In Sidney

Morgenbesser (Ed. ), Philosophy of Science Today. New York: Basic Book; Inc., 1 9 6 7 . P p . 5 5 - 6 7 .

IVHAT IS A SCIENTIFIC THEORY?

Patrick Suppes

Often when we ask what is a seand-so, pie expect a d e a r and definite answer. If, for euample, someone asks me what is a rational number, I may give the simple and precise a r m e r that a rational number is the ratio of two integen. There are other kinds of simple questions for which a prese answer art be given but for which in ordinary t a l k a rather vague anmer b usually given and accepted. Someone reads a b u t nectarines in a book. but has never seen a nectarine, or pasibly has seen nectarines but is not familiar with their English name. Ne may ask me, What is a nectarine? and I would probably apIy, X smoothskinned sort of peach. &&y, this is not a very exact answer, but if my questioner know what peaches are, i t may corne dose to being satisfactory. T h e ]rind o question I want to discuss fits neither one of these patterns. Saentifc theories arenot like rational numbers or nectarines. Certainly they are not like nectarines, for they are not simple phj-sical objects. They are like rational numben in not being phjsical objects, but they are totally unlilre rational numbers in that s c i e n ~ e theories m o t be defined in any ple or direct way in terms o other non-ghysical, a b m c t objects. Gaxl examples of the kind of question we shdl andj-ze ia t h i s chapterare probided by the familiar inquiries: ?\hat is phjsics?,* 7\hat i s psychology?, %hat is science? To none OE these questions do we expect a simple and precise a . u w e r . On theotherhand,there are many interesting things eo be said

WHAT IS A SCIESTIFIC THEORY?

about the sort of thing physics or psychology is. I shall be trying to show in this essay that this is also true of scientific theories.
m

T H E STASDARD SKETCH

The standard sketch of scientific theories-and I emphasize e word sk.et&-runs something like the following. A scientific theory consists of two parts. One part is an abstract logical calculus. In addition to the vocabulary of logic, this calculus includes ehe primitive symbols of the theory, and the logical stpu~cre of the theory is fixed by stating the axioms or postulates of e theory in terns of its primitive symbols. For many theories the primitive symbols will be thought of as theoretical t e m Eke eelectron or particle that are not possible eo relaee in any simple way to observable phenomena. The second part of the theory is a set of mles that assign an mpirical content to the logical calculus by providing what are ally called co-srdinating definitions or empirical intertations for at least some of ehe primitive and defined symbob the calculus. It is always emphasized that the first part alone is not sufficient to define a scientific theory, for without a systematic s dfication of the intended empirical interpretation of the ry, it is not possible in any sense to evaluate the theory as a of science, although it can be studied simply as a piece of ure mathematics. e mast striking thing about this characterization is its schematic nature.Concerningthe first part of a theory, logical calculus, it is unheard of eo find a substantive example theory actually worked out as a logical calculus in the w r i t of most philosophers of science. M u c h handwaving is ined in to demonstrate that t h i s working out of the logical culm is simple in principleand onlya matter of tedious etail, but concrete evidence is seldom given. The sketch of ehe second part of a theory, that is, the coordinating definitions or empirical interpretations of some of the terms, is also highly schematic. A common defense of the reladvely vague schema offered is that the variety of different em-

56

PATRICK SUPPES

pirical interpretations, for example, the many different methods of measuring mass-make a precise characterization difficult. Moreover, as we move from the precisely formulated theory on to the very loose and elliptical sort of experimental language used by almost all scientists, it is difficult to impose a definite pattern on the rules of empirical interpretation. T h e view I want to support in this essay is not that this standard sketch is wrong, but rather that i t is far too simple. Its very sketchiness makes it possible to omt both importane properties s f theories and significantdistinctions &at may be introduced between different Iheories.

