|This is a file in the archives of the Stanford Encyclopedia of Philosophy.|
Sometimes we write or speak a sentence S that expresses nothing either true or false, because some crucial information is missing about what the words mean. If we go on to add this information, so that S comes to express a true or false statement, we are said to interpret S, and the added information is called an interpretation of S. If the interpretation I happens to make S state something true, we say that I is a model of S, or that I satisfies S, in symbols I S. Another way of saying that I is a model of S is to say that S is true in I, and so we have the notion of model-theoretic truth, which is truth in a particular interpretation. But one should remember that the statement S is true in I is just a paraphrase of S, when interpreted as in I, is true; so model-theoretic truth is parasitic on plain ordinary truth, and we can always paraphrase it away.
For example I might say
He is killing all of them,and offer the interpretation that he is Alfonso Arblaster of 35 The Crescent, Beetleford, and that them are the pigeons in his loft. This interpretation explains (a) what objects some expressions refer to, and (b) what classes some quantifiers range over. (In this example there is one quantifier: all of them). Interpretations that consist of items (a) and (b) appear very often in model theory, and they are known as structures. Particular kinds of model theory use particular kinds of structure; for example mathematical model theory tends to use so-called first-order structures, model theory of modal logics uses Kripke structures, and so on.
The structure I in the previous paragraph involves one fixed object and one fixed class. Since we described the structure today, the class is the class of pigeons in Alfonsos loft today, not those that will come tomorrow to replace them. If Alfonso Arblaster kills all the pigeons in his loft today, then I satisfies the quoted sentence today but wont satisfy it tomorrow, because Alfonso cant kill the same pigeons twice over. Depending on what you want to use model theory for, you may be happy to evaluate sentences today (the default time), or you may want to record how they are satisfied at one time and not at another. In the latter case you can relativise the notion of model and write I tS to mean that I is a model of S at time t. The same applies to places, or to anything else that might be picked up by other implicit indexical features in the sentence. For example if you believe in possible worlds, you can index by the possible world where the sentence is to be evaluated. Apart from using set theory, model theory is completely agnostic about what kinds of thing exist.
Note that the objects and classes in a structure carry labels that steer them to the right expressions in the sentence. These labels are an essential part of the structure.
If the same class is used to interpret all quantifiers, the class is called the domain or universe of the structure. But sometimes there are quantifiers ranging over different classes. For example if I say
One of those thingummy diseases is killing all the birds.you will look for an interpretation that assigns a class of diseases to those thingummy diseases and a class of birds to the birds. Interpretations that give two or more classes for different quantifiers to range over are said to be many-sorted, and the classes are sometimes called the sorts.
The ideas above can still be useful if we start with a sentence S that does say something either true or false without needing further interpretation. (Model theorists say that such a sentence is fully interpreted.) For example we can consider misinterpretations I of a fully interpreted sentence S. A misinterpretation of S that makes it true is known as a nonstandard or unintended model of S. The branch of mathematics called nonstandard analysis is based on nonstandard models of mathematical statements about the real or complex number systems; see Section 4 below.
One also talks of model-theoretic semantics of natural languages, which is a way of describing the meanings of natural language sentences, not a way of giving them meanings. The connection between this semantics and model theory is a little indirect. It lies in Tarskis truth definition of 1933. See the entry on Tarskis truth definitions for more details.
The first person has transferred the property to the second person, who thereby holds the property for the benefit of the third person.defines a class of structures which take the form of labelled 4-tuples, as for example (writing the label on the left):
We can extend the idea of model-theoretic definition from a single sentence S to a set T of sentences; Mod(T) is the class of all interpretations that are simultaneously models of all the sentences in T. When a set T of sentences is used to define a class in this way, mathematicians say that T is a theory or a set of axioms, and that T axiomatises the class Mod(T).
Take for example the following set of first-order sentences:
xyz (x + (y + z) = (x + y) + z).Here the labels are the addition symbol +, the minus symbol - and the constant symbol 0. An interpretation also needs to specify a domain for the quantifiers. With one proviso, the models of this set of sentences are precisely the structures that mathematicians know as abelian groups. The proviso is that in an abelian group A, the domain should contain the interpretation of the symbol 0, and it should be closed under the interpretations of the symbols + and -. In mathematical model theory one builds this condition (or the corresponding one for other function and constant symbols) into the definition of a structure.
x (x + 0 = x).
x (x + (-x) = 0).
xy (x + y = y + x).
