Logical Constants How To Pre Final Version

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

1

Logical Constants,

Or How to use Invariance in Order to Complete the Explication of Logical Consequence

1. The problem of logical constants

Given an interpreted language, the problem of logical constants consists in finding a

principled way to draw the line between expressions of the language that are logical (in natural

language, expressions such as “and”, “some”, “not” are obvious candidates) and expressions that

are not (words such as “red”, “cat” and “obvious” clearly seem to fall into that category). Despite

this apparently neutral and general phrasing, it is a problem with many faces. A first source of

variation is to be found in the tools that are used to characterize the would-be logical symbols:

approaches to the problem vary according to whether they are characterized in semantic terms, or

in proof-theoretic or game-theoretic ones. The goals that are pursued are a second cause of

variation. Characterizing logical constants is sometimes one step in a substantial philosophical

agenda. In order to show that mathematical truths are logical, logicists will typically need a sharp

delineation of the realm of logical truths, which in turn requires agreeing upon what the logical

vocabulary is. In such a project, the demarcation itself is largely instrumental. What matters are

the properties that such a selection will bestow upon logical truths, the ultimate goal being to

show that logic and mathematics share some philosophically significant features, e.g. being

knowable a priori. By contrast, the goal could be purely descriptive, a classification of natural
2

language expressions being sought after. Linguists may thus ask whether the class of logical

lexical items constitute a natural kind, with distinctive syntactic, semantic and psycho-linguistic

properties.

Alternatively, the problem of logical constants can be a piece in a conceptual jigsaw

puzzle. Such is the case with the semantic definition of logical consequence that we owe to

Tarski (1936). Given an interpreted language, Tarski’s conceptual analysis explicate logical

consequence as preservation of truth under reinterpretation of the non-logical vocabulary. In

order to put the definition to work, a distinction between logical and non-logical symbols is

needed. Initially, Tarski expressed doubts regarding the possibility to “draw a sharp boundary

between the two groups of terms” (1936, p. 419). However, without such an objective division,

the conceptual value of the semantic explication of logical consequence remains questionable.

Clearly, not any boundary results in a sensible notion of logical consequence. If we do not know

where such a boundary may be drawn, we do not know why our explication works.

Later on though, Tarski himself proposed a principled demarcation.1 This is the well-

known criterion of invariance under permutation, which is meant to single out the logical

expressions of an interpreted language in terms of the semantic values they have. In the present

survey, I shall pursue the limited goal of evaluating this proposal, the objections which have

arisen and the amendments which have ensued. Our take on the problem will thus be to try and

complete Tarski’s semantic analysis of logical consequence by singling out the invariance

properties that are distinctive of logical constants.

Invariance is not the only option when it comes to completing the semantic definition of

logical consequence. Among possible alternatives, it has been suggested that symbols to be

counted as logical are distinguished by their grammatical role (Quine, 1980), by the relationships
3

between their semantic values and the rules which govern their use (Carnap, 1943) or by the

meta-logical properties of the logical systems built around them (Feferman, 2010). Is there

something special with the invariance approach, which would recommend it at least for a first

try? Here are three distinctive features. First, it is pure: only semantic properties are used in order

to characterize the available interpretations for logical constants. By contrast, Quine’s or

Carnap’s approach are not pure, because they appeal to grammatical or proof-theoretical

properties. Second, it is local: a logical constant is to be identified on the basis of the semantic

properties it has, independently of the semantic and non-semantic properties which emerge when

it is used in combination with other logical constants to build a language. By contrast, appealing

to meta-logical properties such as the completeness theorem or Lindström’s theorem does not

give a local characterization. Third, it is intrinsic: logical notions are not contrasted or compared

to other kind of notions. By contrast, requiring that first-order quantifiers express finite

cardinality properties of sets is not intrinsic, because it defines logical notions in terms of their

relationships with other kinds of notions, namely cardinality notions. The invariance approach to

the problem of logical constants is currently the only approach which is pure, local and intrinsic.

This is not to say that the right approach to logical constancy needs to be pure, local and intrinsic.

This is not to say either that no alternative approach could be devised, which would be pure,

local and intrinsic but not invariance-based. But these features help account for the intuitive

appeal of the invariance approach. They certainly make it worth looking into what the prospects

are, which is precisely what the present survey aims at achieving.

