Large Cardinals, Esoteric Quantifiers, and

Very Strong Logics

Beau Madison Mount

29 April 2021

1 / 68
Preliminaries

This project started not from a question intrinsic to set theory, but
from a general interest in the issue of choice of logic: which logics
should an agent adopt for various purposes? Most debates on this
issue consider challenges to classical first-order logic from weaker or
nonclassical systems.1 Here I consider questions involving logics
stronger than first-order logic—in some cases, much stronger.
Very little in this talk is technically novel. Much of the discussion will
involve the philosophical implications of work by Jouko Väänänen,
Juliette Kennedy, and Menachem Magidor that is well-known to set
theorists, although perhaps less well-known to logicians in general.2

1 See
Williamson (2014, 2017); Hjortland (2017); Woods (2019); and many others.
2 Väänänen(1979, 1980, 1985, 2001, 2012, 2014, 2019); Kennedy and Väänänen (2017,
∞); Kennedy (2013, 2015, 2021); Bagaria and Väänänen (2016); Magidor (1971);
Magidor and Väänänen (2011); Kennedy, Magidor, and Väänänen (∞).
2 / 68
Outline

I’m going to set out a series of principles, broadly motivated by

structuralist intuitions about mathematical practice, and claim that
they can be used to motivate a revisionary approach to the question
of logic choice for mathematical theorizing—not only in the
enterprise of formulating theories about set-sized structures, but in
set theory itself. Specifically, I argue that if large cardinals exist, then
these principles mean that some fairly common approaches focussed
on first- and second-order logics turn out to be underpowered. I
discuss two ways forward—Väänänen’s sort logic programme (1979,
2014) and an alternative approach, based on the ‘forced’ absolutely
unrestricted ∃U quantifier discussed by Timothy Williamson (1999).

3 / 68
Outline

I’m not going to argue for one approach or the other. Rather, I’m
going to try to tease out the different pictures of the set theory and
logic that would make one rather than the other appealing. I suggest
that the central choice point depends on one’s attitude toward the use
of higher-order quantification in absolutely general theories (such as
set theory).
Most of this talk will be couched in a universist idiom. I appeal freely
to the ‘true’ V (or sometimes the true W, by which letter I denote the
universal class, including Urelemente). But at the end, I shall discuss
how much sense can be made of these ideas by the pluralist or
multiversist. This is an aspect of the project about which my thoughts
are still very tentative: I suspect that it differs a lot on the precise
pluralist position in question. In general, however, I shall argue that
the sort logic approach will be more congenial to the pluralist than
the ∃U approach.

4 / 68
Some Conventions

ZFCU is formulated in a first-order language with equality, using

primitive elementhood (∈) and sethood (ß) predicates, with full
separation and replacement schemata.
Higher-order systems are formulated in a relational type theory with
logical comprehension and a global well-ordering principle at all
orders; we retain the equality primitive at first order but treat
Xn = Yn at nth order as an abbreviation of ∀Zn+1 (ZX ↔ ZY) (and
similarly for polyadic types). For n ≥ 2, let Extn be the family of
principles ∀Xn = ∀Yn ↔ ∀z(Xz ↔ Yz) and its polyadic analogues.

5 / 68
Some Conventions

ZFC(U)n is axiomatized in the standard way, with second-order

separation and replacement (with parameters up to nth order). Let
ZFC(U)2≡ = ZFC(U)2 and, for n ≥ 3, ZFC(U)n≡ = ZFC(U)n +
{Exti : 2 ≤ i ≤ n − 1}. Let ZFC(U)< ω n
(≡) = {n < ω : ZFC(U)(≡) }.
S

Informally, we use U for the class of Urelemente, V for the class of

pure sets, and W for the universal class; Ū = W \ U. Thus we say the
intended interpretation of ZFCU is hW, ∈, Ū i, that of ZFC is hV, ∈i.

6 / 68
A Weak Structuralist Principle

My starting point is the following principle, which many will find

appealing:
(S1) In that portion of mathematics which is concerned with
making claims about ‘particular’ set-sized structures, what is
most fundamental about a particular structure is that which is
common to all members of its isomorphism class (the shared
‘abstract structure’).
It’s worth emphasizing how minimal (S1) is: it doesn’t concern
algebraic theories or class-sized entities (and thus avoids questions
about structuralism in set theory itself), and it doesn’t commit us to
any metaphysical doctrine about structures. It just expresses a certain
attitude of indifference among isomorphic structures that is
characteristic of mathematical practice (see Burgess 2015).

7 / 68
Another Structuralist Principle

My second principle is the following, which will be slightly more

controversial:
(S2) Abstract structures of the type discussed (S1) are most
perspicuously presented through categorical axiomatizations.
This is a governing idea in (for example) Shapiro (1991, 1997) and
Isaacson (2011), and it draws on the importance of the second-order
categorical axiomatizations of N = hN, 0, 1, <, +, ×i and
R = hR, 0, 1, <, +, ×i. The quasi-categorical axiomatization of set
theory (Zermelo’s result for ZFC2 ) is also intimately related, although
it’s not clear that it explicitly falls under (S1).
Note: there is an important challenge to this idea in Baldwin (2018),
but I’ll set that aside for a moment.

8 / 68
Abstract Model Theory

In order to set out the other principles I shall appeal to, I need to
introduce some distinctions among the roles that a logical system can
play. We’ll start with some concepts from abstract model theory.
For our purposes, a model-theoretic logic L is a quadruple
hΣ, Str[Σ], L [Σ], L [Σ] i, where Σ is a collection of (perhaps
multi-sorted, perhaps higher-order) signatures; for each σ ∈ Σ, Str[σ]
is a collection of sets, which we term structures in the signature σ. We
say that a C ∈ Str[σ ] is a model class if it is closed under
isomorphisms. L [σ ] is the language with signature σ, which we can
identify with a set of sentences. L [σ] ⊆ L [σ ] × Str[σ ]; we define
Mod(φ) = {M : M  φ} and require Mod(φ) to be a model class for
each φ. We further place some obvious requirements on
model-theoretical logics: we require first-order logical constants to
preserve their properties and we require reducts to behave
standardly. See Barwise (1974, 1985) for details.

