Content uploaded by Juliette Cara Kennedy

Author content

All content in this area was uploaded by Juliette Cara Kennedy on Mar 07, 2017

Content may be subject to copyright.

Squeezing Arguments and Strong Logics⇤

Juliette Kennedy Jouko V¨a¨an¨anen

January 19, 2017

Abstract

G. Kreisel has suggested that squeezing arguments, originally for-

mulated for the informal concept of ﬁrst order validity, should be

extendable to second order logic, although he points out obvious ob-

stacles. We develop this idea in the light of more recent advances and

delineate the diﬃculties across the spectrum of extensions of ﬁrst order

logics by generalised quantiﬁers and inﬁnitary logics. In particular we

argue that if the relevant informal concept is read as informal in the

precise sense of being untethered to a particular semantics, then the

squeezing argument goes through in the second order case. Consid-

eration of weak forms of Kreisel’s squeezing argument leads naturally

to reﬂection principles of set theory.

1 Introduction

The foundational project is driven by the idea of modeling mathematical

discourse, more or less globally, by giving an adequate formal reconstruction

of it. Adequacy here, as in the phrase “adequate formal system,” is delivered

by the following: the formalism should be sound, complete, and at least

⇤This paper was written while both authors were participants in the Intensive Research

Program - IRP Large Cardinals and Strong Logics at the Centre de Recerca Matem`atica

during the fall semester of 2016, and while the second author was a participant in the

program “Logical Structures in Computation” at the Simons Institute for the Theory of

Computing at UC Berkeley. The authors thank the CRM and the Simons Institute for

their support. The authors presented this paper to the Working Group in the History and

Philosophy of Logic, Mathematics, and Science at Berkeley in November 2016. We thank

the audience for their questions and comments.

1

to the degree possible, e↵ective and syntactically complete. The formalism

should also be meaning-preserving, relative to a given semantics (ideally).

The idea of foundational formalism, as we called it in [17], is that with

such a system in hand one could reasonably claim that the formalism has

“captured” the informal discourse—whichever way one wishes to express this

idea of “capturing.”

At the same time the idea of considering just the informal mathematical

discourse on its own, so to say in situ, has also attracted interest. This is

implicit in so-called “practice-based” philosophy of mathematics—that prac-

tice is situated, after all, in natural language—while other philosophers take

a more direct interest in natural language. M. Glanzberg, for example, in his

[11], argues that the notion of, e.g., consequence at work in natural language

is to be distinguished from a genuinely logical consequence relation:

The success of applying logical methods to natural language has

led some to see the connection between the two as extremely

close. To put the idea somewhat roughly, logic studies various

languages, and the only special feature of the study of natural

language is its focus on the languages humans happen to speak.

This idea, I shall argue, is too much of a good thing.

Glanzberg is not only alerting us to the pitfalls of conceiving of natural

language, at least in its logical aspects, as a kind of thinly disguised formal

language, a matter of cleaning up the relevant deﬁnitions, concepts and so on.

For Glanzberg, natural language and formal discourse are instead, with re-

spect to their logical sca↵olding anyway, two autonomous domains—“though

the processes of identiﬁcation, abstraction, and idealization can forge some

connections between them.” “Natural language has no logic” is the paper’s

central claim.1

What does it mean to say that natural language has no logic? For

Glanzberg this is to deny the logic in natural language thesis (LNL thesis

henceforth), namely the thesis that:

1Glanzberg construes the term “logic” rather narrowly in the paper. The quote in the

central claim is a deliberate reference to Strawson’s [31]:

Neither Aristotelian nor Russellian rules give the exact logic of any expression

of ordinary language; for ordinary language has no exact logic.

2

A natural language, as a structure with a syntax and a semantics,

thereby determines a logical consequence relation.

This is as opposed to the logics in formal languages thesis (LFL thesis hence-

forth):

Logical consequence relations are determined by formal languages,

with syntactic and semantic structures appropriate to isolate those

relations.. . . Thus, the logics in formal languages thesis holds that

consequence relations are in formal languages, in the sense that

they are deﬁnable from them.

The LFL thesis is uncontroversial, if not trivial. The argument against the

LNL thesis is subtle and turns partly on a critique of the model-theoretic

account of logical consequence.2We will not consider Glanzberg’s critique

here, but will do so below, as the issue becomes relevant to Kreisel’s so-called

squeezing arguments, to which we now turn.

2 Squeezing Arguments and their Critics

Squeezing arguments may be thought of as falling on the other side of the

spectrum of belief in the LNL thesis (albeit tacitly). Introduced by Kreisel

in his 1967 “Informal Rigour and Completeness Proofs” [19], and since taken

up by W. Dean [7], H. Field [9], V. Halbach, P. Smith [30] and others, the

arguments go as follows:

Consider an informally deﬁned mathematical concept I. Formally

deﬁne two concepts Aand Bsuch that falling under the concept

of Ais a suﬃcient condition for falling under the concept of I,and

falling under the concept of Isuﬃces for falling under the concept

of B.ThusA✓I✓B, where the inclusions are understood as

applying to the extensions of the concepts A, B, I .

Now suppose the formal notions Aand Bhave the same exten-

sion. Then by the inclusions A✓I✓Bthe informal concept I

must coincide, again extensionally, with that of Aand B.

2Glanzberg’s arguments against the LNL thesis generally read “logic” as logical conse-

quence.