2. The invariance criterion


4

Let us first have an informal look at invariance under permutation. We assume that a

domain of objects is given. A permutation on that domain is simply a bijective function from the

domain onto itself. The extension of an empirical predicate such as “red” is not invariant under

all permutations: a permutation can swap a red object and a non-red object. In a formal language,

the extension of the quantifier “” consists in all non-empty subsets of the domain – reflecting

the fact that xφ(x) is true iff the set of objects that satisfy φ(x) is not empty. Take an arbitrary

permutation  on the domain.  will map a set A of objects to the set (A)={b / there is aA,

b=(a)}. Now, a set A is not empty iff its image (A) is not empty. Hence, all permutations leave

the extension of “” unchanged. The interpretation of the existential quantifier is invariant under

permutation.

In these two particular cases, the invariance test seems to give the right results, but why

should we think that invariance under permutation is a distinctive feature of logical constants?

Two different reasons might be given. First, invariance under permutation reflects the formality

of logical notions (Sher, 1991). Invariance under permutation basically requires all objects to be

treated on a par. Therefore, invariant notions are formal, in the sense that they are not sensitive to

the identity of objects. Second, invariance under permutation reflects the generality of logic

(Tarski, 1986). The idea is that invariance under groups of permutations can be used as a

measure of generality, and that invariance under the group of all permutations corresponds to

utmost generality. In his Erlanger Program, Klein used groups of transformations to classify

geometries: the fact that topology is more general than geometry is expressed by the fact that

topological notions (e.g. being closed) are invariant under all homeomorphisms of the space onto

itself whereas properly euclidean notions (e.g. being a triangle) are only invariant under
5

transformations of the space that preserve ratios of distance. Logic being our most general theory,

logical notions should be invariant under all transformations, that is under all permutations.

Formally, Tarski’s criterion may be spelled out as a property of operations defined on a

functional hierarchy of typed domains. Let D be a base domain consisting in a set of objects, let

B be the set of truth-values {T,F}. e is the basic type for individuals, so D is of type e. t is the

basic type for truth-values, so B is of type t. Now domains and types are recursively defined. If j

and k are types, (j,k) is also a type. If E and F are domains of type j and k, then FE , the set of

functions from E to F, is a domain of type (j,k). The class of typed domains thus defined from D

and B is called the functional hierarchy over D. Given such a hierarchy, an operation of type l is

simply an object belonging to the domain E of type l. A permutation  on D may be lifted to a

permutation l over any domain E of type l in the hierarchy in the following way:

 e is 

 t is the identity on B,

 (j,k) is defined by (j,k) (f) = { (j(x), k(y) | (x,y)f }, for any f in the domain E of

type (j,k).

Given a functional type hierarchy over a domain D and a permutation  on D, an operation f of

type l is invariant under  iff l (f)=f. This now leads to

Tarski’s criterion: Given a domain D, an operation f in the type hierarchy over D is logical iff it

is invariant under all permutations on D.

The standard use of Tarski’s criterion involves three amendments. First, what we need, strictly

speaking, is a demarcation between logical and non-logical symbols. Moreover, the modern

definition of logical consequence uses varying domains over which symbols in the language are

interpreted. In order to allow for comparisons across domains, invariance under permutations
6

over a domain is replaced by invariance under bijections between domains. 2 Finally, it is

customary to restrict attention to first-order languages, so that only operations of type level at

most 2 are considered.3 This amended version of Tarski’s thesis is known as the Tarski-Sher

thesis, following Sher (1991):

Tarski-Sher thesis: A symbol c of type level at most 2 is logical iff its interpretation is invariant

under bijection.

3. Invariance in Trouble

The Tarski-Sher thesis might be considered as the received view regarding the semantic

characterization of logical constants. And the current attitude towards the received view pretty

much resembles what was the attitude towards the so-called received view in the philosophy of

science when philosophers started casting some doubts on the impressive achievements of

Logical Positivism. Everybody criticizes it but almost everybody uses it in some way or another.

First, let us try and chart the various objections that may be and have been made against our

received view, listing them according to how much they reject in the Tarski-Sher approach.

The most radical criticism simply consists in attacking the relevance of the thesis to our

understanding of logical consequence. Some philosophers argue that demarcating logical

constants is a vain task and that fixed formal criteria such as the invariance criterion are off the

mark when it comes to accounting for the flourishing practice of logicians. Let us call these

skeptical worries.

A different line of criticism is directed at the articulation between Tarski’s invariance

criterion, as a property of operations on a domain and the Tarski-Sher thesis as a demarcation


7

between logical and non-logical symbols. Intensional warnings are issued: the invariance

criterion is extensional, whereas being a logical constant is an intensional property of expressions

that cannot be captured at a purely extensional level.

