Professional Documents
Culture Documents
Theodore Sider
December 4, 2007
Preface
i
PREFACE ii
Another source was Ed Gettier’s 1988 modal logic class at the University of
Massachusetts.
I am also deeply grateful for feedback from colleagues, and from students
in courses on this material. In particular, Marcello Antosh, Josh Armstrong,
Gabe Greenberg, Angela Harper, Sami Laine, Gregory Lavers, Alex Morgan,
Jeff Russell, Brock Sides, Jason Turner, Crystal Tychonievich, Jennifer Wang,
Brian Weatherson, and Evan Williams: thank you.
Contents
Preface i
1 Nature of Logic 1
1.1 Logical consequence and logical truth . . . . . . . . . . . . . . . . . 2
1.2 Form and abstraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Formal logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Correctness and application . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 The nature of logical consequence . . . . . . . . . . . . . . . . . . . 8
1.6 Extensions, deviations, variations . . . . . . . . . . . . . . . . . . . . 10
1.6.1 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6.2 Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6.3 Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.7 Metalogic, metalanguages, and formalization . . . . . . . . . . . . 12
1.8 Set theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Propositional Logic 18
2.1 Grammar of PL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 The semantic approach to logic . . . . . . . . . . . . . . . . . . . . . 21
2.3 Semantics of PL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Natural deduction in PL . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 Sequents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.2 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.3 Sequent proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.4 Example sequent proofs . . . . . . . . . . . . . . . . . . . . . 35
2.5 Axiomatic proofs in PL . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.1 Example axiomatic proofs . . . . . . . . . . . . . . . . . . . . 42
2.5.2 The deduction theorem . . . . . . . . . . . . . . . . . . . . . 47
2.6 Soundness and completeness of PL . . . . . . . . . . . . . . . . . . 47
iii
CONTENTS iv
4 Predicate Logic 71
4.1 Grammar of predicate logic . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Semantics of predicate logic . . . . . . . . . . . . . . . . . . . . . . . 72
4.3 Establishing validity and invalidity . . . . . . . . . . . . . . . . . . . 77
8 Counterfactuals 190
8.1 Natural language counterfactuals . . . . . . . . . . . . . . . . . . . . 191
8.1.1 Not truth-functional . . . . . . . . . . . . . . . . . . . . . . . 191
8.1.2 Can be contingent . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.1.3 No augmentation . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.1.4 No contraposition . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.1.5 Some implications . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.1.6 Context dependence . . . . . . . . . . . . . . . . . . . . . . . 194
8.2 The Lewis/Stalnaker approach . . . . . . . . . . . . . . . . . . . . . 196
8.3 Stalnaker’s system (SC) . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.3.1 Syntax of SC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.3.2 Semantics of SC . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.4 Validity proofs in SC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
8.5 Countermodels in SC . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
8.6 Logical Features of SC . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.6.1 Not truth-functional . . . . . . . . . . . . . . . . . . . . . . . 211
8.6.2 Can be contingent . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.6.3 No augmentation . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.6.4 No contraposition . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.6.5 Some implications . . . . . . . . . . . . . . . . . . . . . . . . . 212
8.6.6 No exportation . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
8.6.7 No importation . . . . . . . . . . . . . . . . . . . . . . . . . . 214
8.6.8 No hypothetical syllogism (transitivity) . . . . . . . . . . . 214
8.6.9 No transposition . . . . . . . . . . . . . . . . . . . . . . . . . . 215
8.7 Lewis’s criticisms of Stalnaker’s theory . . . . . . . . . . . . . . . . 216
8.8 Lewis’s system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
8.9 The problem of disjunctive antecedents . . . . . . . . . . . . . . . 222
References 272
Chapter 1
Nature of Logic
ince you are reading this book, you are probably already familiar with
S some logic. You probably know how to translate English sentences into
symbolic notation—into propositional logic:
You are probably also familiar with some basic techniques for evaluating argu-
ments written out in symbolic notation. You have probably encountered truth
tables, and some form of proof theory (perhaps a “natural deduction” system;
perhaps “truth trees”.) You may have even encountered some elementary model
theory. In short: you had an introductory course in symbolic logic.
What you already have is: literacy in elementary logic. What you will
get out of this book is: literacy in the rest of logic that philosophers tend to
presuppose, plus a deeper grasp of what logic is all about.
So what is logic all about?
1
CHAPTER 1. NATURE OF LOGIC 2
Likewise, we say that “it’s not the case that snow is white and snow is not white”
is a logical truth because it has the form: it’s not the case that φ and not-φ.
We need to think hard about the idea of form. Apparently, we got the
alleged form of Argument A by replacing some words with Greek letters and
leaving other words as they were. We replaced the sentences ‘John is happy’ and
‘Ted is happy’ with φ and ψ, respectively, but left the expressions ‘It’s not the
case that’ and ‘or’ as they were, resulting in the schematic form displayed above.
Let’s call that form, “Form 1”. What’s so special about Form 1? Couldn’t we
make other choices for what to leave and what to replace? For instance, if we
replace the predicate ‘is happy’ with the schematic letter α, leaving the rest
intact, we get this:
And if we replace the ‘or’ with the schematic letter γ and leave the rest intact,
then we get this:
So, what did we mean, when we said that Argument A is logically correct in
virtue of its form? What is Argument A’s form? Is it Form 1, Form 2, or Form
3?
There is no such thing as the form of an argument. When we assign an
argument a form, what we are doing is focusing on certain words and ignoring
others. We leave intact the words we’re focusing on, and we insert schematic
letters for the rest. Thus, in assigning Argument A Form 1, we’re focusing on
the words (phrases) ‘it is not the case that’ and ‘or’, and ignoring other words.
More generally, in (standard) propositional logic, we focus on the phrases
‘if…then’, ‘if and only if’, ‘and’, ‘or’, and so on, and ignore others. We do this
in order to investigate the relations of logical consequence that hold in virtue
of these words’ meaning. The fact that Argument A is logically correct depends
just on the meaning of the phrases ‘it is not the case that’ and ‘or’; it does not
depend on the meanings of the sentences ‘John is happy’ and ‘Ted is happy’.
We can substitute any sentences we like for ‘φ’ and ‘ψ’ in Form 1 and still get
a valid argument.
In predicate logic, on the other hand, we focus on further words: ‘all’ and
‘some’. Broadening our focus in this way allows us to capture a wider range
of logical consequences and logical truths. For example “If Ted is happy then
someone is happy” is a logical truth in virtue of the meaning of ‘someone’, but
not merely in virtue of the meanings of the characteristic words of propositional
logic.
Call the words on which we’re focusing—that is, the words that we leave
intact when we construct the forms of sentences and arguments—the logical
constants. (We can speak of natural language logical constants—‘and’, ‘or’, etc.
for propositional logic; ‘all’ and ‘some’ in addition for predicate logic—as well
as symbolic logical constants: ∧, ∨, etc. for propositional logic; ∀ and ∃ in
addition for predicate logic.) What we’ve seen is that the forms we assign
depend on what we’re considering to be the logical constants.
We call these expressions logical constants because we interpret them in a
constant way in logic, in contrast to other terms. For example, ∧ is a logical
constant; in propositional logic, it always stands for conjunction. There are
fixed rules governing ∧, in proof systems (the rule that from P ∧Q one can
infer P , for example), in the rules for constructing truth tables, and so on.
Moreover, these rules are distinctive for ∧: there are different rules for other
logical constants such as ∨. In contrast, the terms in logic that are not logical
constants do not have fixed, particular rules governing their meanings. For
example, there are no special rules governing what one can do with a P as
CHAPTER 1. NATURE OF LOGIC 5
α is a bachelor
Therefore, α is unmarried
Accordingly, we could treat the predicates ‘is a bachelor’ and ‘is unmarried’
as logical constants, and develop a corresponding logic. We could introduce
special symbolic logical constants for these predicates, we could introduce
distinctive rules governing these predicates in proofs. (The rule of “bachelor-
elimination”, for instance, might allow one to infer “α is unmarried” from “α
is a bachelor”.) As with the choices of the previous paragraph, this choice of
what to treat as a logical constant is also not ruled out by the concept of form.
And it would be more productive than the choices of the last paragraph. Still,
it would be far less productive than the usual choices of logical constants in
predicate and propositional logic. The word ‘bachelor’ doesn’t have as general
application as the words commonly treated as logical constants in propositional
and predicate logic; the latter are ubiquitous.
At least, this remark about “generality” is one idea about what should be
considered a “logical constant”, and hence one idea about the scope of what
is usually thought of as “logic”. Where to draw the boundaries of logic—and
indeed, whether the logic/nonlogic boundary is an important one to draw—is
an open philosophical question about logic. At any rate, in this course, one
thing we’ll do is study systems that expand the list of logical constants from
standard propositional and predicate logic.
CHAPTER 1. NATURE OF LOGIC 6
P
(Q→R)∨(Q→∼S)
P ↔(P ∧Q)
The symbols ∧, ∨, etc., are used to represent the English words ‘and’, ‘or’, and
so on (the logical constants for propositional logic), and the sentence letters
P, Q, etc., are used to represent declarative English sentences.
Why ‘formal’? Because we stipulate, in a mathematically rigorous way,
a grammar for the language; that is, we stipulate a mathematically rigorous
definition of the idea of a sentence of this language. Moreover, since we are
only interested in the logical behavior of the chosen logical constants ‘and’,
‘or’, and so on, we choose special symbols (∧, ∨ . . . ) for these words only; we
use P, Q, R, . . . indifferently to represent any English sentence whose internal
logical structure we are willing to ignore.3
We go on, then, to study (as always, in a mathematically rigorous way) vari-
ous concepts that apply to the sentences in formal languages. In propositional
2
Notes
3
Natural languages like English also have a grammar, and the grammar can be studied
using mathematical techniques. But the grammar is much more complicated, and is discovered
rather than stipulated; and natural languages lack abstractions like the sentence letters.
CHAPTER 1. NATURE OF LOGIC 7
some other English sentences in virtue of ‘and’, ‘or’, etc., iff one can prove the
translation of the former English sentence from the translations of the latter
English sentences in the formal system.
In this book I won’t spend much time on philosophical questions about
which formal systems are correct. My goal is rather to introduce those for-
malisms that are ubiquitous in philosophy, to give you the tools you need to
address such philosophical questions yourself. Still, from time to time, we’ll dip
just a bit into these philosophical questions, in order to motivate our choices
of logical systems to study.
necessarily
it will be the case that
most
it is morally wrong that
1.6.1 Extensions
Here we add to standard logic. We add both:
• new symbols
• new cases of logical consequence and logical truth that we can
model
1.6.2 Deviations
Here we change, rather than add. We retain the same symbols from standard
logic, but we alter standard logic’s proof theory and semantics. We therefore
change what we say about the logical consequences and logical truths that
involve the symbols.
Why do this? Perhaps because we think that standard logicians are wrong
about what the right logic for English is. If we want to correctly model logical
consequence in English, therefore, we must construct systems that behave
differently from standard logic.
For example, in the standard semantics for propositional logic, every for-
mula is either true or false. But some have argued that natural language sen-
tences like the following are neither true nor false:
The king of the United States is bald
Sherlock Holmes weighs more than 178 pounds
Bill Clinton is tall
There will be a sea battle tomorrow
If so, perhaps we should abandon the standard semantics for propositional logic
in favor of multi-valued logic, in which formulas are allowed to be neither true
nor false.
1.6.3 Variations
Here we also change standard logic, but we change the notation without
changing the content of logic. We study alternate ways of expressing the same
thing.
CHAPTER 1. NATURE OF LOGIC 12
∼(P ∧Q)
∼P ∨∼Q
are two different ways of saying the same thing. We will study other ways of
saying what those two sentences say, including:
P |Q
∼∧P Q
In the first case, | is a new symbol for “not both”. In the second case (“Polish
notation”), the ∼ and the ∧ mean what they mean in standard logic; but instead
of going between the P and the Q, the ∧ goes before P and Q. The value of
this, as we’ll see, is that we no longer will need parentheses.
These are really interesting claims! They show that the method of truth tables
and the method of constructing derivations amount to the same thing, as
applied to symbolic formulas of propositional logic. One can establish similar
results for standard predicate logic.
A couple remarks about proving things in metalogic.
First: what do we mean by “proving”? We do not mean: constructing a
derivation in the logical system we’re investigating. We’re trying to construct a
proof about the system. We do this in English, and we do it with informal (though
rigorous!) reasoning of the sort one would encounter in a mathematics book.
Logicians often distinguish the “object language” from the “metalanguage”.
The object language is the language that’s being studied—the language of
propositional logic, for example. Sentences of this object language look like
this:
P ∧Q
∼(P ∨Q)↔R
The metalanguage is the language we use to talk about the object language.
In the case of the present book, the metalanguage is English. Here are some
example sentences of the metalanguage:
those sets that are not members of themselves. For short, R is the set of non-
self-members. Russell asks the following question: is R a member of itself?
There are two possibilities:
/ R. Thus, R is a non-self-member. But R was said to
i) R ∈
be the set of all non-self-members, and so we’d have R ∈ R.
Contradiction.
ii) R ∈ R. So R is not a non-self-member. R, by definition,
/ R. Contradiction.
contains only non-self-members. So R ∈
Thus, each possibility leads to a contradiction. But there are no remaining
possibilities—either R is a member of itself or it isn’t! So it looks like the very
idea of sets is paradoxical.
The modern discipline of axiomatic set theory arose in part to develop a
notion of sets that isn’t subject to this sort of paradox. This is done by imposing
rigid restrictions on when a given “condition” picks out a set. In the example
above, the condition “is a non-self-member” will be ruled out—there’s no set
of all and only the things satisfying this condition. The details of set theory are
beyond the scope of this course; for our purposes, we’ll help ourselves to the
existence of sets, and not worry about exactly what sets are, or how the Russell
paradox is avoided.
Various other useful set-theoretic notions can be defined in terms of the
notion of membership. We say that A is a subset of B (“A ⊆ B”) when every
member of A is a member of B. We say that the intersection of A and B (“A∩B”)
is the set that contains all and only those things that are in both A and B, and
that the union of A and B (“A∪B”) is the set containing all and only those things
that are members of either A or B.
Suppose we want to refer to the set of the so-and-sos—that is, the set
containing all and only objects, u, that satisfy the condition “so-and-so”. We’ll
do this with the term “{u: u is a so-and-so}”. Thus, we could write: “N = {u :
u is a natural number}”. And we could restate the definitions of ∩ and ∪ from
the previous paragraph as follows:
A ∩ B =df {u : u ∈ A and u ∈ B}
A ∪ B =df {u : u ∈ A or u ∈ B}
Sets have members, but they don’t contain them in any particular order. For
example, the set containing me and Bill Clinton doesn’t have a “first” member.
This is reflected in the fact that “{Ted, Clinton}” and “{Clinton, Ted}” are
CHAPTER 1. NATURE OF LOGIC 16
two different names for the same set—the set containing just Clinton and Ted.
But sometimes we need to talk about a set-like thing containing Clinton and
Ted, but in a certain order. For this purpose, logicians use ordered sets. Two-
membered ordered sets are called ordered pairs. To name the ordered pair of
Clinton and Ted, we use: “〈Clinton, Ted〉”. Here, the order is significant, for
〈Clinton, Ted〉 and 〈Ted, Clinton〉 are not the same thing. The three-membered
ordered set of u, v, and w (in that order) is written: 〈u, v, w〉; and similarly for
ordered sets of any finite size. A n-membered ordered set is called an n-tuple.
Let’s even allow 1-tuples: let’s define the 1-tuple 〈u〉 as being the object u itself.
In addition to sets, and ordered sets, we’ll need a further related concept:
that of a function. A function is a rule that “takes in” an object or objects,
and “spits out” a further object. For example, the addition function is a rule
that takes in two numbers, and spits out their sum. As with sets and ordered
sets, functions are not limited to mathematical entities: they can “take in” and
“spit out” any objects whatsoever. We can speak of the father-of function, for
example, which is a rule that takes in a person, and spits out the father of that
person. And later in this book we will be considering functions that take in
and spit out linguistic entities: sentences and parts of sentences from formal
languages.
Each function has a fixed number of “places”: a fixed number of objects
it must take in before it is ready to spit out something. You need to give
the addition function two arguments (numbers) in order to get it to spit out
something, so it is called a two-place function. You only need to give the father-
of function one object, on the other hand, to get it to spit out something, so it
is a one-place function.
The objects that the function takes in are called its arguments, and the object
it spits out is called its value. Suppose f is an n-place function, and u1 . . . un are
n of its arguments; one then writes “ f (u1 . . . un )” for the value of function f as
applied to arguments u1 . . . un . f (u1 . . . un ) is the object that f spits out, if you
feed it u1 . . . un . For example, where f is the father-of function, since Ron is
my father, we can write: f (Ted) = Ron; and, where a is the addition function,
we can write: a(2, 3) = 5.
There’s a trick for “reducing” talk of both ordered pairs and functions to
talk of sets. One first defines 〈u, v〉 as the set {{u}, {u, v}}; one defines 〈u, v, w〉
as 〈u, 〈v, w〉〉, and similarly for n-membered ordered sets, for each positive
integer n. And, finally, one defines an n-place function as a set, f , of n + 1-
tuples obeying the constraint that if 〈u1 , . . . , un , v〉 and 〈u1 , . . . , un , w〉 are both
members of f , then v = w; f (u1 , . . . , un ) is then defined as the object, v, such
CHAPTER 1. NATURE OF LOGIC 17
that 〈u1 , . . . , un , v〉 ∈ f . Thus, ordered sets and functions are defined as certain
sorts of sets. The trick of the definition of ordered pairs is that we put the
set together in such a way that we can look at the set and tell what the first
member of the ordered pair is: it’s the one that “appears twice”. Similarly, the
trick of the definition of a function is that we can take any arguments to the
function, look at the set that is identified with the function, and figure out
what value the function spits out for those arguments. But the technicalities of
these reductions won’t matter for us; I’ll just feel free to speak of ordered pairs,
triples, functions, etc., without defining them as sets.
Chapter 2
Propositional Logic
2.1 Grammar of PL
Modern logic has made great strides by treating the language of logic as a
mathematical object. To do so, grammar needs to be developed rigorously.
(Our study of a new logical system will always begin with grammar.)
If all you want to do is understand the language of logic informally, and be
able to use it effectively, you don’t really need to get so careful about grammar.
For even if you haven’t ever seen the grammar of propositional logic formalized,
you can recognize that things like this make sense:
P →Q
R∧(∼S↔P )
→P QR∼
(P ∼Q∼(∨
But to make any headway in metalogic, we will need more than an intuitive
understanding of what makes sense and what does not; we will need a precise
definition that has the consequence that only the strings of symbols in the first
group “make sense”.
18
CHAPTER 2. PROPOSITIONAL LOGIC 19
Grammatical formulas (i.e., ones that “make sense”) are called well-formed
formulas, or “wffs” for short. We define these by first carefully defining exactly
which symbols are allowed to occur in wffs (the “primitive vocabulary”), and
second, carefully defining exactly which strings of these symbols count as wffs:
Primitive vocabulary:
Definition of wff:
2.3 Semantics of PL
Our semantics for propositional logic is really just truth tables, only presented
a little more carefully than in introductory logic books. Let’s define the notion
of a propositional logic interpretation function, or PL-interpretation:
φ → ψ
∼ φ
1 1 1
1 0
1 0 0
0 1
0 0 1
0 0 0
I’ll show that the first statement is true here; the others are exercises for the
reader. I’ll write out my proof in excessive detail, to make it clear exactly how
the reasoning works:
Let’s reflect on what we’ve done so far. We have defined the notion of
a PL-interpretation, which assigns 1s and 0s to sentence letters of the sym-
bolic language of propositional logic. And we have also defined, for any PL-
interpretation, a corresponding valuation function, which extends the inter-
pretation’s assignment of 1s and 0s to complex wffs of PL. Note that we have
been informally speaking of these assignments as assignments of truth values.
That’s because the assignments of 1s and 0s accurately models the truth values
of statements in English that are represented in the obvious way by PL-wffs.
1
The careful reader will note that here (and henceforth), I treat “VI (α)=0” and “VI (α) is
not 1” interchangeably (for any wff α). (Similarly for “VI (α)=1” and “VI (α) is not 0”.) This is
justified as follows. First, if VI (α) is 0, then it can’t also be that VI (α) is 1—VI was stipulated
to be a function. Second, since it was stipulated that VI assigns either 0 or 1 to each wff, if
VI (α) is not 1, then VI (α) must be 0.