9 / 68
Model-Theoretic Logics and Utility Logics

Most model-theoretic logics are uninteresting (or would be, if we

could even grasp them). But there are some that have been
extensively studied: Lκλ for κ, λ ≥ ω, L(Qα ) (Lωω + the ∃≥ℵα
quantifer), logics with various other generalized quantifiers, and the
higher-order logics Lα .
Not all of these are logics that it would make sense for an agent to
adopt, however. A logic that can be adopted needs tractable syntax
and some kind of (not necessarily complete) associated deductive
calculus. Let’s say that an agent adopts a logic L (with associated
calculus `L ) as a utility logic when she (1) is able and prepared, at
least in principle, to use L and `L to formulate theories about
(set-sized) mathematical objects and structures; (2) takes the
semantics of L, at least when used for the purpose of such theorizing,
to be specified model-theoretically; and (3) has access to some
background framework—a meta-model theory—in which she can
explicitly formulate the semantics of L.

10 / 68
Utility Logics and Home Logics

But not every logic (in the sense of a language and calculus) we use
can be a utility logic. Let’s say that an agent takes a logical system L
as a home logic when she does not take the semantics of L to be given
through model theory, but is nonetheless prepared to use L directly
for general-purpose theory-building. To the extent that she engages
in reflection on L’s semantics at all, she typically does so in a natural
language or in some other non-model-theoretic way. Theories which
are to be interpreted with their quantifiers ranging over non-set-sized
domains will normally be in a home logic; in particular, the
meta-model theory in which an agent describes the semantics of the
utility logics she uses must be couched in a home logic.

11 / 68
Utility Logics and Home Logics

My choice of the term home logic is inspired by this methodological

maxim:
Issues in philosophy often turn on what language we use as our
home language, the language in which we are happy to work, at
least for the time being, without seeing it through the lens of a
meta-language, the language that we treat as basic for explanatory
purpose. (Williamson 2003: 459)
(Väänänen’s notion of an ‘urlogic’ is related, although—as we shall
discuss at the end of the talk—there are some important differences.)

12 / 68
Utility Logics and Home Logics

More than one logical system can be adopted as a home or utility

logic, and one logical system3 can play both roles. For instance, we
standardly develop a semantics for first-order logic (L1 = Lωω ) taken
as a utility logic using set theory as the meta-model theory. But
ordinarily this set theory itself (which we regiment, if we are being
careful, as ZFCU) also uses L1 as a home logic.
We can think of Kreisel’s (1967) squeezing argument as connecting
the home-logical use of L1 and its utility-logical use: it tells us that the
model-theoretic consequence relation 1 and the informal,
non-model-theoretic consequence relation inf 1 coincide with each
other (and with the derivability relation `1 ).

3 In the sense of a collection of languages and a deduction system. I abuse notation

and use the same name for a model-theoretic logic equipped with a deduction system
and the collection of languages and deduction system alone (perhaps interpreted
non-model-theoretically) when no confusion can arise.
13 / 68
Two More Structuralist Principles

Given this distinction, I propose that (S1) and (S2) suggest the
following additional principles:
(S3) For any interesting (set-sized) mathematical structure S , we
should accept a utility logic L strong enough for there to exist a
theory TS that categorically axiomatizes S according to its
model-theoretic consequence relation L .
(S4) We should accept a home logic strong enough to admit a
meta-model theory for every utility logic that we accept, in
which its syntax and semantics can be formulated explicitly.
(The notion of a meta-model theory is related to the technical idea of
‘adequacy for truth’ in abstract model theory (Feferman 1974,
Väänänen 1985), but a much more informal treatment suffices for our
purposes.)

14 / 68
Second-Orderism

Question: how do popular positions measure up to (S3) and (S4)? I’ll

consider two versions of second-orderism.
Utility second-orderism:
Utility logics: L1 , L2 .
Home logic: L1
Meta-model theory: ZFCU
General second-orderism:
Utility logics: L1 , L2 .
Home logic: L1 , L2
Meta-model theory: ZFCU2
We’ll focus first on the utility logics. L2 , as is well-known, permits
categorical axiomatizations of all kinds of structures—N , R, . . . .

15 / 68
Second-Orderism and Quasi-Categoricity

Similarly, where `Inac is the least inaccessible cardinal,

hV`Inac , ∈  V`Inac × V`Inac i can be given a categorical axiomatization in
L2 .
This is related to the fact (although it isn’t strictly a categoricity
result) that we can obtain insight into the intended interpretation of
set theory using Zermelo’s quasi-categoricity theorem, even when we
consider L2 purely as a utility logic and not a home logic. For the key
result can be proved in first-order ZFC(U), since only a set-sized part
of V is considered at any one time:

ZFC ` ∀S (S 2 ZFC2 ↔ ∃κ (κ is inaccessible ∧

S = ∼ hVκ , ∈  Vκ × Vκ i)),

where S ranges over structures in the signature {∈}.

16 / 68
Recaptures

This illustrates an important point: there are often first-order

recaptures of results concerning higher-order set theory that are, in at
least some rough sense, content-preserving. In particular, if
ZFC(U)n ` φ, we can usually find a φ∗ with T ` φ∗ where T is a first
order set theory extending ZFC(U) with appropriate large cardinals,
and φ∗ is an intuitively adequate translation of φ into this framework.
Note in particular that if ZFC2 ` φ and φ is first-order, then
ZFC + LR ` φ, where LR is Lévy’s sentential reflection principle
(Roberts 2020).
But this isn’t always the case. The classic example is Kunen’s (1971)
inconsistency theorem: the intuitive content of the claim that there is
no j : V −→e.e. V is captured neither by the claim that no
j : V −→e.e. V is definable from parameters (Suzuki 1999) nor by the
claim that ZFC ` ∀λ¬∃j : Vλ+2 −→e.e. Vλ+2 .