3

For the informal concept IKreisel took intuitive validity, denoted Val,

understood as truth in all possible structures. This includes set and class-

sized structures, as well as, in principle at least, structures that have no

set-theoretical deﬁnition. Taking formal ﬁrst order provability, denoted DF,

on the left, and taking truth in all set-theoretical structures,3denoted V,

on the right, Kreisel argued as follows: By soundness, DF✓Val . By t he

fact that truth in all structures entails truth in all set-theoretical structures,

Val ✓V. Thus

DF✓Val ✓V.(1)

Invoking the completeness theorem for ﬁrst order logic Kreisel concludes the

following theorem,ashecallsit,for↵any ﬁrst order statement:

Val ↵$V↵and Val ↵$DF↵.

Kreisel’s presentation of the argument has been criticised in the literature.

Smith [30] objects that Kreisel’s somewhat model-theoretic construal of Val

does not obviously capture the pre-theoretic notion in question, validity-in-

virtue-of-form, as Smith prefers to think of Val.

Field’s criticism of the argument in [9] involves the soundness claim,

namely the ﬁrst inclusion 8↵(DF↵!Val ↵):

In chapter 2 I argued that . . . there is no way to prove the sound-

ness of classical logic within classical set theory (even by a rule-

circular proof): we can only prove a weak surrogate. This is

in large part because we cannot even state a genuine soundness

claim: doing so would require a truth predicate. And a deﬁnable

truth predicate, or a stratiﬁed truth predicate, is inadequate for

stating the soundness of classical logic, and even less adequate

for proving it.4

Field goes on to prove soundness—or a weak surrogate of soundness—by

means of a formal truth predicate, applying in restricted cases. Taken as a

repair of Kreisel’s argument one might argue that it ignores the methodology

of the paper, which is heavily semantic (see below). Kreisel’s argument does

3A set-theoretical structure is one whose domain, relations and functions are sets in

the usual sense.

4[9], p. 191.

4

not depend on a proof of soundness in classical set theory. Kreisel is asking

us to take soundness for granted, on the basis of historical experience—or

as Kreisel puts it, intuitive notions standing the test of time. Instead, Field

takes DFas “primary”, in Kreisel’s terminology.5

As for Kreisel’s own “proof” of soundness, extending to ↵i(interpreting

↵as an ith order sentence) for all i, it amounts to arguing that the universal

recognition of the validity of Frege’s rules (DF) at the time, together with

the “facts of actual intellectual experience” accumulated subsequently, should

amount to no more and no less than the acceptance of

8i8↵(DF↵i!Val ↵i)

for us. And though a century of logical history has taught the logician

nothing if not to be extremely suspicious of inclusions such as (1)—suspicions

that Kreisel himself airs at the end of the paper—surely what Kreisel has in

mind here is the idea that DFwas formulated ex post facto,thatisprecisely

so as to guarantee soundness. DFstood for Frege and his contemporaries,

Kreisel claims, and stands for the contemporary mathematician too, as a

completely adequate formalisation of Val.

Returning to our discussion of [19]’s critics, Halbach [13] o↵ers a repair of

squeezing arguments in the form of a formal, syntactic substitutional notion

of logical validity, to be substituted in for Kreisel’s somewhat model-theoretic

notion.6Such a concept of validity is, in Halbach’s view, “closer to rough

and less rigorous deﬁnitions of validity as they are given in introductory logic

courses”:

I put forward the substitutional analysis as a direct, explicit,

formal, and rigorous analysis of logical consequence. The substi-

tutional deﬁnition of logical validity, if correctly spelled out, slots

5From [19], p. 153:

First (e.g. Bourbaki) ‘ultimately’ inference is nothing else but following for-

mal rules, in other words D is primary (though now D must not be regarded

as deﬁned set-theoretically, but combinatorially). This is a specially peculiar

idea, because 99 per cent of the readers, and 90 per cent of the writers of

Bourbaki, don’t have the rules in their heads at all!

6Halbach’s conceptual analysis applies more widely, that is, it is an analysis of the

“natural” concept of logical consequence in terms of substitutional validity ¨uberhaupt, i.e.

not just in connection with squeezing arguments.

5

directly into the place of ‘intuitive validity’ in Kreisel’s squeezing

argument, as will be shown below.7

This is as opposed to the model-theoretic account of consequence, with

its many (in Halbach’s view) drawbacks:

. . . on a substitutional account it is obvious why logical truth im-

plies truth simpliciter and why logical consequence is truth pre-

serving. On the model-theoretic account, valid arguments pre-

serve truth in a given (set-sized) model. But it’s not clear why

it should also preserve simple (‘absolute’) truth or truth in the

elusive ‘intended model’. Truth-preservation is at the heart of

logical validity. Any analysis of logical consequence that doesn’t

capture this feature in a direct way can hardly count as an ade-

quate analysis.8

Under a substitutional account, the connection with set theory is severed, or

such is the claim; and interpretations of logical formulae are now syntactic

objects:

On the model-theoretic account, interpretations are speciﬁc sets;

on the substitutional account they are merely syntactic and (un-

der certain natural assumptions) computable functions replacing

expressions.9

We do not address the question here whether Halbach’s is a reasonable

conceptual analysis of the intuitive notion of logical consequence, the notion

of consequence in itself. What seems clear to us is that the informal concept

of consequence at work in natural mathematical languages is often plainly se-

mantic, and moreover model-theoretic. That when the mathematician draws

inferences in natural language, s/he imagines a situation in which the hy-

pothesis is true—i.e. one has a model for the hypothesis in view—then s/he

argues that the conclusion must hold in that model.

Kreisel states the point rather colorfully—“they don’t have the rules in

their heads at all!”—but what he means is that, e.g. group theorists do

7[13], ibid

8[13], ibid

9[13], ibid

6

not derive theorems directly from the group axioms in practice, rather they

employ the semantic method, i.e. they imagine a group and then show that

the group has the property claimed for it. Analysts do not derive formal

theorems from the axioms of real numbers, the real numbers are taken as

astructure which satisﬁes the axioms, and then theorems are proved about

that structure. In fact this is what mathematicians are trained to do, as

a cursory look at most standard texts demonstrates. For example, Walter

Rudin begins his classic text Principles of Mathematical Analysis [27] with

the existence of the ﬁeld of reals:

Theorem 1.19. There exists an ordered ﬁeld Rwhich has the

least-upper-bound property.