Finally, the class of invariant operations that is obtained on Tarski’s account has been

scrutinized and its adequacy to our intuitions regarding what should count as logical has been

questioned. These are extensional quarrels with Tarski’s invariance: one may be sympathetic to

the invariance approach but deny that invariance under permutation gets the demarcation right.

We shall now consider these various objections in turn, as they will give the occasion to cover

much of the subsequent work on invariance.

4. Skeptical worries and the scope of invariance

The skeptic argues that the way logicians choose the expressions they study is based on

largely pragmatic principles that govern the history of logic. According to Gómez-Torrente, such

principles may include picking up expressions that are “usable and relevant in general reasoning”

(2002, p. 3) or that are “helpful in the clarification of particularly bothersome confusions or

problems in reasoning” (ibid). 4 It seems vain to try to use mathematical criteria in order to

account for such factors. Let us try and push the point.

Consider, as an illustration of the softness of the divide, a determiner like “most”. It is a

device useful for general reasoning, but it is not part of our familiar logical apparatus – “most” is

not definable in first-order logic, there are no standard proof rules for “most” that would be part

of traditional logic teaching. The reason why logicians do not cherish “most” the way they

cherish “some” does not seem to depend on philosophically substantive intuitions. For example,
8

it might just be that “most” is used in common day reasoning but that, as such, it is not of very

much use in mathematical thinking about integers, because deciding what “most integers” means

involves conceptual choices that go significantly beyond our everyday application of “most” to

finite sets. Contemporary logic originates in an attempt to formalize a kind of mathematical

reasoning, where “most” is mostly absent. And maybe this is simply the reason why “most” does

not belong to our standard logical vocabulary. Following the skeptic, this dashes our hopes to

find a mathematically precise criterion mimicking the pragmatic choices made along the history

of logic.

Does the skeptic win so easily? It is worth having a second look at our test case. “Most”

does pass the test of invariance under bijection: whether most As are Bs does not depend on who

the As and the Bs are. “Most” is arguably every bit as formal or as general as “some” and “all”

are. Now if we agree with our skeptic that “most” is not traditional logicians material, we might

reply that accounting for logical constancy is to be thought of as a Carnapian explication.

Agreement with logicality judgments is a desideratum but it may be overridden by other

desiderata, such as the exactness and simplicity of the explication. As it happens, an exact and

simple account of what makes “some” and “all” logical constants takes “most” on board.

But it is not even necessary to concede that much to the skeptic. To the contrary, we might well

say that “most” is a bona fide logical constant, which just happened not to be the focus of logical

studies. After all, the development of generalized quantifier theory has precisely shown that

“most” may be studied on a par with “some” and “all”, and the invariance criterion has been

adopted by formal semanticists as an important semantic universal for natural language

determiners (Barwise and Cooper, 1981). Moreover, semanticists have studied the properties that

might explain which possible determiners get lexicalized in natural languages, helping us to
9

better understand what “most” and the more familiar “all” or “some” have in common. An

example of such a property is conservativity, which says that Q As are Bs iff Q As are As and Bs,

where Q is an arbitrary determiner. Natural language determiners are conservative, and this

ensures that the restriction argument determines what the sentence is about, by restricting the

domain of quantification. Conservativity is not to be seen as a constraint on logical constants, but

rather as one of the properties that shape the logical constants that we use, among all possible

logical expressions. Such properties bridge the gap between the pure analysis of logicality and

the more mundane account of what are our logical apparatus actually is (see Peters and

Westerståhl, 2006, for an up-to-date overview of distinctive properties of determiners).

There is another skeptical worry, more diffuse but certainly quite influential among philosophers

of logic, that needs to be discussed. 5 We live in an age of logical profusion. Extensions of

classical logic show their ability to enlighten our understanding of the way we talk about

knowledge, modality or time. Non-classical logics help us account for phenomena such as

vagueness or presuppositions. Drawing boundaries around classical logic might then be dubbed

an otiose undertaking, whereas understanding the diversity of expressions that successfully come

under logical scrutiny, and the diversity of semantics that prove helpful would be regarded as a

more rewarding task.

However, this take on the agenda for philosophy of logic by no means condemns the

invariance approach to logicality. Quite the contrary. The “invariance under all permutations”

label is actually a bit misleading here. One might think that the approach is not parametric,

because once all permutations are selected as the functions against which invariance is to be

tested, there is no room for differentiating various kinds of logical operations. This is not so. In

the invariance framework, two base domains are distinguished, a domain of objects and a domain
10

of truth values. In virtue of the generality and formality of logic, all permutations on the domain

of objects come into play.