Each mathematical structure is tied to a particular first-order language. A structure contains interpretations of certain predicate, function and constant symbols; each predicate or function symbol has a fixed arity. The collection K of these symbols is called the signature of the structure. Symbols in the signature are often called nonlogical constants, and an older name for them is primitives. The first-order language of signature K is the first-order language built up using the symbols in K, together with the equality sign =, to build up its atomic formulas. (See the entry on classical logic.) If K is a signature, S is a sentence of the language of signature K and A is a structure whose signature is K, then because the symbols match up, we know that A makes S either true or false. So one defines the class of abelian groups to be the class of all those structures of signature +, -, 0 which are models of the sentences above. Apart from the fact that it uses a formal first-order language, this is exactly the algebraists usual definition of the class of abelian groups; model theory formalises a kind of definition that is extremely common in mathematics.
Now the defining axioms for abelian groups have three kinds of symbol (apart from punctuation). First there is the logical symbol = with a fixed meaning. Second there are the nonlogical constants, which get their interpretation by being applied to a particular structure; one should group the quantifier symbols with them, because the structure also determines the domain over which the quantifiers range. And third there are the variables x, y etc. This three-level pattern of symbols allows us to define classes in a second way. Instead of looking for the interpretations of the nonlogical constants that will make a sentence true, we fix the interpretations of the nonlogical constants by choosing a particular structure A, and we look for assignments of elements of A to variables which will make a given formula true in A.
For example let Z be the additive group of integers. Its elements are the integers (positive, negative and 0), and the symbols +, -, 0 have their usual meanings. Consider the formula
v1 + v1 = v2.If we assign the number -3 to v1 and the number -6 to v2, the formula works out as true in Z. We express this by saying that the pair (-3,-6) satisfies this formula in Z. Likewise (15,30) and (0,0) satisfy it, but (2,-4) and (3,3) dont. Thus the formula defines a binary relation on the integers, namely the set of pairs of integers that satisfy it. A relation defined in this way in a structure A is called a first-order definable relation in A. A useful generalisation is to allow the defining formula to use added names for some specific elements of A; these elements are called parameters and the relation is then definable with parameters.
This second type of definition, defining relations inside a structure rather than classes of structure, also formalises a common mathematical practice. But this time the practice belongs to geometry rather than to algebra. You may recognise the relation in the field of real numbers defined by the formula
v12 + v22 = 1.Its the circle of radius 1 around the origin in the real plane. Algebraic geometry is full of definitions of this kind.
During the 1940s it occurred to several people (chiefly Anatolii Maltsev in Russia, Alfred Tarski in the USA and Abraham Robinson in Britain) that the metatheorems of classical logic could be used to prove mathematical theorems about classes defined in the two ways we have just described. In 1950 both Robinson and Tarski were invited to address the International Congress of Mathematicians at Cambridge Mass. on this new discipline (which as yet had no name - Tarski proposed the name model theory in 1954). The conclusion of Robinsons address to that Congress is worth quoting:
[The] concrete examples produced in the present paper will have shown that contemporary symbolic logic can produce useful tools - though by no means omnipotent ones - for the development of actual mathematics, more particularly for the development of algebra and, it would appear, of algebraic geometry. This is the realisation of an ambition which was expressed by Leibnitz in a letter to Huyghens as long ago as 1679.In fact Maltsev had already made quite deep applications of model theory in group theory several years earlier, but under the political conditions of the time his work in Russia was not yet known in the West. By the end of the twentieth century, Robinsons hopes had been amply fulfilled; see the entry on first-order model theory.
There are at least two other kinds of definition in model theory besides these two above. The third is known as interpretation (a special case of the interpretations that we began with). Here we start with a structure A, and we build another structure B whose signature need not be related to that of A, by defining the domain X of B and all the labelled relations and functions of B to be the relations definable in A by certain formulas with parameters. A further refinement is to find a definable equivalence relation on X and take the domain of B to be not X itself but the set of equivalence classes of this relation. The structure B built in this way is said to be interpreted in the structure A.