17 / 68
Recaptures

As we shall see, there are similar failures of recapture for results

about certain strong logics.
(Note that we can always find something first-orderizable: we have
T ` BewZFC2 φ, after all, for any T strong enough to serve as a syntax
theory.) But intuitively this isn’t content-preserving.

18 / 68
The Limits of Second-Orderism

But not all interesting set-sized mathematical structures can be

categorically axiomatized in L2 .
Let’s make some simplifications. For our examples, we’ll focus on
structures consisting only of a set—in effect, pure questions of
cardinality. And we’ll focus on axiomatizability by a sentence. (For
the cases we consider, moving to axiomatizability by a theory in a
denumerable signature would not make a difference.) The spectrum
of a sentence φ (under a specified consequence relation ) is defined
as spect(φ) = {|M | : M ∈ Mod(φ)}. When there is an L-sentence
whose spectrum is {κ }, we say that κ is L-characterizable. (In all
cases, we shall tacitly assume that κ exists.)

19 / 68
The Limits of Second-Orderism

Second-order logic, unsurprisingly, has extensive cardinal-

characterizing power.

Theorem (various; see Väänänen 2012: 106–7 and Garland 1974)

For all α < ω1CK , ℵα and iα are L2 -characterizable. Furthermore, if κ and λ
are L2 -characterizable, then so are κ + , 2κ , ℵκ , iκ , and κ λ . If κ is the least
inaccessible, least Mahlo, least weakly compact, or least Ramsey cardinal, or
κ = 2`Meas for `Meas the least measurable cardinal, then κ is
L2 -characterizable.
But not every cardinal of interest is second-order characterizable.

Theorem (Hanf and Scott 1961)

The cardinal `Meas is not L2 -characterizable.

20 / 68
Higher-Orderism
In itself, this result may not seem like a serious problem. The least
measurable cardinal is L3 -characterizable, and it’s a fairly small
modification to turn utility second-orderism into a more general
utility higher-orderism:
Utility higher-orderism:
Utility logics: L1 , L2 , L3 , . . .
Home logic: L1
Meta-model theory: ZFCU
(Note that accepting Ln for each n is different from accepting the full
simple theory of types (L<ω ) itself, although we could easily add it as
a utility logic too.)
And, just as before, we can pinpoint a relevant quasi-categoricity
result inside ZFCU. Let ZFCM3 be the theory that adds to ZFC3 a Σ21
sentence saying that ON is measurable. We have:

ZFC ` ∀S (S 3 ZFCM3 ↔ ∃κ (κ is measurable ∧

S ∼= hVκ , ∈  Vκ × Vκ i)).
21 / 68
Measurable-Sized Structures

It might be objected that `Meas is far removed from anything of real

mathematical interest. As such, it does not fall under the dictate of
(S3) that we accept a utility logic strong enough to characterize any
interesting set-sized mathematical structure categorically. But this is
not so. Consider how Steel (1996: 5) proceeds in constructing Kc , the
core model containing Woodin cardinals:
Let us fix a measurable cardinal, which we call Ω, for the remainder
of this paper. We shall sometimes think of Ω as the class of all
ordinals; we could have worked in 3rd order set theory + “OR is
measurable”, but opted for a little more room.

22 / 68
Higher-Order Löwenheim-Skolem Properties

Unfortunately, the strategy of going to third, or nth, order can only

get us so far. This results from an important theorem, which states (in
effect) a Löwenheim-Skolem property for logics of finite order:

Theorem (Magidor 1971)

Let `SC be the least supercompact cardinal. If κ ≥ `SC , then κ is not
Ln -characterizable for any n ∈ ω.

We can even extend this into the transfinite:

Theorem (S. Roberts)
If κ ≥ `SC , then κ is not Lα -characterizable for any α < `SC .

23 / 68
Higher-Order Löwenheim-Skolem Properties

A convenient way of stating limitations on the cardinal-

characterizing power of a logic L is through its Löwenheim and Hanf
numbers. The Löwenheim number of L is defined as

l(L ) = sup({min(spect(φ)) : φ ∈ L }),

the Hanf number as

h(L ) = sup({sup(spect(φ)) : φ ∈ L ∧ spect(φ) is bounded}),

(At second order and above we can take L = L = L [∅].)

Theorem (Magidor 1971; see Shapiro 1991: 156)

For Ln , n ∈ ω, we have l(Ln ) < `SC . We also have h(Ln ) < `Ext , where
`Ext is the least extendible cardinal.

24 / 68
Higher-Order Löwenheim-Skolem Properties

What is going on here? Intuitively, the answer is that second- (or

finite-) order logic can give us access only to local properties of the
universe (properties witnessed by a structure of limited rank),
whereas supercompact and extendible cardinals are defined by global
properties of V. More rigorously, we have the following result:

Theorem (Väänänen 2012)

If a model class C = Mod(φ) for a single sentence φ of L2 , then C is
∆ZFC
2 -definable.

The property of being a supercompact cardinal is Π2ZFC ; the property

of being an extendible cardinal is Π3ZFC . So there is no hope of
characterizing the least of these cardinals in L2 (or a language of any
finite order).

25 / 68
Do We Need Categoricity at or above l(L2 )?

So it seems as though, if we take seriously the desiderata expressed

by (S3), we need to adopt a stronger utility logic to characterize
structures such as hV`SC , ∈ V`SC × V`SC i—and merely moving up to a
higher finite (or even tractable transfinite) level won’t do. We need a
very strong logic.
Objection: isn’t this result just a reductio of (S3)? Surely, insofar as
there is any mathematical interest in structures this large, we can just
work with them directly in first-order ZFC.
This objection has some force, but I think it can ultimately be met.
(S2) doesn’t intend to prescribe how mathematicians should actually
work (by restricting them to reasoning from specified axioms); it
merely suggests categorical axiomatizations can provide unique
insight. I think there are at least two cases where this may be a
defensible expectation.