What about the domain of truth values? In the classical setting, it consists of two

elements, True and False, and the only admissible permutation on these elements is the identity.

Why allow all permutations on the first domain and only the identity on the second? The

conceptual motivation is that the difference between True and False is necessary to make

semantic interpretation meaningful. There is no interpreting symbols if we cannot make a

difference between, say, the extension of a concept and its complement. Now this leaves room

for parametrization. First, invariance readily makes sense for semantic settings in which more

than two truth-values are used. In the case of a four-valued semantics with two intermediate

truth-values, for example, one might be willing to consider invariance with respect to all

functions on the domain of truth-values that distinguish top, bottom and intermediate values.

Second, other domains might be introduced, that come up with some structure that needs

to be respected, just like the difference between True and False in the Boolean domain. Thus, in

a modal logic of time, logical operations come up invariant under permutations on the domain of

time instants that respect its ordering structure (van Benthem, 2002). The motivation for treating

invariant operations on time instants as logical is then that semantic evaluation presupposes a

time structure. On a conceptual level, this line of thought has been pursued by MacFarlane

(2000), who elaborates a distinction between intrinsic and extrinsic structure. Intrinsic structure

is necessary to semantic evaluation and invariance should be restricted to permutations over base

domains that respect intrinsic structure. On the mathematical side, classical characterization

results for invariance can be transferred to modal logics when the accessibility relation in Kripke

models is regarded as intrinsic structure (van Benthem and Bonnay, 2008).


11

The overall picture that emerges is that logicality is a graded notion. Pure logic is built

around notions that only presuppose the most basic semantic distinction between truth and falsity.

In the realm of so-called applied logics, by contrast, the notions which are studied satisfy

restricted forms of invariance, that presuppose more semantic structure. This is in keeping with

the original use of invariance in geometry, which was introduced by Klein as a means to classify

the growing diversity of geometries.

5. Intensional warnings

The intensional objections against the Tarski-Sher invariance are based on counter-

examples. One may define symbols such that their interpretation satisfies the invariance criterion

because of the extension they have in every possible world 6 but counting them as logical

constants would yield counter-intuitive results. Here is such a definition, cooked up by McGee

(1996, p. 578):

Hφ =Def (~φ and water is H2O)

Granting that it is necessary that water is H2O, H and ~ have the same extension in every

possible world, so by any criterion depending solely upon extensions in every possible world,

they pass or fail the logicality test together. Now ~ is as logical as one can get, but counting H as

a logical constant is not a very tempting option. The meaning of H just seems not to be of a

purely logical nature. Moreover, as McGee remarks, treating H as logical renders “1=0 or H1=0”

logically true but this cannot be right since it implies that water is H2O.

There are at least three different possible reactions to this problem. First, one might try to

circumvent the objection by going for a notion of interpretation of expressions that would not
12

always make H and ~ co-extensional. Such an analysis is already to be found in McCarthy

(1987), but revenge troubles seem to loom.7 It will be sufficient to find another sentence which is

always true but not logically true.

If modifying the notion of interpretation does not work, one may argue that the problem

is not with the interpretation of H but with how this interpretation is determined. A logical

constant should have an invariant interpretation solely in virtue of its meaning (McGee, 1996, p.

578). H is then excluded because its being invariant depends on water being H2O. However, the

idea of something holding solely in virtue of meaning is notoriously problematic, and revenge

still looms (Gomez-Torrente, 2002, MacFarlane, 2009).

A third more pessimistic reaction might take over. Failures to deal with counter-examples

like H might reveal that the whole approach is misguided. Being logical is an intensional

property of expressions, and tackling the problem at the level of their extensions would be just

wrong. Thus, in his survey on logical constants, MacFarlane concludes his discussion of

permutation invariance by saying that “an adequate criterion […] would operate at the level of

sense” (2009, section 5) and moves on to inferential characterizations of logical constants.

This might be throwing the baby with the bathwater. Defining family of expressions in

terms of the extension they have is common practice, even when membership in the family is an

intensional property. Instead of logical constants and invariant operations, think of defining

zoological expressions as expressions referring to animals. Define “zebra*” by saying that x is a

zebra* iff x is a zebra and water is H2O. “Zebra*” is a strange creature and we would probably

not be willing to consider it as a full-blown zoological expression.