The rest of the book is the investigation of this ﬁeld R. The proof of the ex-

istence of R, i.e. the construction of the reals from the rationals, is relegated

to an appendix.

For another example, Halsey Royden introduces the real numbers in his

Real Analysis [26] with the following:

We thus as su me a s gi ven t he set Rof real numbers, the set Pof

positive real numbers and the functions “+” and “·”onR⇥Rto

Rand assume that these satisfy the following axioms, which we

list in three groups.

As for the reasons behind the mathematicians’ semantic mode of thought,

this is in some sense the moral of G¨odel’s speed-up theorem:

Thus, passing to the logic of the next higher order has the e↵ect,

not only of making provable certain propositions that were not

provable before, but also of making it possible to shorten, by

an extraordinary amount, inﬁnitely many of the proofs already

available.10

This can be interpreted as saying that the semantic method, the method

of establishing logical consequence by considering models, enjoys a so-called

“speed-up” over the method of formal proofs. This may explain why the

model-theoretic notion of logical consequence seemed natural to Tarski and

others.

10[12], p. 397

7

The above objections to [19] are certainly appropriate. Kreisel’s notion

of intuitive validity is clearly overly theorised, in the sense that the intuitive

notion considered is not suﬃciently intuitive or pre-theoretic, per Smith; and

under theorised, per Field and Halbach respectively. At the same time though

one might ask, if the intelligibility of Kreisel’s squeezing argument depends

on replacing the pre-theoretic notion of intuitive consequence by some formal,

syntactic counterpart notion (as in [9] or [13]), what is the point of squeezing

arguments at all? Why not simply analyse formal consequence directly, as

logicians have always done?

For it se em s to us tha t the inte re st o f sq ueezi ng a rgume nts l ie s in thei r

being carried through in such a way as to fulﬁll what was originally claimed

for them in [19], namely to capture an informal, natural language mathemat-

ical notion by “squeezing” it between two formal ones. Of course it can be

argued whether Kreisel himself succeeded in this. The point is that if the aim

was to capture (or “squeeze”) informal notions used in practice—to provide

a conceptual analysis of the informal notion of validity, as it were—then as

we have argued above, Kreisel’s model-theoretic construal of intuitive conse-

quence was the correct and natural one. This is not to question the validity

of the conceptual analyses of logical consequence which have been pursued so

vigorously, especially in the period since Etchemendy’s [8], but rather to ask

whether the concepts emerging from such analyses ought to be slotted in for

the informal notion that appears in Kreisel’s analysis here, i.e. in the origi-

nal squeezing argument. We will return to the issue of “genuine informality”

below.

A ﬁnal objection may concern what is surely the very unnatural restriction

to the ﬁrst order case, that is to propositions of the form ↵1.Theissueis

addressed by Kreisel in the paper, who points to a partial result in this

direction, an analogue of the squeezing argument derivable in the case of

extensions of DFto the !-rule.11

Of interest to us here, and one of the topics of this paper, is the possible

development of squeezing arguments in the direction of inﬁnitary and second

order logics. As it turns out somewhat more is known about completeness

theorems for extensions of ﬁrst order logic than was known in 1967. Going

beyond the !-rule, which Kreisel mentions, completeness theorems have been

obtained for a number of inﬁnitary logics, as well as for logics intermediate

between ﬁrst and second order. The second order logic perspective has also

11The completeness of the !-rule is due to Orey [25].

8

been developed a great deal since the publication of Kreisel’s [19].

As a result of this logical work squeezing arguments of the kind Kreisel

seems to be asking for in [19] may now be available. It is an oddity of the

paper that while Kreisel mentions both the completeness of !-logic and inﬁni-

tary logics, he doesn’t mention completeness theorems that would have been

available already during the writing of [19], namely the Henkin Complete-

ness Theorem (1950) [14] for second order logic with the so-called Henkin

semantics, and the completeness theorem for L(Q1), the extension of ﬁrst

order logic by the quantiﬁer “there are uncountably many” due to Vaught

[36] and published in 1964. We will return to these logics below.

3 Squeezing Arguments and the Logic in Nat-

ural Language Thesis

Before we consider the possibility of expanding squeezing arguments in the

direction of second order and other strong logics, we ask, are the relevant

natural language concepts available for this analysis at all, i.e. even in the

ﬁrst order case? Keeping Glanzberg’s rejection of the LNL thesis in mind,

can one simply extract what one thinks of as the notion of informal validity

at work in the natural language mathematicians use, and devise a squeezing

argument for that notion? In other words, do squeezing arguments require

the LNL thesis?