A simple example, again from standard mathematics, is the interpretation of the group Z of integers in the structure N consisting of the natural numbers 0, 1, 2 etc. with labels for 0, 1 and +. To construct the domain of Z we first take the set X of all ordered pairs of natural numbers (clearly a definable relation in N), and on this set X we define the equivalence relation by
(a,b) (c,d) if and only if a + d = b + c(again definable). The domain of Z consists of the equivalence classes of this relation. We define addition on Z by
(a,b) + (c,d) = (e,f) if and only if a + c + f = b + d + e.The equivalence class of (a,b) becomes the integer a - b.
When a structure B is interpreted in a structure A, every first-order statement about B can be translated back into a first-order statement about A, and in this way we can read off the complete theory of B from that of A. In fact if we carry out this construction not just for a single structure A but for a family of models of a theory T, always using the same defining formulas, then the resulting structures will all be models of a theory T that can be read off from T and the defining formulas. This gives a precise sense to the statement that the theory T is interpreted in the theory T. Philosophers of science have sometimes experimented with this notion of interpretation as a way of making precise what it means for one theory to be reducible to another. But realistic examples of reductions between scientific theories seem generally to be much subtler than this simple-minded model-theoretic idea will allow. See the entry on intertheory relations in physics.
The fourth kind of definability is a pair of notions, implicit definability and explicit definability of a particular relation in a theory. See section 3.3 of the entry on first-order model theory.
Unfortunately there used to be a very confused theory about model-theoretic axioms, that also went under the name of implicit definition. By the end of the nineteenth century, mathematical geometry had generally ceased to be a study of space, and it had become the study of classes of structures which satisfy certain geometric axioms. Geometric terms like point, line and between survived, but only as the primitive symbols in axioms; they no longer had any meaning associated with them. So the old question, whether Euclids parallel postulate (as a statement about space) was deducible from Euclids other assumptions about space, was no longer interesting to geometers. Instead, geometers showed that if one wrote down an up-to-date version of Euclids other assumptions, in the form of a theory T, then it was possible to find models of T which fail to satisfy the parallel postulate. (See the entry on geometry in the 19th century for the contributions of Lobachevski and Klein to this achievement.) In 1899 David Hilbert published a book in which he constructed such models, using exactly the method of interpretation that we have just described.
Problems arose because of the way that Hilbert and others described what they were doing. The history is complicated, but roughly the following happened. Around the middle of the nineteenth century people noticed, for example, that in an abelian group the minus function is definable in terms of 0 and + (namely: -a is the element b such that a + b = 0). Since this description of minus is in fact one of the axioms defining abelian groups, we can say (using a term taken from J. D. Gergonne, who should not be held responsible for the later use made of it) that the axioms for abelian groups implicitly define minus. In the jargon of the time, one said not that the axioms define the function minus, but that they define the concept minus. Now suppose we switch around and try to define plus in terms of minus and 0. This way round it cant be done, since one can have two abelian groups with the same 0 and minus but different plus functions. Rather than say this, the nineteenth century mathematicians concluded that the axioms only partially define plus in terms of minus and 0. Having swallowed that much, they went on to say that the axioms together form an implicit definition of the concepts plus, minus and 0 together, and that this implicit definition is only partial but it says about these concepts precisely as much as we need to know.
One wonders how it could happen that for fifty years nobody challenged this nonsense. In fact some people did challenge it, notably the geometer Moritz Pasch who in section 12 of his Vorlesungen über Neuere Geometrie (1882) insisted that geometric axioms tell us nothing whatever about the meanings of point, line etc. Instead, he said, the axioms give us relations between the concepts. If one thinks of a structure as a kind of ordered n-tuple of sets etc., then a class Mod(T) becomes an n-ary relation, and Paschs account agrees with ours. But he was unable to spell out the details, and there is some evidence that his contemporaries (and some more recent commentators) thought he was saying that the axioms may not determine the meanings of point and line, but they do determine those of relational terms such as between and incident with! Freges demolition of the implicit definition doctrine was masterly, but it came too late to save Hilbert from saying, at the beginning of his Grundlagen der Geometrie, that his axioms give the exact and mathematically adequate description of the relations lie, between and congruent. Fortunately Hilberts mathematics speaks for itself, and one can simply bypass these philosophical faux pas. The model-theoretic account that we now take as a correct description of this line of work seems to have surfaced first in the group around Giuseppe Peano in the 1890s, and it reached the English-speaking world through Bertrand Russells Principles of Mathematics in 1903.