However, how strong is this as an objection against the “animal theory” of zoological

terms? It certainly shows that the animal theory cannot be the whole story, that it needs to be
13

elaborated, say in terms of direct reference, that, one way or the other, the requirement of

reference to animals needs to be applied in such a way that impure cases like our “Zebra*” do

not sneak in. But, arguably, we still get the feeling that reference to animals is at least a starting

point in what would be the good theory of zoological terms. If this feeling is to be trusted, it

suggests that H2O-based objections point at problems in our conception of interpretation and

meaning, as is advocated by the first two options, rather than at problems in the invariance

approach itself, as suggested by MacFarlane’s pessimistic conclusion.

There is another intensional warning against invariance, which does not hinge upon

surreptitiously introducing non-logical content into invariant expressions. A logical symbol c

gets interpreted over varying domains. Now when two domains D and D’ have the same size,

invariance under bijection forces the extension of c to be similar on D and D’. Indeed, c cannot

be interpreted like the existential quantifier on D and like the universal quantifier on D’. Assume

that the interpretation of c gives the value True to a non-empty proper subset A of D. A bijection

from D to D’ will map A to a non-empty proper subset A’ of D’, and invariance under bijection

requires that the interpretation of c on D’ gives A’ the value True. However, if D and D’ do not

have the same size, one and the same logical symbol c may be interpreted completely differently

on D and D’. Since there is no bijection connecting D and D’, invariance under bijection

remains mute. As a consequence, a strange quantifier QS which is interpreted like the universal

quantifier on finite domains of size less than 17 and like the existential quantifier on all other

domains pass the Tarski-Sher test.8 McGee (1996) and Feferman (1999) see this as a serious

problem. According to Feferman, “there is a sense in which the usual operations of the first-order

predicate calculus have the same meaning independently of the domain of individuals over

which they are applied” (1999, p. 38). However, the objection may not be as strong as it seems to
14

be. It is beyond doubt that our usual logical symbols receive similar extensions on domains. But

is it part of the notion of a logical symbol that it exhibits this kind of regularity? Going back to

our example, should we say that QS is a logical constant which is not natural or should we say

that it is not a logical constant because it is not natural? The dispute is not merely verbal. If

regularity is built into logicality, either the invariance criterion needs to be supplemented with

other requirements or it needs to do two jobs at the same time, that is capturing the generality

and formality of logic and making all domains comparable.

There are at least two reasons not to go this way. First, the problem with our quantifier QS

is very similar to the problem with Goodman’s predicate Grue. Grue is an empirical predicate,

but it is not the kind of predicate that we want to use when we make inferences. The same is true

for the quantifier QS and, as a matter of fact, QS may be used to yield exactly the same sort of

inductive fallacy that Grue puts us into. But the fact that Grue has a gerrymandered

interpretation does not make it any less an empirical predicate. Why judge differently about QS?

A complementary response consists in appealing to closure under definability for logical

constants. Arguably, any expression that can be defined by means of purely logical expressions

is itself logical. How would non-logical content sneak in? Any choice of logical vocabulary is

bound to yield definable gerrymandered logical constants – after all, QS is definable in pure first-

order logic with equality. The moral could again be that the quest for logicality ends where

formal linguistics begin: logicians and semanticists using generalized quantifier theory have

identified formal properties that make for smooth interpretations in and across domains (see in

particular van Benthem, 1986), but these properties do not analyze logicality.

6. Extensional quarrels and generalized invariance


15

Which logical constants do we get on account of the Tarski-Sher thesis? The short answer

is that we get a lot of them. Invariance under bijection goes well beyond the familiar limits of

first-order logic. “Most” was already mentioned as a case in point, but much wilder creatures get

admitted. First, any cardinality based notion will pass the test, because bijections preserve

cardinalities, be they finite or infinite. In first order logic with equality, finite cardinality

quantifiers, such as “there are at least six” or “there are exactly seven”, are definable, but infinite

cardinality quantifiers, such as “there are uncountably many” or “there are exactly 17 ” are not.

All these quantifiers are invariant under bijection. The exact scope of invariance can be made

precise by giving a functional completeness result, i.e. by finding a language such that an

operation is definable in that language iff it is invariant. Such a result is known for invariance

under permutation over a fixed domain of arbitrary size: 9 an operation is invariant under

permutation iff it is definable in pure L, with equality, which is the infinitary generalization of

first-order logic where conjunctions over arbitrarily big sets of formulas and quantification over

arbitrarily big sets of variables are allowed (see Bell, 2012, for a precise definition). Finally any

class of first-order structures of a given similarity type which is closed under isomorphism may

be turned into a generalized quantifier which is invariant under bijection. Thus, one may define a

bijection invariant quantifier of type ((e,(e,e)),t) that checks whether the domain equipped with a

distinguished object and a function is isomorphic to <ℕ,0,s>. For any arithmetical formula φ, it is

then easy to find, using that quantifier, another formula which is logically true iff φ is true in the

standard model of arithmetic.