Many researchers in the foundations of logic and mathematics may share

a tacit belief in what one might call the “logic in natural mathematical lan-

guage thesis”. At the more formalist end of the spectrum of foundationalist

views which have been pursued traditionally, one might even attribute a be-

lief in the identity of the notion of logical consequence at work in natural

mathematical languages with the notion deﬁned by the relevant formal lan-

guage. This is just the thought that the logical consequence relation deﬁned

by a suitable, maximally adequate formal language is the correct version of

the logical consequence relation at work in natural language mathematics—

what had been meant by the natural language concept all along. Tarski,

though no formalist, seemed to argue for this or a similar view in his conver-

sations on nominalism with Quine and Carnap at Harvard in 1940-41, when

he remarked that “the di↵erence between logic and mathematics” was that

9

“Mathematics = logic + ✏”.12

Others, with Glanzberg, might see the notion of consequence at work in

the mathematician’s natural language as exact but fundamentally di↵erent

from the formal notion. In that case one might ask, what separates formal

entailment from its counterpart in (mathematical) natural language? We

remarked above on Kreisel’s observation that Frege’s rules gained acceptance

among mathematicians over time. This is to say, presumably, that if the

relevant part of the discourse is formalized, then Frege’s rules would be the

mathematician’s obvious choice of logical rules. Formalization also involves

a choice of a semantics, but in contrast to the rules (DF) it is not clear that

a choice of semantics is determined by the informal practice in the second

order case, which is our interest here. The view taken in this paper is that

in the case of intuitive, informal second order validity, a choice of semantics

is entirely irrelevant to the conceptual analysis of the notion of informal

consequence, or so we will argue below. To be informal in the second order

case is to prescind from a choice of semantics.

We ﬁrst consider the case of strong logics in general.

4 Squeezing Arguments with Completeness

Theorems

Kreisel’s paper is entitled “Informal Rigour and Completeness Proofs,” and

indeed an apparently implicit assumption in [19] is that any time one has

a completeness theorem in hand for a given logic, the associated squeez-

ing argument should go through. More precisely, let Lbe a logic, and let

VLdenote L-validity understood in the standard semantic sense, that is

set-theoretically. Let ValLdenote informal validity (via L), understood in

Kreisel’s sense, that is as referring to truth in all structures. Finally, let

DLdenote derivability in the formal system introduced for L.13 Then one

would expect that a completeness theorem for the logic Ltogether with the

inclusions:

DL✓ValL✓VL

12Tarski is quoted in Carnap’s notebooks. See Mancosu, [24].

13For many Lit is obvious what DLshould be, but this is not always so.

10

should underwrite the extensional equivalence of the concepts L-provability,

informal L-validity and L-validity construed semantically.

Is this plausible? That is, can new squeezing arguments be obtained

from completeness theorems for strong logics? The following completeness

theorems for strong logics are known: in ZFC,completenesstheoremshave

been obtained for L(Q1), the extension of ﬁrst order logic by the quantiﬁer

“there are uncountably many”, due to Vaught [36], as was mentioned; for

L!1!, the logic which is otherwise ﬁrst order, but allowing conjunctions and

disjunctions of countably many formulae, due to Karp [16]; for so-called

coﬁnality logic, the extension of ﬁrst order logic by the quantiﬁer denoted

Qcf

xy(x, y), meaning “deﬁnes a linear order of coﬁnality ”fora regular

cardinal, due to Shelah [29]; and for so-called stationary logic, the extension

of ﬁrst order logic by the quantiﬁer denoted aas(s), meaning “a club of

countable sets ssatisﬁes (s)”, due to Barwise, Kaufmann and Makkai [3].

The last two require the Axiom of Choice for their completeness theorems.

Going beyond ZFC there is the logic L(Q2), the extension of ﬁrst order

logic by the quantiﬁer “there are at least @2many” proved complete by C.C.

Chang [6], as pointed out by G. Fuhrken [10], using the continuum hypothesis

CH.

Another interesting case going beyond ZFC is the extension of ﬁrst or-

der logic by the Magidor-Malitz quantiﬁer [22], deﬁned as follows: M|=

QMM,n

↵x1...x

n(x1,...,x

n)() 9 X✓M(|X|↵^8a1,...,a

n2XM|=

(a1,...a

n)). The completeness theorem for this logic uses the set-theoretical

principle 3, which is stronger than CH, and it is consistent that complete-

ness fails in the absence of 3([1]) .

The squeezing argument for the logic L(Q1), for example, would look

like this: Let DL(Q1)denote the concept of formal provability relative to this

logic. Let VL(Q1)denote the truth of L(Q1)-statements in all set-theoretical

structures. Finally let ValL(Q1)stand for the validity of L(Q1)-statements

relative to all possible structures. Then if the inclusions

DL(Q1)✓ValL(Q1)✓VL(Q1)

hold, the squeezing argument relative to the logic L(Q1) must also hold, by

the completeness theorem for L(Q1).

Strengthening the logic escalates one’s set-theoretic commitments, clearly.

The completeness theorem for ﬁrst order logic is actually equivalent to Weak

K¨onig’s Lemma (WKL), which is also required to prove the completeness

11

both of !-logic and of L!1!. The Axiom of Choice is used for proving com-

pleteness theorems for the coﬁnality and stationary logics, corresponding

to the generalised quantiﬁers Qcf

xy(x, y)andaas(s). The CH is used

for proving the completeness theorem for the logic L(Q2), and ﬁnally 3is

required for proving the completeness of the extension of ﬁrst order logic

obtained from the Magidor-Malitz quantiﬁer. A comprehensive study of the

exact nature of these commitments would seem to be in order, but is not our

concern here.

Countenancing such a hierarchy of commitments is acceptable in some

quarters and unacceptable in others—a matter of deciding whether the rele-

vant completeness theorems “speak for themselves,” to quote Kreisel.14 What

about the soundness claim in this advanced setting? In the ﬁrst order case

we claimed, in the spirit of informal rigour, that DFwas formulated so as to

guarantee soundness—in fact Kreisel’s soundness claim 8i8↵(DF↵i!Val ↵i)

extends to all orders, as we saw. There is no obvious reason why Kreisel’s

argument could not be extended to strong and inﬁnitary logics. In the case

of !-logic, the order is somewhat reversed. That is, formal validity (V) is

considered with respect to !-models, in which the positive integer part is

standard. Thus in the case of !-logic the semantics is designed so as to

underwrite the soundness of the omega-rule.