egin with, there has been a strong tendency on ehe p a n of anyphilosophers to speak of the first p a t of a logicalcalculuspurely in syntactical t e m . %hec definitions provided in thesecond part do not in Iogic provide an adequate semantics for t h uite a art from questions about direct empi nent and natural from a logical standpoint to UR Is of the theory. These modek are highly abstract, mon-linguistic entities,oftenquite remote in theirconception from empirical obsepvations. ]It may well be asked what does the concept of a model have to add to the familiar discussiom of empirical interpretation o I think it i s true to say philosophers find it easier to talk about theories than about models of theories. for h i s are several, but perhaps the most important following: I n the first place, philosophers examples of theories are usually quite simple in character, arad therefore are easy to discuss in a straightforwardlinguisticmanner. %n place, theintroduction of models of a eory inevitably i n a ~ duces a strongermathematicalelement into e discussion. It is a natural thing to talk a b u t theories as linguistic entities-that t,to speak explicitly of the precisely defined set of Sentences sf

57

WEIAT IS A SCIESTIFIC

THEORY?

the theory and the like-when the theories are given in what is called standard formalization. Theories are ordinarily said to have a standard formalization when they are formulated within first-order logic. Roughiy speaking, first-order logic is just the logic of sentential connectives and predicates holdjng for one type of object. Unfortunately, when a theory asumes more than first-order logic, it is neither natural nor simple to formalize it in this fashion. Forexample, if in axiomatizing geometry, we want to define lines as certain sels of points, we must work within a framework that already includes the ideas of set theory. To be sure, it is theoretically possible to axiomatize simultaneousliy geometry and therelevant of set theory, but this is awkward and unduly laborous. Theories of more complicated structure like quantum mechanics, classical thermodynamics, or a modern quantitative version of learning theory, need t o use not only general ideas of set theory but also manyre Formalization of su& theo practical. Theories of this s

observations to the quantitative assertions needed for more elaborate theoretical stages of science. h andysis of how from the qualitative to thequantitative may be p rovided by axiomatizing appropriate algebras of experimentally =&able operations and relations. Given an axiomatized theory of measurement of some empiricalquantity such as mass, & fance, or force, the mathematical t a s t is to prove a representation theorem for models of the theory which establishes, roughly

PAXUCK SUPPES

speaking, that anyempirical model is isomorphic to some numerical model of the theory. The existence of this isomorphism between models justifies the application of numbers to things. IVe cannot literally take a number in our hands and apply i t to a physical object. Tihat we can do is to show that the structure o f a set of phenomena under certain empirical operations is the same as the structure of some set of numbers under arithmetical operations and reiations. T h e definition of isomorphism o& models in the given context makes the intuitiveidea of S Q ~ structure precise. The great significance of finding such an i s 6 morphism of models is that we may then use all our %amiliar knowledge of computational methods, as appliedto the arithmetical model, toinfer facts about the isomorphicempirical h i s central notion of a n model. A linpistic formulation of t empirical model of a theory of measurement being isomorphic to a numerical modelis extremely awkward and tedious to formulate. But in model-theoretic t e m the notion is sinn Ie and in fact represents a direct application of the very gene isomorphism used throughout all do T h e second example of the use cussion of reductionism i n the phi e problem formulated in connection with &e question of reducing one science to another may be f ~ ~ u as~a series a ~ of e ~ l e m using t.he notion of a tion theorem %or&e els of a theory. For instance, esis that psy&oIogy may reduced to physiology would any people appropriestablished if one could show that for any mode1 of a psyogical theory it was possible to within physiological theory. e absence at the present time of yCh010,aY OP IphySiOlOgY question of reduction ical example from phpics is dynamics to statistical mechanics. usually not stated in absolutely satisfactory form from a bgical standpoint, there is no doubt that it is substantially correct and represents one of the great triumphs of matbematicdl phyks.

WHAT IS A SCIESTIFIC THEORY?