CHAPTER 2. PROPOSITIONAL LOGIC 25
A parenthetical remark: now we can see the importance for setting up the
grammar for our system according to precise rules. If we hadn’t, the definition
of ‘truth value’ given here would have been impossible. In this definition we
defined truth values of complicated formulas based on their form. For example,
if a formula has the form (φ→ψ), then we assigned it an appropriate truth value
based on the truth values of φ and ψ. But suppose we had a formula in our
language that looked as follows:
P →P →P
and suppose that P has truth value 0. What is the truth value of the whole? We
can’t tell, because of the missing parentheses. For if the parentheses look like
this:
(P →P )→P
then the truth value is 0, whereas if the parentheses look like this:
P →(P →P )
the truth values of sentence letters not contained in a formula do not affect
the truth value of the formula. We need the latter assumption to be sure that
we only need to check a finite number of cases—namely, the permutations
of truth values of the contained sentence letters—to see whether a formula is
valid. After all, there are infinitely many interpretation functions (since there
are infinitely many sentence letters in the language of PL), and a valid formula
must be true in each one.
These two assumptions are obviously true, but it would be good to prove
them. I’ll prove the first assumption here (the second may be proved by a
similar method), and take this opportunity to introduce an important technique
for metalanguage proofs: proof by induction.
Proof that every wff contains a finite number of sentence letters: In this
sort of proof by induction, we’re trying to prove a statement of
the form: every wff has property P. The property P in this case is
having a finite number of different sentence letters. In order to do this,
we must show two separate statements:
base case: we show that every atomic sentence has the property. This
is obvious—atomic sentences are just sentence letters, and each of
them contains one sentence letter, and thus finitely many different
sentence letters.
induction step: we begin by assuming that if formulas φ and ψ have
the property, then so will the complex formulas one can form from
φ and ψ by the rules of formation, namely ∼φ and φ→ψ. So, we
assume that φ and ψ have finitely many different sentence letters;
and we show that the same must hold for ∼φ and φ→ψ. That’s
obvious: ∼φ has as many different sentence letters as does φ; since
φ, by assumption, has only finitely many, then so does ∼φ. As for
φ→ψ, by hypothesis, φ and ψ have finitely many different sentence
letters, and so φ→ψ has, at most, n + m sentence letters, where
n and m are the number of different sentence letters in φ and ψ,
respectively.
We’ve shown that every atomic formula has the property having a
finite number of different sentence letters; and we’ve shown that the
property is inherited by complex formulas built according to the
recursion rules. But every wff is either atomic, or built from atomics
CHAPTER 2. PROPOSITIONAL LOGIC 28
1 P →(Q→R)
2 P ∧Q
3 P 2, ∧E
4 Q 2, ∧E
5 Q→R 1, 3 →E
6 R 4, 5 →E
7 (P ∧Q)→R 2-6, →I
or like this:
CHAPTER 2. PROPOSITIONAL LOGIC 29
1. P →(Q→R) Pr.
2. show (P ∧Q)→R CD
3. P ∧Q As.
4. show R DD
5. P 3, ∧E
6. Q 3, ∧E
7. Q→R 1, 5 →E
8. R 6, 7→E
2.4.1 Sequents
Natural deduction systems model the kind of reasoning one employs in everyday
life. How does that reasoning work? In its simplest form, one reasons in a
step-by-step fashion from premises to a conclusion, each step being sanctioned
by a rule of inference. For example, suppose that one already knows the premise
that P ∧(P →Q) is true. One can then reason one’s way to the conclusion that
Q is also true, as follows:
1. P ∧(P →Q) premise
2. P from line 1
3. P →Q from line 1
4. Q from lines 2 and 3
CHAPTER 2. PROPOSITIONAL LOGIC 30
In this kind of proof, each step is a tiny, indisputably correct, logical inference.
Consider the moves from 1 to 2 and from 1 to 3, for example. These are
indisputably correct because a conjunctive statement clearly logically implies
either of its conjuncts. Likewise for the move from 2 and 3 to 4: it is clear
that a conditional statement, plus its antecedent, together imply its consequent.
Natural deduction systems consist in part of simple general principles like these
(“a conjunctive statement logically implies either of its conjuncts”); they are
known as rules of inference.
In addition to rules of inference, ordinary reasoning employs a further
technique: the use of assumptions. In order to establish a conditional claim “if P
then Q”, one would ordinarily i) assume P , ii) reason one’s way to Q, and then
iii) on that basis conclude that the conditional claim “if P then Q” is true. Once
P ’s assumption is shown to lead to Q, the conditional claim “if P then Q” may
be concluded. Another example: to establish a claim of the form “not-P ”, one
would ordinarily i) assume P , ii) reason one’s way to a contradiction, and iii)
on that basis conclude that “not-P ” is true. Once P ’s assumption is shown to
lead to a contradiction, “not-P ” may be concluded. The first sort of reasoning
is called conditional proof, the second, reductio ad absurdum.
When one reasons with assumptions, one writes down statements that one
does not know to be true. When you write down P as an assumption, with the
goal of proving the conditional “if P then Q”, you do not know P to be true.
You’re merely assuming P for the sake of establishing the conditional “if P
then Q”. Outside the context of this proof, the assumption need not hold; once
you’ve reasoned your way to Q on the basis of the assumption of P , and so
concluded that the conditional “if P then Q” is true, you stop assuming P . To
model this sort of reasoning formally, we need a way to keep track of how the
conclusions we establish depend on the assumptions we have made. Natural
deduction systems in introductory textbooks tend to do this geometrically (by
placement on the page), with special markers (e.g., ‘show’), and by drawing
lines or boxes around parts of the proof once the assumptions that led to those
parts are no longer operative. We will do it differently: we will keep track of the
dependence of conclusions on assumptions by writing down explicitly, for each
conclusion, which assumptions it depends on. We will do this by constructing
our derivations out of sequents.
A sequent looks like this:
Γ`φ
CHAPTER 2. PROPOSITIONAL LOGIC 31
2.4.2 Rules
The first step in developing our system is to write down sequent rules. A sequent
rule is a permission to move from certain sequents to certain other sequents. Our
goal is to construct rules with the following feature: if the “from” sequents are
all logically correct sequents, then any of the “to” sequents will be guaranteed
to be a logically correct sequent. Call such sequent rules “logical-correctness
preserving”.
Consider, as an example, the first rule of our system “∧ introduction”, or
“∧I” for short. We picture this sequent rule thus:
Γ`φ ∆`ψ
∧I
Γ,∆ ` φ∧ψ
CHAPTER 2. PROPOSITIONAL LOGIC 32
Above the line go the “from” sequents; below the line go the “to”-sequents.
(The comma between Γ and ∆ in the “to” sequent simply means that the
premises of this sequent are all the members of Γ plus all the members of ∆.
Strictly speaking it would be more correct to write this in set-theoretic notation
as: Γ ∪ ∆ ` φ∧ψ.) Thus, ∧I permits us to move from the sequents Γ ` φ and
∆ ` ψ to the sequent Γ, ∆ ` φ∧ψ. For any sequent rule, we say that any of the
“to” sequents (Γ, ∆ ` φ∧ψ in this case) follows from the “from” sequents (in this
case Γ ` φ and ∆ ` ψ) via the rule.
It seems intuitively clear that ∧I preserves logical correctness. For if some
assumptions Γ logically imply φ, and some assumptions ∆ logically imply ψ,
then (since φ∧ψ intuitively follows from φ and ψ taken together) the conclusion
φ∧ψ should indeed logically follow from all the assumptions together, the ones
in Γ and the ones in ∆.
Our next sequent rule is ∧E:
Γ ` φ∧ψ
∧E
Γ`φ Γ`ψ
This lets one move from the sequent Γ ` φ∧ψ to either the sequent Γ ` φ or
the sequent Γ ` ψ (or both). This, too, appears to preserve logical correctness.
If the members of Γ imply the conjunction φ∧ψ, then (since φ∧ψ intuitively
implies both φ and ψ individually) it must be that the members of Γ imply φ,
and they must also imply ψ.
The rule ∧I is known as an introduction rule for ∧, since it allows us to move
to a sequent whose major connective is the ∧. Likewise, the rule ∧E is known
as an elimination rule for ∧, since it allows us to move from a sequent whose
major connective is the ∧. In fact our sequent system contains introduction and
elimination rules for the other connectives as well: ∼, ∨, and → (let’s forget
the ↔ here.) We’ll present those rules in turn.
First ∨I and ∨E:
Γ`φ Γ ` φ∨ψ ∆1 ,φ ` χ ∆2 ,ψ ` χ
∨I ∨E
Γ ` φ∨ψ Γ ` ψ∨φ Γ,∆1 ,∆2 ` χ
Let’s think about what ∨E means. Remember the intuitive meaning of a sequent:
its conclusion is a logical consequence of its premise. Another (related) way to
think of it is that Γ ` φ means that one can establish that φ if one assumes the
members of Γ. So, if the sequent Γ ` φ∨ψ is logically correct, that means we’ve
got the disjunction φ∨ψ, assuming the formulas in Γ. Now, suppose we can
CHAPTER 2. PROPOSITIONAL LOGIC 33
φ`φ RA
Note that this is the one sequent rule when no “from” sequent is required,
since there are no sequents above the line. The rule permits us to move from
no sequents at all to a sequent of the form φ ` φ. (Strictly, this sequent should
be written “{φ} ` φ”.) Call any such sequent an “assumption sequent”. Clearly,
any assumption sequent is a logically correct sequent, since clearly φ can be
proved if we assume φ itself.
CHAPTER 2. PROPOSITIONAL LOGIC 34
1. P ∧Q ` P ∧Q As
2. P ∧Q ` P 1, ∧E
3. P ∧Q ` Q 1, ∧E
4. P ∧Q ` Q∧P 2, 3 ∧I
Though it isn’t strictly required, we write a line number to the left of each
sequent in the series, and to the right of each line we write the sequent rule
that justifies it, together with the line or lines (if any) that contained the “from”
sequents required by the sequent rule in question. (The rule of assumptions
requires no “from” sequents, recall.)
(It’s important to distinguish what we’re now calling proofs, namely, sequent
proofs, from the kinds of informal arguments I gave in section 2.3, and will
give elsewhere in this book. Sequent proofs (and also the axiomatic proofs we
will introduce in section 2.5) are formalized object-language proofs. The sentences
in sequent proofs are sentences in the object language; they are wffs of PL.
Moreover, we gave a rigorous definition of what a sequent proof is. Moreover,
sequent proofs are restrictive in that only the system’s official rules may be
used. For contrast, consider the argument I gave in section 2.3 that any PL-
valuation assigns 1 to φ∧ψ iff it assigns 1 to φ and 1 to ψ. That argument was
an informal metalanguage proof. The sentences in the argument were sentences
of English, and the argument used informal (i.e., not formalized) techniques of
reasoning. “Informal” doesn’t imply lack of rigor. The argument was perfectly
rigorous: it conforms to the standards of good argumentation that generally
prevail in mathematics. We’re free to use any reasonable pattern of reasoning,
for example “universal proof” (to establish something of the form “everything
is thus-and-so”, we consider an arbitrary thing and show that it is thus-and-so).
We may “skip steps” if it’s clear how the argument is supposed to go. In short,
CHAPTER 2. PROPOSITIONAL LOGIC 35
(Note that it would be equivalent to define a sequent proof as any line in any
sequent proof, because at any point in a sequent proof one may simply stop
adding lines, and the proof up until that point counts as a legal sequent proof.)
So, for example, the sequent proof given above establishes that P ∧Q ` Q∧P is
a provable sequent. (We call a sequent proof, whose last line is Γ ` φ, a sequent
proof of Γ ` φ.)
Given the notion of a provable sequent, we can now offer sequent-proof-
theoretic notions of logical truth and logical consequence:
The symbol ∅ stands for the “empty set”—the set containing no members.
Thus, a logical truth here is understood as a formula that is provable from no
assumptions at all.
1. P ∧Q ` P ∧Q As
2. P ∧Q ` P 1, ∧E
3. P ∧Q ` Q 1, ∧E
4. P ∧Q ` Q∧P 2, 3 ∧I
Notice the strategy. We first use the rule of assumptions to enter the premise
of the sequent we’re trying to prove: P ∧Q. We then use the rules of inference
to infer the consequent of that sequent: Q∧P . Since our initial assumption of
P ∧Q was dependent on the formula P ∧Q, our subsequent inferences remain
dependent on that same assumption, and so the final formula concluded, Q∧P ,
remains dependent on that assumption.
Let’s write our proofs out in a simpler way. Instead of writing out entire
sequents, let’s write out only their conclusions. We can indicate the premises
of the sequent using line numbers; the line numbers indicating the premises
of the sequent will go to the left of the number indicating the sequent itself.
Rewriting the previous proof in this way yields:
1 (1) P ∧Q As
1 (2) P 1, ∧E
1 (3) Q 1, ∧E
1 (4) Q∧P 2, 3 ∧I
1. P →Q ` P →Q As
2. Q→R ` Q→R As
3. P ` P As
4. P →Q, P ` Q 1,3 →E
5. P →Q, Q→R, P ` R 2,4 →E
6. P →Q, Q→R ` P →R 5,→I
It can be rewritten in the simpler style as follows:
CHAPTER 2. PROPOSITIONAL LOGIC 37
1 (1) P →Q As
2 (2) Q→R As
3 (3) P As
1,3 (4) Q 1, 3 →E
1,2,3 (5) R 2, 4 →E
1,2 (6) P →R 5, →I
Let’s think about this example. We’re trying to establish P →R on the ba-
sis of two formulas, P →Q and Q→R, so we start by assuming the latter two
formulas. Then, since the formula we’re trying to establish is a conditional,
we assume the antecedent of the conditional, in line 3. We then proceed, on
that basis, to reason our way to R, the consequent of the conditional we’re
trying to prove. (Notice how in lines 4 and 5, we add more line numbers on
the very left. Whenever we use → E, we increase dependencies: when we infer
Q from P and P →Q, our conclusion Q depends on all the formulas that P and
P →Q depended on, namely, the formulas on lines 1 and 3. Look back to the
statement of the rule → E: the conclusion ψ depends on all the formulas that φ
and φ→ψ depended on: Γ and ∆.) That brings us to line 5. At that point, we’ve
shown that R can be proven, on the basis of various assumptions, including P .
The rule →I (that is, the rule of conditional proof) then lets us conclude that
the conditional P →R follows merely on the basis of the other assumptions;
that rule, note, lets us in line 6 drop line 3 from the list of assumptions on
which P →R depends.
Next let’s establish an instance of DeMorgan’s Law, ∼(P ∨Q) ` ∼P ∧∼Q:
1 (1) P ∨Q As
2 (2) ∼P As
3 (3) Q As (for use with ∨E)
4 (4) P As (for use with ∨E)
5 (5) ∼Q As (for reductio)
4,5 (6) ∼Q∧P 4,5 ∧I
4,5 (7) P 6, ∧E
2,4,5 (8) P ∧∼P 2,7 ∧I
2,4 (9) ∼∼Q 8, RAA
2,4 (10) Q 9, DN
1,2 (11) Q 1,3,10 ∨E
The basic idea of this proof is to use ∨ E on line 1 to get Q. That calls,
in turn, for showing that each disjunct of line 1, P and Q, leads to Q. Showing
that Q leads to Q is easy; that was line 3. Showing that P leads to Q took lines
4-10; line 10 states the result of that reasoning, namely that Q follows from P
(as well as line 2). I began at line 4 by assuming P . Then my strategy was to
establish Q by reductio, so I assumed ∼Q in line 5. At this point, I basically
had my contradiction: at line 2 I had ∼P and at line 4 I had P . (You might
think I had another contradiction: Q at line 3 and ∼Q at line 5. But at the end
CHAPTER 2. PROPOSITIONAL LOGIC 39
Lines 4, 6 and 7 in the proof employ this trick: initially, at line 4, P only depends
on 4, but then by line 7, P also depends on 5. That way, the move from 8 to 9
by RAA is justified.
MP φ→ψ
φ
ψ
Wff φ is provable from set of wffs Γ (“Γ ` φ”) iff φ is the last line of
some proof from Γ.
The symbol ` may be subscripted with the name of the system in question.
Thus, for our axiom system for PL below, we may write: `PL . (We’ll omit this
subscript when it’s clear which axiomatic system is at issue.)
Here is an axiomatic system for propositional logic:3
φ→(ψ→φ)
This string of symbols isn’t itself an axiom because it isn’t a wff; it isn’t a wff
because it contains Greek letters, which aren’t allowed in wffs (since they’re
not on the list of PL primitive vocabulary). φ and ψ are variables of our
metalanguage; you only get an axiom when you replace these variables with
wffs. P →(Q→P ), for example, is an axiom; it results from A1 by replacing φ
with P and ψ with Q. (Note: since you can put in any wff for these variables,
and there are infinitely many wffs, there are infinitely many axioms.)
A few points of clarification about how to construct axioms from schemas.
First point: you can stick in the same wff for two different Greek letters. Thus
you can let both φ and ψ in A1 be P , and construct the axiom P →(P →P ).
(But of course, you don’t have to stick in the same thing for φ as for ψ.) Sec-
ond point: you can stick in complex formulas for the Greek letters. Thus,
3
See Mendelson (1987, p. 29).
CHAPTER 2. PROPOSITIONAL LOGIC 42
1. P →(Q→P ) (A1)
2. P →(Q→P ))→((P →Q)→(P →P )) (A2)
3. (P →Q)→(P →P ) 1,2 MP
The existence of this proof shows that (P →Q)→(P →P ) is a theorem. And,
building on the previous proof, we can construct a proof of P →P from {P →Q}:
1. P →(Q→P ) (A1)
2. (P →(Q→P ))→((P →Q)→(P →P )) (A2)
3. (P →Q)→(P →P ) 1,2 MP
4. P →Q member of {(P →Q)}
5. P →P 3, 4 MP
CHAPTER 2. PROPOSITIONAL LOGIC 43
then shows us how to construct a proof of φ→χ . I call this the MP technique
because its effect is that you can do modus ponens “within the consequent of
the conditional φ→”. Here’s how the MP technique works:
It’s usually more useful to think in terms of proof schemas, rather than
proofs, because they can go into our toolkit, if they have general applicability.
The proof schema we just constructed, for example, shows that anything of the
form (φ→ψ)→(φ→(χ →ψ)) is a theorem. As it happens, this is a fairly intuitive
theorem schema. Think of it as the principle of “weakening the consequent”:
χ →ψ is logically weaker than ψ, so if φ leads to ψ, φ must also lead to χ →ψ.
That sounds like a pattern that might well recur, so let’s put it into the toolkit,
under the label “weakening the consequent”. If we’re ever in the midst of a
proof and could really use a line of the form (φ→ψ)→(φ→(χ →ψ)), then we
can simply write that line down, and annotate on the right “weakening the
consequent”. Given the proof sketch above, we know that we could always
CHAPTER 2. PROPOSITIONAL LOGIC 45
1. [φ→(ψ→χ )]→[(φ→ψ)→(φ→χ )] A2
2. [(φ→ψ)→(φ→χ )]→[ψ→(φ→χ )] strengthening the antecedent
3. [φ→(ψ→χ )]→[ψ→(φ→χ )] 1,2 transitivity
` (∼ψ→∼φ)→(φ→ψ) (“contraposition”):
1. (∼ψ→∼φ)→[(∼ψ→φ)→ψ] A3
2. [(∼ψ→φ)→ψ]→(φ→ψ) strengthening the antecedent
3. (∼ψ→∼φ)→(φ→ψ) 1, 2 transitivity
1. ∼φ→(∼ψ→∼φ) A1
2. (∼ψ→∼φ)→(φ→ψ) contraposition
3. ∼φ→(φ→ψ) 1, 2 transitivity
CHAPTER 2. PROPOSITIONAL LOGIC 47
proof are valid. And now, consider the result of applying modus
ponens. That means that the new line we’ve added to the proof
is some formula ψ, which we’ve inferred from two earlier lines
that have the forms φ→ψ and φ. We must show that ψ is a valid
formula, i.e., that VI (ψ)=1 for every PL-interpretation I . By the
inductive hypothesis, all earlier lines in the proof are valid, and
hence both φ→ψ and φ are valid. Thus, VI (φ)=1 and VI (φ→ψ)=1.
But if VI (φ)=1 then VI (ψ) can’t be 0, for if it were, then VI (φ→ψ)
would be 0, and it isn’t. Thus, VI (ψ)=1.
