
in presenting group theory we just provide the axioms for groups, and not
for all the logic paraphernalia that precedes them. Then, the problem turns
to the axiomatizability of this class of structures; as Suppes says,

One of the simplest ways of providing such an extrinsic charac-

terization is simply to define the intended class of models of the
theory. To ask if we can axiomatize the theory is then just to
ask if we can state a set of axioms such that the models of these
axioms are precisely the models in the defined class. [24, loc.cit]

That is, axiomatization is not to be avoided by the semantic approach, it

is rather required by the semantic approach. The nature of the axioms, in the
sense of whether they are stated in some formal language (which, for our later
purposes we may take as being some first or higher order language) or directly
in some informal set theory, is not an issue settled automatically by the simple
fact that we adopt the semantic approach. Suppes himself is ambivalent
as to which option one should adopt. In some places, Suppes speaks as if
models were taken in the Tarskian sense, namely, as structures “satisfying a
theory” (see Suppes [21, p. 289]). In other places (see [25, sec. 2]), however,

6
Suppes recognizes that if by “semantic view” we mean the stipulation of a
set of structures satisfying a set of sentences of a sharply designed formal
language, then his view is not the semantic view, since axiomatization by a
set theoretical predicate in Suppesian sense does not require satisfaction of a
set of sentences in the Tarskian sense (see also Muller [16, sec.6]). Our point
is that both views are allowable for the semantic approach, and constitute
merely distinct versions of this view (see our discussion below).
Before we go ahead, let us pause for a minor consideration of the first
alleged opposition between the syntactic and the semantic approaches. The
point until now has been that a shift to the semantic approach would allow us
to avoid axiomatizing the whole of mathematics required for the characteri-
zation of a scientific theory from rock bottom. Now, as we have mentioned,
it is not true that the adherents of the syntactical view adopted really a first
order language to advance their axiomatizations (it is enough to take a look
at the examples advanced in Carnap [3]). Indeed, they took Russell’s type
theory as their standard logic (with Carnap studying diverse alternative log-
ical systems), and just as it happens in ZFC, all of the mathematics required
for a scientific theory may be developed inside type theory, although in a
more cumbersome way than inside ZFC. So, why couldn’t they just claim as
Suppes does that the math is already available, and proceed directly to the
specific axioms? Despite the fact the working inside type theory is not really
something we are used to, there is nothing objectionable to that move. This
consideration strengthens our position, according to which it is not the use
of axioms or of a particular language that distinguishes the semantic and the
syntactic view, but rather further consideration which are taken into account
in the next point4 .
For a second point, we must recall that besides an axiomatization of a
theory, the syntactic approach had to include some coordination rules to
provide an empirical interpretation of the theory. The vocabulary had to
be divided in two parts: the theoretical and the observational vocabulary.
The axioms of the theory had to be formulated by employing the theoretical
vocabulary, and the meanings of terms referring to unobservable items had to
be attributed by coordinating rules that related them with the observational
consequences expressed in the observational vocabulary. There is a bunch
4
We do not deny, of course, that working inside informal set theory is much easier and in
accordance to common practice. Our point is only that the idea that one must axiomatize
the mathematics of the theory together with the theory is based on a misconception of
the logical positivists’ approach, in particular of the logical bases they employed.

7
of problems related with the distinction of the vocabulary and the crude
interpretation of observational terms.
Here comes one of Suppes strongest claims in favor of the semantic ap-
proach. Indeed, one of the main advantages of dealing with models rather
than with a linguistic formulation concerns the relationship between theory
and experimental data. The raw correspondence between interpreted terms
and experience based on the observational vocabulary is the main target
of Suppes in many occasions (in particular, see [22], [23, pp. 62-63]). As
we have mentioned, the syntactic approach, employed mainly for the anti-
metaphysical purposes of the Logical Positivists, had a rather crude view on
these relationships.
According to Suppes [23, loc. cit.], [24], the relationship of a theory with
phenomena is not direct, for there is much elaboration and idealization to
make the phenomena apt to quantitative treatment. We proceed by the elab-
oration of data structures, which gradually evolve to more and more elaborate
structures. Theories do not face experience directly, rather they are related
to data structures (see the account in Suppes [22]). The results taken from
brute data must be labored upon to the construction of a data structure,
which would be shown to be isomorphic to some arithmetical structure in
order to “attribute numbers to things”, which on its turn enables us to go
from qualitative information to quantitative concepts. It is data structures
which are then embedded in some model of the theory. What is relevant here
is that the characterization of theories in terms of models, or set-theoretical
structures, is in no way problematical just because they are taken as axiom-
atized by some set of sentences. Indeed, Suppes requires that there is some
characterization of the theory in such a way as to obtain its models (set-
theoretical structures), but for some purposes it does not matter that the
characterization is provided for in some logic (for instance, first-order logic),
or rather inside set theory directly.
Let us make this point clear: models play a central role in characterizing
theories due to their relationships with data models. The crude account of
these relationships by the syntactical view is seen by Suppes as its main
defect and accounts for its unrealistic and sketchy nature. It is not the use
of logical notation by itself that is problematical. That is, there is nothing
to prevent one from axiomatizing a theory in some formal system of first- or
higher-order logic, but it should be noticed that the whole of mathematics
employed by the theory will have to be axiomatized together, rendering the
process tedious and artificial. In the case of highly mathematical theories

8
of physics, for instance, the set theoretical characterization of structures is
to be preferred because it avoids a whole bunch of distracting work, given
that the underlying mathematics is taken as already available. In any case,
however, the same set theoretical structures are present in both kinds of
axiomatization.
We may quote Suppes in defending the use of models ([22, p.252]; note
also the mentioned ambivalence as to the sense of “model” being employed):
To provide complete mathematical flexibility I shall speak of the-
ories axiomatized within general set theory by defining an ap-
propriate set theoretical predicate (e.g. “is a group”) rather than
theories axiomatized directly within first-order logic as a formal
language. [. . .] But whichever sense of formalization is used, es-
sentially the same logical notion of model applies.
Working inside set theory spares us from a detailed treatment of the
whole mathematical basis required by the theory, but is not mandatory for
the view Suppes is advocating. The essential is that we must be able to
provide the set-theoretical models needed to account for the relationship of
the theory with the data models (which come themselves as structures). So,
the first problem, axiomatization in first-order logic, is really a defect of the
syntactical approach as long as it is artificial and, in most times, does not
reflect scientific practice. However, the real problem comes when to that
axiomatization one adjoins the idea that an interpretation should be given
through some coordinate rules to relate theory with experience, along with a
division in the vocabulary between observational and theoretical terms. That
is how the second great difficulty of the syntactical approach is supposed to
be solved by the focus on models.
That is how both issues relate in the Suppesian approach: axiomatiza-
tion through set theoretical predicates makes it easier for us to furnish models
that may fit data structures. Data and theory are related in very complex
ways. The axiomatization, as we have been discussing, is an essential step in
gathering the adequate structures. To present axioms one must employ one
language or another, but the required structures must be somehow charac-
terized (and depending on the language employed, as we shall discuss in the
next section, distinct characterizations may be originated).
So, there are two ingredients involved in the discussion: the use of set
theory in an axiomatization and the use of models to account for the re-
lation with data and (indirectly) with phenomena. In the next section we

9
investigate how axiomatization may proceed according to two distinct fronts.
On one possible path, we explicit Suppes approach to gather the structures
by employing a set theoretical predicate in which the language of set theory
itself is employed. On the other front, a specific language is designed, and
the axioms for a theory are written in that language. In this second case,
although there is no escape from axiomatizing the underlying mathematics
of a theory, greater rigor is achieved and some interesting results from meta-
mathematics may be employed. Our main point, as we shall defend in the
concluding section, is that distinct kinds of analysis may be provided for by
both methods. Distinct kinds of insight may be achieved in any kind of ap-
proach, so that the one to be used will depend on our aims and purposes, that
is, by pragmatic criteria. We shall propose, however, that both approaches
are part of the semantic approach.