26 / 68
Do We Need Categoricity at or above l(L2 )?

First, if the programme of inner model theory succeeds in

constructing an inner model containing a supercompact, the Ω
assumed as background for building the model (à la Steel’s
construction of Kc ) might itself be very large. It does not seem
implausible that we might, for some purposes, like an ‘internal’
treatment of the construction in the same way that Steel considers
(but ultimately rejects) working in ZFCM3 . The background axioms
for such a construction might have to be couched in a logic stronger
than Ln for finite n.

27 / 68
Do We Need Categoricity at or above l(L2 )?

Second, Vopěnka’s Principle is equivalent to a number of claims of

interest to category theorists (Adámek and Rosicky 1994). For that
reason, we can conceive of wanting to work in a category-theoretical
framework whose intended models involve extremely large
cardinals. The property of being a Vopěnka cardinal (a Vκ such that
hVκ , Vκ +1 i  Vopěnka’s Principle) is, however, second-order
characterizable. Nonetheless we might think that, without true
extendibles and more, these are unnatural models (much as worldly
cardinals that are not inaccessible are unnatural models of ZFC—a
“poor-man’s inaccessible cardinal” per Hamkins).

28 / 68
Sort Logic

Väänänen has proposed (1979, 2014) has proposed an extremely

strong—but nonetheless finitarily presented and
model-theoretic—logic that avoids the expressive limitations of
ordinary higher-order logic.
He terms this logic sort logic (LSort ), and he claims that “[s]ort logic
[ . . . ] is a many-sorted extension of second-order logic. In an exact
sense it is the strongest logic that there is” and that “[a]s the strongest
logic sort logic is an ultimate yardstick of definability in
mathematics” (2014: 185).

29 / 68
Sort Logic

The syntax of LSort is as follows:

Every natural number is a basic sort. Every basic sort is a sort;
every finite nonempty sequence of basic sorts is a sort. The
closure of a nonempty set ∆ of sorts, cl(∆), is the smallest set ⊇ ∆
generated from basic sorts by closing under finite sequences.
For every basic sort i ∈ ω, the vocabulary of every L Sort contains
denumerably many first-order variables of sort i, xi1 , xi2 , . . . . For
every sort τ = hi1 , . . . , in i, the vocabulary of every L Sort contains
denumerably many predicate variables of sort τ, X1τ , X2τ , . . . .
A signature σ consists of predicate constants, each associated
with a sort. Predicate constants and predicate variables of sort τ
are terms of sort τ. If τ = hi1 , . . . , in i and T is a term of type τ,
then Tt1 · · · tn is well-formed only when the sort of t1 is i1 , . . . ,
and the sort of tn is in .

30 / 68
Sort Logic

The vocabulary of L Sort [σ ] contains, in addition to these

variables, the constants in σ, and the usual logical constants, the
special sort quantifier ∃˜ .
Formation rules are standard, except that ∃˜ Xhi1 ,...,in i φ is
well-formed only when φ contains no free variables and no
constants, other than X, of any sort im (1 ≤ m ≤ n) or of any sort
τ such that τ ∩ cl({i1 , . . . , in }) 6= ∅.

31 / 68
Sort Logic

We define multi-sort structures standardly; given a collection Γ of

basic sorts, a structure S over cl(Γ) specifies a domain for each sort
in Γ and an interpretation of nonlogical constants, if such there are, in
the sorts in cl(Γ). The notion of the expansion of a structure S1 over
cl(Γ) to a structure S2 over cl(∆) for some ∆ ) Γ is straightforward,
as is the notion of an assignment β over S2 ’s extending an
assignment α over S1 .
Semantic clauses for Sort are standard for a multi-sorted
second-order theory, except for the ∃˜ quantifier. We set
S Sort
α ∃˜ X φ, where the collection of sorts appearing in φ is Γ
and the sort of X is hi1 , . . . , in i, just in case there exist an S 0 over
cl(Γ ∪ {i1 , . . . , in }) extending S and an assignment α0 over S 0
extending α such that S 0 Sort α0 ∃X φ.

32 / 68
Sort Logic: An Example
How does this work in practice? Let’s take a simple example: say that
we want to write a sentence φ in the signature {≺} such that
S Sort φ just in case S is order-embeddable into hR, <i. We
formulate the axioms for a linear order in sort 1, with ≺ as a
sort-h1, 1i constant, and call their conjunction A; we formulate the
second-order axioms for the real ordered field in a relational
signature ({⊕, ⊗, <}) in sort 2; we call their conjunction B. Let B∗ be
B[R1/⊕][R2/⊗][R3/<]. Our sentence is

h2,2i h2,2i h2,2i

A ∧ ∃˜ Qh1,2i ∃R1 ∃R2 ∃R3 (B∗ ∧
∀x1 ∃!y2 (Qxy ∧ ∀z1 (Qzy → x = z))∧
∀x11 ∀x12 ∀y21 ∀y22 (x1 < x2 ∧ Qx1 y1 ∧ Qx2 y2 → R3 y1 y2 )),

i.e.

A ∧ ∃˜ f h1,2i ∃R1 ∃R2 ∃R3 (B∗ ∧ f : h[λx1  x = x], <i ,→ [hλy2  y = y], R3 i).

33 / 68
Sort Logic: Metatheory
So how much structure-characterizing power does sort logic have?
Väänänen claims that it is “the strongest logic”, and he supports it
with the following theorem:

Theorem
For any model class C , C = Mod(φ) for a φ ∈ LSort just in case C is
ZFC-definable.

From this it follows trivially that

Corollary
A cardinal κ is LSort -characterizable just in case it is ZFC-definable.

Furthermore,
Corollary (Väänänen 1979: 393)
l(LSort ) = h(LSort ) = sup({κ : κ is ZFC-definable}).