We’ve shown that axioms are valid, and that modus ponens pre-
serves validity. So, by induction, every time one adds to a proof,
one adds a valid formula. So the last line in a proof is always a valid
formula. Thus, every theorem is valid. QED.
Notice the general structure of this proof: we first showed that every axiom
has a certain property, and then we showed that the rule of inference preserves
the property. Given the definition of ‘theorem’, it followed that every theorem
has the property. We chose our definition of a theorem with just this sort of
proof in mind.
Remember that this is a proof in the metalanguage, about propositional
logic. It isn’t a proof in any system of derivation.
Completeness is harder to prove, and I won’t prove it here.
One nice thing about soundness is that it lets us establish facts of unprov-
ability. Soundness says “if ` φ then φ”. Equivalently, it says: “if 2 φ then
0 φ”. So, to show that something isn’t a theorem, it suffices to show that it
isn’t valid. Consider, for example, the formula (P →Q)→(Q→P ). There exist
PL-interpretations in which the formula is false, namely, PL-interpretations in
which P is 0 and Q is 1. So, (P →Q)→(Q→P ) is not valid (since it’s not true
in all PL-interpretations.) But then soundness tells us that it isn’t a theorem
either. In general: if we’ve established soundness, then in order to show that a
formula isn’t a theorem, all we need to do is find an interpretation in which it
isn’t true.
One could also prove soundness and completeness for the natural deduction
system of section 2.4.4 The soundness proof, for instance, would proceed by
4
Note: if we allow sequents with infinite premise-sets, then to secure completeness we will
need to add a rule for adding dependencies: from Γ ` φ conclude Γ ∪ ∆ ` φ, where ∆ is any set
of wffs. The trick described above for adding dependencies (using RA, ∧I and ∧E) only allows
CHAPTER 2. PROPOSITIONAL LOGIC 50
(“operations”) (the rules of formation, recall, say that if φ and ψ are formulas,
then so are (φ→ψ) and ∼φ.) So, to show that all formulas have feature F, we
must merely show that a) all the sentence letters have feature F, and b) show
that if φ and ψ both have feature F, then both (φ→ψ) and ∼φ also will have
feature F.
If you’re ever proving something by induction, it’s important to identify
which sort of inductive proof you’re constructing.
Chapter 3
s promised, we will not stop with the standard logics familiar from in-
A troductory textbooks. In this chapter we examine some philosophically
important variations and deviations from standard propositional logic.
f (1) = 0
f (0) = 1
This is a called one-place function because it takes only one truth value as
input.
In fact, we have a name for this truth function: negation. And we express
that truth function with our symbol ∼. So the negation truth function is one
we can symbolize in propositional logic. More carefully, what we mean by
saying “We can symbolize truth function f in propositional logic” is this:
52
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 53
Think of this truth function as “not both”. Unlike the negation and conjunc-
tion truth functions, we don’t have a single symbol for this truth function.
Nevertheless, it can be symbolized in propositional logic: by the following
sentence:
∼(P ∧Q)
All truth functions (of any finite number of places) can be symbol-
ized in propositional logic using just the ∧, ∨, and ∼
I’ll begin by illustrating the idea of the proof with an example. Suppose we
want to symbolize the following three-place truth-function:
f( 1, 1, 1 )=0
f( 1, 1, 0 )=1
f( 1, 0, 1 )=0
f( 1, 0, 0 )=1
f( 0, 1, 1 )=0
f( 0, 1, 0 )=0
f( 0, 0, 1 )=1
f( 0, 0, 0 )=0
Since this truth function returns the value 1 in just three cases (rows two,
four, and seven), what we want is a sentence containing three sentence letters,
P1 , P2 , P3 , that is true in exactly those three cases: (a) when P1 , P2 , P3 take on the
three truth values in the second row (i.e., 1, 1, 0) , (b) when P1 , P2 , P3 take on
the three truth values in the fourth row (1, 0, 0) , and (c) when P1 , P2 , P3 take
on the three truth values in the seventh row (0, 0, 1) . Now, we can construct a
sentence that is true in case (a) and false otherwise: P1 ∧P2 ∧∼P3 . We can also
construct a sentence that’s true in case (b) and false otherwise: P1 ∧∼P2 ∧∼P3 .
And we can also construct a sentence that’s true in case (c) and false otherwise:
∼P1 ∧∼P2 ∧P3 . But then we can simply disjoin these three sentences to get the
sentence we want:
We’ll again use the method of induction. We want to show that the assertion is
true for all sentences. So we first prove that the assertion is true for all sentence
with no connectives (i.e., for sentences containing just sentence letters.) This is
the base case, and is very easy here, since if φ has no connectives, then obviously
φ is just the sentence letter P itself, in which case, clearly, if P is true then so is
φ. Next we assume the inductive hypothesis:
This is easy to show. We want to show that the assertion holds for φ∧ψ—that
is, if P is true then so is φ∧ψ. But we know by the inductive hypothesis that
the assertion holds for φ and ψ individually. So we know that if P is true, then
both φ and ψ are true. But then, we know from the truth table for ∧ that
φ∧ψ is also true in this case. The reasoning is exactly parallel for φ→ψ: the
inductive hypothesis tells us that whenever P is true, so are φ and ψ, and then
we know that in this case φ→ψ must also then be true, by the truth table for
→. Therefore, by induction, the result is proved.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 57
What’s the point? This notation eliminates the need for parentheses. With the
usual notation, in which we put the connectives between the sentences they
connect, we need parentheses to distinguish, e.g.:
(P ∧Q)→R
P ∧(Q→R)
But with polish notation, these are distinguished without parentheses; they
become:
→∧P QR
∧P →QR
respectively.
with $100, 000 is such a person. They seem neither definitely rich nor definitely
not rich. So there’s pressure to say that the statement “this person is rich” is
capable of being neither definitely true nor definitely false. It’s vague.
Others say we need a third truth value for statements about the future. If
it is in some sense “not yet determined” whether there will be a sea battle
tomorrow, then (it is argued) the sentence:
is neither true nor false. In general, statements about the future are neither
true nor false if there is nothing about the present that determines their truth
value one way or the other.4
Yet another case in which some have claimed that bivalence fails concerns
failed presupposition. Consider this sentence:
In fact, I have never beaten a dog. I don’t even have a dog. So is it true that
I stopped beating my dog? Obviously not. But on the other hand, is this
statement false? Certainly no one would want to assert its negation: “Ted has
not stopped beating his dog”. The sentence presupposes that I was beating a dog;
since this presupposition is false, the question of the sentence’s truth does not
arise: the sentence is neither true nor false.
For a final challenge to bivalence, consider the sentence:
‘Sherlock Holmes’ doesn’t refer to a real entity. Further, Sir Arthur Conan
Doyle does not specify in his Sherlock Holmes stories whether Holmes has
such a mole. Either of these reasons might be argued to result in a truth value
gap for the displayed sentence.
It’s an interesting philosophical question whether any of these arguments
for bivalence’s failing are any good. But we won’t take up that question. Instead,
we’ll look at the formal result of giving up bivalence. That is, we’ll introduce
some non-bivalent formal systems. We won’t ask whether these systems really
model English correctly.
4
An alternate view preserves the “openness of the future” as well as bivalence: both ‘There
will be a sea battle tomorrow’ and ‘There will fail to be a sea battle tomorrow’ are false. This
combination is not contradictory, provided one rejects the equivalence of “It will be the case
tomorrow that ∼φ” and “∼ it will be the case tomorrow that φ”.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 60
These systems all give different truth tables for the Boolean connectives.
The original truth tables give you the truth values of complex formulas based
on whether their sentence letters are true or false (1 or 0) . The new truth
tables need to take into account cases where the sentence letters are # (neither
1 nor 0) .
1 if VI (φ) = 1 or VI (ψ) = 1
iii) VI (φ∨ψ) = 0 if VI (φ) = 0 and VI (ψ) = 0
# otherwise
1 if VI (φ) = 0
v) VI (∼φ) = 0 if VI (φ) = 1
# otherwise
So basically, the classical bit of each truth table is what you’d expect; but
everything gets boring if any constituent formula is a #.
One way to think about these tables is to think of the # as indicating nonsense.
The sentence “The sun is purple and blevledgekl;rz”, one might naturally think,
is neither true nor false because it is nonsense. It is nonsense even though it
has a part that isn’t nonsense.
3.3.4 Supervaluationism
Recall the guiding thought behind the strong Kleene tables: if a formula’s
classical truth values fix a particular truth value, then that is the value that the
formula takes on. There is a way to take this idea a step further, which results
in a new and interesting way of thinking about three-valued logic.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 64
According to the strong Kleene tables, we get a classical truth value for
φ
ψ, where
is any connective, only when we have “enough classical
information” in the truth values of φ and ψ to fix a classical truth value for
φ
ψ. Consider φ∧ψ for example: if either φ or ψ is false, then since
falsehood of a conjunct is classically sufficient for the falsehood of the whole
conjunction, the entire formula is false. But if, on the other hand, both φ and
ψ are #, then neither φ nor ψ has a classical truth value, we do not have enough
classical information to settle on a classical truth value for φ∧ψ, and so the
whole formula is #.
But now consider a special case of the situation just considered, where φ is
P , ψ is ∼P , and P is #. According to the strong Kleene tables, the conjunction
P ∧∼P is #, since it is the conjunction of two formulas that are #. But there is a
way of thinking about truth values of complex sentences according to which
the truth value ought to be 0, not #: no matter what classical truth value P were
to take on, the whole sentence P ∧∼P would be 0—therefore, one might think,
P ∧∼P ought to be 0. If P were 0 then P ∧∼P would be 0∧∼0—that is 0; and if
P were 1 then P ∧∼P would be 1∧∼1—0 again.
The general thought here is this: suppose a sentence φ contains some
sentence letters P1 . . . Pn that are #. If φ would be false no matter how we assign
classical truth values to P1 . . . Pn —that is, no matter how we precisified φ—then
φ is in fact false. Further, if φ would be true no matter how we precisified it,
then φ is in fact true. But if precisifying φ would sometimes make it true and
sometimes make it false, then φ in fact is #.
The idea here can be thought of as an extension of the idea behind the
strong Kleene tables. Consider a formula φ
ψ, where
is any connective.
If there is enough classical information in the truth values of φ and ψ to fix on a
particular classical truth value, then the strong Kleene tables assign φ
ψ that
truth value. Our new idea goes further, and says: if there is enough classical
information within φ and ψ to fix a particular classical truth value, then φ
ψ
gets that truth value. Information “within” φ and ψ includes, not only the
truth values of φ and ψ, but also a certain sort of information about sentence
letters that occur in both φ and ψ. For example, in P ∧∼P , when P is #, there
is insufficient classical information in the truth values of P and of ∼P to settle
on a truth value for the whole formula P ∧∼P (since each is #). But when we
look inside P and ∼P , we get more classical information: we can use the fact
that P occurs in each to reason as we did above: whenever we turn P to 0, we
turn ∼P to 1, and so P ∧∼P becomes 0; and whenever we turn P to 1 we turn
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 65
P ∧Q
P ∧∼P
5
I am counting #, in addition to 1 and 0, as a “truth value”.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 67
P ∨Q # P →Q #
P ∨∼P 1 P →P 1
3.4 Intuitionism
Intuitionism in the philosophy of mathematics is a view according to which
there are no mind-independent mathematical facts. Rather, mathematical facts
and entities are mental constructs that owe their existence to the mental activity
of mathematicians constructing proofs. This philosophy of mathematics leads
intuitionists to a distinctive form of logic: intuitionist logic.
Let P be the statement: The sequence 0123456789 occurs somewhere in the
decimal expansion of π. How should we think about its meaning? For the classical
mathematician, the answer is straightforward. P is a statement about a part of
mathematical reality, namely, the infinite decimal expansion of π. Either the
sequence 0123456789 occurs somewhere in that expansion, in which case P is
true, or it does not, in which case P is false and ∼P is true.
For the intuitionist, this whole picture is mistaken, premised as it is on the
reality of an infinite decimal expansion of π. Our minds are finite, and so only
the finite initial segment of π’s decimal expansion that we have constructed so
far is real. The intuitionist’s alternate picture of P ’s meaning, and indeed of
meaning generally (for mathematical statements) is a radical one.6
The classical mathematician, comfortable with the idea of a realm of mind-
independent entities, thinks of meaning in terms of truth and falsity. As we saw,
she thinks of P as being true or false depending on the facts about π’s decimal
expansion. Further, she explains the meanings of the propositional connectives
in truth-theoretic terms: a conjunction is true iff each of its conjuncts are true;
6
One intuitionist picture, anyway, on which see Dummett (1973). What follows is a crude
sketch. It does not do justice to the actual intuitionist position, which is, as they say, subtle.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 68
a negation is true iff the negated formula is false; and so on. Intuitionists, on
the other hand, reject the centrality of truth to meaning, since truth is tied
up with the rejected picture of mind-independent mathematical reality. For
them, the central semantic concept is that of proof. They simply do not think
in terms of truth and falsity; in matters of meaning, they think in terms of the
conditions under which formulas have been proved.
Take P , for example. Intuitionists advise us: don’t think in terms of what
it would take for P to be true. Think, rather, in terms of what it would take
to prove P . And the answer is clear: we would need to actually continue our
construction of the decimal expansion of π to a point where we found the
sequence 0123456789.
What, now, of ∼P ? Again, thinking in terms of proof, not truth: what
would it take for ∼P to be proved? The answer here is less straightforward.
Since P said that there exists a number of a certain sort, it was clear how it
would have to be proved: by actually exhibiting (calculating) some particular
number of that sort. But ∼P says that there is no number of a certain sort; how
do we prove something like that? The intuitionist’s answer: by proving that the
assumption that there is a number of that sort leads to a contradiction. In general, a
negation, ∼φ, is proved by proving that φ leads to a contradiction.7
Similarly for the other connectives: the intuitionist explicates their meanings
by their role in generating proof conditions, rather than truth conditions. φ∧ψ
is proved by separately giving a proof of φ and a proof of ψ; φ∨ψ is proved
by giving either a proof of φ or a proof of ψ; φ→ψ is proved by exhibiting a
construction whereby any proof of φ can be converted into a proof of ψ.
Likewise, the intuitionist thinks of logical consequence as the preservation
of provability, not the preservation of truth. For example, φ∧ψ logically implies
φ because if one has a proof of φ∧ψ, then one has a proof of φ; and conversely,
if one has proofs of φ and ψ separately, then one has the materials for a proof
of φ∧ψ. So far, so classical. But ∼∼φ does not logically imply φ, for the
intuitionist. Simply having a proof of ∼∼P —a proof that the assumption that
0123456789 occurs nowhere in π’s decimal expansion leads to a contradiction—
wouldn’t give us a proof of P , since proving P would require exhibiting a
particular place in π’s decimal expansion where 0123456789 occurs.
Likewise, intuitionists do not accept the law of the excluded middle, φ∨∼φ,
as a logical truth. To be a logical truth, according to an intuitionist, a sentence
7
Given the contrast with the classical conception of negation, a different symbol (often
“¬”) is sometimes used for intuitionist negation.
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 69
Γ ` ∼∼φ
Γ`φ
Γ ` φ∧∼φ
Γ`ψ
Note that EF can be proved in the original system of chapter 2.4: simply use
RAA and then DNE. So, intuitionist logic results from a system for classical
logic by simply dropping one rule (DNE) and adding another rule that was pre-
viously provable (EF). It follows that every intuitionistically provable sequent
is also classically provable (because every intuitionistic proof can be converted
to a classical proof).
Notice how dropping DNE blocks proofs of various classical theorems the
intuitionist wants to avoid. The proof of ∅ ` P ∨∼P in chapter 2.4, for instance,
used DNE. Of course, for all we’ve said so far, there might be some other way
to prove this sequent. Only when we have a semantics for intuitionistic logic,
and a soundness proof relative to that semantics, can we show that this sequent
CHAPTER 3. VARIATIONS AND DEVIATIONS FROM PL 70
1 (1) ∼∼∼P As
2 (2) P As (for reductio)
2 (3) ∼∼P 2, DN (accepted version)
1,2 (4) ∼∼P ∧ ∼∼∼P 1,3 ∧I
1 (5) ∼P 4, RAA
Note that you can’t use this sort of proof to establish ∼∼P ` P . Given the
way RAA is stated, its application always results in a formula beginning with
the ∼.
Chapter 4
Predicate Logic
et’s now turn from propositional logic to the “predicate calculus” (PC),
L as it is sometimes called. As with propositional logic, we’re going to
formalize predicate logic. We’ll first do grammar, and then move to semantics.
We won’t consider proof theory at all.1
i) logical: →, ∼, ∀
ii) nonlogical:
a) for each n > 0, n-place predicates F , G . . ., with or without
subscripts
b) variables x, y . . . with or without subscripts
c) individual constants (names) a, b . . ., with or without sub-
scripts
1
Proof systems for predicate logic, in both axiomatic and natural-deduction form, are
straightforward, and can be found in standard logic textbooks.
71
CHAPTER 4. PREDICATE LOGIC 72
iii) parentheses
No symbol of one type is a symbol of any other type. Let’s call any variable or
constant a term.
We’ll call formulas that are wffs in virtue of clause i) “atomic” formulas. When
a formula has no free variables, we’ll say that it is a closed formula, or sentence;
otherwise it is an open formula.
We have the same defined logical terms: ∧, ∨, ↔. We also add the following
definition of the existential quantifier:
F x don’t have truth values at all. The variable ‘x’ doesn’t stand for any one
thing, and so ‘F x’ doesn’t have a truth value.
The solution to this problem is due to the Polish logician Alfred Tarski. It
begins with a new conception of a configuration, that of a model:
As before, we can give recursive clauses for the truth values of negations
and conditionals. φ→ψ, for example, will be true iff either φ is false or ψ is
true.
But this becomes tricky when we try to specify the truth value of ∀xF x. It
should, intuitively, be true if and only if ‘F x’ is true, no matter what we put
in in place of ‘x’. But this is vague. Do we mean “whatever name (constant)
we put in place of ‘x”’? No, because we don’t want to assume that we’ve got a
name for everything in the domain, and what if F x is true for all the objects we
have names for, but false for one of the nameless things! Do we mean, “true no
matter what object from the domain we put in place of ‘x”’? No; objects from
the domain aren’t part of our primitive vocabulary, so the result of replacing ‘x’
with an object from the domain won’t be a formula!2
Tarski’s solution to this problem goes as follows. Initially, we don’t consider
truth values of formulas absolutely. Rather, we let the variables refer to certain
things in the domain temporarily. Then, we’ll say that ∀xF x will be true iff
for all objects u in the domain D: F x is true while x temporarily refers to u.
We implement this idea of temporary reference by defining the notion of a
variable assignment:
The variable assignments give the “temporary” meanings to the variables; when
g (x) = u, then u is the temporary denotation of x.
We need a further bit of notation. Let u be some object in D, let g be some
variable assignment, and let α be a variable. We then define “g u/α ” to be the
variable assignment that is just like g , except that it assigns u to α. (If g already
assigns u to α then g u/α will be the same function as g .)
Note the following important fact about variable assignments: g u/α , when
applied to α, must give the value u. (Work through the definitions to see that
this is so.) That is:
g u/α (α) = u
One more bit of apparatus. Given any model M (=〈D,I〉), and given any
variable assignment, g , and given any term (i.e., variable or name) α, we define
the denotation of α, relative to M and g , “[α]M ,g ” as follows:
2
Unless the domain happens to contain members of our primitive vocabulary!
CHAPTER 4. PREDICATE LOGIC 75
i) for any n-place predicate Π and any terms α1 …αn , VM , g (Πα1 …αn )
=1 iff 〈[α1 ]M ,g …[αn ]M ,g 〉 ∈ I (Π)
ii) for any wffs φ, ψ, and any variable α:
a) VM ,g (∼φ)=1 iff VM , g (φ)=0
b) VM ,g (φ→ψ)=1 iff either VM , g (φ)=0 or VM ,g (ψ)=1
c) VM ,g (∀αφ)=1 iff for every u ∈ D, VM ,gu/α (φ)=1
(In understanding clause i), recall that the one tuple containing just u, 〈u〉 is
just u itself. Thus, in the case where Π is F , some one place predicate, clause i)
says that VM , g (F α)=1 iff [α]M ,g ∈ I (F ).)