The case of second order logic in this regard is also striking, in that the

Henkin semantics is formulated speciﬁcally so as to guarantee not soundness

but completeness. We will now take up the question of whether squeezing

arguments can be obtained in the second order case.

5 Squeezing for Second Order Logic

The mathematician’s informal discourse very naturally includes second order

concepts—quantifying over functions and relations and so forth—so it is rea-

sonable to ask for a squeezing argument for informal second order validity.

14Kreisel used this phrase in discussing the independence of the CH in the paper (p.

140):

The present conference showed beyond a shadow of doubt that several recent

results in logic, particularly the independence results for set theory, have left

logicians bewildered about what to do next: in other words, these results do

not ‘speak for themselves’ (to these logicians).

12

But if a logic has a completeness theorem, then if the proof system of the

logic is e↵ective in the sense that the set of axioms and rules are recursive

and proofs are ﬁnite, then the set of valid sentences is recursively enumerable.

By [34, Theorem 1] the set of valid sentences of second order logic is actually

⇧2-complete in the Levy hierarchy. Thus on simple grounds of complexity

no reasonable completeness theorem can exist for second order logic.

Does this mean one shouldn’t pursue a squeezing argument for second

order ↵?Kreiselhimselftooktheviewthat“Forhigherorderformulaewe

do not have a convincing proof of 8↵2(V↵2$Val↵2) though one would

expect one.” We will now argue that a squeezing argument for second order

↵is available, once one incorporates the concept of validity with respect to

general models.

Before addressing this point, recall that the Henkin semantics is deﬁned

simply so that in the so-called general models, the second order variables of

a given formula are thought of as ranging over a ﬁxed subset of the power-set

of the domain. This subset of the power set may be a proper subset but it has

to satisfy the axioms of second order logic, including the full comprehension

axioms. In case the domain of quantiﬁcation is actually the full power set,

one refers to the model as “full” or “standard”, and the associated semantics

as full or standard semantics.

Taking the deﬁnition of Henkin semantics into account, a squeezing ar-

gument for informal validity of second order ↵would be the following:

Let Ddenote the usual axiom system of second order logic, already given

in [15]. As above, let Val↵2mean that ↵2is informally true in all structures,

including class-sized structures and including, in principle at least, structures

that have no set-theoretical deﬁnition. Now let V↵2mean that ↵2is true in

all set-theoretical structures. The unproblematic implications are:

D↵2!Val ↵2!V↵2.

Note that if the completeness theorem held for second order logic, we

could conclude straightway that Val↵2$V↵2and Val↵2$D↵2,asbefore.

Now denote by V0↵2the statement that ↵2is valid with respect to set-

theoretically deﬁned general models (satisfying—as we have assumed—the

full comprehension axioms). Consider the following implications:

D↵2!Val ↵2!V0↵2!V↵2.(2)

Suppose we assume (2). Then by Henkin’s [14] proof of

V0↵2!D↵2

13

Figure 1: Varieties of validity in second order logic.

together with (2) we would obtain:

Val ↵2$D↵2$V0↵2.

The ﬁrst implication of (2) is clear, modulo the soundness claim, and so

is the last, for trivial reasons. What about the middle implication Val ↵2!

V0↵2?If↵2is informally valid in all structures, why is it that general models

should count as such structures? If the second order variables of ↵2are

thought of as ranging over the full power set of the domain in question, why

is it the case that these second order variables can be regarded as ranging over

the subsets in a general model? Is there a principled way to distinguish non-

standard general models from standard (full) structures, from the “informal

practice” point of view?

To analyse the situation in more detail, let us write Vc|=↵2if ↵2is

valid in both class-sized structures and set structures. Let us also write

V0

c|=↵2if ↵2is valid in both class-sized general models and set-sized general

models. Finally, let us write Val0

c↵2if ↵2is informally true in all structures,

full or general, including class-sized structures. Figure 1 depicts the trivial

implications. The possible implication Val↵2!V0↵2is the problematic one.

So what distinguishes Val ↵2from Val0↵2?Weclaimthatontheinfor-

mal level it is impossible to see a di↵erence between a standard model and

14

a general model. It is true that if we consider a general model in isolation,

from outside, so to speak, it is easy to imagine that something is missing

from the model, in order for it to count as a standard model. For example,

if we consider an inﬁnite general model with a countable set of relations as

the range of second order variables, we know that the model is not standard.

There may be other ways of seeing the non-standardness from the outside.

We may, for example “see” that a general model of arithmetic has an ele-

ment with inﬁnitely many predecessors. The position taken here is that it is

contrary to the idea of informal validity that one should be able to survey

the situation from outside.

One might still think15 that there really is an informal concept of a general

model, encapsulated by the thought: “All the sets I need are there and if

some are missing, they do not change anything”. This would seem to be

di↵erent from the informal concept of a standard model, encapsulated by

the thought: “All the sets are there and no set, whether I need it or not,

is missing”. If this is the case it is conceivable that for some ↵2we make

the judgement that it is informally valid only in standard models, not in

all general models. However, while it is crystal clear what the di↵erence is

between standard models and general models in the technical, logical sense,

it is a di↵erent matter to see the di↵erence on the informal level.

We go further and claim that on the informal level, the di↵ erenc e is

not discernable. The reason for this is (essentially) that the general models

“know” all the deﬁnable sets and relations (by the Comprehension Axioms)

and they are the ones we refer to in mathematical practice.16

A similar line is articulated in [35], in which the second author has argued

that from the point of view of mathematical practice, when we actually use

second order logic we do not and in fact cannot see a di↵erence between

ordinary (“full”, or “standard”) models and general models.

I will argue in this paper that if second-order logic is used in

formalizing or axiomatizing mathematics, the choice of semantics

is irrelevant: it cannot meaningfully be asked whether one should

use Henkin semantics or full semantics. This question arises only

if we formalize second-order logic after we have formalized basic

15We are indebted to A. Blass for suggesting this line of thought.