34 / 68
Sort Logic: Philosophical Implications

So what if we add LSort as a utility logic? We now can meet the (S3)
desideratum for any ZFC-definable structure.
But how does this affect (S4)? The Sort relation is not
ZFC-definable—at most a ΣSortn (where n is the depth of sort
quantifiers) fragment of it can be defined at a time (in essentially the
same way that a ΣZFC
n partial truth predicate can be defined in ZFC).
So, if we accept LSort as a utility logic, we cannot rest content with
ZFCU as our background meta-model theory. We have two options:
we can move to a stronger meta-model theory in a stronger home
logic, or we strengthen the meta-model theory without strengthening
the home logic by adding new expressive resources to the
metalanguage.

35 / 68
Sort Logic: Philosophical Implications

The first option is taken by Väänänen (1980): he explicitly indicates

he’s working in Morse–Kelley class theory. In (2014), however, it
seems to be suggested that the second approach is favoured
(although the official approach can be understood as restricting
oneself to fragments of sort logic to stay within ZFC). What happens
if we make our meta-model theory ZFC(U) extended with axioms for
a truth-in-a-model predicate for sort logic (still formulated in L1 )?

36 / 68
Sort Logic: Philosophical Implications

Let ZFCSL be ZFC formulated in the language with the new primitive
(with full schemata for replacement and specification) and MTSort be
an axiomatization of Sort (such as in Väänänen 2014: 178).

Theorem
ZFCSL + MTSort ` Con(ZFC).

In fact, we have something stronger:

Theorem
ZFCSL + MTSort is mutually interpretable with ZFCT + CT,

where ZFCT is ZFC in a language with a satisfaction predicate and CT

is the axiomatic compositional theory of truth.

37 / 68
The Alternative: ∃U

Before assessing further the philosophical consequences of adopting

sort logic as a utility logic, I’ll present an alternative due to Timothy
Williamson (1999). The governing idea is that the a quantifier ∃U is
introduced whose range over the universal domain is held fixed,
regardless of the interpretation:
The use of domains, whether sets, classes or properties, assumes
that what individuals the truth-value of a quantified sentence de-
pends on is not a logical matter; it varies between interpretations.
Rather than allowing generalizations in the object-language to de-
pend for their truth-values on the same individuals as generaliza-
tions in the meta-language, [the quantifier axioms] explicitly re-
strict them with the qualification ‘∈ dom(I )’ . But then the meta-
language has the resources to define the unrestricted analogues of
those quantifiers (identified by the subscript ‘U’), simply by drop-
ping that qualification [. . . ]. (Williamson 1999: 326)

38 / 68
The Alternative: ∃U

Like ∃˜ , ∃U is what we might term a breakout operator. A breakout

operator has the following property: in order to determine whether a
sentence involving the operator holds on a structure S , we have to
look at some (set- or class-sized) S 0 * S . With ∃˜ , this S 0 is a
structure extending S with new sorts. With ∃U , it’s the entire
universe W.
Breakout operators are not entirely unexampled elsewhere in logic.
Consider the ‘fixedly’ operator in modal logic on ‘pointed’ models
(ones with a designated actual world). Where M = hW, w@ , I i,
M  Fφ just in case, for all M 0 = hW, w, I i, M 0  φ.

39 / 68
First-Order ∃U

In his (1999), Williamson considers first-order ∃U : let’s call first-order

logic with this quantifier L1,1U . How much expressive power does
this add? Not much. With a first-order ∃U , we can only say that W
has a certain finite size or (using infinitely many sentences) that it is
infinite; since the fact that W is infinite is true simpliciter, and thus
true in every model, we can specify no new model classes.
Furthermore, let 1,1U be the L1,1U model-theoretic consequence
relation (as defined in a second-order meta-model theory) and let
`1,1U be the result of supplementing any standard first-order
deductive system with rules and axioms for ∃U exactly matching
those for ∃, the schema ∃v φ → ∃U v φ, and the axiom ∃U ≥n x x = x for
each n ∈ ω.
Theorem (Rayo and Williamson 2003)
ZFCU2 `2 (Γ 1,1U φ ↔ Γ `1,1U φ).

40 / 68
Higher-Order ∃U

What about higher-order versions of the ∃U quantifier? Let’s use

Lm,nU for the result of supplementing Lm with the ∃U quantifier up to
order n.
Here, however, issues about the meta-model theory and the home
logic it is couched in come to the fore. Rayo and Williamson (2003)
work in ZFCU2 to get their metatheorem about the completeness of
`1,1U for 1,1U , but it would have been possible to get the same result
in ZFCUT + CT.
For higher-order ∃U , we need to know what ‘everything’ means in a
higher-order context. We cannot say that second-order ∃U ranges
over the power set of the domain, since W is not a set.

41 / 68
Higher-Order ∃U

The natural approach is to work in a higher-order meta-model theory

and take the values of the nth-order entities in the truth conditions
for ∃U to be nth-order entities in the sense of the home logic.
We could do this for Lm,nU in ZFCUmax({m,n}) + CT, but it’s more
natural to follow Rayo and Williamson’s lead (and that of Linnebo
and Rayo 2012 and Button and Walsh 2018) and move up a level,
directly defining the m,nU truth-in-a-model relation in
ZFCUmax({m,n})+1 .

42 / 68
Higher-Order ∃U

But this introduces a subtlety: it’s very natural to treat the ordinary ∃
as a restriction of ∃U , but if we want the results of doing this to match
up with the treatment of higher-order quantification in ordinary Ln in
terms of power sets, we need to worry about how much
extensionality we have in the home logic once n ≥ 3.
Things get messy very quickly, so for the purposes of this talk I’m
going to assume we have full extensionality in the home logic: our
meta-model theories will be extensions of ZFCUn≡ .