3 Structures and models

We shall be interested in presenting two basic methods of axiomatization, two
approaches to characterize a class of structures. One of them is recommended
by Suppes himself, and proceeds by employing the language of set theory
and all of its resources in the description of the axioms. The another was
advanced by da Costa and Chuaqui (see [7]), and the axioms are presented in
a conveniently designed logical vocabulary. As we shall see, both approaches
differ in strength and technical resources available to them; both serve to
distinct purposes accordingly inside the semantic approach. More important
than that, both deserve to be called “semantic approach” and are indeed so
called (sometimes confusingly) in the literature. We intend to clarify matters
by distinguishing them and by presenting their main features.

3.1 Structures
We shall be working inside first-order Zermelo-Fraenkel set theory with the
axiom of choice, ZFC (but of course this framework could be modified if nec-
essary). Our first immediate goal is to present a general theory of structures
that should be rigorous enough for our attempt in designing both versions
of the semantic approach we shall investigate in this section. We basically
follow da Costa and Rodrigues in the definitions [8].
We begin by defining the concept of type:

10
Definition 3.1 (Types) The set τ of types is the least set satisfying the
following conditions:
(i) i ∈ τ

(ii) If a0 , a1 , . . . , an−1 ∈ τ , then ha0 , a1 , . . . , an−1 i ∈ τ , with 1 ≤ n < ω.

As usual, ω stands for the set of all natural numbers. Next we introduce
the concept of order of a type in τ .

Definition 3.2 (Order) If a ∈ τ , the order of a (henceforth ord(a)), is

defined as follows:
(i) ord(i) = 0

(ii) ord(ha0 , a1 , . . . , an−1 i) = max{ord(a0 ), ord(a1 ), . . . , ord(an−1 )} + 1.

In the next definition, the usual set-theoretical operations of power set P

and cartesian product × are being employed.

Definition 3.3 Given a non-empty set D, we define a function t, having τ

as its domain, as follows:
(i) t(i) = D

(ii) If a0 , a1 , . . . , an−1 ∈ τ , then t(ha0 , a1 , . . . , an−1 i) = P(t(a0 ) × t(a1 ) ×

. . . × t(an−1 )).

Thus, following the usual extensional set-theoretical definition of a rela-

tion as a set of ordered pairs, for each type a, the mapping t gives us the
set of all relations of that type, built from the elements of the domain D.
Obviously, as the orders of the types increase, the objects attributed to them
by t increase in complexity too. We say that the elements of t(a) have type a.
For example, the elements of type i are, by definition, the elements of D, and
are called individuals. Then, hii is the type of the properties of individuals,
hi, ii is the type of the binary properties of individuals, hhiii is the type of
the unary properties of unary properties of individuals, that is, sets of sets
of individuals, and so on.
S
Definition 3.4 (Scale) The scale based on D is the set (range t(D)),
denoted by ε(D).

11
Definition 3.5 (Cardinal) The cardinal κD associated to ε(D) is defined
as
κD = sup{|D|, |P(D)|, |P 2 (D)|, . . .}.

Definition 3.6 (Structure) A structure A based on D is an ordered pair

of the form
A = hD, Rι i
where Rι is a sequence of elements of ε(D) (with ι ranging over a set of
indices), and we suppose that the domain of this sequence has cardinal strictly
less than κD . We say that κD is the cardinal associated with A, and that ε(D)
is the scale associated with A.

As we said before, each element of ε(D) has a certain type, for it belongs
to t(a) for some a ∈ τ . Now, the order of a relation is defined as the order of
its type. The order of A, denoted ord(A), is the order of the greatest of the
types of the relations of the family Rι , if there is one, and if there is no such
relation, we put ord(A) = ω.
If a structure has order κ, we say that it is an order–κ structure. Groups
are examples of order–1 structures, for they contain only order–1 relations
(actually, depending on the formalization, they contain just one order–1 re-
lation of type hi, i, ii, for the binary operation on the domain can be seen
as a ternary relation). There are some simple examples of structures which
are not order–1 structures such as well-orderings, ciclic groups, Dedekind-
complete ordered fields, and topological spaces. Empirical theories, whose
structures are highly abstract and employ heavy mathematics are also not
order–1 structures, and this is quite important to note, for in consequence,
the models of such theories are not order-1 structures (typical of standard
model theory) as it seems to be assumed by many philosophers.5

3.2 The language of a structure

Of course we aim at to speak of structures and of all the objects of a scale. To
do that in an adequate and general way, we consider two basic higher-order
languages, termed (cf. [8]) Lωωω (R) (or simply Lω (R)) and Lωωκ (R).
5
Really, it seem to us that when philosophers speak of models of scientific theories,
they do not realize that these structures are in general not order–1 structures. It is com-
mon to find philosophers speaking about applying Löwenheim-Skolem or the Compactness
theorems to scientific theories in the semantic approach.

12
 In general, an infinitary language Lηµκ , with κ < µ being infinite cardi-
nals (or ordinals) and 1 ≤ η ≤ ω, enables us to consider conjunctions and
disjunctions of n ≤ µ formulas and blocks of quantifiers with m < κ many
quantifiers. The superscript η indicates the order of the language (first-order,
second order, etc.). In both cases R is the set of the non-logical constants
of the language. Thus, in Lωµω (R) (ω < µ) we may have infinitely many con-
junctions and disjunctions of formulae, but blocks of quantifiers with finitely
many quantifiers only. Lωωω (R) is a higher-order language, suitable for type
theory (higher-order logic). Standard first-order languages are of the kind
L1ωω , so is L∈ , the basic language of ZFC. Put in a more precise way,

Definition 3.7 (Order of a language) A language Lnµκ , with 1 ≤ n < ω,

is called a language of order n. A language Lωµκ is said to be or order ω.

A language of order n contains only types of order t ≤ n and quantifica-

tion of variables of types having order ≤ n − 1 [8, p.8].
In order to exemplify how we can define a higher-order language using a
first-order language (such as L∈ ), let us sketch the language Lωω1 ω (R), but we
could consider whatever infinitary language Lωµκ , provided that the involved
cardinals exist in ZFC (for instance, we couldn’t use an inaccessible cardinal).
The primitive symbols of Lωω1 ω (R) are the following ones:
V W
(i) Sentential connectives: ¬, ∧, ∨, →, , and .