So far we have defined the notion of truth in a model relative to a variable
assignment. But what we really want is a notion of truth in a model, period—that
is, absolute truth in a model. (We want this because we want to define, e.g., a
valid formula as one that is true in all models.) So, let’s define absolute truth in
a model in this way:
It might seem that this is too strict a requirement—why must φ be true relative
to each variable assignment? But in fact, it’s not too strict at all. The kinds of
formulas we’re really interested in are formulas without free variables (we’re
interested in formulas like F a, ∀xF x, ∀x(F x→Gx); not formulas like F x,
CHAPTER 4. PREDICATE LOGIC 76
∀x Rxy, etc.). And if a formula has no free variables, then if there’s even a
single variable assignment relative to which it is true, then it is true relative to
every variable assignment. (And so, we could just as well have defined truth in
a model as truth relative to some variable assignment.) I won’t prove this fact,
but it’s not too hard to prove; one would simply need to prove (by induction)
that, for any wff φ, if variable assignments g and h agree on all variables free
in φ, then VM ,g (φ)=VM ,h (φ).
Now we can give semantic definitions of the core logical notions:
Since our new definition of the valuation function treats the propositional
connectives → and ∼ in the same way as the propositional logic valuation did,
it’s easy to prove that it also treats the defined connectives ∧, ∨, and ↔ in the
same way:
Moreover, we can also prove that the valuation function treats ∃ as it should
(given its intended meaning):
VM ,g (∃αφ)=1 iff there is some u ∈ D such that VM , gu/α (φ)=1
The claim in step iv) that I (a) ∈ D comes from the definition of an inter-
pretation function: the interpretation of a constant is always a member of the
domain. Notice that “I (a)” is a term of our metalanguage; that’s why, when
I’m given that “for any u ∈ D” in step iii), I can set u equal to I (a) to obtain
step iv).
One further example; let’s establish ∀x∀yRxy→∀xRx x:
We’ve seen how to establish that particular formulas are valid. How do
we show that a formula is invalid? We need to simply exhibit a single model
in which the formula is false. (The definition of validity specifies that a valid
formula is true in all models; therefore, it only takes one model in which a
formula is false to make that formula invalid.) So let’s take one example; let’s
show that the formula (∃xF x∧∃xGx)→∃x(F x∧Gx) isn’t valid. To do this, we
must produce a model in which this formula is false. All we need is a single
model, since in order for the formula to be valid, it must be true in all models.
My model will contain letters in its domain:
D = {u, v}
I (F ) = {u}
I (G) = {v}
It is intuitively clear that the formula is false in this model. In this model,
something has F (namely, u), and something has G (namely, v), but nothing
has both.
One further example: let’s show that ∀x∃yRxy 2 ∃y∀xRxy. We must show
that the first formula does not semantically imply the second. So we must come
up with a model in which the first formula is true and the second is false. It helps
to think about natural language sentences that these formulas might represent.
If R symbolizes “respects”, then the first formula says that “everyone respects
someone or other”, and the second says that “there is someone whom everyone
respects”. Clearly, the first can be true while the second is false: suppose that
each person respects a different person, so that no one person is respected by
everyone. A simple case of this occurs when there are just two people, each of
whom respects the other, but neither of whom respects him/herself:
CHAPTER 4. PREDICATE LOGIC 79
(
•h •
Here is a model based on this idea:
D = {u, v}
I (R) = {〈u, v〉, 〈v, u〉}
Chapter 5
5.1 Identity
“Standard” predicate logic is usually taken to include the identity sign (“=”).
“a=b ” means that a and b are one and the same thing.
80
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 81
That is, the sentence α = β is true iff the terms α and β refer to the same
object.
As an example, let’s show that the formula ∀x∃y x=y is valid. We need
to show that in any model, and any variable assignment g in that model,
Vg (∀x∃y x=y) = 1. So:
The most obvious sentences that may be symbolized with ‘=’ are those that
explicitly concern identity, such as:
(It will be convenient to abbreviate ∼α=β as α6=β. Thus, the second symbol-
ization can be rewritten as: ∀x(M x→x6= s).) But many other sentences involve
the concept of identity in subtler ways.
Consider, for example, “Every lawyer hates every other lawyer”. The ‘other’
signifies nonidentity; we have, therefore:
∀x(Lx→∀y[(Ly∧x6=y)→H xy])
Consider next “Only Ted can change grades”. This means: “no one other
than Ted can change grades”, and may therefore be symbolized as:
∼∃x(x6= t ∧C x)
This says that there are two different objects, x and y, each of which are di-
nosaurs. To say “There are at least three dinosaurs” we say:
Indeed, for any n, one can construct a sentence φn that symbolizes “there are
at least n F s”:
φn : ∃x1 . . . ∃xn (F x1 ∧ · · · ∧F xn ∧ δ)
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 83
where δ is the conjunction of all sentences “xi 6= x j ” where i and j are integers
between 1 and n (inclusive) and i < j . (The sentence δ says in effect that no
two of the variables x1 . . . xn stand for the same object.)
Since we can construct each φn , we can symbolize other sentences involving
number as well. To say that there are at most n F s, we write: ∼φn+1 . To say
that there are between n and m F s (where m > n), we write: φn ∧∼φ m+1 . To
say that there are exactly n F s, we write: φn ∧∼φn+1 .
These methods for constructing sentences involving number will always
work; but one can often construct shorter numerical symbolizations by other
methods. For example, to say “there are exactly two dinosaurs”, instead of
saying “there are at least two dinosaurs, and it’s not the case that there are at
least three dinosaurs”, we could say instead:
i) logical: →, ∼, ∀
ii) nonlogical:
a) for each n > 0, n-place predicates F , G, . . ., with or with-
out subscripts
b) for each n > 0, n-place function symbols f , g ,…, with
or without subscripts
c) variables x, y, . . ., with or without subscripts
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 86
The definition of a wff, actually, stays the same; all that needs to change is the
definition of a “term”. Before, terms were just names or variables. Now, we
need to allow for f (a), f ( f (a)), etc., to be terms. This is done by the following
recursive definition of a term:
Definition of terms
i) D is a non-empty set
ii) I is a function such that:
a) if α is a constant then I (α) ∈ D
b) if Π is an n-place predicate, then I (Π) = some set of
n-tuples of members of D.
c) If f is an n-place function symbol, then I (f ) is an
n-place (total) function defined on D.
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 87
(“Total” simply means that the function must yield an output for any n members
of D.)
The definition of a valuation function stays the same; all we need to do is
update the definition of denotation to accommodate our new complex terms.
Since we now can have arbitrarily long terms (not just names or variables), we
need a recursive definition:
Note the recursive nature of this definition: the denotation of a complex term
is defined in terms of the denotations of its smaller parts. Let’s think carefully
about what clause iii) says. It says that, in order to calculate the denotation of
the complex term f (α1 ,…,αn ) (relative to assignment g ), we must first figure
out what I ( f ) is—that is, what the interpretation function I assigns to the
function symbol f . This object, the new definition of a model tells us, is an
n-place function on the domain. We then take this function, I ( f ), and apply
it to n arguments: namely, the denotations (relative to g ) of the terms α1 . . . αn .
The result is our desired denotation of f (α1 ,…,αn ).
It may help to think about a simple case. Suppose that f is a one-place
function symbol; suppose our domain consists of the set of natural numbers;
suppose that the name a denotes the number 3 in this model (i.e., I (a) = 3),
and suppose that f denotes the successor function (i.e., I ( f ) is the function,
successor, that assigns to any natural number n the number n + 1.) In that case,
the definition tells us that:
[ f (a)] g = I ( f )([a] g )
= I ( f )(I (a))
= successor(3)
=4
Here’s a sample metalanguage proof that makes use of the new definitions.
As mentioned earlier, ‘George W. Bush’s father was a politician’ logically
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 88
implies ‘ ‘Someone’s father was a politician’. Let’s show that these sentences’
symbolizations stand in the relation of semantic implication. That is, let’s show
that P f (c) ∃xP f (x)—i.e., that in any model in which P f (c) is true, ∃xP f (x)
is true:
i) Suppose that P f (c) is true in a model 〈I , D〉—i.e., Vg (P f (c)) =
1 (where V is the valuation for this model), for each variable
assignment g .
ii) Suppose for reductio that ∃xP f (x) is false in this model; i.e.,
for some variable assignment g , Vg (∃xP f (x)) = 0
iii) By line i), Vg (P f (c)) = 1, and so [ f (c)] g ∈ I (P ). [ f (c)] g is
just I ( f )([c] g ), and [c] g is just I (c). So I ( f )(I (c)) ∈ I (P ).
iv) By ii), for every object u ∈ D,Vgu/x (P f (x)) = 0.
v) I (c) ∈ D. So, by line iv), VgI (c))/x (P f (x)) = 0, and hence,
[ f (x)] gI (c)/x ∈
/ I (P )
vi) [ f (x)] gI (c)/x is just I ( f )([x] gI (c))/x ), and [x] gI (c)/x is just gI (c)/x (x)—
i.e., I (c).
vii) So I ( f )(I (c)) ∈
/ I (P ), which contradicts line iii)
Primitive vocabulary:
i) logical: →, ∼, ∀, ι
ii) nonlogical:
a) for each n > 0, n-place predicates F , G, . . ., with or with-
out subscripts
b) variables x, y, . . ., with or without subscripts
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 90
Notice how we needed to combine the recursive definitions of term and wff
into a single recursive definition of wffs and terms together. The reason is that
we need the notion of a wff to define what counts as a term containing the ι
operator (clause ii); but we need the notion of a term to define what counts as
a wff (clause iii). The way we accomplish this is not circular. The reason it isn’t
is that we can always decide, using these rules, whether a given string counts as
a wff or term by looking at whether smaller strings count as wffs or terms. And
the smallest strings are said to be wffs or terms in non-circular ways.
predicate, and hence that atomic sentences containing empty descriptions are
always false. Here’s how the semantics looks (“PC+DD”—“predicate calculus
plus definite descriptions”):
i) D is a non-empty set
/D
ii) E ∈
iii) I is a function such that:
a) if α is a constant then I (α) ∈ D
b) if Π is an n-place predicate, then I (Π) = some set of
n-tuples of members of D
(As with the grammar, we need to mix together the definition of denotation
and the definition of the valuation function. The reason is that we need to
define the denotation of definite descriptions using the valuation function (in
clause v), but we need to define the valuation function using the concept of
denotation in clause vi. As before, this is not circular.)
An alternate approach would appeal to three-valued logic. We could leave
the denotation of ιxφ undefined if there is no object in the domain such that φ.
We could then treat any atomic sentence that contains a denotationless term as
being neither true nor false—i.e., #. We would then need to update the other
clauses to allow for #s, perhaps using the Kleene tables, perhaps some other
truth tables. I won’t explore this further.
as:
That is, “there is something such that: i) it is a black cat, ii) nothing else is a
black cat, and iii) it is happy”.
This method for symbolizing sentences containing ‘the’ is called “Russell’s
theory of descriptions”, in honor of its inventor Bertrand Russell, the 19th and
20th century philosopher and logician.1 The general idea is to symbolize: “the
φ is ψ” as “∃x[φ(x) ∧ ∀y(φ(y)→x=y) ∧ ψ(x)]. This method can be iterated so
as to apply to sentences with two or more definite descriptions, such as:
which becomes, letting ‘E’ stand for ‘is eight feet tall’ and ‘T ’ stand for ‘is
twenty feet long’:
? Or does it mean:
i) there’s no president
ii) there is more than one president
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 94
P ιxF x b
But given Russell’s method, we can symbolize the whole thing without using
either function symbols or the ι:
We can get rid of all function symbols this way, if we want. Here’s the method:
• In any sentence containing the term “ f (α1 . . . αn )”, replace each occur-
rence of this term with “the x such that R(x, α1 . . . αn )”.
Instead of introducing a function symbol s(x, y) for “the sum of x and y”, let’s
introduce a predicate letter R(z, x, y) for “z is a sum of x and y”. We then use
Russell’s method to symbolize the whole sentence thus:
The end of the formula (beginning with ∃w) says “the product of y and z is
identical to x”—that is, that there exists some w such that w is a product of y
and z, and there is no other product of y and z other than w, and w = x.
So we should require the following of RQ . Consider any set, D, and any one-
one function, f , from D onto another set D 0 . Then, if a subset X of D bears
RQ to D, the set f [X ] must bear RQ to D 0 . ( f [X ] is the image of X under
function f —i.e., {u : u ∈ D 0 and u = f (v), for some v ∈ D}. It is the subset of
D 0 onto which f “projects” X .)
Everything is material
Something is spiritual
Here, the predicates (verb phrases) ‘is material’ and ‘is spiritual’ correspond
to the open sentences of logic; it is to these that ‘everything’ and ‘something’
attach.
But in fact, monadic quantifiers in natural language are atypical. ‘Every’
and ‘some’ typically occur as follows:
The quantifiers ‘every’ and ‘some’ attach to two predicates. In the first, ‘every’
attaches to ‘[is a] student’ and ‘is happy’; in the second, ‘some’ attaches to ‘[is
a] fish’ and ‘[is] tasty’. In these sentences, we may think of ‘every’ and ‘some’ as
binary quantifiers. (Indeed, one might think of ‘everything’ and ‘something’ as
the result of applying the binary quantifiers ‘every’ and ‘some’ to the predicate
‘is a thing’.) A logical notation can be introduced which exhibits a parallel
structure, in which ∀ and ∃ attach to two open sentences. In this notation, the
form of quantified sentences is:
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 99
(∀α:φ)ψ
(∃α:φ)ψ
The first is to be read: “all φs are ψ”; the second is to be read “there is a φ that
is a ψ”. The clauses for these new binary quantifiers in the definition of the
valuation function for a PC-model are:
VM , g ((∀α:φ)ψ) = 1 iff φM ,g ,α ⊆ ψM ,g ,α
VM ,g ((∃α:φ)ψ) = 1 iff φM ,g ,α ∩ ψM ,g ,α 6= ∅
That is, (the α:φ)ψ is true iff i) there is exactly one φ, and ii) every φ is a
ψ. This truth condition, notice, is exactly the truth condition for Russell’s
symbolization of “the φ is a ψ”; hence the name the.
As with the introduction of the monadic quantifiers ∃n , the introduction of
the binary existential and universal quantifiers, and of the, does not increase the
expressive power of first order logic, for the same effect can be achieved with
monadic quantifiers. (∀α:φ)ψ, (∃α:φ)ψ, and (the α:φ)ψ become, respectively:
∀α(φ→ψ)
∃α(φ∧ψ)
∃α(φ ∧ ∀β(φβ →β=α) ∧ ψ)
(where φβ is φ with free αs changed to βs.) But, as with the monadic quantifiers
∃∞ and most, there are binary quantifiers one can introduce that genuinely
increase expressive power. For example, most occurrences of ‘most’ in English
are binary, e.g.:
2
Westerståhl (1989, p. 29).
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 101
(where gU/π is the variable assignment just like g except in assigning U to π.)
Notice that, as with the generalized monadic quantifiers, no alteration to the
definition of a PC-model is needed. All we need to do is change grammar and
the definition of the valuation function.
Second-order logic is different from first-order logic in many ways. For
instance, one can define the identity predicate in second-order logic:
x=y =df ∀X (X x↔Xy)
This can be seen to work correctly as follows. A one-place second-order variable
X gets assigned a set of things. Thus, the atomic sentence X x says that the
object (currently assigned to) x is a member of the set (currently assigned to)
X . Thus, ∀X (X x↔Xy) says that x and y are members of exactly the same
sets. But since x and only x is a member of {x} (i.e., x’s unit set), that means
that the only way for this to be true is for y to be identical to x.
More importantly, the metalogical properties of second-order logic are dra-
matically different from those of first-order logic. For instance, the axiomatic
method cannot be fully applied to second order logic. One cannot write down a
set of axioms for second-order logic that are both sound and complete—unless,
that is, one resorts to cheap tricks like saying “let every valid wff be an axiom”.
This trick is “cheap” because one would have no way of telling what an axiom
is. (More precisely, the resulting set of axioms would fail to be “recursive”.3 )
Second-order logic also allows us to express claims that cannot be expressed
in first-order logic. Consider the “Geach-Kaplan sentence”:4
(GK) Some critics admire only one another
It can be shown that there is no way to symbolize (one reading of) the sentence
using just first-order logic and predicates for ‘is a critic’ and ‘admires’. The
sentence (on the desired reading) says that there is a group of critics such that
the members of that group admire only other members of the group, but one
cannot say this in first-order logic. However, the sentence can be symbolized
in second-order logic:
3
For a rigorous statement and proof of this and other metalogical results about second-order
logic, see Boolos and Jeffrey (1989, chapter 18).
4
The sentence and its significance were discovered by Peter Geach and David Kaplan. See
Boolos (1984).
CHAPTER 5. EXTENSIONS OF PREDICATE LOGIC 102
(GK2 ) “symbolizes” (GK) in the sense that it contains no predicates other than
C and A, and for every model 〈D, I 〉, the following is true:
(GK1 ) doesn’t “symbolize” (GK) in the sense of satisfying (*) in every model,
for in some models the two-place predicate ∈ doesn’t mean set-membership.
Nevertheless, if we just restrict our attention to models 〈D, I 〉 in which ∈ does
mean set-membership (restricted to the model’s domain, of course—that is,
I (∈) = {〈u, v〉 : u, v ∈ D and u ∈ v}), and in which D contains each subset of
I (C ) as a member, then (GK1 ) will indeed satisfy (*). In essence, the difference
between (GK1 ) and (GK2 ) is that it is hard-wired into the definition of truth in
a model that second-order predications Xy express set-membership, whereas
this is not hard-wired into the definition of the first-order predication y ∈ z.5
5
For more on second-order logic, see Boolos (1975, 1984, 1985).
Chapter 6
2φ: “It is necessary that φ”, “Necessarily, φ”, “It must be that φ”
3φ: “It is possible that φ”, “Possibly, φ”, “It could be that φ”, “It
can be that φ”, “It might be that φ”
103
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 104
worlds in which I am ten feet tall, and so on. “Complete” means simply that no
detail is left out—possible worlds are completely specific scenarios. There is no
possible world in which I am “somewhere between 10 and 11 feet tall” without
being some particular height.1 Likewise, in any possible world in which I am
exactly 10 feet, six inches tall (say), I must have some particular weight, must
live in some particular place, and so on. One of these possible worlds is the
actual world—this is the complete and possible scenario that in fact obtains.
The rest of them are merely possible—they do not obtain, but would have
obtained if things had gone differently.
In terms of possible worlds, we can think of our modal operators thus:
It is necessarily true that all bachelors are male; in every possible world, every
bachelor is male. There might have existed a talking donkey; some possible
world contains a talking donkey. Possible worlds provide, at the very least,
a vivid way to think about necessity and possibility. How much more than a
vivid guide they provide is an open philosophical question. Some maintain that
possible worlds are the key to the metaphysics of modality, that what it is for a
proposition to be necessarily true is for it to be true in all possible worlds.2
Whether this view is defensible is a question beyond the scope of this book;
what is important for present purposes is that we distinguish possible worlds as
a vivid heuristic from possible worlds as a concern in serious metaphysics.
Our first topic in modal logic is the addition of the 2 and the 3 to proposi-
tional logic; the result is modal propositional logic (“MPL”). A further step will be
be modal predicate logic (chapter 9).
Primitive vocabulary:
1
This is not to say that possible worlds exclude vagueness.
2
Sider (2003) presents an overview of this topic.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 105
Definition of wff:
The 2 is the only new primitive connective. But just as we were able to
define ∧, ∨, and ↔, we can define new nonprimitive modal connectives:
B→2U
But since I am in fact a bachelor, it would follow from this symbolization that
the proposition that I am unmarried is necessarily true. But clearly I am not
necessarily a bachelor—I could have been married! The sentence is not saying
that if I am in fact a bachelor, then the following is a necessary truth: I am
married. It is rather saying that, necessarily, if I am a bachelor then I am
married:
2(B→U )
I can come to the party, but I can’t stay late. (“can” = “is not inconve-
nient”)
Humans can travel to the moon, but not Mars. (“can” = “is achievable
with current technology”)
Objects can move almost as fast as the speed of light, but nothing can travel
faster than light. (“can” = “is consistent with the laws of nature”)
Objects could have traveled faster than the speed of light (if the laws
of nature had been different), but no matter what the laws had been,
nothing could have traveled faster than itself. (“can” = “metaphysical
possibility”)
You can borrow but you can’t steal. (“can” = “morally acceptable”)
So when representing English sentences using the 2 and the 3, one should
keep in mind that these expressions can be used to express different strengths
of necessity and possibility. (Though we won’t do this, one could introduce
different symbols for different sorts of possibility and necessity.)