NTRIXSIC VERSUS EXTRISSIC

uite apart from the two applications just mentioned of the concept of a model of atheory, we may bring this concept to ear direcely on the question of characterizing a scientific theory. e contrast I wish eo draw is between intrinsic and extrinsic raceerization. T h e f o r m u l a ~ o n of a theory as culus or, to put ie in terms that H prefer, as a standardformalization, gives an intrinsiccharac is is certainly not the Q+ approach. For instance, a natural ueseion to ask within ehe context of logic is if a certain theory cala be axiomatized with standard formalization, nt-order logic. I n ordes to f o w u h t e such a question in a precise a m e r , it is necessary eo have some exerinsic way of characterizne of the simplest ways of zation is simply to define eory. T o ask if we can axiomatize ehe theory if we can staee a set of U ~ O H ~such S that the recisely the models in thedefined

of a theory formulated both e x t h nsider ehe extrinsicformulation of

s isomorphicto some fragrea1 n u m b e n T h eexthinsic tion of a theory usuallyfollows thesortgivenfor te a particular model of ehe calless-thanrelation)andthen models of ehe theory in relation to
sic characterization is plow to formulate of models \cithout nly to the intrinsic

15th the present case the solution gh even it is not easily formulated intrinsic axioms are just those for a

PATRICK SUPPE5

simple ordering plus the axiom that the ordering must contain in its domain a countable subset, dense w i t h respect to the ordering in question. A casual inspection of scientific theories suggests that the usual formulations are intrinsic rather than extrinsic in character, and the question of extrinsic formulations usually therefore that arisesonly in pure mathematics. This would also seem to be a ahat an happy result,forourphilosophicalintuitionissurely intrinsic characterization is in general to be preferred to an exproblem or" intrinsic axiomatization of a sdenorecomplicated and considerablymoresubtle would indicate. Fortunately, i t is preckely by tion of the cliass of m d e I s of thetheory t h a t roper perspective and formuIaEed Ie consideration of its exact soluve one simple example. The axiom for classical n i a are ordinarilystatedin such a way that a em, as a frame of reference, k tacitly assumed. ne effect of this is that relationships deducible from t necessarily invariant with respect eo Galilean transformations. $Ve can view the tacit assumption of a frame of nce as an extrinsic aspect of the familiar characterizations theoryy.From &e standpoint of the models of the theory, e d3cuIty in &e standardaxiomatizations of mechanics k a t a large number of formally distinct modeIs may be used to express the same mechanical faces. Each of these different models resents the tacit choice of a different Lnme of reference, but models representing the same Enechanid facts are related by dean transformations. It is thus fair to sa at in this instance difference between models related by dean transformaGons does not have any theoretical significance, and it may be ded as adefect of the asioms that els elrist. It is important to redize elsrelated by Galileantransformacio Int usually made under the heading s f e m p i r i d interpretations of the theory. Ia is a conceptual

WHAT IS A SCIEhTEIC THEORY?

theoretical side of physics. I have introduced this example here in order to provide a simple instance of how the explicit consideration of models can lead to a more subtle discussion of the nature of a scientific theory. It is certainly possible from a p h i b sophical standpoint to maintainthatparticle mechanics as a scientific theory should be expressed only in t e m of Galilean invariant relationships, and that the customary f 0 ~ ~ l a tare i o ~ defective in this respect.

I turn now to the second part of theories mentioned above. %tis m e that in the foregoing discussion we have been using the word theory to refer only to the first part of theories-that is, tothe axiomatization of the theory or the expression of the theory as a logical calculus-but as I emphasized at the beginning, the necessity of providingempiricalinterpretation of a theory is just as important as the development of the formal side of the theory. My central point on this aspect of theories is that story is much more complicated than the familiar remarks ut cwrdinating definitions and empirical interpretations of theories would indicate. T h e kind of coordinating definitions often described by philosophers have their place inpopular philosophical expositions of theories, but in the actual practice o f testing scientific theories a more elaborate and more sophisticated f o w a l machinery for relating a theory to data is required. T h e concrete experience that scientists label anexperiment cannot itselfbe connected to a theory in any complete sense. That experience must be put through a conceptual grinder that in many cases is excessively coarse. Once the experience is passed the inder,often in the form of thequitefragmentary of complete experiment, the experimental data n canonical form and constitute a model of the experiment. It is &is model of the experiment rather than a model of the theory for which direct coordinating definitions are provided. It is also characteristic that the model of the ex62