(ii) Quantifiers: ∀ and ∃

(iii) For each type t, a family of variables of type t whose cardinal is ω.

(iv) Primitive relations: for any type t, a collection of constants of that

type (possibly some of them may be empty). The collection of these
constants form the set R. For instance, for groups, R comprises constants
for S and ◦, while for vector spaces it comprises constants for V , F , +
and · at least.

(v) Parentheses: left and right parentheses (‘(’ and ‘)’), and comma (‘,’).

(vi) Equality: =t of type t = ht1 , t2 i, with t1 and t2 of the same type.

Variables and constants of type t are terms of that type. If T is a term

of type ht1 , . . . , tn i and T1 , . . . , Tn are terms of types t1 , . . . , tn respectively,

13
then T (T1 , . . . , Tn ) is an atomic formula. If T1 and T2 are terms of the same
type t, then T1 =ht1 ,t2 i T2 is an atomic formula. We shall write T1 = T2
V If α,
for this last formula, leaving the type of the identity relation implicit.
W αi are formulas (i = 1, . . . ), then ¬α, α ∧ β, α ∨ β, α → β, αi , and
β,
αi are formulas. Then, we are able to write formulas with denumerably
many conjunctions and disjunctions. Furthermore, if X is a variable of type
t, then ∀Xα and ∃Xα are also formulas (and only finite blocks of quantifiers
are allowed). These are the only formulas of the language. The concepts of
free and bound variables and other syntactic concepts can be introduced as
usual.
Now let A = hD, Rι i be a structure, where Rι ∈ R, that is, the primitive
relations of the structure are chosen among the constants of our language
Lωω1 ω (R). Then, Lωω1 ω (R) can be taken as a language for A = hD, Rι i, pro-
vided that κD = ω (recall that κD is the cardinal associated to A). Still
working within (say) ZFC, we can define an interpretation of Lωω1 ω (R) in
A = hD, Rι i in an obvious way, so as what we mean by a sentence S of
Lωω1 ω (R) (a formula without free variables) being true in such a structure in
the Tarskian sense, in symbols,

A |= S. (1)

In the same vein, we can define the notion of validity. A sentence S is

valid, and we write
|= S,
if A |= S for every structure A.
Important to emphasize that we are describing the language Lωω1 ω (R)
using the resources of some set theory such as ZFC. This way, we can speak
of denumerable many variables, for instance, in a precise way. In this sense,
any symbol of Lωωκ (R), as we have remarked already, can be seen as a name
for a set. Thus, ‘(’ (the left parenthesis), for example, names a set (for a
detailed description of formal languages inside set theory, see [1, sec. 3]).
Now let A = hD, Rι i be a structure, while rng(Rι ) denotes the range of
Rι . Remember that Rι stands for a sequence of relations of the scale ε(D),
that is, it is a mapping from a finite ordinal into a collection of relations
in the scale. Thus rng(Rι ) stands for just the set of these relations. So,
Lωωω (rng(Rι )) (= Lω (rng(Rι ))) the basic language of the structure (it is not
the only language to speak of the structure, for other stronger languages
encompassing it could be used instead). In this case, we can interpret a

14
sentence containing constants in rng(Rι ) in A = hD, Rι i, and to define the
notion of truth for sentences of this language according to this structure in
an obvious way.
To summarize, given a certain structure, we may construct within ZFC
a formal language that “speaks” of the elements of the scale based on the
domain of the structure. The above definitions are quite general, but show
that we can construct very strong languages within ZFC for the purposes of
talking about certain structures.

4 Da Costa-Chuaqui approach
The approach we shall call da Costa-Chuaqui (CCA for short) is a first step
towards a rigorous presentation of how a certain class of structures may stand
for a theory. That is, given an informal scientific theory, we may employ the
methods provided by this approach to see the theory as a class of models.
We begin with this approach because it is not the original version intended
by Suppes (not, at least, taking into account his latter claims; but check the
ambivalence mentioned in section 2), nor by latter philosophers endorsing the
semantic view. CCA really adds some rigor to the semantic view by requiring
that the postulates of the theory be formulated in a precisely stated formal
language, as described in the previous subsection 3.2.
To put the approach in a nutshell, following the original presentation in
da Costa and Chuaqui [7], to present a theory we proceed as follows:

• We start with a structure built in ZFC according to 3.1 (or any other
set theory which may happen to be the framework we are working
in). This structure is supposed “to structure” some particular field of
knowledge we are interested in. For instance, we can suppose a typical
group structure of the form hG, ◦i or the McKinsey, Sugar and Suppes’
structure for a classical particle mechanics, P = hP, s, m, f , gi (cf. [24,
p.320]).

• we search for an appropriate formal language for that structure;

• in this language, we formulate the axioms which will have the structure
as a model in the Tarskian standard sense of model.6
6
We need to call the reader’s attention to a particular exception here; suppose we think

15
 Note that here we try to characterize a class of models, although they are
not in general order-1 structures; as a consequence, the languages employed
to stipulate the axioms are in general not first-order languages. It is a char-
acteristic of this approach that different languages may be used for this task,
each one of them with its costs and benefits, for there are infinitely many
formal languages that could be used as suitable for some given structure, cf.
[8].
For a simple example of a mathematical theory, one may consider the
mathematical theory of groups. As usually stated, a group is a structure
A = hG, ◦i, where G 6= ∅ and the binary operation ◦ satisfies the usual
postulates of group theory, that is (the quantifers range on G):

G1 ∀x∀y∀z(((x ◦ y) ◦ z) = (x ◦ (y ◦ z)))

G2 ∃z(∀x(x ◦ z = z ◦ x = x) ∧ ∀x∃y(x ◦ y = y ◦ x = z))

Here we only quantify over individuals of D, but the use of a second order
language can be advisable to deal with subgroups. A set theoretical predicate
for group theory in da Costa-Chuaqui style is a sentence in the language of
set theory, augmented by the symbols we are using now (in order to form
the set R of the non-logical symbols of the language of group theory), of the
following kind (for a detailed treatment, see Arenhart and Moraes [1]):

Definition 4.1 (Group) A group is a structure X satisfying the following

set theoretical predicate:

G(X) ⇐⇒ ∃G∃ ◦ (X = hG, ◦i ∧ (∀x, y, z ∈ X)(G1 ∧ G2))

So, for instance, we know that hR, +i is a group because when ◦ is in-
terpreted as denoting + and the quantifiers range over R, we can prove (in
ZFC) that hR, +i |= G1 and hR, +i |= G2 in the Tarskian sense. For further
examples and developments, see da Costa and Chuaqui [7], Arenhart and
Moraes [1], Krause, Arenhart and Moraes [14]. Now, one may simply think
that this only makes explicit what was really intended by the semantic ap-
proach since the beginnings. However, considering the discussion in section
of a structure of the form hV, ∈i suitable for ZFC proper. In this case, the models would
not be sets of ZFC as we are suggesting. This difficulty can be surmounted by assuming a
stronger theory in the metalevel, but we shall leave this point to be discussed at another
time, given that the theories we are emphasizing here have models which are sets of ZFC.