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 108
2 φ
? 1
0 0
There’s nothing we can put in this slot in the truth table, since when φ is true,
sometimes 2φ is true and sometimes it is false.
Our challenge is clear: we need a semantics for the 2 and the 3 other than
the method of truth tables.
6.3.1 Relations
Before we investigate how to overcome this challenge, a digression is necessary.
Recall our discussion of ordered sets from section 1.8. In addition to their use
in constructing models, ordered sets are also useful for constructing relations.
We take a binary (2-place) relation to be a set of ordered pairs. For example,
the taller-than relation may be taken to be the set of ordered pairs 〈u, v〉 such
that u is taller than v. The less-than relation for positive integers is the set
of ordered pairs 〈m, n〉 such that m is a positive integer less than n, another
positive integer. That is, it is the following set:
of predicate logic; but here I’m speaking the metalanguage, not displaying
sentences of a formalized language.)
Some definitions. Let R be any binary relation.
In other words, the domain of R is the set of all things that bear R to something;
the range is the set of all things that something bears R to; and R is over A iff
the members of the ’tuples in R are all drawn from A.
Let R be any binary relation over A. Then we define the following notions
with respect to A:
What about the truth values for complex formulas that contain ∧, ∨, ↔, and
3? Given the definition of these defined connectives in terms of the primitive
connectives, it is easy to prove that the following derived conditions hold:
i) VM (φ∧ψ, w) = 1 iff VM (φ, w) = 1 and VM (ψ, w) = 1
ii) VM (φ∨ψ, w) = 1 iff VM (φ, w) = 1 or VM (ψ, w) = 1
iii) VM (φ↔ψ, w) = 1 iff VM (φ, w) = VM (ψ, w)
iv) VM (3φ, w) = 1 iff for some v ∈ W , R wv and VM (φ, v) = 1
So far, we have introduced a general notion of an MPL model, and have
defined the notion of a wff’s being true at a world in an MPL model. Next, let
us consider how to define validity.
Remember that our overall strategy is C. I. Lewis’s: we want to construct
different modal systems, since it isn’t obvious which formulas ought to count as
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 114
logical truths. The systems will be named: K, D, T, B, S4, S5. Each system will
come with its own definition of a model. As a result, different formulas will
come out valid in the different systems. For example, as we’ll see, the formula
2P →22P is going to come out valid in S4 and S5, but not in the other systems.
Here are the definitions:
Notice that for each system, the valid formulas are defined as the formulas
that are valid in every model in which the accessibility relation has a certain
formal feature. The systems differ from from one another by what that formal
feature is. For T it is reflexivity: a formula is T-valid iff it is valid in every
model in which the accessibility relation is reflexive. For S4 the formal feature
is reflexivity + transitivity. Other systems correspond to other formal features.
As before, we’ll use the notation for validity. But since we have many
modal systems, if we claim that a formula is valid, we’ll need to indicate which
system we’re talking about. Let’s do that by subscripting with the name of
the system; thus, “T φ” means that φ is T-valid.
It’s important to get clear on the status of possible-worlds lingo here. Where
〈W , R, I 〉 is a model, we call the members of W “worlds”, and we call R the
“accessibility” relation. Now, there is no question that “possible worlds” is a
vivid way to think about necessity and possibility. But officially, W is nothing
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 115
but a nonempty set, any old nonempty set. Its members needn’t be the kinds
of things metaphysicians call possible worlds: they can be numbers, people,
bananas—whatever you like. Similarly, R is just defined to be any old binary
relation on R; it needn’t have anything to do with the metaphysics of modality.
Officially, then, the possible-worlds talk we use to describe our models is just
talk, not heavy-duty metaphysics. Still, models are usually intended to model
something—to depict some aspect of the dependence of truth on the world.
So if modal sentences of English containing ‘necessarily’ and ‘possibly’ aren’t
made true by anything like possible worlds, it’s hard to see why possible worlds
models would shed any light on their meaning, or why truth-in-all-possible-
worlds-models would be a good way of modeling (genuine) validity for modal
statements. At any rate, this philosophical issue should be kept in mind. Back,
now, to the formalism.
Thus, K 2(P ∨∼P ). Note that similar reasoning would establish K φ, for
any propositional-logic tautology φ. The reason is this: within any world, the
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 116
OK, we’ve shown that the formula (32(P →Q)∧2P )→3Q is valid in T.
Suppose we were interested in showing that this formula is also valid in S4.
What more would we have to do? Nothing! To be S4-valid is to be valid in
every S4-model; but a quick look at the definitions shows that every S4-model
is a T-model. So, since we already know that the the formula is valid in all
T-models, we already know that it must be valid in all S4-models (and hence,
S4-valid), without doing a separate proof.
Think of it another way. To do a proof that the formula is S4-valid, we need
to do a proof in which we are allowed to assume that the accessibility relation is
both transitive and reflexive. And the proof above did just that. We didn’t ever
use the fact that the accessibility relation is transitive—we only used the fact
that it is reflexive (in line 9). But we don’t need to use everything we’re allowed
to assume.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 117
In contrast, the proof above doesn’t establish that this formula is, say, K-valid.
To be K-valid, the formula would need to be valid in all models. But some
models don’t have reflexive accessibility relations, whereas the proof we gave
assumed that the accessibility relation was reflexive. And in fact the formula
isn’t in fact K-valid, as we’ll show how to demonstrate in the next section.
Consider the following diagram of systems:
= S5 `@@
||| @@
| @@
||
S4 aB ~>
B
BB
BB ~~~
B ~~
TO
DO
An arrow from one system to another indicates that validity in the first system
implies validity in the second system. For example, if a formula is D-valid, then
it’s also T-valid. The reason is that if something is valid in all D-models, then,
since every T-model is also a D-model (since reflexivity implies seriality), it
must be valid in all T-models as well.
S5 is the strongest system, since it has the most valid formulas. (That’s
because it has the fewest models—it’s easier to be S5-valid because there are
fewer potentially falsifying models.)
Notice that the diagram isn’t linear. That’s because of the following. Both B
and S4 are stronger than T; each contains all the T-valid formulas. But neither
B nor S4 is stronger than the other—each contains valid formulas that the
other doesn’t. (They of course overlap, because each contains all the T-valid
formulas.) S5 is stronger than each; S5 contains all the valid formulas of each.
These relationships between the systems will be exhibited below.
Suppose you are given a formula, and for each system in which it is valid,
you want to give a semantic proof of its validity. This needn’t require multiple
semantic proofs—as we have seen, one semantic proof can do the job. To prove
that a certain formula is valid in a number of systems, it suffices to prove that it
is valid in the weakest possible system. Then, that very proof will automatically
be a proof that it is valid in all stronger systems. For example, a proof that a
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 118
6.3.4 Countermodels
We have a definition of validity for the various systems, and we’ve shown how
to establish validity of particular formulas. Now we’ll investigate establishing
invalidity.
Let’s show that the formula 3P →2P is not K-valid. A formula is K-valid if
it is valid in all K-models, so all we must do is find one K-model in which it
isn’t valid. What follows is a procedure for doing this:4
r 3P →2P
4
This procedure is from Cresswell and Hughes (1996).
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 119
Now, since the box represents a world, we should have some way of representing
the accessibility relation. What worlds are possible, relative to r; what worlds
does r “see”? Well, to represent one world (box) seeing another, we’ll draw
an arrow from the first to the second. However in the case of this particular
model, we don’t need to make this world r see anything. After all, we’re trying
to construct a K-model, and the accessibility relation of a K-model doesn’t
even need to be serial—no world needs to see any worlds at all. So, we’ll forget
about arrows for the time being.
0
3P →2P
r
1 0 0
3P →2P
r
Enter asterisks
When we assign a truth value to a modal formula, we thereby commit ourselves
to assigning certain other truth values to various formulas at various worlds.
For example, when we make 3P true at r, we commit ourselves to making P
true at some world that r sees. To remind ourselves of this commitment, we’ll
put an asterisk (*) below 3P . An asterisk below indicates a commitment to there
being some world of a certain sort. Similarly, since 2P is false at r, this means
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 120
that P must be false in some world P sees (if it were true in all such worlds,
then by the semantic clause for the 2, 2P would be true at r). We again have a
commitment to there being some world of a certain sort, so we enter an asterisk
below 2P as well:
1 0 0
r 3P →2P
∗ ∗
1 0 0
r 3P →2P
∗ ∗
??
??
??
??
??
a 1 0
b
P P
What I’ve done is added two more worlds to the diagram: a and b. P is true in
a, but false in b. I have thereby satisfied my obligations to the asterisks on my
diagram, for r does indeed see a world in which P is true, and another in which
P is false.
The set of worlds We first must specify the set of worlds, W . W is simply
the set of worlds I invoked:
W = {r, a, b}
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 121
But what are r, a, and b? Let’s just take them to be the letters ‘r’, ‘a’, and ‘b’. No
reason not to—the members of W , recall, can be any things whatsoever.
That is, we write out the set of all ordered pairs 〈w1 , w2 〉 such that w1 “sees”
w2 .
I (P, a) = 1
I (P, b) = 0
for all other sentence letters α and worlds w, I (α, w) = 0
down is reflexive. (In our case, we were only trying to construct a K-model, and
so for us this step is trivial.) Secondly, make sure that the formula in question
really does come out false at one of the worlds in your model.
Simplifying models
Sometimes a model can be simplified. Consider the diagram of the final version
of the model above:
1 0 0
r 3P →2P
∗ ∗
??
??
??
??
??
a 1 0
b
P P
We needn’t have used three worlds in the model. When we discharged the first
asterisk, we needed to put in a world that r sees, in which P is true. But we
needn’t have made that a new world—we could have simply have made P true in
r. Of course we couldn’t haven’t done that for both asterisks, because that would
have made P both true and false at r. So, we could make one simplification:
1 1 0 0
r 3P →2P
∗ ∗
0
0
b
P
The official model would then look as follows:
W : {r, b}
R : {〈r, r〉, 〈r, b〉}
I (P, r) = 1; otherwise, everything false
accessibility relation (i.e., some “D-model”). Well, we haven’t quite done this,
since the model above does not have a serial accessibility relation. But we can
easily change this, as follows:
1 1 0 0
r 3P →2P
∗ ∗
0
0
b
0 P
Official model:
W : {r, b}
R : {〈r, r〉, 〈r, b〉, 〈b, b〉}
I (P, r) = 1; otherwise, everything false
That was easy—adding the fact that b sees itself didn’t require changing any-
thing else in the model.
Suppose we want now to show that 3P →2P isn’t T-valid. Well, we’ve
already done so! Why? Because we’ve already produced a T-model in which
this formula is false. Look back at the most recent model. Its accessibility
relation is reflexive. So it’s a T-model already. In fact, that accessibility relation
is also already transitive, so it’s already an S4-model.
So far we have established that 2K,D,T,S4 3P →2P . What about B and S5?
It’s easy to revise our model to make the accessibility relation symmetric:
1 1 0 0
r 3P →2P
∗ ∗
0 O
0
b
0 P
Official model:
W : {r,b}
R: {〈r,r〉,〈r,b〉,〈b,b〉,〈b,r〉}
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 124
= S5 `@@
||| @@
| @@
||
S4 aB ~>
B
BB
BB ~~~
B ~~
TO
DO
K
An arrow from one system to another, recall, indicates that validity in the first
system implies validity in the second. The arrows also indicate facts about
invalidity, but in reverse: when an arrow points from one system to another,
then invalidity in the second system implies invalidity in the first. For example,
if a wff is invalid in T, then it is invalid in D. (That’s because every T-model is
a D-model; a countermodel in T is therefore a countermodel in D.)
When our task is to discover which systems a given formula is invalid in,
usually only one countermodel will be needed—a countermodel in the strongest
system in which the formula is invalid. But there is an exception involving B
and S4. Suppose a given formula is valid in S5, but we discover a model showing
that it isn’t valid in B. That model is automatically a T, D, and K-model, so we
know that the formula isn’t T, D, or K-valid. But we don’t yet know about that
formula’s S4-validity. If it is S4-invalid, then we will need to produce a second
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 125
Above asterisks Let’s try to get a countermodel for 32P →23P in all the
systems in which it is invalid, and a semantic validity proof in all the systems in
which it is valid. We always start with countermodelling before doing semantic
validity proofs, and when doing countermodelling, we start by trying for a
K-model. After the first few steps, we have:
1 0 0
r 32P →23P
∗ ∗
CC
{{ CC
{{{ CC
{{ CC
{{ CC
{
} { C!
1 0
2P 3P
a b
At this point, we’ve got a true 2, and a false 3. Take the first: a true 2P . This
doesn’t commit us to adding a world in which P is true; rather, it commits us
to making P true in every world that a sees. Similarly, a zero over a 3, over
3P in world b in this case, commits us to making P false in every world that
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 126
1 0 0
r 32P →23P
∗ ∗
@@
~~ @@
~~~ @@
~~ @@
~~ @@
~ @@
∗ ~
~ ∗
a 1 b 0
2P 3P
Now, how can we discharge these asterisks? In this case, when trying to
construct a K-model, we don’t need to do anything. Since a, for example,
doesn’t see any world, then automatically P is true in every world it sees; the
statement “for every world, w, if Raw then V(P, w) = 1” is vacuously true.
Same goes for b—P is automatically false in all worlds it sees. So, we’ve got a
K-model in which 32P →23P is true.
Now let’s turn the model into a D-model. Every world must now see at
least one world. Let’s try:
1 0 0
r 32P →23P
∗ ∗
@@
~~ @@
~~~ @@
~~ @@
~~ @@
~ @@
∗ ~
~ ∗
a 1 b 0
2P 3P
1 0
c d
P P
0 0
I added worlds c and d, so that a and b would each see at least one world.
(Further, worlds c and d each had to see a world, to keep the relation serial.
I could have added still more worlds that c and d saw, but then they would
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 127
themselves need to see some worlds…So I just let c and d see themselves.) But
once c and d were added, discharging the upper asterisks in worlds a and b
required making P true in c and false in d (since a sees c and b sees d).
Let’s now try for a T-model. This will involve, among other things, letting
a and b see themselves. But this gets rid of the need for worlds c and d, since
they were added just to make the relation serial. I’ll try:
1 0 0
r 32P →23P
∗ ∗
0~ @@
~~ @@
~ @@
~~ @@
~~~ @@
~ @@
∗ ~
~ ∗
a 1 1 b 0 0
2P 3P
0 0
When I added arrows, I needed to make sure that I correctly discharged the
asterisks. This required nothing of world r, since there were no top asterisks
there. There were top asterisks in worlds a and b; but it turned out to be easy
to discharge these asterisks—I just needed to let P be true in a, but false in b.
Notice that I could have moved straight to this T-model—which is itself a
D-model—rather than first going through the earlier mere-D-model. However,
this won’t always be possible—sometimes you’ll be able to get a D-model, but
no T-model.
At this point let’s verify that our model does indeed assign the value 0 to
our formula 32P →23P . First notice that 2P is true in a (since a only sees
one world—itself—and P is true there). But r sees a. So 32P is true at r. Now,
consider b. b only sees one world, itself, and P is false there. So 3P must also
be false there. But r sees b. So 23P is false at r. But now, the antecedent of
32P →23P is true, while its consequent is false, at r. So that conditional is
false at r. Which is what we wanted.
Onward. Our model is not a B-model, since a, for example, doesn’t see r,
despite the fact that r sees a. So let’s try to make this into a B-model. This
involves making the relation symmetric. Here’s how it looks before I try to
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 128
1 0 0
r 32P →23P
∗ ∗
0~? @_ @
@@
~~~ @@
~~ @@
~~~ @@
~ @@
∗ ~
~ ∗
a 1 1 b 0 0
2P 3P
0 0
Now I need to make sure that all top asterisks are discharged. For example,
since a now sees r, I’ll need to make sure that P is true at r. However, since
b sees r too, P needs to be false at r. But P can’t be both true and false at r.
So we’re stuck, in trying to get a B-model in which this formula is false. This
suggests that maybe it is impossible—that is, perhaps this formula is true in all
worlds in all B-models—that is, perhaps the formula is B-valid. So, the thing
to do is try to prove this: by supplying a semantic validity proof.
So, let 〈W , R, I 〉 be any model in which R is reflexive and symmetric, let
V be its valuation function, and let w be any member of W ; we must show that
V(32P →23P, w) = 1.
W = {r, a, b}
R = {〈r, r〉, 〈a, a〉, 〈b, b〉, 〈r, a〉, 〈r, b〉}
I (P, a) = 1, all others false
∗
1 0 0 0
32P →3232P
r
Notice how commitments ∗
to specific truth values for 1 1 0
2P 232P
different formulas are recorded a
by placing the formulas
side by side in the box ∗
0
∗
0 0 1
32P
b
P
∗
0
0
c
P
0
Official model:
W = {r, a, b, c}
R = {〈r, r〉, 〈a, a〉, 〈b, b〉, 〈c, c〉, 〈r, a〉, 〈r, b〉, 〈a, b〉, 〈b, c〉}
I (P, b) = 1, all else 0
Now consider what happens when we try to turn this model into a B-model.
World b must see back to world a. But then the false 32P in b conflicts with the
true 2P in a. So it’s time for a validity proof. In constructed this validity proof,
we can be guided by failed attempt to construct a countermodel (assuming all
of our choices in constructing that countermodel were forced). In the following
proof that the formula is B-valid, I chose variables for worlds that match up
with the countermodel above:
i) Suppose for reductio that V(32P →3232P, r ) = 0, in some
world r in some B-model 〈W , R, I 〉
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 131
We now have a T-model for the formula, and a proof that it is B-valid. The
B-validity proof shows the formula to be S5-valid; the T-model shows it to
be K- and D-invalid. We still don’t yet know about S4. So let’s return to the
T-model above, and see what happens when we try to make its accessibility
relation transitive. World a must then see world c, which is impossible since
2P is true in a and P is false in c. So we’re ready for a S4-validity proof (the
proof looks like the B-validity proof at first, but then diverges):
Daggers
There’s another kind of step in constructing models. When we make a condi-
tional false, we’re forced to enter certain truth values for its components: 1 for
the antecedent, 0 for the consequent. But consider making a disjunction true. A
disjunction can be true in more than one way. The first disjunct might be true,
or the second might be true, or both could be true. So we have a choice for
how to go about making a disjunction true. Similarly for making a conditional
true, a conjunction false, or a biconditional either true or false.
When one has a choice about which truth values to give the constituents
of a propositional compound, it’s best to delay making the choice as long as
possible. After all, some other part of the model might force you to make one
choice rather than the other. If you investigate the rest of the countermodel,
and nothing has forced your hand, you may need then to make a guess: try one
of the truth value combinations open to you, and see whether you can finish
the countermodel. If not, go back and try another combination.
To remind ourselves of these choice points, we will place a dagger (†) un-
derneath the major connective of the formula in question. Consider, as an
example, constructing a countermodel for the formula 3(3P ∨2Q)→(3P ∨Q).
Throwing caution to the wind and going straight for a T-model, we have after
a few steps:
∗
1 0 0 0 0 0
r
3(3P ∨2Q)→(3P ∨ Q)
∗
0
1 0
a 3P ∨2Q P
†
0
We still have to decide how to make 3P ∨2Q true in world a: which disjunct
to make true? Well, making 2P true won’t require adding another world to
the model, so let’s do that. We have, then, a T-model:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 133
∗
1 0 0 0 0 0
r
3(3P ∨2Q)→(3P ∨ Q)
∗
0
∗
1 1 1 0
3P ∨2Q
a
P
†
0
W : {r, a}
R : {〈r, r〉, 〈a, a〉, 〈r, a〉}
I (Q, a) = 1, all else 0
OK, let’s try now to upgrade this to a B-model. We can’t simply leave
everything as-is while letting world a see back to world r, since 2Q is true
in a and Q is false in r. But there’s another possibility. We weren’t forced to
discharge the dagger in world a by making 2Q true. So let’s explore the other
possibility; let’s make 3P true:
∗
1 0 0 0 0 0
r
3(3P ∨2Q)→(3P ∨ Q)
∗
0 O
1 1 0
a 3P ∨2Q P
∗ †
0 O
1
b
P
0
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 134
W : {r, a, b}
R : {〈r, r〉, 〈a, a〉, 〈b, b〉, 〈r, a〉, 〈a, r〉, 〈a, b〉, 〈b, a〉}
I (P, b) = 1, all else 0
Summary of steps
Here, then, is a final list of the steps for constructing countermodels:
When we defined the notion of validity, what we defined was the notion of
a valid formula. (That’s because validity is defined in terms of truth in a model,
which itself was defined only for formulas.) So it’s not, strictly speaking, correct
to apply the notions of validity or invalidity to schemas.