The upshot of invariance under permutation is to conflate logic and mathematics. This

conclusion is embraced by Sher, who argues that the difference between logic and mathematics
16

is a difference of perspective rather than a difference in nature (Sher, 2008). Mathematics studies

formal structures, and needs logical tools to do so. Logic uses formal structures in reasoning, and

mathematics is needed in the first place to understand the properties of these formal structures.

Thus, treating “there are uncountably many” as a logical constant is just a natural step in the co-

evolution of logic and mathematics. Once the underlying properties of uncountable sets are

mathematically well-understood, the quantifier may be added to our reasoning tool-kit. Note that

this does not amount to a vindication of logicism. Mathematics has not been reduced to an

epistemically or ontologically more fundamental theory. Rather, according to Sher, gaining a

better understanding of what logic is has made us realized that, from the start, mathematics and

logic were one and the same thing considered under different guises.

One might wish to resist this conclusion and take the conflation of logic and mathematics

as a sign that something is wrong with the Tarski-Sher thesis. As argued by Feferman (1999,

2010), it is not satisfactory to use set-theoretic notions that are not robust in the semantics of

logical constants. Logic is more basic than set-theory and should be immune to disputes raging in

the far reaching branches of set-theory that deal with higher infinites. The meaning of the

quantifier “there are uncountably many” depends on the exact extent of the set-theoretic universe,

in a way in which the quantifier “there is at least one” does not. A non-empty set living in a

model of ZFC will always remain non-empty even when we shift to a smaller or bigger model.

By contrast, an uncountable set may become countable when we shift to a bigger model of ZFC

(because a bijection between that set and the set of natural numbers did not exist in the smaller

model but exists in the bigger one). Moreover, it might well be that the confusion of logic and

mathematics is merely the unwelcome by-product of a partial conceptual analysis of logicality. If

it is decided right from start that there is nothing more to logic than being formal, or if utmost
17

generality is equated with invariance under permutation, then it is no surprise that logic and

mathematics end up being the same (Bonnay, 2008). After all, mathematic may be characterized

as the study of structures identified up to isomorphism. Invariance under permutation is the

hallmark of mathematics, at least as much as it is the hallmark of logic. Maybe this is the sign

that what has been analyzed is being mathematical rather than being logical?

The desire to break down the conflation of logic and mathematics has fostered interest in

invariance criteria different from Tarski’s. Independently of the success or failure of these

attempts, they have debunked a very bad reason to accept Tarski’s invariance criterion, namely

the belief that it is the only option on the table. Because there is no built-in structure which is

required be preserved, one might get the wrong impression that there is no going beyond

invariance under permutation or bijection. This is not so. To see this, it is sufficient to realize that

any notion of similarity between structures can be turned into an invariance criterion (Bonnay,

2008). In order to see this, it is easier not to interpret, say, a unary quantifier Q by means of a

function ||Q||M of type <<e,t>,t> for every domain M, but rather by a class ||Q|| of first-order

structures, namely the class of structures of the form <M,P>, with P a subset of M, such that

||Q||M (P)=True. Invariance under bijection amounts to requiring that the class ||Q|| is closed

under isomorphisms and Tarski’s approach can be reinterpreted in the following way. Two

structures are logically similar iff they are isomorphic, and the interpretation of logical constants

should not distinguish between logically similar structures. Extensional quarrels with the Tarski-

Sher thesis suggest that logical similarity is more coarse-grained than similarity as existence of

an isomorphism. As a case in point, consider rejecting the claim that infinite cardinality

quantifiers are logical. On the invariance account, this is not possible if two structures <M,P>

and <M,P’> are deemed as logically dissimilar whenever they are not isomorphic – as we have
18

seen before, this is the reason why all cardinality quantifiers, including “there are exactly 17 ”

end up being invariant and hence logical according to Tarski’s criterion. But consider two

structures <M,P> and <M,P’> where P and P’ are two infinite subsets of M of different

cardinalities. They are not isomorphic. But one might still be willing to consider them as

logically similar, because distinguishing between higher infinites is not the business of logic

alone. Such thoughts vindicate the search for notions of logical similarity that do not require the

existence of a full-blown isomorphism.