16
2 we have seen that there is an ambivalent relation between the semantical
tradition and the use of language to characterize a set theoretical predicate.
One of the reasons to prefer not to use formal languages in characterizing
the intended class of models is that simplicity is lost. Apart from that, there
are other main points advanced against the use of formal languages which
may threaten the semantical approach as presented in the da Costa-Chuaqui
style. Now we advance some of them and argue that they are not definitive
against this version of the semantic approach.
First of all, let us consider once again the example presented above con-
cerning group theory (we discuss group theory for the sake of simplicity). It
will be easily recognizable that structures as A = hG, ◦, 0,−1 i subjected to
the following list of axioms are also examples of groups, once we consider the
following axioms:

G’1 ∀x∀y∀z(((x ◦ y) ◦ z) = (x ◦ (y ◦ z)))

G’2 ∀x(x ◦ 0 = 0 ◦ x = x)

G’3 ∀x(x ◦ x−1 = x−1 ◦ x = 0)

The set theoretical predicate, in this case is the following one:

Definition 4.2 (Group’) A group is a structure X satisfying the following

predicate:

G(X) ⇐⇒ ∃G∃ ◦ ∃−1 ∃0(X = hG, ◦,−1 , 0i ∧ (∀x, y, z ∈ X)(G0 1 ∧ G0 2 ∧ G0 3)).

It is easy to notice that the structures satisfying G1−G2 and G0 1−G0 3 are
distinct, for they have distinct signatures. The classes of structures collected
by both predicates are distinct, despite the fact that both are conceived as the
same theory. So, how can we say that the same theory is being characterized
in both cases? This difficulty may be pointed as an artifact of our rigorous
characterization of a set theoretical predicate in terms of a formal language,
and may be seen as betraying the spirit of the semantic approach. Can we
rest assured that this is enough for us to abandon CCA approach? Not
really. Muller [16] noticed the difficulty and provided a particular solution:
instead of considering the models of a particular formulation of the axioms,
we may provide models for all the theories formulated in intertranslatable
lexicons, so that we have an as near as possible formulation of the semantic

17
view as being language-free. The theories are the classes of all models of the
distinct intertranslatable lexicons. That is supposed to provide for both a
class of models insensitive to the particular formulation of the axioms while
still retaining the Tarskian account of models, which requires some set of
sentences which the models are said to model.7
Another way of doing this without having to look for translations of
languages consists in employing a very interesting tool which has not been
explored by philosophers of science yet: the general notions of definability
and invariance developed in da Costa and Rodrigues [8]. Those are clearly
model theoretic notions, and allow for generalizations of some tools of model
theory for higher order languages and structures (recall that usual model
theory deals with order–1 structures only). If we consider infinitary higher
order languages as the language of our structure, as we mentioned in 3.2, we
shall call, following da Costa and Rodrigues [8], a concept definable in this
language as being definable in the wide sense. Now, two structures A and
B are called equivalent, written A ≡ B iff the primitive relations of each of
them are definable in terms of the primitive relations of the other. So, we
may think of two structures with distinct signatures as presenting the same
theory only when they are equivalent.
Of course, this answer does not give us a language-free characterization of
a theory (neither does Muller’s solution mentioned above). However, perhaps
the dream of a language free characterization of a theory is not achievable;
CCA’s approach does not hide the fact that a language must be employed to
characterize a theory, and distinct languages may do the job. Notice that we
do not identify a theory with its class of models. Since we may have distinct
classes of models characterizing intuitively the same theory, we only claim
that the set theoretical predicate “presents” a theory, and as the discussion
above has shown, a theory may be presented in various ways. We shall have
more to say about this relation of an informal theory, as used by scientists,
with the relevant set theoretical predicate characterizing it in the concluding
section. We believe that this kind of foundational work is relevant from a
philosophical point of view, for it may help us in enlighten the structure of
7
Perhaps we can also consider “language free” structures in the case when the class of
models os a theory does not constitute an axiomatizable class. In this case, perhaps he
most we can ask, then, following Suppes, may be the search for a Representation Theorem
in order to characterize the models of the theory, up to isomorphisms [20, chap.12]. But
this is a guess, for by employing infinitary languages, we may axiomatize any class of
structures.

18
scientific theories.
Another criticism that may be posed concerns the existence of non-
standard models. Indeed, according to Suppe’s account to the syntactic
view, once a theory is formulated in axiomatic form, the Löwenheim-Skolem
theorem may be employed to provide for highly unintended models of the
theory (see Suppe [19, p. S104]). In practice, however, we only pick the
intended ones. Unintended models made the analysis of confirmation, expla-
nation, etc., difficult for the syntactic approach. Is this not an objection also
for the CCA approach? To discuss it, there are some points that should be
made clear. First of all, the semantics in the syntactic approach was per-
formed by the coordination rules, the somehow “direct” empirical interpreta-
tion. Models according to model theory are mathematical entities. Secondly,
Löwenheim-Skolem theorem works only for order-1 structures, and most in-
teresting scientific theories are of a higher order. As a third point, we shall
see in the next section that even the Suppesian approach, in formulating a
theory directly inside some set theory, may be plagued by unintended models
and some further metamathematical artifacts. So, if we are going to handle
with those issues anyway, this should not be pointed as a defect, but should
rather be recognized and studied looking for some profitable results from a
logico-foundational point of view.
So, we may keep the idea that models are structures in the standard
Tarskian sense.8 Thomson-Jones [26] called such models truth makers, mean-
ing that they are employed to make a set of sentences true. According to him
[26, pp.533-534], the semantic view should not consider models as truth mak-
ers, but only as informally interpreted mathematical structures representing
the systems under investigation. There are two main criticisms addressed to
the use of models in the above Tarskian sense. First of all, in taking into ac-
count the language characterizing the models, we distance ourselves from the
scientific practice. And, as we mentioned before, one of the main troubles of
the syntactic approach is precisely its artificiality and distance from scientific
practice. The second criticism concerns the use of set theory to characterize
the structures that will be the models. Models in the Tarskian approach are
set theoretical structures; employing such structures we must trust set the-
ory and confine ourselves to the tools and limits of the particular set theory
in question. Such an approach, however, is claimed to be distracting when
it comes to deal with the major concerns for the philosophers of science:
8
Regarded the remark we have made concerning theories like ZFC proper.