However, it’s often interesting to show that every instance of a given schema
is valid. (Instances are schemas are formulas, and so the notion of validity can
be properly applied to them.) It’s easy, for example, to show that every instance
of the schema 2(φ→φ) is valid in each of our modal systems. (Let φ be any
MPL-wff, and take any world w in any model. Since the rules for evaluating
propositional compounds within possible worlds are the classical ones, φ→φ
must be true at w, no matter what truth value φ has at w. Hence 2(φ→φ) is
true in any world in any model, and so is valid in each system.)
There is, therefore, a kind of indirect notion of schema-validity: validity
of all instances. How about the invalidity of schemas? Here we must take
great care. In particular, the notion of a schema, all of whose instances are
invalid, is not a particularly interesting notion. Take, for instance, the schema
3φ→2φ. We showed earlier that a certain instance of this schema, namely
3P →2P is invalid in each of our systems. However, the schema 2φ→3φ also
has plenty of instances that are valid in various systems. The following formula,
for example, is an instance of 3φ→2ψ, and can easily be shown to be valid in
each of our systems:
3(P →P )→2(P →P )
Thus, even intuitively terrible schemas like 3φ→2φ have some valid instances.
(For an extreme example of this, consider the schema φ. Even this has some
valid instances: P →P , for one.) So it’s not interesting to inquire into whether
each instance of a schema is invalid. What is interesting is to inquire into
whether a given schema has some instances that are invalid. We can show, for
example, that the schema 3φ→2φ has some invalid instances (3P →2P , for
one), and hence is in this way unlike the schema 2(φ→φ).
So when dealing with schemas, it will often be of interest to ascertain
whether each instance of the schema is valid; it will rarely (if ever) be of interest
to ascertain whether each instance of the schema is invalid.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 137
6.4.1 System K
Our first system, K, is the weakest system—i.e., the system with the fewest
theorems.
Let’s investigate what one can prove in K. The simplest sort of distinctively
modal proof consists of first proving something from the PL axioms, and then
necessitating it, as in the following proof of 2((P →Q)→(P →P ))
1. P →(Q→P ) (A1)
2. P →(Q→P ))→((P →Q)→(P →P )) (A2)
3. (P →Q)→(P →P ) 1,2 MP
4. 2((P →Q)→(P →P )) 3, NEC
Using this technique, we can prove anything of the form 2φ, where φ is
provable in PL. And, since the PL axioms are complete (I mentioned but did
not prove this fact in chapter ??), that means that we can prove 2φ whenever
φ is a tautology—i.e., a valid wff of PL. But constructing proofs from the PL
axioms is a pain in the neck!—and anyway not what we want to focus on in
this chapter. So let’s introduce the following time-saving shortcut. Instead of
writing out proofs of tautologies, let’s instead allow ourselves to write any PL
tautology at any point in a proof, annotating simply “PL”.5 Thus, the previous
proof could be shortened to:
1. (P →Q)→(P →P ) PL
2. 2((P →Q)→(P →P )) 1, NEC
2P →2P PL
Why am I making such a fuss about this? Didn’t I just say in the previous
paragraph that we can write down any tautology at any time, with the annotation
“PL”? Well, strictly speaking, 2P →2P isn’t a tautology. A tautology is a valid
5
How do you know whether something is a tautology? Figure it out any way you like: do a
truth table, or a natural deduction derivation—whatever.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 140
wff of PL, and 2P →2P isn’t even a wff of PL (since it contains a 2). But it
is the result of beginning with some PL-tautology (Q→Q, in this case) and
uniformly changing sentence letters to chosen modal wffs (in this case, Qs to
2P s); hence any proof of the PL tautology may be converted into a proof of
it; hence the “PL” annotation is just as justified here as it is in the case of a
genuine tautology. So in general, MPL wffs that result from PL tautologies in
this way may be written down and annotated “PL”.
Back to investigating what we can prove in K. As we’ve seen, we can prove
that tautologies are necessary—we can prove 2φ whenever φ is a tautology.
One can also prove in K that contradictions are impossible. For instance,
∼3(P ∧∼P ) is a theorem of K:
1. ∼(P ∧∼P ) PL
2. 2∼(P ∧∼P ) 1, NEC
3. 2∼(P ∧∼P )→∼∼2∼(P ∧∼P ) PL
4. ∼∼2∼(P ∧∼P ) 2, 3, MP
1. ∼(P ∧∼P ) PL
2. 2∼(P ∧∼P ) 1, NEC
3. ∼∼2∼(P ∧∼P ) 2, PL
So far our modal proofs have only used necessitation and the PL axioms.
What about the K-axioms? The point of the K-schema is to enable “distribution
of the 2 over the →”. That is, if you ever have the formula 2(φ→ψ), then you
can always move to 2φ→2ψ as follows:
2(φ→ψ)
2(φ→ψ)→(2φ→2ψ) K axiom
2φ→2ψ MP
Distribution of the 2 over the →, plus the rule of necessitation, combine
to give us a powerful proof strategy. Whenever one can prove the conditional
φ→ψ, then one can prove the modal conditional 2φ→2ψ as well, as follows.
First prove φ→ψ, then necessitate it to get 2(φ→ψ), then distribute the 2
over the arrow to get 2φ→2ψ. This procedure is one of the core K-strategies,
and is featured in the following proof of 2(P ∧Q)→(2P ∧2Q):
1. (P ∧Q)→P PL
2. 2[(P ∧Q)→P ] NEC
3. 2[(P ∧Q)→P ]→[2(P ∧Q)→2P ] K axiom
4. 2(P ∧Q)→2P 3,4 MP
5. 2(P ∧Q)→2Q Insert steps similar to 1-4
6. 2(P ∧Q)→(2P ∧2Q) 4,5, PL
Notice that the preceding proof, like all of our proofs since we introduced
the time-saving shortcuts, is not a K-proof in the official defined sense. Lines 1,
5, and 6 are not axioms, nor do they follow from earlier lines by MP or NEC;
similarly for line 6.6 So what kind of “proof” is it? It’s a metalanguage proof:
an attempt to convince the reader, by acceptable standards of rigor, that some
real K-proof exists. A reader could use this metalanguage proof as a blueprint
for constructing a real proof. She would begin by replacing line 1 with a proof
from the PL axioms of the conditional (P ∧Q)→P . (As we know from chapter
??, this could be a real pain in the neck!—but the completeness of PL assures us
that it is possible.) She would then replace line 5 with lines parallel to lines 1-4,
but which begin with a proof of (P ∧Q)→Q rather than (P ∧Q)→P . Finally,
in place of line 6, she would insert a proof from the PL axioms of the sen-
tence (2(P ∧Q)→2P )→[(2(P ∧Q)→2Q)→(2(P ∧Q)→(2P ∧2Q))], and then
use modus ponens twice to infer 2(P ∧Q)→(2P ∧2Q).
Another example: (2P ∨2Q)→2(P ∨Q):
6
A further (even pickier) reason: the symbol ∧ isn’t allowed in wffs; the sentences in the
proof are mere abbreviations for official MPL-wffs.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 142
1. P →(P ∨Q) PL
2. 2(P →(P ∨Q)) 1, NEC
3. 2(Q→(P ∨Q)) PL, NEC
4. 2P →2(P ∨Q) 2, K
5. 2Q→2(P ∨Q) 3, K
6. (2P ∨2Q)→2(P ∨Q) 4,5 PL
Here I’ve introduced another time-saving shortcut that I’ll use more and more
as we progress: doing two (or more) steps at once. Line 3 is really short for:
One further comment about this last proof: it illustrates a strategy that is
common in modal proofs. We were trying to prove a conditional formula
whose antecedent is a disjunction of two modal formulas. But the modal
techniques we had developed didn’t deliver formulas of this form. They only
showed us how to put 2s in front of PL-tautologies, and how to distribute 2s
over →s. They only yield formulas of the form 2φ and 2φ→2ψ, whereas the
formula we were trying to prove looks different. To overcome this problem,
what we did was to use the modal techniques to prove two conditionals, namely
2P →2(P ∨Q) and 2Q→2(P ∨Q), from which the desired formula, namely
(2P ∨2Q)→2(P ∨Q), follows by propositional logic. The trick, in general, is
this: remember that you have PL at your disposal. Simply look for one or
more modal formulas you know how to prove which, by PL, imply the formula
you want. Assemble the desired formulas, and then write down your desired
formula, annotating “PL”. In doing so, it may be helpful to recall PL inferences
like the following:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 143
φ→ψ
φ→(ψ→χ )
ψ→φ
(φ∧ψ)→χ
φ↔ψ
φ→ψ φ→χ
φ→χ ψ→χ
φ→(ψ∧χ ) (φ∨ψ)→χ
φ→ψ φ→∼ψ
∼φ∨ψ ∼(φ∧ψ)
The next example illustrates our next major modal proof technique: com-
bining two 2 statements to get a single 2 statement. Let us construct a K-proof
of (2P ∧2Q)→2(P ∧Q):
1. P →(Q→(P ∧Q)) PL
2. 2[P →(Q→(P ∧Q))] NEC
3. 2P →2(Q→(P ∧Q)) 2, K
4. 2(Q→(P ∧Q))→[2Q→2(P ∧Q)] K axiom
5. 2P →[2Q→2(P ∧Q)] 3,4 PL
6. (2P ∧2Q)→2(P ∧Q) 5, PL
(If you wanted to, you could skip step 5, and just go straight to 6 by propositional
logic, since 6 is a propositional logical consequence of 3 and 4; I put it in for
perspicuity.)
The general technique illustrated by the last problem applies anytime you
want to move from several 2 statements to a further 2 statement, where the in-
side parts of the first 2 statements imply the inside part of the final 2 statement.
More carefully: it applies whenever you want to prove a formula of the form
2φ1 →(2φ2 → · · · (2φn →2ψ) . . . ), provided you are able to prove the formula
φ1 →(φ2 → · · · (φn →ψ) . . . ). (The previous proof was an instance of this because
it involved moving from 2P and 2Q to 2(P ∧Q); and this is a case where one
can move from the inside parts of the first two formulas (namely, P and Q), to
the inside part of the third formula (namely, P ∧Q)—by PL.) To do this, one
begins by proving the conditional φ1 →(φ2 → · · · (φn →ψ) . . . ), necessitating it to
get 2[φ1 →(φ2 → · · · (φn →ψ) . . . )], and then distributing the 2 over the arrows
repeatedly using K-axioms and PL to get 2φ1 →(2φ2 → · · · (2φn →2ψ) . . . ).
One cautionary note in connection with this last proof. One might think to
make it more intuitive by using conditional proof:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 144
But this is not a legal proof, since our axiomatic system allows neither assump-
tions nor conditional proof.
In fact, our decision to omit conditional proof was not at all arbitrary. Given
our rule of necessitation, we couldn’t add conditional proof to our system. If we
did, proofs like the following would become legal:
Thus, P →2P would turn out to be a K-theorem. But we don’t want that: after
all, a statement P might be true without being necessarily true.
Once we have a soundness proof (section 6.5), we’ll be able to show that
P →2P isn’t a K-theorem. But as we just saw, one can construct a K-proof
from {P } of 2P (recall the notion of a proof from a set Γ, from section 2.5.)
It follows that the deduction theorem (section 2.5.2), which says that if there
exists a proof of ψ from {φ}, then there exists a proof of φ→ψ, fails for K (it
likewise fails for all the modal systems we will consider.) So there will be no
conditional proof in our axiomatic modal systems. (Of course, to convince
yourself that a given formula is really a tautology of propositional logic, you
may sketch a proof of it to yourself using conditional proof in some standard
natural deduction system for nonmodal propositional logic; and then you may
write that formula down in one of our axiomatic MPL proofs, annotating “PL”.)
Back to techniques for constructing proofs in K. The following proof of
22(P ∧Q)→22P illustrates a technique for proving formulas with “nested”
modal operators:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 145
1. (P ∧Q)→P PL
2. 2(P ∧Q)→2P 1, NEC, K
3. 2[2(P ∧Q)→2P ] 2, NEC
4. 22(P ∧Q)→22P 3, K
Notice in line 3 that we necessitated something that was not a PL theorem.
That’s ok; we’re allowed to necessitate any K-theorems, even those whose proofs
were distinctly modal. Notice also how this proof contains two instances of our
basic K-strategy. This strategy involves obtaining a conditional, necessitating
it, then distributing the 2 over the →. We did this first using the conditional
(P ∧Q)→P ; that led us to a conditional, 2(P ∧Q)→2P . Then we started the
strategy over again, using this as our initial conditional.
So far we have no techniques dealing with the 3, other than eliminating it
by definition. It will be convenient to derive some shortcuts. For one, there
are the following theorem schemas, which may collectively be called “modal
negation”, or “MN” for short:
`K ∼2φ→3∼φ `K 3∼φ→∼2φ
`K ∼3φ→2∼φ `K 2∼φ→∼3φ
I’ll do one of these; the rest can be an exercise:
1. ∼∼φ→φ PL
2. 2∼∼φ→2φ 1, NEC, K
3. ∼2φ→∼2∼∼φ 2, PL
The final line, 3, is the definitional equivalent of ∼2φ→3∼φ.
It will also be worthwhile to know that an analog of the K axiom for the 3
is a K-theorem:
“K3”: 2(φ→ψ)→(3φ→3ψ)
K3 is, by definition of the 3, the same formula as:
2(φ→ψ)→(∼2∼φ→∼2∼ψ):
How are we going to construct a K-proof of this theorem? In a natural de-
duction system we would use conditional proof and reductio ad absurdum,
but these strategies are not available to us here. What we must do instead is
look for a formula we know how to prove in K, which is PL-equivalent to the
formula we want to prove. Here is such a formula:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 146
2(φ→ψ)→(2∼ψ→2∼φ)
This is equivalent, given PL, to what we want to show; and it looks like the
result of necessitating a tautology and then distributing the 2 over the → a
couple times—just the kind of thing we know how to do in K. Here, then, is
the desired proof of K3:
1. (φ→ψ)→ (∼ψ→∼φ) PL
2. 2(φ→ψ)→2(∼ψ→∼φ) 1, NEC, K
3. 2(∼ψ→∼φ)→(2∼ψ→2∼φ) K
4. 2(φ→ψ)→(2∼ψ→2∼φ) 2,3 PL
5. 2(φ→ψ)→(∼2∼φ→∼2∼ψ) 4,PL
In doing proofs, let’s also allow ourselves to refer to earlier theorems proved,
rather than repeating their proofs. The importance of K3 may be illustrated
by the following proof of 2P →(3Q→3(P ∧Q)):
1. P →[Q→(P ∧Q)] PL
2. 2P →2[Q→(P ∧Q)] 1, NEC, K
3. 2[Q→(P ∧Q)]→[3Q→3(P ∧Q)] K3
4. 2P →[3Q→3(P ∧Q)] 2,3, PL
In other words, one must swap one of the other φi s (I arbitrarily chose φn )
with φ3 . What one obtains at the end will therefore have the modal statements
out of order:
But that problem is easily solved; this is equivalent in PL to what we’re trying
to get. (Recall that φ→(ψ→χ ) is logically equivalent in PL to ψ→(φ→χ ).)
Why do we need to save K3 for last? The strategy of successively distribut-
ing the box over all the nested conditionals comes to a halt as soon as the K3
theorem is used. Let me illustrate with an example. Suppose we wish to prove
`K 3P →(2Q→3(P ∧Q)). We might think to begin as follows:
1. P →(Q→(P ∧Q)) PL
2. 2[P →(Q→(P ∧Q))] 1, Nec
3. 3P →3(Q→(P ∧Q)) K3
4. ?
But neither K nor K3 gets us this. The remedy is to begin the proof with a
different conditional:
6.4.2 System D
System D results from adding a new axiom schema to system K:
System D
Notice that since D includes all the K axioms, we retain all the K-theorems.
The addition of the D-schema just adds more theorems. In fact, all of our
systems will build on K in this way, by adding new axioms to K. (K is so weak
that it isn’t of much independent interest; we study it because it’s the common
basis of all of the other systems we’re studying.)
An example of what we can do now is prove that tautologies are possible:
1. P ∨∼P PL
2. 2(P ∨∼P ) 1, NEC
3. 2(P ∨∼P )→3(P ∨∼P ) D
4. 3(P ∨∼P ) 2,3 MP
Like K, system D is very weak. As we will see later, we can’t prove 2φ→φ
in D. Therefore, D doesn’t seem to be a correct logic for metaphysical, or
nomic, or technological necessity, for surely, if something is metaphysically,
nomically, or technologically necessary, then it must be true. (If something is
true in all metaphysically possible worlds, or all nomically possible worlds, or
all technologically possible worlds, then surely it must be true in the actual
world, and so must be plain old true.) But perhaps there is some interest in
D anyway; perhaps D is a correct logic for moral necessity. Suppose we read
2φ as “One ought to make φ be the case”, and, correspondingly, read 3φ as
“One is permitted to make φ be the case”. Then the fact that 2φ→φ cannot
be proved in D would be a virtue, for from the fact that something ought be
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 149
done, it certainly doesn’t follow that it is done. The D-axiom, on the other
hand, would correspond to the principle that if something ought to be done
then it is permitted to be done, which does seem like a logical truth. But I won’t
go any further into the question of whether D in fact does give a correct logic
for moral necessity. That’s philosophy, not logic.
6.4.3 System T
T is the first system we have considered that has any plausibility of being a
correct logic for a wide range of concepts of necessity (metaphysical necessity,
for example):
Recall that in the case of K, we proved a theorem schema, K3, which was
the analog for the 3 of the K-axiom schema. Let’s do the same thing here; let’s
prove a theorem schema T3, which is the analog for the 3 of the T axiom
schema:
T3: φ→3φ
1. 2∼φ→∼φ T
2. φ→∼2∼φ 1, PL
2 is just the definition of φ→3φ. Thus, we have established that for every
wff φ, `T φ→3φ. So let’s allow ourselves to write down formulas of the form
φ→3φ, annotating simply “T3”.
Notice that instances of the D-axioms are now theorems: 2φ→φ is a T
axiom, we just proved that φ→3φ is a theorem; and from these two by PL we
can prove 2φ→3φ. Thus, T is an extension of D: every theorem of D remains
a theorem of T. (Since D was an extension of K, T too is an extension of K.)
6.4.4 System B
Our systems so far don’t allow us to prove anything interesting about iterated
modalities, i.e., sentences with consecutive boxes or diamonds. Which such
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 150
B3: φ→23φ:
1. 32∼φ→∼φ B
2. φ→∼32∼φ 1, PL
3. ∼32∼φ↔2∼2∼φ MN
4. ∼2∼φ→3φ PL (since 3 abbreviates ∼2∼)
5. 2∼2∼φ→23φ 4, NEC, K, MP
6. φ→23φ 2, 3, 5, PL
6.4.5 System S4
The characteristic axiom of our next system, S4, is a different principle gov-
erning iterated modalities:
7
This is # 41 from the handout “MPL theorems”.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 151
6.4.6 System S5
Here, instead of the B or S4 schemas, we add the S5 schema to T:
S53: 3φ→23φ
1. 32∼φ→2∼φ S5
2. ∼3φ→2∼φ MN
3. 3∼3φ→32∼φ 2, NEC, K3, MP
4. ∼23φ→3∼3φ MN
5. 2∼φ→∼3φ MN
6. 3φ→23φ 4,3,1,5, PL
Next, note that the B and S4 axioms are now derivable as theorems. The B
axiom, 32φ→φ, is trivial:
1. 32φ→2φ S5
2. 2φ→φ T
3. 32φ→φ 1,2 PL
And now the S4 axiom, 2φ→22φ. This is a little harder. I used the B3
theorem, which we can now appeal to since the theoremhood of the B-schema
has been established.
1. 2φ→232φ B3
2. 32φ→2φ S5
3. 2(32φ→2φ) 2, Nec
4. 232φ→22φ 3, K, MP
5. 2φ→22φ 4, 1, PL
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 153
But note that (χ1 β →χ2 β ) is just the same formula as (χ1 →χ2 )β . So
we’ve shown what we wanted to show.