16Note that the deﬁnable sets taken on their own are not suﬃcient as they do not satisfy

the Comprehension Axiom. One needs a little “blurring” around the edges, otherwise one

can diagonalise out of the class.

15

mathematical concepts needed for semantics. A choice between

the Henkin second-order logic and the full second-order logic as a

primary formalization of mathematics cannot be made;theyboth

come out the same.17

For example, let us consider Bolzano’s Theorem:

Theorem 1. (Bolzano) Every continuous real function on [0,1] which has

a negative value at 0and a positive value at 1assumes the value 0at some

point of (0,1).

For the proof, by the second order comprehension axiom one can instan-

tiate a universal second order quantiﬁer at X={x|f(x)<0}.ThesetXis

even ﬁrst order deﬁnable, with fas parameter. This is a paradigm example:

we operate on sets deﬁnable from existing sets. Of course, principles such

as the Axiom of Choice force us to introduce also non-deﬁnable sets, but

they do not exist because “all” sets exists but because we assume—and the

general models are assumed to satisfy—the Axiom of Choice.

We now tur n to t he i ssue of s et v s class -s ized mo del s . Con si de r the weaker

claim that

V↵2!Vc↵2(3)

that is, the claim that second order formulas valid relative to set-theoretical

structures are also valid relative to class-sized structures. In other words,

we ask, is it true that if a second order sentence has a class-sized model, it

also has a set-sized model? This cannot be proved from the axioms of von

Neumann-G¨odel-Bernays class theory (NGB), as the following “Zermelian”

argument shows: Let ↵2be the second order sentence which says that the

universe of the model is an inaccessible cardinal. Let be the least inaccessi-

ble and let Mdenote the cumulative hierarchy up to .ThenhM,P(M)i

is a model of NGB satisfying ↵2. But no set-sized model, in the sense of

hM,P(M)i,satisﬁes↵2.

Paul Bernays [5] formulated more or less exactly (3), albeit in dual form,

as a reﬂection principle, and observed that it implies the existence of inac-

cessible cardinals.18 In fact, (3) implies a parameter-free version of so-called

Levy’s Schema [20], which says that every deﬁnable closed unbounded class

Cof ordinals contains a regular cardinal. In the original Levy’s Schema the

17[35], p. 505. Emphasis ours.

18We are indebted to A. Blass for pointing this out.

16

deﬁnition of Cis allowed to have parameters. Since the class of all cardi-

nals is deﬁnable without parameters, we obtain from (3) a proper class of

inaccessible cardinals. Bernays goes on to formulate (3) with second order

parameters and arrives at what became later to be known as indescribable

cardinals.19 L. Tharp [33] showed that the parametrized principle implies

that for every n,theclassof⇧

1

n-indescribable cardinals is a proper class.

This gave immediately a proper class of e.g. weakly compact cardinals. For

an analysis and discussion of the situation we refer to Tait [32, Ch.6].

Thus we cannot expect a proof of V↵2!Vc↵2,atleastwithoutad-

ditional axioms. On the other hand, the assumption (3) formulated in a

reasonable class theory (such as NGB) seems plausible. By a result of Scott

[28], it is true in the above hM,P(M)i, assuming that is not only weakly

compact, but even measurable. In fact, it suﬃces to assume that is ⇧1

<!-

indescribable, hence (3) is consistent with V=L, assuming the consistency

of a ⇧1

<!-indescribable cardinal.

What about (3) for sentences ↵in other extensions of ﬁrst order logic than

second order logic? For ﬁrst order logic this is an immediate consequence of

the Levy Reﬂection Principle. For extended logics of the form L(Q↵)wecan

use translation to ﬁrst order set theory and get the analogue of (3) as for ﬁrst

order logic. The same is true for L(QMM,n

↵), L(Qcf

), and the extension of ﬁrst

order logic by the H¨artig-quantiﬁer Ixy(x) (y),meaning: the cardinality

of the set of elements xsatisfying (x) is the same as the cardinality of

the set of elements ysatisfying (y). For these powerful logics, unlike for

second order logic, the analogue of the small part of the squeezing argument

represented by (3) is simply provable in ZFC. The situation with stationary

logic is more complicated. We leave the status of (3) open, if ↵2is taken to

be a formula of so-called stationary logic rather than second order logic. The

Open Question is, whether it is provable in ZFC or not.

Attempts to formulate higher order reﬂection with higher order parame-

ters leading to larger large cardinals than (!) have failed (see [18]). However,

a di↵erent approach, due to P. Welch, to a very strong reﬂection principle

with second order parameters, called the Global Reﬂection Principle,givesa

proper class of Woodin cardinals [38, 37].

19In our model theoretic context second order parameters would correspond to adding

generalised quantiﬁers to second order logic.

17

6 L¨owenheim-Skolem Theorems

Kreisel asks for a convincing proof of 8↵2(V↵2$Val ↵2), on its face im-

possible as we saw. Short of such a proof, Kreisel then asks a more speciﬁc

question, which can be answered. Stating the L¨owenheim-Skolem Theorem

for ﬁrst order logic in the form 8↵8>!(V!+1↵1$V↵1), what is the ana-

logue to !for second order formulae?20

First we recall some deﬁnitions. Given a logic L,wesaythatLhas

L¨owenheim-Skolem number if is the least cardinal such that for all vo-

cabularies ⌧such that the cardinality of ⌧is ,ifasentencein the

vocabulary ⌧of the logic has a model M,thenithasamodelNof size

.IncaseNcan be taken to be a submodel of Mthen is called the

L¨owenheim-Skolem-Tarski (LS T )numberofthelogic.