19
themes such as the relation between theory and phenomena, confirmation of
theories, and scientific explanation. Really, does the use of a particular set
theory has importance in discussing these issues? What if we use a radically
different set theory such as Rosser-Quine’s NF system? No answer to this
point was, as far as we know, taken into account till today.
Our answer to the second point above will bear also on the first one. Ac-
cording to our view, when it comes to the study of the foundations of science
one must be as explicit as possible relatively to the logical basis and all kinds
of presuppositions. So, to make clear which set theory (or, more generally,
which mathematics) is being employed is a requirement for a successful foun-
dational study. More than that, the argument may be turned around. As
Suppes himself suggested (see [25]), the choice of an adequate formal appara-
tus may be relevant in getting closer to actual scientific practice. Suppes loc.
cit. suggests that constructive mathematics should be employed in the study
of data structures to reflect the finite character of experimental practice.
That is, the structures used to characterize mature scientific theories use in
general lots of tools of mathematics (continuous functions, infinite quantities
of all kinds), while the structures reflecting actual experiments are discrete
and finitistic (see also Suppes [23, p. 63]). So, the idea that a definite math-
ematical framework limits our investigation may be reverted to the benefit
of the philosopher of science in both directions: clearing foundational issues
and keeping in touch with actual experimental practice.
Of course, this brings to light our answer to the first objection. The
idea that we may become closer to actual experimentation by the use of
some distinct mathematical tool reveals that the use of a specific language
and logical apparatus is not in disagreement with the impetus of the semantic
approach to be in touch with what scientists do. By its very nature, the study
of the foundations of science requires some tools; it is a kind of second-order
investigation which may be seen as describing and investigating more or less
properly an activity performed by scientists. The claim that restriction to set
theoretical structures is a limitation gets things wrong, in our view. Every
approach to models that allows for mathematical structures will have to make
clear which is the underlying framework, otherwise we fall in vagueness and
hand-waving discourse. Issues that interest philosophers of science may be
more properly addressed once those topics are clearly settled.
For one example of the proficuous interplay between foundational study
in the semantic approach and scientific practice, we may mention that the use
of set theoretical structures inside some clearly formulated set theory helps

20
in making a very important philosophical issue clearer, namely, the complex
relation between data and theory. To recall very briefly, we mention the
fact that to go from qualitative data to quantitative theorizing requires the
idea that models are mathematical structures representing the systems under
study. One does not simply embed experiments in some model of the theory,
but gradually goes from qualitative finitistic structures to more complex ones.
This fact helps us also in getting a clear picture in the hotly debated issue of
scientific representation. To attribute to a model of the theory a direct role in
accounting for the results of experiments may also lead us to forget the fact
that representation is a relation holding (if it does hold at all) between target
systems and data structures. It is not the theory which represents directly
the behavior of the entities, but rather the data structures which reflect
the (highly idealized) behavior of the entities in particular experiments. Of
course, we do not have space here to enter into the details and intricacies of
the debate concerning scientific representation, but we just point to that issue
to make clear how foundational rigor may help in such kind of discussion.
So, it seems to us that CCA is a useful account of the semantic view. Its
main force, however, is still not explored. Higher order model theory may
provide some helpful insights into the structure of scientific theories. It is not
possible for us, in the scope of this work, to develop further into those topics,
but we agree with the general tone of [13]: philosophers of science have the
mathematical tools to study structures and models, we have only ourselves
to blame if we turn our backs to the work. We now present an alternative
account of the semantic approach, the one favored by Suppes when dealing
with axiomatization.

5 The Suppesian approach

The Suppesian approach, as we mentioned before, differs from CCA by not
having to employ a formal language to talk about the structures we wish to
axiomatize: we do it directly in the set theory in which we built the structures
themselves. Of course, most of the times this set theory is ZFC. In this case,
despite the fact that we still want to characterize a class of structures as
“models” of a theory, this time we cannot take those structures as models
in the Tarskian sense, or in Thomson-Jones’ truth making sense: there is
simply no set of sentences to be made true by the structures according to an
appropriate interpretation. Instead of that, the postulates are written in the

21
language of set theory increased by some constants to allow us to talk about
the relevant structure.
As Suppes has emphasized, it is ‘more economic’ to work as the present
approach suggests, that is, presupposing a set theory (and its underlying
logic), say ZFC, and not going to metamathematics (to the syntactical ap-
proach). Thus, it seems to be more interesting to assume a mathematics, say
first-order ZFC set theory (but, as we have said, we could use other theories
as well) and work within it. Suppes, as it is well known, works in an informal
set theory, so it is not necessary to explicit neither the logic nor the set the-
ory; in his approach, the things we need can be rescued by adding necessary
axioms to the set theory enrolled. Then, although he says that “[a]s far as
I can see, most problems of central importance to the philosophy of science
can be discussed in full detail by accepting something like a standard formu-
lation of set theory, without questioning the foundations of mathematic” [24,
p.1]. But most does not mean all, and there are interesting questions to be
dealt with if we pay attention to the commitments assumed when we opt for
a certain metamathematical basis, as we hope our discussion will make clear
below.
To make the distinction clearer, let us once again present the case of group
theory. As mathematicians usually characterize it, a group is an ordered pair
hG, ◦i, and axioms (G1) and (G2) above hold (see definition 4).
Now, before we proceed and present the usual set theoretical predicate for
groups, let us recall how mathematicians prove, for instance, that hR, +i is a
group. That is very simple; once the set of real numbers is constructed inside
set theory (with real numbers taken as Dedekind cuts, for instance), and
addition is properly defined, one proves that the operation + has properties
G1 and G2. That is, one establishes the following theorems of ZFC:

• ZFC ` ∀x, y, z ∈ R, (((x + y) + z) = (x + (y + z)));

• ZFC ` ∀x ∈ R(x+0 = 0+x = x)∧∀x ∈ R, ∃y ∈ R(x+y = y +x = 0)).

Notice that in this case there is no need for an attribution of set theoretical
entities to linguistic entities; there is simply no formal language designed to
talk about groups. Instead of that, the derivations above are performed
directly in the set theory, in this case increased by the new constant symbols
R, +, and 0. Really, there is no “satisfaction relation” holding between
elements of the structure and formulas of a specific language. All the magic

22
is performed inside ZFC proper. Of course, that also means that there are
no specific model theoretic tools to deal with those structures either. What
holds for groups holds also for more complicated theories, such as classical
and quantum mechanics, so as for whatever theory whose models are set
theoretical structures.
A major difference from the CCA approach concerns the availability of
all the mathematics required in order to do science. We do not need to ax-
iomatize the Dedekind complete field in order to speak about the field of real
numbers, for we have them available from the underlying set theory (once
again, provided by some particular construction, such as Dedekind cuts or
Cauchy sequences). The same goes for vector spaces, Hilbert spaces, Eu-
clidean geometry and the whole of mathematics required to do most of our
current scientific theories. That fact may prompt the idea that mathemati-
cal structures in the Suppesian approach are already interpreted structures
satisfying some postulates. Consider the simplicity of Suppes’ style classical
particle mechanics (CPM). A system of particle mechanics is a 5-tuple9

P = hP, T, m, s, f, gi

where P is a non-empty set (the ‘particles’), T is an interval in the set of real

numbers (say expressing an interval of time), m is a function from P to R+
so that if p ∈ P , then m(p) is the mass of p, s is a function from P × T in R3 ,
so that s(p, t) is a vector expressing the position of the particle p at time t,
f is a function with domain P × T × I, where I is a set of positive integers,
so that f (p, t, i) is a vector representing the forces acting on p at t, g is a
vector representing the external forces acting on a particle at a given time t.
All these concepts are subjected to kinematical and dynamical axioms which
below:

Definition 5.1 A Newtonian particle mechanics is a structure

P = hP, T, s, m, f, gi

satisfying the following axioms:

Kinematical axioms

P1 P is a finite non-empty set;

9
Here we take a simplified version of this structure; see [20, chap. 12]. In his [24],
Suppes provides an alternative formulation.