β
Finally, take the third case. We must show that `S 2χ1 ↔2χ1 .
β
This follows from the inductive hypothesis `S χ1 ↔χ1 . For the
β
inductive hypothesis implies `S χ1 →χ1 , by PL; and then, using
β
NEC and a K-axiom, we have `S 2χ1 →2χ1 . A parallel argument
β
establishes `S 2χ1 →2χ1 ; and then the desired conclusion follows
by using PL. That completes the inductive proof.
2∼φ↔∼3φ
3∼φ↔∼2φ
Call these “the duals equivalences”.9 Given the duals equivalences, we can swap
∼3φ and 2∼φ, or ∼2φ and 3∼φ, within any theorem of any system, and
the result will also be a theorem of that system. So we can “move” ∼s through
series of modal operators at will. For example, it’s easy to show that each of the
following is a theorem of each system S:
9
Given the duals equivalences, the 2 is related to the 3 the way the ∀ is related to the
∃ (since ∀x∼φ↔∼∃xφ, and ∃x∼φ↔∼∀xφ are logical truths). This shared relationship,
which holds between the 2 and the 3, and between the ∀ and the ∃, is called “duality”; 2 and
3 are said to be duals, as are ∀ and ∃. This logical analogy would be neatly explained by a
metaphysics according to which necessity just is truth in all worlds and possibility just is truth
in some worlds!
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 155
(Note how this greatly simplifies the process of establishing the existence of
theorems such as K3, T3, B3, S43, and S53!)
Our second meta-theorem concerns only system S5:
That is, whenever a formula has a string of modal operators in front, it is always
equivalent to the result of deleting all the modal operators except the innermost
one. For example, 223232232323φ and 3φ are provably equivalent in
S5; i.e., 223232232323φ↔3φ is a theorem of S5). This follows from
the fact that the following equivalences are all theorems of S5:
a) 32φ↔2φ
b) 22φ↔2φ
c) 23φ↔3φ
d) 33φ↔3φ
The left-to-right direction of a) is just S5; the right-to-left is T3; b) is T and
S4; c) is T and S53; and d) is S43 and T3. Thus, by repeated applications of
these equivalences, using substitution of equivalents, we can reduce strings of
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 156
Our study of modal logic has reversed that of history. We began with
semantics, because that is the more intuitive approach. Historically (as we
noted earlier), the axiomatic systems came first, in the work of C. I. Lewis.
Given the uncertainty over what formulas ought to be counted as axioms, modal
logic was in disarray. The discovery by the teenaged Saul Kripke in the late
1950s of the possible-worlds semantics we studied in section ??, and of the
correspondence between simple constraints (reflexivity, transitivity, etc.) on
the accessibility relation in his models and Lewis’s axiomatic systems, was a
major advance in the history of modal logic.
The soundness and completeness theorems have practical as well as the-
oretical value. First, once we’ve proved soundness, we will for the first time
have a method for establishing that a given formula is not a theorem: construct
a countermodel for that formula, thus establishing that the formula is not valid,
and then conclude via soundness that the formula is not a theorem. Second,
given completeness, if we want to know that a given formula is a theorem, it
10
The modal reduction formula, the duals equivalences, and substitution of equivalents
together let us “reduce” strings of operators that include ∼s as well as modal operators. Simply
use the duals equivalents to drive any ∼s in the string to the far right hand side, then use the
modal reduction theorem to eliminate all but the innermost modal operator.
11
The proofs of soundness and completeness in this and the next section are from Cresswell
and Hughes (1996).
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 157
suffices to show that it is valid. Since semantic validity proofs are comparatively
easy to construct, it’s nice to be able to use them rather than axiomatic proofs.
Let’s begin with soundness. We’re going to prove a general theorem, which
we’ll use in several soundness proofs. First we’ll need a piece of terminology.
Where Γ is any set of modal wffs, let’s call “K+Γ” the axiomatic system that
consists of the same rules of inference as K (MP and NEC), and which has
as axioms the axioms of K (instances of the K- and PL- schemas), plus the
members of Γ. Here, then, is the theorem:
Modal systems of the form K+Γ are commonly called normal. Normal
modal systems contain all the K-theorems, plus possibly more. What Theorem
6.1 gives us is a method for constructing a soundness proof for any normal
system. Since all the systems we have studied here (K, D, etc.) are normal, this
method is sufficiently general for us. Here’s how the method works for system
T. System T has the same rules of inference as K, and its axioms are all the
axioms of K, plus the instances of the T-schema. In the “K+Γ” notation, T = K
+ {2φ→φ : φ is an MPL wff}. To establish soundness for T, all we need to do
is show that every instance of the T-schema is valid in all reflexive models; for
we may then conclude by Theorem 6.1 that every theorem of T is valid in all
reflexive models. This method can be applied to each of our systems: for any
system, S, to establish S’s soundness it will suffice to show that the S’s “extra-K”
axioms are valid in all of the S-models.
To prove Theorem 6.1, we will first prove two lemmas:
Theorem 6.1 follows from the lemmas: assume that every wff in Γ is valid in a
given MPL-model M , and consider any theorem φ of K+Γ. That theorem is
a last line in a proof in which each line is either an axiom K+Γ, or follows from
earlier lines in the proof by MP or NEC. But axioms of K+Γ are either PL
axioms, K axioms, or members of Γ. The first two classes of axioms are valid in
all MPL-models, by Lemma 6.2, and so are valid in M ; and the final class of
axioms are valid in M by hypothesis. Thus, all axioms in the proof are valid
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 158
First MP. Let φ and φ→ψ be valid in model 〈W , R, I 〉; we must show that
ψ is also valid in that model. That is, where V is this model’s valuation, and
w is any member of W , we must show that V(ψ, w) = 1. Since φ and φ→ψ
are valid in this model, V(φ→ψ, w) = 1, and V(φ, w) = 1; but by the truth
condition for →, V(ψ, w) must also be 1.
Next NEC. Suppose φ is valid in model M . We must show that 2φ is
valid in M , i.e., that 2φ is true at each world in M , i.e., that for each world,
w, φ is true at every world accessible from w. But since φ is valid in M , φ is
true in every world in M , and hence is true at every world accessible from w.
6.5.1 Soundness of K
We can now construct soundness proofs for the individual systems. I’ll do this
for some of the systems, and leave the verification of soundness for the other
systems as exercises.
First K. In the “K+Γ” notation, K is just K+∅, and so it follows immediately
from Theorem 6.1 that every theorem of K is valid in every MPL-model. So
K is sound.
6.5.2 Soundness of T
T is K+Γ, where Γ is the set of all instances of the T-schema. So, given Theorem
6.1, to show that every theorem of T is valid in all T-models, it suffices to show
that all instances of the T-schema are valid in all T-models:
6.5.3 Soundness of B
B is K+ Γ, where Γ is the set of all instances of the T- and B- schemas. Given
Theorem 6.1, it suffices to show that every instance of the B-schema and every
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 160
valid in the canonical model for T. But then the italicized statement tells us
that every T-valid formula is a theorem of T. So we would have established
completeness for T.
The trick for constructing canonical models will be to let the worlds in these
models be sets of formulas (remember, worlds are allowed to be anything we
like). And we’re going to construct the interpretation function of the canonical
model in such a way that a formula will be true at a world iff the formula is a
member of the set that is the world. Working out this idea will occupy us for
awhile.
Let’s think for a bit about this definition. As promised, we have defined
the members of W to be maximal S-consistent sets of wffs. And note that all
maximal S-consistent sets of wffs are included.
Accessibility is defined using the “2− ” notation. Think of this operation
as “stripping off the boxes”: to arrive at 2− (∆) (“the box-strip of set ∆”),
begin with set ∆, discard any formula that doesn’t begin with a 2, line up
the remaining formulas, and then strip one 2 off of the front of each. The
definition of accessibility, therefore, says R ww 0 iff for each wff 2φ that is a
member of w, the wff φ is a member of w 0 .
The definition of accessibility in the canonical model says nothing about
formal properties like transitivity, reflexivity, and so on. As a result, it is not
true by definition that the canonical model for S is an S-model. T-models,
for example, must have reflexive accessibility relations, whereas the definition
of the accessibility relation in the canonical model for T says nothing about
reflexivity. As we will eventually see, the canonical model for each system S
turns out to be an S-model, but this fact must be proven; it’s not built into the
definition of a canonical model.
An atomic wff (sentence letter) is defined to be true at a world iff it is a
member of that world. Thus, for atomic wffs, truth and membership coincide.
What we really need to know, however, is that truth and membership coincide
for all wffs, including complex wffs. Proving this turns out to be a big task,
which will occupy us for several sections. We’ll need first to assemble some
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 163
6.2a if `S φ then φ ∈ Γ
6.2b if φ ∈ Γ and `S φ→ψ then ψ ∈ Γ
going through this list one-by-one, at each point adding either φi or ∼φi .
Here’s how we do this more carefully. Let’s begin by defining an infinite
sequence of sets, Γ0 , Γ1 , . . . :
i) Γ0 is ∆
ii) Γn+1 is Γn ∪ {φn+1 } if that is S-consistent; otherwise Γn+1 is
Γn ∪ {∼φn+1 }
Note the recursive nature of the definition: the next member of the sequence
of sets, Γn+1 is defined as a function of the previous member of the sequence,
Γn .
Next let’s prove that each member in this sequence—that is, each Γi —is an
S-consistent set. We do this inductively, by first showing that Γ0 is S-consistent,
and then showing that if Γn is S-consistent, then so will be Γn+1 .
Obviously, Γ0 is S-consistent, since ∆ was stipulated to be S-consistent.
Next, suppose that Γn is S-consistent; we must show that Γn+1 is S-consistent.
Look at the definition of Γn+1 . What Γn+1 gets defined as depends on whether
Γn ∪{φn+1 } is S-consistent. If Γn ∪{φn+1 } is S-consistent, then Γn+1 gets defined
as that very set Γn ∪ {φn+1 }, and so of course is S-consistent. So we’re ok in
that case.
The remaining possibility is that Γn ∪ {φn+1 } is S-inconsistent. In that case,
Γn+1 gets defined as Γn ∪{∼φn+1 }. So must show that in this case, Γn ∪{∼φn+1 }
is S-consistent. Suppose for reductio that it isn’t. The conjunction of some
finite subset of its members must therefore be provably false in S. Since Γn was
S-consistent, the finite subset must contain ∼φn+1 , and so there exist ψ1 . . . ψ m ∈
Γn such that `S ∼(ψ1 ∧ · · · ∧ψ m ∧∼φn+1 ). Furthermore, since Γn ∪ {φn+1 } is S-
inconsistent, it too contains a finite subset that is provably false in S. Since Γn
is S-consistent, the finite subset must contain φn+1 , so there exist χ1 . . . χ p ∈ Γn
such that `S ∼(χ1 ∧ · · · ∧χ p ∧φn+1 ). But notice that ∼(ψ1 ∧ · · · ∧ψ m ∧χ1 ∧ · · · ∧χ p )
is a PL-semantic-consequence of the formulas ∼(ψ1 ∧ · · · ∧ψ m ∧∼φn+1 ) and
∼(χ1 ∧ · · · ∧χ p ∧φn+1 ). It follows that `S ∼(ψ1 ∧ · · · ∧ψ m ∧χ1 ∧ · · · ∧χ p ) (each of
our modal systems “contains PL”: given the completeness of the PL axioms,
one can always move within an MPL axiomatic proof from some formulas to a
PL-semantic-consequence of those formulas.) Since ψ1 . . . ψ m and χ1 . . . χ p are
all members of Γn , this contradicts the fact that Γn is S-consistent.
We have shown that all the sets in our sequence Γi are S-consistent. Let
us now define Γ to be the union of all the sets in the infinite sequence—i.e.,
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 166
{φ : φ ∈ Γi for some i }. We must now show that Γ is the set we’re after: that i)
∆ ⊆ Γ, ii) Γ is maximal, and iii) Γ is S-consistent.
Any member of ∆ is a member of Γ0 (since Γ0 was defined as ∆), hence is a
member of one of the Γi s, and hence is a member of Γ. So ∆ ⊆ Γ.
Any wff of MPL is in the list somewhere—i.e., it is φi for some i. But by
definition of Γi , either φi or ∼φi is a member of Γi ; and so one of these is a
member of Γ. Γ is therefore maximal.
Suppose for reductio that Γ is S-inconsistent; there must then exist ψ1 . . . ψ m
∈ Γ such that `S ∼(ψ1 ∧ · · · ∧ψ m ). By definition of Γ, each of these ψi ’s are
members of Γ j , for some j . Let k be the largest such j . Note next that, given
the way the Γi ’s are constructed, each Γi is a subset of all subsequent ones.
Thus, all of the ψi ’s are members of Γk , and thus Γk is S-inconsistent. But that
can’t be—we showed that all the Γi ’s are S-consistent. QED.
6.6.6 “Mesh”
Our ultimate goal is to show that in canonical models, a wff is true at a world
iff it is a member of that world. If we’re going to be able to show this, we’d
better be able to show things like this:
We’ll need to be able to show (2) and (3) because it’s part of the definition of
truth in any MPL-model (whether canonical or not) that 2φ is true at w iff φ
is true at each world accessible from w, and that 3φ is true at w iff φ is true at
some world accessible from w. Think of it this way: (2) and (3) say that the
modal statements that are members of a world w in a canonical model “mesh”
with the members of the other worlds in that canonical model. This sort of
mesh had better hold if truth and membership are going to coincide.
(2) we know to be true straightaway, since it follows from the definition of
the accessibility relation in canonical models. The definition of the canonical
model for S, recall, stipulated that w 0 is accessible from w iff for each wff 2φ
in w, the wff φ is a member of w 0 . (3), on the other hand, doesn’t follow
immediately from our definitions; we’ll need to prove it. Actually, it will be
convenient to prove something slightly different which involves only the 2:
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 167
This proof establishes that `S ∼(2ψ1 ∧ · · · ∧2ψn ∧∼2φ). But since 2ψ1 …2ψn ,
and ∼2φ are all in ∆, this contradicts ∆’s S-consistency (2ψ1 …2ψn are mem-
bers of ∆ because ψ1 …ψn are members of 2− (∆).)
We’ve established (*): 2− (∆) ∪ {∼φ} is S-consistent. It therefore has a
maximal S-consistent extension, Γ, by Theorem 6.3. Since 2− (∆) ∪ {∼φ} ⊆ Γ,
we know that 2− (∆) ⊆ Γ and that ∼φ ∈ Γ. Γ is therefore our desired set; the
proof of Lemma 6.4 is complete.
Proof of Theorem 6.5: We’ll use induction. The base case is when φ has zero
connectives—i.e., φ is a sentence letter. In that case, the result is immediate:
by the definition of the canonical model, I (φ, w) = 1 iff φ ∈ w; but by the
definition of the valuation function, VM (φ, w) = 1 iff I (φ, w) = 1.
Now the inductive step. We suppose (ih) that the result holds for φ, ψ, and
show that it holds for ∼φ, φ→ψ, and 2φ as well:
i) ∼φ ∈ w iff φ ∈
/ w (6.1a)
ii) φ ∈
/ w iff φ is not true at w (ih)
iii) φ is not true at w iff ∼φ is true at w (truth cond. for ∼)
iv) so, ∼φ is true at w iff ∼φ ∈ w (i, ii, iii)
Completeness of D
Let us show that in the canonical model for D, the accessibility relation, R, is
serial. Let w be any world in that model. We showed above that 3(P →P ) is
a theorem of D, and so is a member of w by 6.2a, and so is true at w by 6.5.
Thus, by the truth condition for 3, there must be some world accessible to w
in which P →P is true; and hence there must be some world accessible to w.
Now for D’s completeness. Let φ be D-valid. It is then valid in all D-
models, i.e., all models with a serial accessibility relation. But we just showed
that the canonical model for D has a serial accessibility relation. φ is therefore
valid in that model, and hence by 6.6, `D φ.
Completeness of T
All we need to do is to prove that the accessibility relation in the canonical model
for T is reflexive; given that, every T-valid formula is valid in the canonical
model for T, and hence by 6.6, every T-valid formula is a T-theorem.
Let w be any world in the canonical model for T. For any φ, `T 2φ→φ;
thus, by 6.2b, it follows that for any φ, if 2φ ∈ w then φ ∈ w. But this is the
definition of R ww.
Completeness of B
We must show that the accessibility relation in the canonical model for B is
reflexive and symmetric. Reflexivity can be demonstrated in the same way as it
was for T, since every T-theorem is a B-theorem.
Now for symmetry: in the canonical model for B, suppose that R wv. We
must show that R v w—that is, that for any 2ψ in v, ψ ∈ w. So, suppose that
2ψ ∈ v. By 6.5, 2ψ is true at v; since R wv, by the definition of 3 it follows
that 32ψ is true at w, and hence is a member of w by 6.5. Since `B 32ψ→ψ,
by 6.2b, ψ ∈ w.
CHAPTER 6. PROPOSITIONAL MODAL LOGIC 171
Completeness of S4
We must show that the accessibility relation in the canonical model for S4 is
reflexive and transitive. Again, reflexivity can be demonstrated as it was for
T; transitivity remains. Suppose R wv and R v u. We must show R w u—that
is, for any 2ψ ∈ w, ψ ∈u. If 2ψ ∈ w, since `S4 2ψ→22ψ, by 6.2b we have
22ψ ∈ w. By 6.5, 22ψ is true at w; hence by the truth condition for 2, 2ψ is
true at v; again by the truth condition for 2, ψ is true at u; by 6.5, ψ ∈ u.
Completeness of S5
We must show that the accessibility relation in the canonical model for S5 is
reflexive, symmetric, and transitive. But since each T, B, and S4 theorem is
an S5 theorem, the proofs of reflexivity, symmetry, and transitivity from the
previous three sections apply here.
Chapter 7
s we have seen, possible worlds are useful for giving a semantics for propo-
A sitional modal logic. Possible worlds are useful in other areas of logic as
well. In this chapter we will briefly examine two other uses for possible worlds:
semantics for tense logic, and semantics for intuitionist propositional logic.
172
CHAPTER 7. VARIATIONS ON MPL 173
P ∃xD x
And according to Prior, P ∃xD x is not to be analyzed as saying that there exist
dinosaurs located in the past. For him, there is no further analysis of P ∃xD x.
Prior’s attitude toward P is like everyone else’s attitude toward the ∼: no one
thinks that ∼∃xU x, “there are no unicorns”, is to be analyzed as saying that
there exist unreal unicorns. Further, Prior can represent the fact that I am
now, but have not always been, an adult, without adding argument places for
times to predicates. Symbolizing ‘is an adult’ with ‘A’, and ‘Ted’ with ‘t ’, Prior
would write: At ∧ P∼At (“Ted is an adult, but it was the case that: Ted isn’t an
adult”). For Prior, the sentence At (“Ted is an adult”) is a complete statement,
but nevertheless can alter its truth value.
Gφ iff ∼F∼φ
Hφ iff ∼P∼φ
Thus one could start with just two of them, G and H say, and define the others.
And one could use these two to define further tense operators, for example A
and S, for “always” and “sometimes”
Aφ iff Hφ∧Gφ
Sφ iff ∼H∼φ∨∼G∼φ (i.e., iff Pφ∨Fφ)
There are further tense operators that are not definable in terms of those
we’ve been considering so far. There are, for example, metrical tense operators,
which concern what happened or will happen at specific temporal distances in
the past or future:
Whether we take the tense operators as including the present moment will
affect what kind of logic we develop for them. For example, the “2-like”
operators G and H will obey the T-principle (Gφ and Hφ will imply φ) if they
are interpreted as including the present moment, but not otherwise.
CHAPTER 7. VARIATIONS ON MPL 176
We’ll need to update the definition of the valuation function. The clauses
for the propositional connectives remain the same; what we need to add is
clauses for the tense operators. Let’s take just G and H as primitive; here are
the clauses:
VM (Gφ, t ) = 1 iff for every t 0 such that t ≤ t 0 , VM (φ, t 0 ) = 1
VM (Hφ, t ) = 1 iff for every t 0 such that t 0 ≤ t , VM (φ, t 0 ) = 1
If we define F and P as ∼G∼ and ∼H∼, respectively, then we get the following
derived clauses:
VM (Fφ, t ) = 1 iff for some t 0 such that t ≤ t 0 , VM (φ, t 0 ) = 1
VM (Pφ, t ) = 1 iff for some t 0 such that t 0 ≤ t , VM (φ, t 0 ) = 1
Call an MPL-model, thought of in this way, a “PTL-model” (for “Priorean
Tense Logic”). And say that a wff is PTL-valid iff it is true in every time in every
PTL-model. Given our discussion of system K from chapter 6, we already
know a lot about PTL-validity. The truth condition for the G is the same as
the truth condition for the 2 in MPL. Thus, for each K-valid formula φ of
MPL, there is a PTL-valid formula of tense logic: simply replace each 2 in
φ with G. Replacing 2s with Gs in the K-valid formula 2(P ∧Q)→2P results
in the PTL-valid formula G(P ∧Q)→GP , for example. Similarly, the result of
replacing 2s with Hs in a K-valid formula also results in a PTL-valid formula.