Let L2denote second order logic. We can now state Magidor’s result [21],

which answers Kreisel’s question: is the the least supercompact cardinal if

and only if =LST (L2).

In fact there is now a whole range of logics calibrated by large cardi-

nals, in the sense that the assumption of the cardinal is equivalent to or

implies a L¨owenheim-Skolem-Tarski theorem for the logic. For the cases al-

ready mentioned the results are as follows: for coﬁnality logic, corresponding

to the generalised quantiﬁer Qcf

xy(x, y), the LST number is @1.21 For

stationary logic, corresponding to the quantiﬁer aas(s), the LST number

is consistently @1, assuming the consistency of a supercompact cardinal,

22

but the LST number of stationary logic can also be the ﬁrst supercompact

cardinal.23

Finally, the interesting case of the H¨artig quantiﬁer: It is now known

that if the LST number LS T (I) of this logic exists, then there is a weakly

inaccessible cardinal and LST (I)isatleasttheleastweaklyinaccessiblecar-

dinal. It is consistent relative to the consistency of a supercompact cardinal

that LST (I) is the ﬁrst weakly inaccessible, and also consistent that it is the

ﬁrst supercompact.24

A general approach to strong logics and reﬂection principles they give rise

20In Kreisel’s notation V↵1denotes the assertion “↵1is true is the cumulative hierarchy

up to ”.

21Shelah, [29].

22See Ben-David, [4].

23Magidor, unpublished.

24[23].

18

to is in J. Bagaria et al. [2], where a close connection is established between

LST numbers of strong logics and so-called structural reﬂection principles in

set theory.

Just as in the completeness theorems, and the ensuing squeezing argu-

ments, obtaining L¨owenheim-Skolem type theorems may require principles

that go beyond Weak K¨onig’s Lemma (WKL), suﬃcient in the case of ﬁrst

order logic.

7 Squeezing very simple concepts

Consider the concept Wof ﬁnite words in a given vocabulary X.Intuitively

we construct a word by placing letters from Xone after another a ﬁnite

number of times. What does this mean? We can use a squeezing argument

to shed light on this question. As an analogue of derivability consider the

concept Dof starting from the empty word and then adding one letter from

Xat a time to the end of any word we already have. As an analogue of set

theoretic validity we take the concept Cof being a member of every closed

set, where a set Ais called closed if the empty word is in A,allone-letter

words are in A, and the concatenation ww0of any two words w, w0of Aare

in A.Clearly,

D✓W✓C.

The ﬁrst “✓” is intuitively obvious because adding one letter to the end of a

word certainly yields another word. The second is less obvious but one can

run an informal induction on the length of the word to see that if Ais closed,

then the word is in A. It is a mathematical fact that

C✓D,

because Dis one of the sets that Cis the intersection of. Hence

D=W=C,

and the informal concept of a ﬁnite word is squeezed between two (exten-

sionally) identical exact concepts. Although everything in this squeezing ar-

gument in on a very elementary level, it is noteworthy that strictly speaking

the inclusion C✓Dis based on an impredicative argument.

19

Similarly, we may consider the concept Fof a ﬁnite set. Intuitively we call

asetﬁnite if we can use some natural number to list the elements of the set.

On the other hand, natural numbers can be identiﬁed with ﬁnite ordinals.

Thus there is a certain amount of circularity in the concept of ﬁniteness. So

what does “ﬁnite” exactly mean? Let us take as Dthe concept of starting

from the empty set and then adding one element at a time to get more sets.

Let Cbe the concept of belonging to every ideal class i.e. to every class

which contains the empty set, all singletons and is closed under unions of

any two elements of the class. Clearly,

D✓W✓C.

The ﬁrst “✓” is again intuitively obvious because adding one element to

a ﬁnite set certainly preserves the set ﬁnite. For the second one can use

informal induction on the ﬁnite size of the set to see that if Ais an ideal

class, then the set is in A. It is a mathematical fact that

C✓D,

because Dis one of the classes that Cis the intersection of. Hence

D=W=C,

and the informal concept of a ﬁnite set is squeezed between two (extension-

ally) identical exact concepts of class theory.

8 Conclusion

Do squeezing arguments capture the mathematician’s informal discourse,

even as it strays beyond ﬁrst order talk, quantifying over relations and func-

tions, and making implicit use of inﬁnitary rules? This is diﬃcult enough to

argue for in the ﬁrst order case. Nevertheless, we hope to have reinforced

Kreisel’s original argument in [19] that squeezing arguments have a general

role in the conceptual analysis of informal mathematical concepts. Moreover,

we have pointed out and given evidence to the claim that the circumstance

that the two sides of the squeeze (extensionally) agree is based in general on

a non-trivial mathematical fact.

In particular, we hope to have shown for strong logics that if we refashion

the relevant informal concepts appropriately (here validity), we can, so to say,

20

ﬁltrate the informal discourse involving those concepts through a hierarchy

of set-theoretic commitments ranging from Weak K¨onig’s Lemma (WKL) up

to 3.

We also s aw th at vari ous str ategi es p resent t he ms el ves i n th e secon d or der

case, that go beyond what Kreisel suggests, if intuitive second order validity

is understood in the right way.

References

[1] Uri Abraham and Saharon Shelah. A 2

2well-order of the reals and

incompactness of L(QMM). Ann. Pure Appl. Logic, 59(1):1–32, 1993.

[2] Joan Bagaria and Jouko V¨a¨an¨anen. On the symbiosis between model-

theoretic and set-theoretic properties of large cardinals. J. Symb. Log.,

81(2):584–604, 2016.

[3] Jon Barwise, Matt Kaufmann, and Michael Makkai. Stationary logic.