43 / 68
Higher-Order ∃U

This prevents inequivalence between power-set-based and

higher-order definitions of truth-in-a-model for the ordinary ∃, and it
prevents both ∃ and ∃U from getting access to information about the
potentially nonextensional higher-order structure of the universe and
using it. To make the notion of cardinal characterizability (and more
generally, model class characterizability) for Lm,nU tractable, we need
two more assumptions in the meta-model theory.
The universal injection axiom (UIA) is the second-order axiom stating
that ∃f : W −→inj V (i.e., the universe is injectable into the pure sets).
The Urelement definability axiom (UDA) is the second-order axiom
stating that ∃f : V −→inj W ∨ (∃x ∈ V ∃f : U −→bij [λz  z ∈ x] ∧ x is
ZFC-definable). For n ≥ 2, let ZFCXn≡ be ZFCUn≡ + UIA + UDA.

44 / 68
ZFCXn≡

We adopt the ZFCXn≡ family (or ZFCX< ω

≡ ) as our meta-model theories.
In this framework, we have the following nice simplifying result:

Theorem
ZFCXn≡+1 ` ∀x ∈ V (x is ZFCn -definable ↔ x is ZFCUn -definable).

45 / 68
L2,2U : An Example
How would we characterize the model class we discussed earlier
({hS, ≺i : hS, ≺i is order-embeddable into hR, <i}) in L2,2U ? Again,
let A be the conjunction of the axioms for a linear order in the
signature {≺} and let B be the conjunction of the second-order
axioms for the real ordered field in the signature {⊕, ⊗, <}: this time
there is no distinction of sorts A and B. Again, let B∗ be
B[R1/⊕][R2/⊗][R3/<]. Define φ† as the result of replacing ∃ with ∃U and
∀ with ∀U in φ. Our sentence is:

A ∧ ∃U Q∃U R1 ∃U R2 ∃U R3 ((B∗ )† ∧
∀x∃U !y (Qxy ∧ ∀z(Qzy → x = z))∧
∀x1 ∀x2 ∀U y1 ∀U y2 (x1 < x2 ∧ Qx1 y1 ∧ Qx2 y2 → R3 y1 y2 )),

i.e.

A ∧ ∃U f ∃U R1 ∃U R2 ∃U R3 (B∗ ∧ f : h[λx  x = x], <i ,→ [hλU y  y = y], R3 i).

46 / 68
Ln,nU : Metatheory

As we can see, sort logic and higher-order ∃U are practically very

similar. But the unrestricted second-order quantifier means that L2,2U
can exploit the full power of the second-order quantifier over W.
More generally, we have the following theorem for the Ln,nU case:

Theorem
For n ≥ 2, ZFCXn≡+1 ` for every model class C , C = Mod(φ) for some
single sentence φ of Ln,nU if and only if C is is ZFCn -definable.

(Lm,nU with m 6= n requires consideration of a few different cases. In

some circumstances, Ln,2U does as well as Ln,nU .)

47 / 68
Ln,nU : Metatheory
For those who are interested, this is a sketch of how the left-to-right
direction goes for the simplest case (characterizing a cardinal κ). Let κ
be defined by some ZFCn -formula φ. Let ZFC2 be the conjunction of
the ZFC2 axioms, and let φ(X,R) be the result of replacing ∈ by R in φ
and relativing all quantifiers to a one-place predicate X. Define φ† as
the result of replacing ∃ with ∃U and ∀ with ∀U in φ. Define φ∗ as the
following L n,nU [∅] sentence:

0 0
∃U V, ε(ZFC2(V,ε)† ∧ ∀U V 0 , ε0 (ZFC2(V ,ε )† → ∃U f : V 0 −→inj† V )∧
†
∃U X(∀U y(Xy ↔ φ(V,ε) (y) ∧ ∃Z(∀xZx ∧ ∃U f : X −→bij† Z))).

(Note that −→inj† /bij† is to be formalized in higher-order terms using

∃U /∀U .)
Theorem
ZFCXn≡+1 ` ∀S (S n,nU φ∗ ↔ |S | = κ ).
48 / 68
Metatheory and Recapture

Note that the result proved in ZFCXn≡+1 about the expressive power of
Ln,nU does not have a recapture in first-order set theory that preserves
its intuitive content. We can of course show that
Theorem
ZFC(U) ` ∀S (S  ZFCXn≡+1 → S  “∀C (C is a model class →
(∃φ ∈ Ln,nU C = Mod(φ) ↔ C is ZFCn -definable”)),

(or a version guaranteed to be nonvacuous in ZFC(U)+ large

cardinals) but this doesn’t get the message across: it only tells us how
set-sized ZFCXn≡+1 models without access to a true breakout ∃U
believe that the ersatz ∃U they can define behaves, not how the real
∃U behaves.

49 / 68
Two Pictures

Let’s move back to larger issues. I’ve suggested that, if we take

(S1)–(S4) seriously and accept large cardinals, we can find ourselves
needing to strengthen both our utility logics and our meta-model
theory. There are a number of options here, but these seem to be the
two best-motivated packages:
Sort-logic programme:
Utility logics: L1 , Ln for finite n, LSort .
Home logic: L1
Meta-model theory: ZFCUSL + MTSort
∃U programme:
Utility logics: L1 , Ln for finite n, Lm,nU for finite m, n
Home logic: L<ω
Meta-model theory: ZFCU<ω or some extension thereof such as
ZFCX<ω

50 / 68
Two Pictures

Question: why, on the ∃U programme, is the home logic still L<ω and
not something like L<ω,<ωU (or at least each Ln,nU separately)? If
we’re going to endorse ∃U and give it higher-order truth-conditions,
why not take it back into the home logic as well?
Answer: we already have! The quantifiers of our home logic were
absolutely unrestricted all along. In the home logic, ∃ just is ∃U : the
point of the new quantifier was just to vouchsafe some of the
expressive power of the home logic to the utility logics. (For similar
reasons, the proponent of the sort logic programme would gain no
expressive power by adopting LSort as a home logic—if it’s even clear
what that would mean.)