The first to tap into the resources of invariance generalized beyond Tarski’s criterion was

Feferman (1999), who considers replacing bijections with arbitrary surjective functions. Making

room for functions that are not injective allows for different objects to be identified. As a result,

cardinality notions and identity are no more deemed to be logical. Infinite cardinality quantifiers

are excluded, but finite cardinality quantifiers of the form “there are exactly n” have also been

dropped out. Feferman’s proposal yields exactly the operations which are definable in first-order

logic without equality, but this is actually due to seemingly arbitrary syntactic restrictions on the

syntactic types that are allowed for logical constants. Without these restrictions, one gets

functional completeness with respect to pure L, without equality, for quantifiers restricted to

definable subsets of the domain.10 If getting L, with equality is getting too many, getting L,

without equality is probably getting too many as well, because of the possibility to define a

congruence relation mimicking the real identity.

Feferman’s move is not the only available one. In particular, discarding infinite

cardinality quantifiers does not imply discarding identity. Invariance under potential

isomorphisms, which are finite approximations of isomorphisms, emerges in that context as a


19

conceptually well-motivated candidate (Bonnay, 2008), because of the set-theoretic robustness of

potential isomorphisms11 and because it is the most general notion of invariance that preserves

the logicality of the existential quantifier.

Feferman (2010) has two objections. First, robustness should not be required of the

invariance criterion itself, but rather of the resulting operations. Second, the motivation in terms

of generality presupposes that all logical notions have to be invariant under the same notion of

invariance, but families of invariance properties could be considered instead, as is familiar from

Ehrenfreucht-Fraïssé games for first-order logic. As a consequence, the best option might rather

be to supplement the invariance criterion. Granting that invariance under bijection applies to

mathematics as well as to logic, proper logical constraints would be added on invariant

operations in order to single out the class of logical operations. Feferman (2010) thus suggests to

require, on top of invariance under bijection, either absoluteness with respect to the Kripke-

Platek set theory (KP) minus the axiom of infinity or absoluteness with respect to KP and

recursive enumerability of logical validities. Both suggestions yield exactly the operations which

are definable in pure first-order logic.12

7. Conclusion

Innocence has been lost on the way of a careful examination of the Tarski-Sher thesis.

Much of the appeal for the thesis stems from its elegant simplicity. As we carefully consider

objections, we get a better understanding of the issues involved, and a simple solution to all these

issues seems more and more out of reach. Addressing skeptical worries and making room for the

variety of applied logics forces us to recognize the need for a distinction between semantic
20

structure, which even logical operations need to respect, and non-semantic structure. Drawing

appropriate lessons from the intensional warnings against the shift from Tarski’s invariance

criterion to the Tarski-Sher thesis, we realize that characterizing logical operations is not

sufficient to characterize what it is to be a logical expression. Finally, extensional quarrels reveal

the need for a deeper conceptual analysis of logicality, invariance under permutation being just

one candidate among many. This is not to say that the project has failed. As we get wiser, we

may simply have to accept that there is no cheap way to solve a foundational issue such as the

demarcation of logical constants: what we need is to carefully lay down all the assumptions that

need to be made in order for a particular solution to be justified.


21

Works Cited

Barwise, Jon & Cooper, Robin. "Generalized Quantifiers and Natural Language." Linguistics

and Philosophy 4 (1981): 159-219.

Bell, John. "Infinitary Logic." (ed), Edward N. Zalta. The Stanford Encyclopedia of Philosophy

(Spring 2012 Edition). 2012.

Bonnay, Denis and Fredrik Engström. "Invariance and Definability, with and without equality."

manuscript, 2013.

Bonnay, Denis. "Logicality and Invariance." Bulletin of Symbolic Logic 14.1 (2008): 29-68.

—. Qu'est-ce qu'une constante logique ? Thèse de doctorat. Université Paris I, 2006.

Carnap, Rudolf. Formalization of Logic. Cambridge, Mass.: Harvard University Press, 1943.

Etchemendy, John. The concept of logical consequence. Stanford: CSLI, 1999.

Feferman, Solomon. "Logic, logics and logicism." 40.1 (1999): 31-54.

—. "Set-theoretical invariance criteria for logicality." Notre Dame Journal of Formal Logic 51.1

(2010): 3-20.

—. "Which quantifiers are logical? A combined semantical and inferential criterion." Synthese

(2012): to appear.

Gomet-Torrente, Mario. "The Problem of Logical Constants." Bulletin of Symbolic Logic 8.1

(2002): 1-37.

Krasner, Marc. "Généralisation abstraite de la Théorie de Galois." Algèbre et théorie des

nombres, colloque CNRS numéro 34. CNRS éditions, 1950. 163-168.