23
 P2 T is an interval of real numbers;

P3 for p ∈ P , sp is twice differentiable in T ;

Dynamical axioms

P4 for p ∈ P , m(p) is a positive real number;

P5 for p, q ∈ P , t ∈ T , f (p, q, t) = −f (q, p, t) ;

P6 for p, q ∈ P , t ∈ T , s(p, t) × f (p, q, t) = −s(q, t) × f (q, p, t);

P7 for p ∈ P and t ∈ T , m(p)D2 sp (t) = q∈P f (p, q, t) + g(p, t).

P

The fact that we intuitively call P a set of particles, m a mass function

and p a function attributing position to particles at every instant of time,
should not allow us to forget that this is just an intended interpretation; what
we really have here are abstract sets. Thomson-Jones [26] calls this kind of
presentation description fitter, in opposition to the previous one (CCA) which
he calls, as we mentioned, truth maker models. The structures modeling the
kinematical and dynamical axioms are not making a set of sentences true, but
they are just fitting the description provided by the axioms. This is in close
connection with our claim that in the Suppesian style axiomatization there
is no formal language that gets interpreted in the structures in question,
but rather the whole demonstration that the structure fits the description
provided by the axioms, something that is performed inside the framework
of set theory proper, using the deductive resources of this theory along with
the particular features of the structure under scrutiny.
We may make this point even clearer by recalling an objection by Thomson-
Jones against use of models in the Tarskian truth-maker sense. Recall from
a previous section that the objection ran roughly as follows: since in model
theory we always operate inside some set theory, the structures used as mod-
els are set theoretical entities, that is, they are limited to the sets the set
theoretical apparatus is able to provide.10 The objection advanced that the
mentioned fact is really is a limitation to the semantic approach, since one
should not be restricted to the resources of one specific set theory. Now, as
we are developing it, this objection could be made also against the Suppesian
approach, since the structures taken into account here are also set theoretical
10
Once more, some care must be taken when the theory is the set theory proper.

24
entities. The mere fact that structures are being taken as description fitters,
rather than truth makers, does not free us from building them inside some
framework, a framework which must be strong enough to provide the math-
ematics required to develop scientific theories. So, the semantic approach
will have to live with the fact that there is an underlying mathematical ap-
paratus which must be recognized (and this is a central contention of this
paper). Rather than being a limitation, as we mentioned in our answer to
this objection in the previous paragraph, this fact may sometimes help us
getting closer to scientific practice, as in the case of the finitistic structures
required for a rigorous account of the data structures.
So, we should not forget that the intended interpretation of the struc-
tures, such as that one for classical particle mechanics, is really inside a set
theory, which on its turn may be seen from a completely formal point of view
as a first-order theory. Everything done in the set theoretical predicate for
mechanics could in principle be rewritten only with ∈. To get a clear idea
where we are working in is a necessary condition for foundational studies.
Otherwise, how should we know what kind of tools we do have at our dis-
posal? We shall soon exemplify how the apparatus we work in influences the
discussion in foundations of science too.
The point then is that understanding structures as descriptions fitters
does not exempt them from being set theoretical entities (nor entities in some
mathematical framework). As we mentioned, this should not be seen as a
limitation, even more because it is widely recognized that set theory is a very
powerful framework, probably exceeding the needs of actual science (except,
of course, the requirements of set theory itself, when considered as a branch
of mathematics). The fact that set theory is the standard framework for the
Suppesian approach is responsible for its great power(for further discussion
see also [14])).
Indeed, this enormous strength may be exemplified by a trivial example:
as is well-known, we may easily present the Suppes predicate for well-orders
as usually done by mathematicians. A well-order is a pair hD, Ri such that
R is a partial order over the non-empty set D and every non-empty subset of
D has an R-least element. As an example of a well-ordered structure, take
hω, ∈i. So, the class defined by the set theoretical predicate for well-orderings
will have that structure as an element. Could we leave that structure out?
Yes, it is really simple: just state another axiom requiring that the structures
satisfying the set theoretical predicate should be well-orderings distinct from
hω, ∈i. Of course, more interesting cases may be conceivable, but this already

25
shows some of the resources that the Suppesian approach affords thanks to
the use of the language of set theory.
The great expressive power of this language led Suppe to a little exag-
gerated praise of the Suppesian approach (loc. cit.). According to Suppe’s
interpretation of Suppes’ approach to scientific theories, instead of present-
ing syntactic axioms we just select the intended models of the theory and
work only with them. This avoids the objection presented in the previous
section against the syntactic approach according to which a formalization
in a formal first-order language is always prey to non-standard models (but
recall our remarks on the scope of applicability of Löwenheim-Skolem and
other standard model theory theorems). That is, according to tis view, we
escape non-standard models by selecting precisely the intended ones:

In practice we do pick just intended models (else there would

not be a problem of unintended models), and so we can proceed
directly to the specification of these intended models without re-
course to syntactic axiomatization.

However, one has to deal with unintended models in the Suppesian ap-
proach too. As one may easily notice, most models of a set theoretical predi-
cate in the present approach will be mathematical models, having no relation
to empirical reality. It is not difficult to find models for the postulates of par-
ticle mechanics above as structures that can be built within ZFC. Although
we do not present them here, it is easy to acknowledge that they are just
mathematical structures without having necessarily any connection to a field
of knowledge in reality. But other models may constitute (as we believe) a
map of parts of empirical reality (this is discussed in [24]). Of course this pre-
sentation makes use of mathematics (ZFC set theory if we wish, and classical
logic). But they are not explicitly stated, becoming implicit. Furthermore, if
one is to block those unintended purely mathematical models somehow (say,
by imposing new axioms which ban the unintended models), then the great
advantage of Suppes’ approach against the syntactical one gets diminished,
since we are once again having to worry ourselves with the tricks required to
keep only the intended models (see once again Suppe loc. cit).
The models of CPM are structures like P above, and are set-theoretical
constructs, really structures in ZFC. The models, then, within this schema,
are set-theoretical structures and what we can infer from them will depend on
the underlying logic and on the axioms of set theory. It is not simply the case