PATRICK SLTPES

periment isof a relatively different logical type from that s f the model of the theory. It is common for the models of a theory to contain continuous functions or infinite sequences, but for the model of the experiment to be highly discrete and finitistic in character. The assessment of the reIation between the model ofehe experimentand some designated model o theory is a characteristic fundamental problem of mode tistical methodology. about this meth for present purposes t place, it is itse 1 and theoretical in t has been a typical function of this methodology to develop a n elaborate theory of experimentation that intercedes between any fundamentalsaentific theory and raw experimental experience. I t is not possible in rather eIaborate hierar d beween the fundamental scientific ts presumed to support it. My only explicit the existence of this hierarchy ere is no simple procedure for giving c or a theory. It is even a bowderization c o a d i n a t h g definitions are given to establish the proper CORnections between models of the theory and models of t h e experiment in the sense of &e canonical f o w of the data just tioned. T h e elaborate estimating ple, for theoretical pammeeers theory bom models of the experiment are not adequately covered by a reference to dinating definitions. someone asks, What is a scientific eoq? it seem to me ere is no simple response to be given. Are we to indude as the.theory the ~ ~ e ~ ~ ~ ~ statistid ~ o r methodology ~ e d ~ u tfor ia the theory? If we are eo take seriously thestandard daims e ccwrdinating definitions are part of the theory, seem inevitable that we must also include i n am led description of theories a ~ e for designing ~ ~ ments, estimating parameters and testing goodness-of-fit of the theory. It does not seem to me important to give definitions of the form: X is a scientific theory if, and

WHAT PS A SCIESXIFIC THEORY?

onIr if, so-and-so.1t;hat is important is to recognize thatthe existence of a hierarchy of theories arising from the methodology of experimentation for testing the fundamental theory is an essential ingredient of any sophisticated scientific discipline.

There is one view Qf scientific hleories which is undoubtedly of considerableimportanceandwhich I have not yet mentioned. This is the view that theories are to be looked at from an instruentalviewpoint. T h e most importantfunction of a theory, according to this view, is not to organize or assert statements that are true or false but to furnish material principles of inference &at may be used in inferring oneset of facts from mother. Thus, i n ehe familiar syllogism all men are mortal; Socsates is a man; therefore,Socrates is mortal,themajorpremise aE1 menare mortal,according to this instruinentalviewpoint, is converted ineo a principle of inference. And the syllogism S OW plas only &e mise Socrates is a man. m a logical standpoint it is clear that th% is a fairly trivial ove, and the question naturally arises if there is anything more 1 to besaidabouttheinstrumentalviewpoint. ose interestingargumentforclaiming khat th ore than a verbal difference between these two ways of looking theories or laws is ehe argument that when theoriesarereded as principles of inference rather than as major premises, ongerconcerned direcely to establish their truth or to evaluatetheir usefulness ininferringnew statect. It i s characteristic of discussions in this vein by hilosophers that I-IQ genuinely original formal notions have f these discussions to displace the dassical semantical oeions of truth and validity. T o talk, for instance, about laws having different jobs than statements of fact is trivial unless some systematic semantical notions are introduced to replace ehe seanda d analysis. From another direction there has been one concerted serious effort to provide a formal framework for the evduafion of