Mod(T) Mod()expresses that every structure of signature K which is a model of T is also a model of . This is known as the model-theoretic consequence relation, and it is written for short as
The double use of is a misfortune. But in the particular case where L is first-order, the completeness theorem (see the entry on classical logic) tells us that T holds if and only if there is a proof of from T, a relation commonly written
Since and express exactly the same relation in this case, model theorists often avoid the double use of by using for model-theoretic consequence. But since what follows is not confined to first-order languages, safety suggests we stick with here.
Before the middle of the nineteenth century, textbooks of logic commonly taught the student how to check the validity of an argument (say in English) by showing that it has one of a number of standard forms, or by paraphrasing it into such a form. The standard forms were syntactic and/or semantic forms of argument in English. The process was hazardous: semantic forms are almost by definition not visible on the surface, and there is no purely syntactic form that guarantees validity of an argument. For this reason most of the old textbooks had a long section on fallacies - ways in which an invalid argument may seem to be valid.
In 1847 George Boole changed this arrangement. For example, to validate the argument
All monarchs are human beings. No human beings are infallible. Therefore no infallible beings are monarchs.Boole would interpret the symbols P, Q, R as names of classes:
P is the class of all monarchs.Then he would point out that the original argument paraphrases into a set-theoretic consequence:
Q is the class of all human beings.
R is the class of all infallible beings.
(P Q), (Q R = 0) (R P = 0)(This example is from Stanley Jevons, 1869. Booles own account is idiosyncratic, but I believe Jevons example represents Booles intentions accurately.) Today we would write x (Px Qx) rather than P Q, but this is essentially the standard definition of P Q, so the difference between us and Boole is slight.
Insofar as they follow Boole, modern textbooks of logic establish that English arguments are valid by reducing them to model-theoretic consequences. Since the class of model-theoretic consequences, at least in first-order logic, has none of the vaguenesses of the old argument forms, textbooks of logic in this style have long since ceased to have a chapter on fallacies.
But there is one warning that survives from the old textbooks: If you formalise your argument in a way that is not a model-theoretic consequence, it doesnt mean the argument is not valid. It may only mean that you failed to analyse the concepts in the argument deeply enough before you formalised. The old textbooks used to discuss this in a ragbag section called topics (i.e. hints for finding arguments that you might have missed). Here is an example from Peter of Spains 13th century Summulae Logicales:
There is a father. Therefore there is a child. ... Where does the validity of this argument come from? From the relation. The maxim is: When one of a correlated pair is posited, then so is the other.Hilbert and Ackermann, possibly the textbook that did most to establish the modern style, discuss in their section III.3 a very similar example: If there is a son, then there is a father. They point out that any attempt to justify this by using the symbolism
x S(x) x F(x)is doomed to failure. A proof of this statement is possible only if we analyze conceptually the meanings of the two predicates which occur, as they go on to illustrate. And of course the analysis finds precisely the relation that Peter of Spain referred to.
On the other hand if your English argument translates into an invalid model-theoretic consequence, a counterexample to the consequence may well give clues about how you can describe a situation that would make the premises of your argument true and the conclusion false. But this is not guaranteed.
One can raise a number of questions about whether the modern textbook procedure does really capture a sensible notion of logical consequence. For example in Booles case the set-theoretic consequences that he relies on are all easily provable by formal proofs in first-order logic, not even using any set-theoretic axioms; and by the completeness theorem (see the entry on classical logic) the same is true for first-order logic. But for some other logics it is certainly not true. For instance the model-theoretic consequence relation for some logics of time presupposes some facts about the physical structure of time. Also, as Boole himself pointed out, his translation from an English argument to its set-theoretic form requires us to believe that for every property used in the argument, there is a corresponding class of all the things that have the property. This comes dangerously close to Freges inconsistent comprehension axiom!