26
that we may pick whichever models we want. Let us just mention another
example to illustrate this fact. In non-relativistic quantum mechanics, posi-
tion and momentum operators are quite relevant. In the Hilbert space L2 (R)
of the equivalence classes of square integrable functions, they are unbounded
operators. Just to recall, if A is a linear operator, then A is unbounded if for
any M > 0 there exists a vector α such that kA(α)k > M kαk. Otherwise, A
is bounded. But R. Solovay proved that if ZF (ZFC with the axiom of choice
dropped) is consistent, and if DC stands for the axiom of dependent choices,
then it follows that ZF (without the axiom of choice) plus DC has a model
in which each subset of the real numbers is Lebesgue measurable.11
Let us call ‘Solovay’s axiom’ (AS) the statement that “Any subset of R is
Lebesgue measurable”. Then, in the theory ZF+DC+AS (termed ‘Solovay’s
set theory’), it can be proven that any linear operator is bounded —see [15].
Thus, if we use Solovay’s theory instead of standard ZFC to build quantum
structures, the mathematics of quantum mechanics would be different from
the standard one. In particular, we are not sure how to deal with position
and momentum operators.
Another example also involving Solovay is the following one. One of the
fundamental theorems in functional analysis is Gleason’s theorem (it does
not matter for our purposes here its exact formulation), which is relevant in
quantum mechanics. The theorem implies the existence of certain probabil-
ity measures in separable Hilbert spaces. Solovay obtained a generalization
of the theorem also for non-separable spaces, but it was necessary to assume
the existence of a gigantic orthonormal basis whose cardinal is a measurable
cardinal (see [4]). But the existence of such cardinals cannot be proven in
ZFC set theory (supposed consistent). Thus, in order to get the generaliza-
tion, we need to extend ZFC. That is, as far as a physical theory demands
mathematics, perhaps new mathematical frameworks will be required (but
this is not a novelty, for it has ever been this way!)
These examples show that for certain considerations, it is extremely im-
portant to consider the mathematical framework we are working in. What
counts as allowable models and the mathematics available in the construction
of the relevant set theoretical predicate may vary according to the resources
of the underlying set theory. Furthermore, one may come to the conclusion
that some particular set theory is more convenient for some purposes than
another. From the philosophical point of view, this is particularly true for
11
DC is not sufficient to prove that there are nonmeasurable subsets of the real numbers.

27
expressing metaphysical views.
When other examples of such situations are taken into account, it may
be the case that we arrive at a conclusion that ZFC seems no more adequate
to be the ground of our science (or of some particular scientific theory). For
instance, this would be the case if we find true contradictions in the world, as
the dialetheists believe. Another example would be to strongly belief Birkhoff
and von Neumann, who claimed that quantum mechanics would demand a
non standard (‘quantum’) logic. In this particular setting, let us recall that
the realm of quantum logic was erected from their seminal work, but al-
though it has profound connections with quantum physics, quantum logics
progressed as the algebraic study of certain lattices, without direct appeal to
the foundations of quantum theory proper (for instance, see [9]. As is well
known, some attempts to develop quantum mechanics from a non-classical
logic were advanced, for instance, by Reichenbach [17]). Another specific
case is that which shows that the standard quantum formalism is compatible
with two alternative metaphysical packages concerning the nature of quan-
tum objects; the first one take them as individuals as in the classical sense,
as entities that can be individuated and considered on a pair with their “clas-
sical” cousins. Quantum objects of this sort can be treated inside ZFC, for
sets are collections of such entities by definition. But there is another meta-
physical package that comes from the forerunners of quantum theory, a view
that sees them as non-individuals, as entities for which the standard notion
of identity would not apply. These non-individuals can also be treated in-
side ZFC (which confirms the strength of this language), but at the expenses
of losing the metaphysical assumption about their non-individuality. In this
case, it has been said, a different mathematical framework would be suitable,
namely, quasi-set theory (see [12] for a full discussion on this topic; see also
[10], [11]).
In situations such as those above, we need to acknowledge that there are
different alternatives suggesting the use of deviant logics and mathematics
in the realm of science.

6 Semantic view, foundations, and real science

As we have been discussing, there are two alternative ways to approach the
semantic view. Both have their attractions and their defects. We believe that
we have made clear that, according to our view on both types of semantic

28
approach, the use of language is not an inessential part of the semantic view;
in both approaches one should recognize that a class of models does not
come by itself, but it must be characterized by some set of axioms; and these
axioms have an impact on the nature of the class of models being defined.
Some may believe that, by allowing language to play such a fundamental
role, we have betrayed the spirit of the semantic approach. However, we
believe that only by putting the cards explicitly on the table we may have
a clear picture of what is going on when relevant issues in philosophy of sci-
ence are discussed. Indeed, recall that a view about the nature of theories is
deeply relevant for the successful consecution of most philosophical purposes
concerning science nowadays (to mention but a few: realism/anti-realism
debate, structural realism and the nature of structure, issues concerning sci-
entific representation, and so on); with so much at stake, it is only natural to
try to have a most rigorous account of scientific theories. We believe our ex-
position has made it clear that there are many possibilities to be investigated,
but the linguistic aspect must not be ignored. Now that some skepticism is
being raised concerning the success of the semantic approach as the most ad-
equate rendering of scientific theories, we must provide rigorous foundations
for hand-waving slogans and verify the tenability of the view (for criticism
of the semantic view, see [26], [13]).
The first difficulty for those attempts (at least as we have proposed them)
come from a recent trend in de-emphasizing formal methods in philosophy of
science. In some parts of the paper we have been concerned with criticisms
as to the relation of some particular kinds of axiomatization with actual sci-
ence. We have tried to diminish the weight of such criticism by making it
clear how attention to axiomatization may be helpful in the study of ac-
tual science and may enlighten some aspects of it, in particular by pointing
(with Suppes [23], [25]) that the use of constructive mathematics may be
helpful to reflect the finitistic nature of the data structures obtained in scien-
tific practice. That point is relevant in this situation because it has become
part of the philosophical canon to dismiss a particular proposal by arguing
that it does not reflect real science, mainly because the proposal in question
amounts to what may be regarded as philosophically irrelevant “overformal-
ization”. Really, that position seems to be historically justified by recalling
that the Logical Positivists’ taste for logical analysis and the whole concep-
tual machinery required by their anti-metaphysical prejudices has created a
general aversion to formalization and some forms of logical studies (see also
the concluding remarks in Contessa [6] for doubts on the usefulness of formal

29
apparatus on philosophy of science; curiously, one of the criticisms Contessa
advances to the particular approach he is examining in that paper is that it
is not rigorous enough!).
That historical point notwithstanding, we would like also to argue that
there is much in formalization that, despite the appearance of not being re-
lated to actual scientific practice, may be allowed by a scientifically respectful
philosophy of science. One of those points, as we have emphasized above,
comes from the use of diverse logical tools to understand the structure of
scientific theories. Really, as philosophers of science we should be allowed to
apply freely the techniques of logic and formal sciences in general for the in-
vestigation of topics of philosophical interest (even when that does not reflect
“actual scientific practice”). Distinct models of ZFC, set theories employing
weakened or restricted forms of the axiom of choice or even non-standard
set theories should be investigated as frameworks for the semantic approach.
That involves a great amount of formal work, but it should not be dismissed
as being far away from any kind of scientific practice. Indeed, such formal
investigation and detailed scrutinizing is what foundational work amounts to
after all. Some idealization in the philosophical analysis of science is not a
sin against scientifically engaged philosophy.
More than looking for the right formal apparatus to reflect some aspect of
actual scientific practice, we must be aware that some topics of great philo-
sophical interest involve taking a greater than usual care with the underlying
framework. One of such cases concerns quantum mechanics. There are di-
verse aspects of the theory which require a discussion of the mathematical
basis of the theory. To mention a few of them, we should point to (i) the
attempts of capturing the Bohrian complementarity relation by some three
valued logic, (ii) the attempts of modeling the non-distributive character of
some logical operators by orthomodular lattices, (iii) the attempts of get-
ting a clear picture as to the nature of individuality and non-individuality
of quantum entities (see [12], [10]). So, a look for the right tools may be of
fundamental importance, and formalization in philosophy of science should
not be feared.
For a second point of difficulty concerning the consecution of the project
of endowing the semantic approach with rigorous foundation, we mention
some consequences of the sloganeering that has characterized the approach
most of the times until now. Most attempts to apply the semantic approach
in philosophical projects have tended to identify a theory with a (supposed)
linguistically-free class of models. So, theories were seen as classes of models