Ann. Math. Logic, 13(2):171–224, 1978.

[4] Shai Ben-David. On Shelah’s compactness of cardinals. Israel J. Math.,

31(1):34–56, 1978.

[5] Paul Bernays. Zur Frage der Unendlichkeitsschemata in der axiomatis-

chen Mengenlehre. In Essays on the foundations of mathematics,pages

3–49. Magnes Press, Hebrew Univ., Jerusalem, 1961.

[6] C. C. Chang. A note on the two cardinal problem. Proc. Amer. Math.

Soc., 16:1148–1155, 1965.

[7] Walter Dean. Squeezing feasability. In Pursuit of the Universal (Lecture

Notes in Computer Science), pages 78–88. Springer, 2016.

[8] John Etchemendy. The concept of logical consequence. Harvard Univer-

sity Press, Cambridge, MA, 1990.

[9] Hartry Field. Saving truth from paradox. Oxford University Press, Ox-

ford, 2008.

[10] Gebhard Fuhrken. Languages with added quantiﬁer “there exist at least

@↵”. In Theory of Models (Proc. 1963 Internat. Sympos. Berkeley),pages

121–131. North-Holland, Amsterdam, 1965.

21

[11] Michael Glanzberg. Logical consequence and natural language. In Foun-

dations of Logical Consequence, pages 71–120. Oxford University Press,

2015.

[12] Kurt G¨odel. Collected works. Vol. I. The Clarendon Press, Oxford Uni-

versity Press, New York, 1986. Publications 1929–1936, With a preface

by Solomon Feferman, Edited by Feferman, John W. Dawson, Jr., War-

ren Goldfarb, Charles Parsons and Robert M. Solovay.

[13] V. Halbach. The substitutional analysis of logical consequence.

https://mdetlefsen.nd.edu/assets/201030/halbach.consequence.pdf.

[14] Leon Henkin. Completeness in the theory of types. J. Symbolic Logic,

15:81–91, 1950.

[15] D. Hilbert and W. Ackermann. Grundz¨uge der theoretischen Logik.

Berlin: J. Springer (Die Grundlehren der mathematischen Wis-

senschaften Bd. 27). VIII, 120 S. (1928)., 1928.

[16] Carol R. Karp. Languages with expressions of inﬁnite length. North–

Holland Publishing Co., Amsterdam, 1964.

[17] Juliette Kennedy. On formalism freeness: Implementing G¨odel’s 1946

Princeton bicentennial lecture. Bull. Symbolic Logic, 2013.

[18] Peter Koellner. On reﬂection principles. Ann. Pure Appl. Logic, 157(2-

3):206–219, 2009.

[19] G. Kreisel. Informal rigour and completeness proofs. In Proceedings

of the International Colloquium in the Philosophy of Science, London,

1965, Vol. 1. Edited by Imre Lakatos, pages 138–157. North-Holland

Publishing Co., Amsterdam, 1967.

[20] Azriel L´evy. Axiom schemata of strong inﬁnity in axiomatic set theory.

Paciﬁc J. Math., 10:223–238, 1960.

[21] M. Magidor. On the role of supercompact and extendible cardinals in

logic. Israel J. Math., 10:147–157, 1971.

[22] Menachem Magidor and Jerome Malitz. Compact extensions of L(Q).

Ia. Ann. Math. Logic, 11(2):217–261, 1977.

22

[23] Menachem Magidor and Jouko V¨a¨an¨anen. On L¨owenheim-Skolem-

Tarski numbers for extensions of ﬁrst order logic. J. Math. Log.,

11(1):87–113, 2011.

[24] Paolo Mancosu. Harvard 1940–1941: Tarski, Carnap and Quine on

a ﬁnitistic language of mathematics for science. Hist. Philos. Logic,

26(4):327–357, 2005.

[25] Steven Orey. On !-consistency and related properties. J. Symb. Logic,

21:246–252, 1956.

[26] H. L. Royden. Real analysis. Macmillan Publishing Company, New

York, third edition, 1988.

[27] Walter Rudin. Principles of mathematical analysis. McGraw-Hill Book

Co., New York-Auckland-D¨usseldorf, third edition, 1976. International

Series in Pure and Applied Mathematics.

[28] Dana Scott. Measurable cardinals and constructible sets. Bull. Acad.

Polon. Sci. S´er. Sci. Math. Astronom. Phys., 9:521–524, 1961.

[29] Saharon Shelah. Generalized quantiﬁers and compact logic. Trans.

Amer. Math. Soc., 204:342–364, 1975.

[30] Peter Smith. Squeezing arguments. Analysis (Oxford), 71(1):23–30,

2011.

[31] P. F. Strawson. On referring. Mind, 9:320–344, 1950.

[32] William Tait. The provenance of pure reason. Logic and Computation

in Philosophy. Oxford University Press, New York, 2005. Essays in the

philosophy of mathematics and its history.

[33] Leslie H. Tharp. On a set theory of Bernays. J. Symbolic Logic, 32:319–

321, 1967.

[34] Jouko V¨a¨an¨anen. Second-order logic and foundations of mathematics.

Bull. Symbolic Logic, 7(4):504–520, 2001.

[35] Jouko V¨a¨an¨anen. Second order logic or set theory? Bull. Symbolic

Logic, 18(1):91–121, 2012.

23

[36] R. L. Vaught. The completeness of logic with the added quantiﬁer “there

are uncountably many”. Fund. Math., 54:303–304, 1964.

[37] Philip Welch. Global reﬂection principles, 2012. Isaac Newton Institute

preprint series NI12051-SAS.

[38] Philip Welch and Leon Horsten. Reﬂecting on absolute inﬁnity. Journal

of Philosophy, 113(2):89–111, 2016.

24