51 / 68
Two Pictures: Universism and Pluralism

Which of these two pictures should one favour? Let’s start with how
pluralists and universists should approach this decision. The natural
background understanding of quantification in the ∃U programme is
generality-absolutist; although this doesn’t entail universism, it’s
certainly a more natural fit. Nonetheless some kinds of moderate
pluralists can perhaps accept ∃U on the interpretation that it
quantifies over everything ‘at the current stage’: with potentialist
modal operators one can get appropriate modal translations of the
metatheorems. This is easiest to motivate on a potentialist logic that
includes S4.3, although perhaps one could justify it in an S4.2
setting.4

4 See Hamkins and Linnebo (∞) for potentialist modal logics.

52 / 68
Two Pictures: Universism and Pluralism

Nonetheless, in general pluralists will probably find the sort-logic

programme a more comfortable fit. The breakout quantifier ∃˜
‘reaches out’ only to a set-sized chunk of the universe outside the
current model at any one time, unlike ∃U . (Although note that we
need sentences with universal or negated existential sort quantifiers
as well to make use of the full power of sort logic.)

53 / 68
Two Pictures: The Status of the Home Logic

But there is another issue that is orthogonal to the universism/

pluralism debate. The proponent of sort logic can remain within a
first-order home logic and first-order ZFC (extended with a new
predicate) as a meta-model theory. The ∃U programme requires at
least absolutely unrestricted second-order quantification in the home
logic, and it fits best with quantification at all finite orders. There are
a number of ways to understand this resource—pluralist (which
validates the simplifying extensionality assumptions), primitivist,
and perhaps others. But someone sceptical of higher-order logic in
the home logic may not be happy with these moves.

54 / 68
Two Pictures: The Status of the Home Logic

Of course, there’s no time to canvass the debate about the status of

higher-order quantifiers (understood as properly basic and not
defined in a model-theoretic logic) here. But there are reasons to think
that Väänänen, at least, is suspicious of non-L1 resources that aren’t
ultimately confined to model-theoretically specified utility logics.
We have used set theory to analyze second order characterizable
structures and found set-theoretical concepts such as the cardinal-
ity of the model and the Levy-hierarchy useful tools. One can re-
define these tools in the second order setting, but one would have
to rely heavily on the concept of universal truth, and having done
that, there is a temptation to see the whole of second order logic as
the ∆ 2 -fragment of a more powerful framework, namely set theory.
(Väänänen 2012: 116)

55 / 68
Two Pictures: The Status of the Home Logic

Whatever is proved from ZFC can be turned into an informal ar-

gument in second-order logic, e.g., in ZFC2 . On the other hand, if
something can be informally argued in ZFC2 , it seems very likely,
that behind the informal argument is an argument that can be for-
malized in ZFC. In this respect there is very little difference be-
tween first-order set theory and second-order logic. [ . . . ] On the
other hand, if we try to analyse why we are not able to decide e.g.,
Continuum Hypothesis, on the basis of ZFC2 , it seems very plau-
sible to develop a theory about what the second-order quantifiers
range over. The first-order set theory ZFC is exactly such a theory,
and it is indeed the strongest currently available tool for investigat-
ing formalizations of second order logic. (Väänänen 2001: 519)

56 / 68
Two Pictures: The Status of the Home Logic

It seems as though there is a divide—although one whose content is

hard to make precise—between Väänänen’s picture and that of what
might be called the logic-first approach, in which one assesses
candidate home logics and subject-matter-specific theories in them
separately. It is perhaps significant that Väänänen doesn’t stress the
distinction between what I am terming home logic and what I am
terming meta-model theory: he uses the term ‘urlogic’ for ZFC itself,
not L1 (2001: 512–14).

57 / 68
Two Pictures: The Status of the Home Logic

A similar, although perhaps more radical view, can be found in

Kennedy’s recent work on formalism-freeness:
Recent debates concerning the comparative virtues of second order
logic vs. set theory, for example, depart from Tarski’s explicit cos-
mopolitanism in decrying the entanglement of set theory with sec-
ond order logic—insofar as it is admitted to exist at all. Whereas
from the symbiosis point of view, we will see (indeed prove) that
second order logic is actually symbiotic with the power set opera-
tion. Or to put it another way: it is useless to try to separate second
order logic from set theory. (Kennedy 2021: 149)

58 / 68
Conclusion

Here the issues are very complex; there are a great many further areas
for research and a number of loose ends left unresolved.
I shall conclude simply by suggesting that, to the extent that
consideration of very strong logics and their relationship to large
cardinal axioms can help make precise what is at stake in debates
over logic adoption, the metaphysical commitments of various logics,
and ‘formalism-freeness’ and related ideas, they have an important
role to play in the philosophy of set theory and logic more generally.

59 / 68
Works Cited

Adámek, Jiří, and Jiří Rosicky (1994). Locally Presentable and

Accessible Categories. Cambridge: Cambridge University Press.
Bagaria, Joan, and Jouko Väänänen (2016). “On the Symbiosis
between Model-Theoretic and Set-Theoretic Properties of Large
Cardinals.” Journal of Symbolic Logic 81: 584–604.
Barwise, K. Jon (1974). “Axioms for Abstract Model Theory.”
Annals of Mathematical Logic 7: 221–65.
—————— (1985). “Model-Theoretic Logics: Background and
Aims.” In K. Jon Barwise and Solomon Feferman, eds.,
Model-Theoretic Logics (New York: Springer), 3–25.
Baldwin, John T. (2018). Model Theory and the Philosophy of
Mathematical Practice. Cambridge: Cambridge University Press.
Burgess, John (2015). Rigor and Structure. Oxford: Oxford
University Press.