In 1936 Alfred Tarski proposed a definition of logical consequence for arguments in a fully interpreted formal language. His proposal was that an argument is valid if and only if: under any allowed reinterpretation of its nonlogical symbols, if the premises are true then so is the conclusion. Tarski assumed that the class of allowed reinterpretations could be read off from the semantics of the language, as set out in his truth definition. He left it undetermined what symbols count as nonlogical; in fact he hoped that this freedom would allow one to define different kinds of necessity, perhaps separating logical from analytic. One thing that makes Tarskis proposal difficult to evaluate is that he completely ignores the question we discussed above, of analysing the concepts to reach all the logical connections between them. The only plausible explanation I can see for this lies in his parenthetical remark about
the necessity of eliminating any defined signs which may possibly occur in the sentences concerned, i.e. of replacing them by primitive signs.This suggests to me that he wants his primitive signs to be by stipulation unanalysable. But then by stipulation it will be purely accidental if his notion of logical consequence captures everything one would normally count as a logical consequence.
Historians note a resemblance between Tarskis proposal and one in section 147 of Bernard Bolzanos Wissenschaftslehre of 1837. Like Tarski, Bolzano defines the validity of a proposition in terms of the truth of a family of related propositions. Unlike Tarski, Bolzano makes his proposal for propositions in the vernacular, not for sentences of a formal language with a precisely defined semantics.
On all of this section, see also the entry on logical consequence.
For example if L is a first-order language with identity, whose signature consists of 1-ary predicate symbols, and L is the language whose sentences consist of the four syllogistic forms (All A are B, Some A are B, No A are B, Some A are not B) using the same predicate symbols, then L is reducible to L, because the syllogistic forms are expressible in first-order logic. (There are some quarrels about which is the right way to express them; see the entry on the traditional square of opposition.) But the first-order language L is certainly not reducible to the language L of syllogisms, since in L we can write down a sentence saying that exactly three elements satisfy P(x), and there is no way of saying this using just the syllogistic forms. Or moving the other way, if we form a third language L by adding to L the quantifier Qx with the meaning There are uncountably many elements such that ..., then trivially L is reducible to L, but the downward Loewenheim-Skolem theorem shows at once that L is not reducible to L.
These notions are useful for analysing the strength of database query languages. We can think of the possible states of a database as structures, and a simple Yes/No query becomes a sentence that elicits the answer Yes if the database is a model of it and No otherwise. If one database query language is not reducible to another, then the second can express some query that cant be expressed in the first.
So we need techniques for comparing the expressive strengths of languages. One of the most powerful techniques available consists of the back-and-forth games of Ehrenfeucht and Fraïssé between the two players Spoiler and Duplicator; see the entry on logic and games for details. Imagine for example that we play the usual first-order back-and-forth game G between two structures A and B. The theory of these games establishes that if some first-order sentence is true in exactly one of A and B, then there is a number n, calculable from , with the property that Spoiler has a strategy for G that will guarantee that he wins in at most n steps. So conversely, to show that first-order logic cant distinguish between A and B, it suffices to show that for any finite n, Duplicator has a strategy that will guarantee she doesnt lose G in the first n steps. If we succeed in showing this, it follows that any language which does distinguish between A and B is not reducible to the first-order language of the structures A and B.
These back-and-forth games are immensely flexible. For a start, they make just as much sense on finite structures as they do on infinite; many other techniques of classical model theory assume that the structures are infinite. They can also be adapted smoothly to many non-first-order languages.
In 1969 Per Lindström used back-and-forth games to give some abstract characterisations of first-order logic in terms of its expressive power. One of his theorems says that if L is a language with a signature K, L is closed under all the first-order syntactic operations, and L obeys the downward Loewenheim-Skolem theorem for single sentences, and the compactness theorem, then L is reducible to the first-order language of signature K. These theorems are very attractive; see Chapter XII of Ebbinghaus, Flum and Thomas for a good account. But they have never quite lived up to their promise. It has been hard to find any similar characterisations of other logics. Even for first-order logic it is a little hard to see exactly what the characterisations tell us. But very roughly speaking, they tell us that first-order logic is the unique logic with two properties: (1) we can use it to express arbitrarily complicated things about finite patterns, and (2) it is hopeless for discriminating between one infinite cardinal and another.