MacFarlane, John. "Logical Constants." (ed), Edward N. Zalta. The Stanford Encyclopedia of

Philosophy (Fall 2009 edition). 2009.


22

—. What does it mean to say that logic is formal? PhD Dissertation, University of Pittsburgh,

2000.

McCarthy, Tim. "Modality, Invariance and Logical Truth." Journal of Philosophical Logic 16

(1987): 423-443.

McGee, Vann. "Logical Operations." Journal of Philosophical Logic 25 (1996): 567-580.

Peters, Stanley and Westerståhl. Quantifiers in Language and Logic. Oxford University Press,

2006.

Quine, W.V. "Grammar, Truth and Logic." Kanger, S. and Öman and S. Philosophy and

Grammar. Dordrecht: Reidel, 1980. 17-28.

Read, Stephen. "Formal and Material Consequence." Journal of Philosophical Logic 23 (1994):

247-265.

Sher, Gila. "A characterization of logical constants is possible." Theoria 18.2 (2003): 189-198.

—. The Bounds of Logic. Cambridge: MIT Press, 1991.

Tarski, Alfred. "Ueber den Begriff der logischen Folgerung." Actes du Congrès International de

Philosophie Scientifique. 1936. 1-11.

—. "What are Logical Notions?" History and Philosophy of Logic (1986): 143-154.

van Benthem, Johan & Bonnay, Denis. "Modal Logic and Invariance." Journal of Applied and

Non-Classical Logic 18.2-3 (2008): 153-173.

van Benthem, Johan. "Logical Constants, the variable fortunes of an elusive notion." W. Sieg, R.

Sommer & C. Talcott. Reflections on the Foundations of Mathematics: Essays in Honor

of Solomon Feferman. Lecture Notes in Logic (15). Natick, MA: AK Peters, Ltd, 2002.

420-440.

Warmbrod, Ken. "Logical Constants." Mind 108 (1999): 503-538.


23

1
The proposal was made in a lecture given in 1966, edited by John Corcoran and

published posthumously as Tarski (1986). The history of invariance under permutation and

logicality starts well before 1966, see a brief history in Bonnay (2006, chapter 3, section 1).
2
Adapting the definition of invariance to invariance under bijection is straightforward,

we skip the details here.


3
The type level of a symbol may be identified with the type level of the operations

interpreting it, which is defined as follows. e and t are of type 0 and the type of (j,k) is one more

than the maximum of the levels of j and k.


4
Skeptics are numerous. Warmbrōd (1999) voices similar skeptical worries. Etchemendy

(1990) and Read (1994) are also skeptics, but of a different kind. Gomez-Torrente and

Warmbrōd argue that it is unlikely that a formal characterization of logical constants is possible,

given how logical constants were selected through the history of logical systems. Etchemendy

and Read’s arguments take a different perspective. Both try to give reasons to think that no

demarcation can get logical consequence right, because of some mismatch between the

demarcation task and our understanding of logical consequence.


5
This skeptical worry is harder to locate in the literature, but see van Benthem (2002) for

a plea in favor of diversity in the context of a discussion of logicality and invariance.


6
For the sake of simplicity, it is now assumed that interpreted symbols get an extension

over domains that represent possible worlds. However, the objection does not essential hinge on

the use of a “representational” rather than an “interpretational” semantics for the definition of

logical consequence (in the terms of Etchemendy, 1990).


24

7
Sher’s own answer is based on the idea that a logical constant is identified with its

extension, though it is not so clear how this solves the problem (see Gómez-Torrente, 2002 and

Sher, 2003 for a discussion).


8
The issue is raised about the extension of QS. We still wish to classify this as an

intensional warning, because the objection has it that there is no homogenous sense for a logical

constant receiving such a denotation. Admittedly, this shows up with its reference being

gerrymandered, but being gerrymandered is not an objection per se: the objection operates at the

level of sense rather than at the level of denotation.


9
The result is commonly attributed to McGee (1996) but it was essentially proven much

earlier by Krasner (1950).


10
A claim without restriction to definable sets was made by Bonnay (2006) and referred

back to in the subsequent literature. However, the original proof was shown to be incorrect by

Fredrik Engström. Bonnay and Engström (2013) provide a correct proof for the claim restricted

to definable subsets.
11
Being potentially isomorphic is absolute with respect to KP (Feferman, 2010).
12
Feferman has recently advocated a very different approach, which identifies logical

constants in terms of how the inference rules that characterize them are defined (2012).

You might also like