60 / 68
Works Cited

Button, Tim, and Sean Walsh (2018). Philosophy and Model Theory.
Oxford: Oxford University Press.
Feferman, Solomon (1974). “Applications of Many-Sorted
Interpolation Theorems.” In Leon Henkin et al., eds., Proceedings
of the Tarski Symposium (Providence, R.I.: American Mathematical
Society), 205–24.
Garland, Stephen J. (1974). “Second-Order Cardinal
Characterizability.” In Dana S. Scott and Thomas Jech, eds.,
Axiomatic Set Theory (2 vols.; Providence, R.I.: American
Mathematical Society, 1971–74), 2.127–46.
Hamkins, Joel, and Øystein Linnebo (∞). “The Modal Logic of
Set-Theoretic Potentialism and Potentialist Maximality
Principles.” Forthcoming, Review of Symbolic Logic.

61 / 68
Works Cited

Hanf, William, and Dana Scott (1961). “Classifying Inaccessible

Cardinals.” Notices of the American Mathematical Society 8: 445.
Hjortland, Ole Thomassen (2017). “Anti-Exceptionalism about
Logic.” Philosophical Studies 174: 631–58.
Isaacson, Daniel (2011). “The Reality of Mathematics and the
Case of Set Theory.” In Zsolt Novák and András Simonyi, eds.,
Truth, Reference, and Realism (Budapest: Central European
University Press), 1–76.
Kennedy, Juliette (2013). “On Formalism Freeness:
Implementing Gödel’s 1946 Princeton Bicentennial Lecture.”
Bulletin of Symbolic Logic 19: 351–93.

62 / 68
Works Cited

—————— (2015) “On the Logic without Borders Point of

View.” In Åsa Hirvonen, Juha Kontinen, Roman Kossak, and
Andrés Villaveces., eds., Logic without Borders (Berlin: De
Gruyter), 1–13.
—————— (2021). Gödel, Tarski, and the Lure of Natural
Language: Logical Entanglement, Formalism Freeness. Cambridge:
Cambridge University Press.
——————, Menachem Magidor, and Jouko Väänänen (∞).
“Inner Models from Extended Logics: Part I.” Forthcoming,
Journal of Mathematical Logic.
—————— and Jouko Väänänen (2017). “Squeezing
Arguments and Strong Logics.” In Hannes Leitgeb, Ilkka
Niiniluoto, Päivi Seppälä, and Elliot Sober, eds., Logic,
Methodology, and Philosophy of Science: Proceedings of the Fifteenth
International Conference (London: College Publications), 61–81.

63 / 68
Works Cited

——————, —————— (∞) “Logicality and Model Classes.”

Typescript.
Kreisel, Georg (1967). “Informal Rigour and Consistency
Proofs.” In Irme Lakatos, ed., Problems in the Philosophy of
Mathematics (Amsterdam: North-Holland), 138–71.
Kunen, Kenneth (1971). “Elementary Embeddings and Infinitary
Combinatorics.” Journal of Symbolic Logic 36: 407–13.
Linnebo, Øystein, and Agustín Rayo (2012). “Hierarchies
Ontological and Ideological.” Mind 121: 269–308.
Magidor, Menachem (1971). “On the Role of Supercompact and
Extendible Cardinals in Logic.” Israel Journal of Mathematics 10:
147–157.

64 / 68
Works Cited

—————— and Jouko Väänänen (2011). “On

Löwenheim-Skolem-Tarski Numbers for Extensions of
First-Order Logic.” Journal of Mathematical Logic 11: 87–113.
Rayo, Agustín, and Timothy Williamson (2003). “A
Completeness Theorem for Unrestricted First-Order Languages.”
In In Jc Beall, ed., Liars and Heaps (Oxford: Oxford University
Press), 331–55.
Roberts, Sam (2020). “Classless.” Analysis 80: 76–83.
Shapiro, Stewart (1991). Foundations without Foundationalism: A
Case for Second-Order Logic. Oxford: Oxford University Press.
—————— (1997). Philosophy of Mathematics: Structure and
Ontology. Oxford: Oxford University Press.

65 / 68
Works Cited

Steel, John R. (1996). The Core Model Iterability Problem. Berlin:

Springer Verlag.
Suzuki, Akira (1999). “No Elementary Embedding from V into V
is Definable from Parameters.” Journal of Symbolic Logic 64:
1591–94.
Väänänen, Jouko (1979). “Abstract Logic and Set Theory I:
Definability.” In Maurice Boffa, Dirk van Dalen, and Kenneth
McAloon, eds., Logic Colloquium ’78: Proceedings of the Colloquium
Held in Mons, August 1978 (Amsterdam: North-Holland),
391–421.
—————— (1980). “Results on Sort Logic (Abstract).” Journal
of Symbolic Logic 45: 410.
—————— (1985). “Set-Theoretic Definability of Logics.” In K.
Jon Barwise and Solomon Feferman, eds., Model-Theoretic Logics
(New York: Springer), 599–644.

66 / 68
Works Cited

—————— (2001). “Second-Order Logic and Foundations of

Mathematics.” Bulletin of Symbolic Logic 7: 504–20.
—————— (2012). “Second-Order Logic or Set Theory?”
Bulletin of Symbolic Logic 18: 91–121.
—————— (2014). “Sort Logic and Foundations of
Mathematics.” In Chitat Chong, Qi Feng, Theodore A. Slaman,
and W. Hugh Woodin, eds., Infinity and Truth (Singapore: World
Scientific), 171–86.
—————— (2019). “Second-Order and Higher-Order Logic.”
Stanford Encyclopedia of Philosophy.
Williamson, Timothy (1999). “Existence and Contingency.”
Proceedings of the Aristotelian Society (Supplement) 73: 181–203;
corr. rpt. in Proceedings of the Aristotelian Society 100: 321–43.

67 / 68
Works Cited

—————— (2003). “Everything.” Philosophical Perspectives 17:

415–65.
—————— (2014). “Logic, Metalogic, and Neutrality.”
Erkenntnis 79: 211–31.
—————— (2017). “Semantic Paradoxes and Abductive
Methodology.” In Bradley Armour-Garb, ed., Reflections on the
Liar (Oxford: Oxford University Press), 325–46.
Woods, Jack (2019). “Logical Partisanhood.” Philosophical Studies
176: 1203–24.

68 / 68