These two properties (1) and (2) are just the properties of first-order logic that allowed Abraham Robinson to build his nonstandard analysis. The background is that Leibniz, when he invented differential and integral calculus, used infinitesimals, i.e. numbers that are greater than 0 and smaller than all of 1/2, 1/3, 1/4 etc. Unfortunately there are no such real numbers. During the nineteenth century all definitions and proofs in the Leibniz style were rewritten to talk of limits instead of infinitesimals. Now let R be the structure consisting of the field of real numbers together with any structural features we care to give names to: certainly plus and times, maybe the ordering, the set of integers, the functions sin and log, etc. Let L be the first-order language whose signature is that of R. Because of the expressive strength of L, we can write down any number of theorems of calculus as sentences of L. Because of the expressive weakness of L, there is no way that we can express in L that R has no infinitesimals. In fact Robinson used the compactness theorem to build a structure R that is a model of exactly the same sentences of L as R, but which has infinitesimals. As Robinson showed, we can copy Leibnizs arguments using the infinitesimals in R, and so prove that various theorems of calculus are true in R. But these theorems are expressible in L, so they must also be true in R.
Since arguments using infinitesimals are usually easier to visualise than arguments using limits, nonstandard analysis is a helpful tool for mathematical analysts. Jacques Fleuriot in his recent PhD thesis automated the proof theory of nonstandard analysis and used it to mechanise some of the proofs in Newtons Principia.
In cognitive science the difference between models and modelling has become blurred. A central question of cognitive science is how we represent facts or possibilities in our minds. If one formalises these mental representations, they become something like models of phenomena. But it is a serious hypothesis that in fact our mental representations have a good deal in common with simple set-theoretic structures, so that they are models in the model-theoretic sense too. In 1983 two influential works of cognitive science were published, both under the title Mental Models. The first, edited by Dedre Gentner and Albert Stevens, was about peoples conceptualizations of the elementary facts of physics; it belongs squarely in the world of modelling of phenomena. The second, by Philip Johnson-Laird, is largely about reasoning, and makes considerable use of model-theoretic semantics in our sense. Researchers in the Johnson-Laird tradition tend to refer to their approach as model theory, and to see it as allied in some sense to what we have called model theory. (The book by Alan Garnham in the references is a recent work in this line.)
Pictures and diagrams seem at first to hover in the middle ground between theories and models. In practice model theorists often draw themselves pictures of structures, and use the pictures to think about the structures. On the other hand pictures dont generally carry the labelling that is an essential feature of model-theoretic structures. There is a fast growing body of work on reasoning with diagrams, and the overwhelming tendency of this work is to see pictures and diagrams as a form of language rather than as a form of structure. For example Eric Hammer and Norman Danner (in the book edited by Allwein and Barwise, see the Bibliography) describe a model theory of Venn diagrams; the Venn diagrams themselves are the syntax, and the model theory is a set-theoretical explanation of their meaning.
The model theorist Yuri Gurevich introduced abstract state machines (ASMs) as a way of using model-theoretic ideas for specification in computer science. According to the Abstract State Machine website (see Other Internet Resources below),
One uses a specification methodology to describe a system by means of a particular syntax and associated semantics. If the semantics of the specification methodology is unclear, descriptions using the methodology may be no clearer than the original systems being described. ASMs use classical mathematical structures to describe states of a computation; structures are well-understood, precise models.The book of Staerk et al. in the Bibliography is an example of ASMs at work.
Today you can make your name and fortune by finding a good representation system. There is no reason to expect that every such system will fit neatly into the syntax/semantics framework of model theory, but it will be surprising if model-theoretic ideas dont continue to make a major contribution in this area.
Table of Contents
First published: November 9, 2001
Content last modified: November 9, 2001