PATRICK SUPPES

theories which replaces the classical concept of truth.Fl'hat I have in mind is modern statistical decision theory. It is typical of statisticaldecisiontheory to talkaboutactionsratherthan statements. Once the focus is shifted from statements to actions, it seems quite natural to replace the concept of tmth by that of expected loss or risk. It is appropriate to ask if a statement is true, but it does not make much sense to ask if it is risky. On e other hand, it is reasonable to ask how risky an action is, but t to ask if i t is true. It is apparent that statisticaldecision eory., when taken literally, projects a more radica1 h u u r n e n t a l view of theories than does the view already sketched. Theories are not regarded even as principles of inference but as methods of organizing evidence to dedde a\..hich one of several actions to take. $%en theories are regarded as principles of inference, it i s a straightforavard matter to return to the classical view and to connect a theory as a principle of inference with t h e concept of a theory as a m e major pre se in an argument. The connection between the classical view and the view of theories as f an action is certainly more hseruments leading eo the taking s remote and indirect. gh many examples of applications of the ideas of se decision theory have been worked out in recent literat u e on ehe foundations of statistics, Lbese examples in no case ea1 with complicated scientific theories, and P have seen no serious discussion of the treatment of scientifictheories from ehe dpoine of statistical decision theory. Again, it is fair to say when we wane to ealk about ehe evaluation of a sophisticated scientifictheory,disciplineslike statistic2 decisiontheory have oe yet offered any genuine alternative to the semantical notions of truth and validity. I n face, evena a u a l inspection of the of statistical decision theory shows that in spiee of &e talorientation of thefundamental ideas, formaldevelopment of the theory is wholly dependentonthestandard semantical notions and inno sense replace them. $ n a t I mean by this is that in concentrating on the taking of an action as the terminal state of an inquiry the decision theorists havefound it necessary to use seandardsemantical notions in escribing evidence, their OIM theory, and so for

WHAT IS A SCIESTIFIC

THEORY?

I cannot recall a single discussion by decision theorists in which particdar observation statements are treated in terms of utility rather than in ternls of their truth orfalsity. It seems apparent that statistical decision theory does not at the present time offer agenuinelycoherent or deeply original new view of scientific theories. Perhaps future developments of decision theory will proceed in t h i s direction. Be that as it may, there is one still more radical instrumental view that I would like to discuss as the final point to be covered in this essay. As I have already noted; it is characteristic of many instrumental analyses to distinguishthe s t a t u of theories from the s t a t u of particular assertions of fact. It is the point of a more radical instrumental, behavioristic view of the use of language to challenge this distinction and to look at the entire use of language, induding the statement of theories as well as of particular mat& e r sof fact, f r o m a behavioristic viewpoint. According to this view of the matter, all uses of language are with strong emphasis on the language users. It is e semantical analysis of modern logic gives a very inadequate account even of the cognitive uses of language, bec a k e it does not explicitly consider the production and reception of linguistic stimuli by speakers, writers, listeners, and readers. It is plain&atforthe behaviorist -an ultimatelymeaningful answer to the question What is a scientific theory? cannot be given in t e m of the kinds of concepts considered earlier. An adequate and complete answer can be given only in terms of an explicit and detailedconsideration of both theproducers and consumers of the theory. There is much that is attractive in this behaviorist way of looking at theories or language in general. What it lacks at present, however, is sufficient scientific depth and dehiteness to serve as a genuine alternative to the precise notions of modern logic and semantics. Moreover, mu& of the language of models and theories discussed earlier in his chapter is surely so approximately correct that any behaviorist revision of our way of looking at theories must yield the ordinary talk about models and theories as a first approximation. It is a matter for the future to see whether or not the behaviorists approach w i l l deepen our understanding of the nature of scientific theories.

66

PATRICK SL'PPES

In current perspective, themethodsandconcepts of modern logic provide a satisfactory a n d p o w ~ f u set l of tools for analyzing the detailed structure of scientific theories. l\That would seem to be needed for the presentis deeper and more detailed application of these tools to the job of analysis. I have tried to indicate what I think are some of the more fruitful directions for future investigation.