30
properly. This is also claimed to be in accordance with scientific practice
(recall Suppe loc. cit.). This trend have led to two kinds of related objections
to the semantic approach. The first one concerns obvious difficulties with
the identity of scientific theories, a problem that resemble the difficulties
that plagued the syntactic approach. Let us call this the Identity Problem.
The second main difficulty concerns the central role attributed to theory in
scientific practice and in philosophical investigations of science. Let us call
this the Theory Monopoly Problem.
As we deal with the Identity Problem in section 4, the difficulty appears
that distinct kinds of structures may be models of what we would (by in-
dependent criteria) call the same theory. So, the approach, by making the
strict identification of a theory with a class of models seems to take as dis-
tinct theories those classes of structures which should represent the same
theory (see the example of group theory in section 4). A more liberal view
of the role of language may help us in avoiding such troubles (see the related
concerns in Muller [16] and Halvorson [13]). Furthermore, in our discussion
throughout the paper we have not held that a theory is to be identified with a
class of structures, but rather that a theory is represented by a class of struc-
tures. That is, one may investigate some relevant aspects of the theory by
dealing with the class of models. Notice, this is much more akin to scientific
practice, where scientists know, at least in most of the times, when they are
dealing with the same theory. Our point is that there is an intuitive body of
knowledge which may be rendered rigorous by diverse classes of models, all
of them presenting the same theory. So, the strong thesis attributed to the
semantic view, the one claiming that a theory is a class of models should be
weakened to contemplate distinct presentations of what would intuitively be
regarded as the same theory.
As for the Monopoly of Theory Problem, we must be aware that scientific
theorizing and working with the development of theories is only part of the
real scientific enterprise. There is much more to science than that, of course.
Scientists conduct experiments, rely on intuitive thinking and lucky guesses
and hypothesis, on tacit knowledge and a lot of other practices that do not
always involve theorizing. Most of those aspects of science, in particular a lot
of social aspects, are not involved in the schema provided by the semantic
view. One should recognize that the semantic view concerns the logical
aspect of the structure of theories, but leave most other aspects of science
untouched. Otherwise, we shall fall prey to the criticism that the semantic
view does not provide for a complete account of science. However, as we have

31
been mentioning, only some aspects are to be touched on by the semantic
view, and we must take a humility stance toward that fact, recognizing that
the scientific practice involve other aspects as well.

References
[1] Arenhart, J. B.; Moraes, F. T. F., Structures, languages and models: a
unifying approach. Forthcoming in Logique et Analyse, 2013.

[2] Chakravartty, A., The semantic or model-theoretic view of theories and

scientific realism. Synthese, 127, pp. 325-345, 2001.

[3] Carnap. R., Introduction to Logic and its Appliacations. Dover, 1958.

[4] Chernoff, P.R., Andy Gleason and quantum mechanics, Notices of the
American Mathematical Society, November, pp.1253-1259, 2009.

[5] Church, A., Introduction to Mathematical Logic. Vol. 1, Princeton:

Princeton Un. Press, 1956.

[6] Contessa, G., Scientific models, partial structures and the new received
view of theories. Studies in History and Philosophy of Science, 37, pp.
370-377, 2006.

[7] da Costa, N. C. A.; Chuaqui, R., On Suppes’ Set Theoretical Predicates.

Erkenntnis, 29, pp. 95-112, 1988.

[8] da Costa, N. C. A. and Rodrigues, A. A. M., Definability and Invariance.

Studia Logica, 82, pp. 1-30, 2007.

[9] Dalla Chiara, M.L., Giuntini, R. & Greechie, R. Reasoning in Quantum

Theory: Sharp and Unsharp Quantum Logics, Kluwer Ac. Press (Trends
in Logic, vol.22), 2004.

[10] Domenech, G., Holik, F. and Krause, D. Q-spaces and the foundations
of quantum mechanics, Foundations of Physics 38 (11), 969-994, 2008.

[11] Domenech, G., Holik, F., Kniznik, L, and Krause, D., No Labeling
Quantum Mechanics of Indiscernible Particles, International J. Theo-
retical Physics DOI 10.1007/s10773-009-0220-x, 2010.

32
[12] French, S., and Krause, D., Identity in Physics. A historical, philosoph-
ical and formal analysis. Oxford: Oxford University Press, 2006.

[13] Halvorson, H., What scientific theories could not be. Philosophy of sci-
ence, 79(2), pp. 183-206, 2012.

[14] Krause, D.; Arenhart, J. R. B.; Moraes, F. T. F., Axiomatization and

models of scientific theories. Foundations of Science, 16, pp. 363-382,
2011.

[15] Maitland Wright, J. D., All operators on a Hilbert space are bounded,
Bulletin of the Americal Mathematical Society 79 (6), 1247, 1973.

[16] Muller, F. A., Reflections on the revolution at Stanford. Synthese,

183(1), pp. 87-114, 2011.

[17] Reichenbach, H., Philosophic Foundations of Quantum Mechanics. New

York, 1998.

[18] Suppe, F., The Structure of Scientific Theories. 2nd. ed., Urbana and
Chicago: U, Illinois Press, 1977.

[19] Suppe, F., Understanding scientific theories: an assesment of develop-

ments, 1969-1998. Philosophy of science, 67 Supplement, pp. S102-S115,
2000.

[20] Suppes, P., Introduction to Logic. Toronto: van Nostrand, 1957.

[21] Suppes, P., A Comparison of the Meaning and Uses of Models in Math-
ematics and the Natural Sciences, Synthese 12, pp. 287-301, 1960.

[22] Suppes, P., Models of data. In: E. Nagel, P. Suppes and A. Tarski (eds.),
Logic, Methodology and Philosophy of Science: Proceedings of the 1960
International Congress, Stanford: Stanford Un. Press, 1962, pp. 252-
261.

[23] Suppes, P., What is a scientific theory? In: Sidney Morgenbesser (ed.),
Philosophy of Science Today, New York: Basic Books, 1967, pp. 55-67.

[24] Suppes, P., Representation and Invariance of Scientific Structures. Stan-

ford: CSLI Pu., 2002.

33
[25] Suppes, P., Future develpment of scientific structures closer to experi-
ments: response to F. A. Muller. Synthese, 183(1), pp. 115-126, 2011.

[26] Thomson-Jones, M., Models and the semantic view. Philosophy of Sci-
ence,73, pp. 524-535, 2006.

34