But there are further cases of PTL-validity that depend on the interaction
between different tense operators, and hence have no direct analog in MPL.
For example, we can demonstrate that PTL φ→GPφ:
i) Suppose for reductio that in some PTL-model M (= 〈T , ≤
, I 〉) and some t ∈ T , VM (φ→GPφ, t ) = 0. (I henceforth
drop the subscript M .)
ii) So V(φ, t ) = 1 and …
iii) …V(GPφ, t ) = 0.
iv) Given iii), by the truth condition for G: for some t 0 ∈ T , t ≤ t 0
and V(Pφ, t ) = 0
v) Given iv), by the (derived) truth condition for P: for every
t 00 ∈ T , if t 00 ≤ t 0 then V(φ, t 00 ) = 0
vi) letting t 00 in v) be t , given that t ≤ t 0 (iv), we have: V(φ, t ) = 0,
contradicting ii).
Similarly, one can show that PTL φ→HFφ.
CHAPTER 7. VARIATIONS ON MPL 178
Gφ→GGφ
Hφ→HHφ
There are other interesting constraints on ≤ that one might impose. One
might impose reflexivity, for example. This is natural to impose if we are
construing the tense operators as including the present moment; not otherwise.
Imposing reflexivity validates the “T-schemas” Gφ→φ and Hφ→φ.
One might also impose “connectivity” of some sort. Connective principles
disallow, in various ways, “incomparable” pairs of times—pairs of times neither
CHAPTER 7. VARIATIONS ON MPL 179
of which bears the ≤ relation to the other. The strongest sort disallows all
incomparable pairs:
A weaker sort merely disallows incomparable pairs when each member of the
pair is after or before some one time:
Weak connectivity disallows “branches” in the temporal order but allows dis-
tinct timelines wholly disconnected from one another; strong connectivity
insures that all times are part of a single non-branching structure. Each sort
validates every instance of the following schemas:
G(Gφ→ψ)∨G(Gψ→φ)
H(Hφ→ψ)∨H(Hψ→φ)
There are other constraints one might impose, for example anti-symmetry
(no distinct times bear ≤ to each other), density (between any two times there is
another time), or eternality (there exists neither a first nor a last time). In some
cases, imposing a constraint validates an interesting schema being validated.
Further, some constraints are more philosophically controversial than others.
CHAPTER 7. VARIATIONS ON MPL 180
Notice that one should not impose symmetry on ≤. Obviously if one time
is at least as early as another, then the second time needn’t be at least as early
as the first. Moreover, imposing symmetry would validate the “B” schemas
FGφ→φ and PHφ→φ; but these clearly ought not to be validated. Take the
first, for example: it doesn’t follow from it will be the case that it is always going to
be the case that I’m dead that I’m (now) dead.
So far we have been interpreting the tense operators as including the present
moment. That led us to call the temporal ordering relation in our models “≤”,
and require that it be reflexive. What if we instead interpreted the tense
operators as not including the present moment? We would then call the
temporal ordering relation “<”, and think of it as the earlier-than relation; and
we would no longer require that it be reflexive. Indeed, it would be natural to
require that it be irreflexive: that it never be the case that t < t .
We have considered only the semantic approach to tense logic. What of a
proof-theoretic approach? Given the similarity between tense logic and modal
logic, it should be no surprise that axiom systems similar to those of section
6.4 can be developed for tense logic. Moreover, the techniques developed in
sections 6.5-6.6 can be used to give soundness and completeness proofs for
tense-logical axiom systems, relative to the possible-worlds semantics that we
have developed in this section.
semantics for propositional modal logic, and so it will include valuation func-
tions that assign the values 1 and 0 to formulas relative to the members of a
set W . But the idea is to now think of the members of W as stages in the
construction of proofs, rather than as possible worlds, and to think of 1 and 0 as
“proof statuses”, rather than truth values. That is, we are to think of V(φ, w) = 1
as meaning that formula φ has been proved at stage w.
Let us treat the ∧ and the ∨ as primitive connectives. Here is Kripke’s
semantics for intuitionist propositional logic. (To emphasize the different way
we are regarding the “worlds”, we rename W “S ”, for stages in the construction
of proofs, and we will use the variables s, s 0 , etc., for its members.)
Given any such model, we can define the corresponding valuation function
thus:
Note that the truth conditions for the → and the ∼ at stage s no longer
depend exclusively on what s is like; they are sensitive to what happens at
stages accessible from s. Unlike the ∧ and the ∨, → and ∼ are not “truth
functional” (relative to a stage); they behave like modal operators.
Let us think intuitively about these models. We are to think of each member
of S as a stage in the construction of mathematical proofs. At any stage, one
CHAPTER 7. VARIATIONS ON MPL 182
has come up with proofs of some things but not others. When V assigns 1 to a
formula at a stage, that means intuitively that as of that state of information,
the formula has been proven. The assignment of 0 means that the formula has
not been proven thus far (though it might nevertheless in the future.)
The holding of the accessibility relation R represents which future stages
are possible, given one’s current stage. If s 0 is accessible from s , that means that
s 0 contains all the proofs in s, plus perhaps more. Given this understanding,
reflexivity and transitivity are obviously correct to impose, as is the heredity
condition, since (on the somewhat idealized conception of proof we are oper-
ating with) one does not lose proved information when constructing further
proofs. But the accessibility relation will not in general be symmetric: for
sometimes one will come across a new proof that one did not formerly have.
Let’s also think through why the truth conditions for →, ∧, ∨ and ∼ are
intuitively correct. Intuitionists, recall, associate with each propositional con-
nective, a conception of what proofs of formulas built using that connective
must be like:
of ψ, then there could never be a possible future in which one has a proof of
φ but not one of ψ. Conversely, if one lacks such a method, then it should be
possible one day to have a proof of φ without being able to convert it into a
proof of ψ, and thus without then having a proof of ψ.
We can now define intuitionist semantic consequence and validity in the
obvious way:
Q P →Q
Take any model and any stage s; assume that V(Q, s) = 1 and V(P →Q, s) = 0.
Thus, for some s 0 , R s s 0 and V(P, s 0 ) = 1 and V(Q, s 0 ) = 0. But this violates
heredity.
Next:
P →Q ∼Q→∼P (contraposition):
2 P ∨∼P :
0 0 0
r P ∨∼P
∗
0
∗
a 1
P
0
S : {r, a}
R : {〈r, r〉, 〈a, a〉, 〈r, a〉}
I (P, a) = 1, all other atomics 0 everywhere
(I’ll skip the official models from now on.) So, 2 P ∨∼P . Given the soundness
of the intuitionist proof system (proved below), it follows that ∅ ` P ∨∼P is
indeed an unprovable sequent.
Next example:
∼∼P 2 P :
CHAPTER 7. VARIATIONS ON MPL 185
∗
1 0 0
r
∼∼P P
∗
0
∗
1 0
a
P ∼P
∗
0
∗
1 0 0 0 0 0
r
∼(P ∧Q) ∼P ∨∼Q
∗ ∗
0 t JJ
tt JJ
tt JJ
tt JJ
tt JJ
zt
t J%
∗ ∗
a 1 0 0 b 1 0 0
P P ∧Q Q P ∧Q
0 0
Next example:
Suppose V(∼P ∨∼Q, s ) = 1 and V(∼(P ∧Q), s) = 0. Given the latter, for some
s 0 , R s s 0 and V(P ∧Q, s 0 ) = 1. So, V(P, s 0 ) = 1 and V(Q, s 0 ) = 1. Given the
CHAPTER 7. VARIATIONS ON MPL 186
We begin thus:
∗
1 0 0 0
r
P →(Q∨R) (P →Q)∨(P →R)
∗ ∗
0
∗
1 0 0 0
r
P →(Q∨R) (P →Q)∨(P →R)
∗ ∗
0rr LLL
rrrrr LLL
LLL
r
rrrrr LLL
LLL
∗ x r
r LL
&
a 1 0 b P R
0
P Q
0
Proof: for the first, suppose φ ψ, and suppose for reductio that V(φ→ψ, s) =
0. Then for some s 0 (that s sees), V(φ, s 0 ) = 1 and V(ψ, s 0 ) = 0—contradicts
φ ψ.
As for the second, suppose φ→ψ, and suppose for reductio that V(φ, s ) = 1
while V(ψ, s ) = 0. By φ→ψ, V(φ→ψ, s ) = 1. By reflexivity, either V(φ, s) = 0
or V(ψ, s ) = 1. Contradiction.
Counterfactuals1
here are certain conditionals in natural language that are not well-
T represented either by propositional logic’s material conditional or by
modal logic’s strict conditional. In this chapter we consider “counterfactual”
conditionals—conditionals that (loosely speaking) have the form:
If it had been that φ, then it would have been that ψ
For instance:
If I had struck this match, it would have lit
The counterfactuals that we typically utter have false antecedents (hence
the name), and are phrased in the subjunctive mood. They must therefore
be distinguished from English conditionals phrased in the indicative mood.
Counterfactuals are generally thought to semantically differ from indicative
conditionals. A famous example: the counterfactual conditional ‘If Oswald
hadn’t shot Kennedy, someone else would have’ is false (assuming that certain
conspiracy theories are false and Oswald was acting alone); but the indicative
conditional ‘If Oswald didn’t shoot Kennedy then someone else did’ is true
(we know that someone shot Kennedy, so if it wasn’t Oswald, it must have been
someone else.) The semantics of indicative conditionals is an important topic
in its own right, but we won’t take up that topic here.
We represent the counterfactual with antecedent φ and consequent ψ thus:
φ2→ψ
1
This section is adapted from my notes from Ed Gettier’s fall 1988 modal logic class.
190
CHAPTER 8. COUNTERFACTUALS 191
What should the logic of this new connective be, if it is to accurately represent
natural language counterfactuals?
∼P 2 P 2→Q
Q 2 P 2→Q
For consider: I did not strike the match; but it doesn’t logically follow that
if I had struck the match, it would have turned into a feather. So if 2→ is to
represent ‘if it had been that…, it would have been that…’, ∼P should not
semantically imply P 2→Q. Similarly, George W. Bush (somehow) won the last
United States presidential election, but it doesn’t follow that if the newspapers
had discovered beforehand that Bush had an affair with Al Gore, he would
still have won. So our semantics had better not count P 2→Q as a semantic
consequence of Q either.
These implications hold for the material conditional, however:
∼P P →Q
Q P →Q
it would be true that if Oswald hadn’t shot Kennedy, someone else would
have. Thus, our logic should allow counterfactuals to be contingent statements.
Just because a counterfactual is true, it should not follow logically that it is
necessarily true; and just because a counterfactual is false, it should not follow
logically that it is necessarily false. Our semantics for 2→, that is, should have
the following features:
One reason this is important is that it shows an obstacle to using the strict
conditional ⇒ to represent natural language counterfactuals. For remember
that P ⇒Q is defined as 2(P →Q). As a result:
8.1.3 No augmentation
The → and the ⇒ obey the argument form augmentation:
P →Q P ⇒Q
(P ∧R)→Q (P ∧R)⇒Q
So, our next desideratum is that the corresponding argument should not hold
good for 2→ (that is, P 2→Q 2 (P ∧R)2→Q.)
CHAPTER 8. COUNTERFACTUALS 193
8.1.4 No contraposition
→ and ⇒ obey contraposition:
φ→ψ φ⇒ψ
∼ψ→∼φ ∼ψ⇒∼φ
But counterfactuals do not. Suppose I’m on the firing squad, and we shoot
someone dead. My gun was loaded, but so were those of the others. Then the
premise of the following argument is true, while its consequent is false:
P 2→Q
P →Q
P P 2→Q
P 2→Q ∼Q
Q ∼P
The reason is, of course, that modus ponens and modus tollens are valid for
the →. (Note that it’s not inconsistent to say that modus tollens holds for the
2→ and also that contraposition fails.)
Another implication: the strict conditional should imply the counterfactual:
P ⇒Q
P 2→Q
To see that these implications should hold, consider first the argument from
the strict conditional to the counterfactual conditional. Surely, if P entails Q,
then if P were true, Q would be as well. As for the counterfactual implying the
CHAPTER 8. COUNTERFACTUALS 194
material, suppose that you think that if P were true, Q would also be true. Now
suppose that someone tells you that P is true, but that Q is false. Wouldn’t
you then need to give up your original claim that if P were to be true, then
Q would be true? It seems so. So, the statement P 2→Q isn’t consistent with
P ∧∼Q—that is, it isn’t consistent with the denial of P →Q.
Who was right? If Ted Williams had played at the time the story was told,
would he or wouldn’t he have hit over .300?
Clearly, there’s no one answer. The first respondents were imagining
Williams playing as a young man. Understood that way, the answer is no
doubt: yes, he would have hit over .300. But Berra took the question a different
way: he was imagining Williams hitting as he was then: a 74 year old man.
Berra took the others off guard, by deliberately (?—this is Yogi Berra we’re
talking about…) shifting how the question was construed, but he didn’t make a
semantic mistake in so doing. It’s perfectly legitimate, in other circumstances
anyway, to take the question in Berra’s way. (Imagine Williams talking to him-
self five years after he retired: “These punks today! If I were playing today,
I’d still hit over .300!”) Counterfactual sentences can mean different things
depending on the conversational context in which they are uttered.
Another example:
If Syracuse were in Louisiana, Syracuse winters would be warm.
True or false? It might seem true: Louisiana is in the south. But wait—perhaps
Louisiana would include Syracuse by extending its borders north to Syracuse’s
actual latitude.
Would Syracuse be warm in the winter? Would Williams hit over .300?
No one answer is correct, once and for all. Which answer is correct depends
CHAPTER 8. COUNTERFACTUALS 195
8.3.1 Syntax of SC
The primitive vocabulary of SC is that of propositional modal logic, plus the
connective 2→. Here’s the grammar:
Definition of SC-wffs:
8.3.2 Semantics of SC
First, let’s define two formal properties for relations: strong connectivity and
anti-symmetry. Let R be a binary relation over set A:
least as close to itself as every other. Given C3, it has the further implication
that every world is closer to itself than every other. (We define “x is closer to w
than y is” (x ≺w y) to mean x w y and not: y w x.) C5 is called the “limit”
assumption: according to it, for any formula φ and any base world w, there
is some world that is a closest world to w in which φ is true (that is, unless φ
isn’t true at any worlds at all). This rules out the following possibility: there
are no closest φ worlds, only an infinite chain of φ worlds, each of which is
closer than the previous. Certain of these assumptions have been challenged,
especially C3 and C5. We will consider those issues below.
Note that the valuation function for a model was built directly into the
model. This is in contrast to our earlier definitions of models, in which we
equipped models with interpretation functions of different sorts, and then
defined the resultant valuation functions accordingly. The valuation function is
built into the definition of a SC-model because the limit assumption constrains
SC-models, and the limit assumption is stated in terms of the valuation function.
Given our definitions, we can define validity and semantic consequence:
Here is another validity proof, for the formula {[P 2→Q]∧[(P ∧Q)2→R]}
→[P 2→R]. (This formula is worth taking note of, because it is valid despite
its similarity to the invalid formula {[P 2→Q]∧[Q2→R]}→[P 2→R]):
8.5 Countermodels in SC
Our other main concern will be countermodels. As before, we’ll be able to use
failed attempts at constructing countermodels as guides to our validity proofs.
First, consider the truth condition for φ2→ψ:
/. 1 1 1 0 0
-,
In keeping with the advice I gave a moment ago, let’s “take care of” the false
counterfactual first: let’s make P 2→R false in r. This means that we need to
have R be false in the nearest-to-r P world. I’ll indicate this as follows:
/.1 0 1
-,
()P Q*+
a
O R
no P
/. 0 1 1 1 0 0
-,
/.1 0 1
-,
()P Q*+
O
a
R no Q
no P
/. 0 1 0 1 1 0 0
-,
()[(P 2→Q)∧(Q2→ R)]→(P 2→ R)*+
r
CHAPTER 8. COUNTERFACTUALS 204
This doesn’t work, because when we make b the nearest-to-r Q world, this
contradicts the fact that we’ve got Q true at a nearer world—namely, a.
Likewise, we can’t make the nearest-to-r Q world be tied with a. Anti-
symmetry would make this world be a. But we need to make R true at the
nearest Q world, whereas R is already false in a.
But the final possibility works out just fine—let the nearest Q world be
closer than a:
/.1 0 1
-,
()P Q*+
a
O R
/.1 -,
()Q P*+
1 0
no P b
O
R
no Q
/. 0 1 0 1 1 0 0
-,
()[(P 2→Q)∧(Q2→ R)]→(P 2→ R)*+
r
Notice that I made P false in b, since I said “no P ” of all worlds nearer to r
than a. Here’s the official model:
W = {r, a, b}
r = {〈b, a〉 . . . }
V(P, a) = V(Q, a) = V(Q, b) = V(R, b) = 1; all other atomics false
everywhere else.
A couple comments on the official model. I left out a lot in giving the similarity
relation for this model. First, I left out some of the elements of r —fully
written out, it would be:
r = {〈r,b〉,〈b,a〉,〈r,a〉,〈r,r〉,〈a,a〉,〈b,b〉}
I left out 〈r,b〉 because it gets included automatically given the “base” assumption
(C4). Also, the element 〈r,a〉 is required to make r transitive. The elements
〈r,r〉, 〈a,a〉, and 〈b,b〉 were entered to make that relation reflexive. Why must it
be reflexive? Because reflexivity comes from strong connectivity. (Let w and x
be any members of W ; we get (x w x or x w x) from Strong connectivity of
CHAPTER 8. COUNTERFACTUALS 205
w , and hence x w x.) My plan was to write out enough of r so that the rest
could be inferred, given the definition of an SC-model.
Secondly, this isn’t a complete writing out of itself; it is just r . To be
complete, we’d need to write out a , and b . But in this problem, these latter
two parts of don’t matter, so it’s permissible to omit them. We’ll do a problem
below where we need to consider more of than simply r .
Suppose one has a formula, for example:
(P 2→R)→((P ∧Q)2→R)
(This is the formula corresponding to the inference pattern of augmentation.)
Suppose we ask: is it SC-valid? The best way to approach this is as follows.
First try to found a countermodel. If we succeed, then the formula is not valid.
If we have trouble finding a countermodel, then we can try giving a semantic
validity proof (for if the formula is in fact valid, then no countermodel exists).
In this particular case, the formula is invalid, as the following countermodel
shows:
/.1 1 1 0
-,
()P ∧Q R*+
a
O
/.1 -,
()P Q*+
1 0
no P ∧Q b
O
R
/. 0 1 -,
no P
0 0
()(P 2→ R)→[(P ∧Q)2→ R)*+
r
Ok, I began with the false subjunctive: (P ∧Q)2→R. This forced the existence
of a nearest P ∧Q world, in which R was false. But since P ∧Q was true there,
P was true there; this ruled out the true P 2→R in r being vacuously true. So I
was forced to consider the nearest P world. It couldn’t be farther out than a,
since P is true in a. It couldn’t be a, since R was already false there. So I had to
put it nearer than a. Notice that I had to make Q false at b. Why? Well, it was
in the “no P ∧Q zone”, and I had made P true in it. Here’s the official model:
W = {r, a, b}
r = {〈b, a〉 . . . }
CHAPTER 8. COUNTERFACTUALS 206
()P ∼Q*+
a 1 1 0
O
no P
/.1 0 0 -,
At world a, we’ve got Q being both true and false. A word about how we got
to that point. I noticed that I had to make two counterfactuals true: P 2→Q
and P 2→∼Q. Now, this isn’t a contradiction all by itself. Remember that
counterfactuals are vacuously true if their antecedents are impossible. So if P
were impossible, then both of these would indeed be true, without any problem.
But 3P has to be true at r. This rules out those counterfactual’s being vacuously
true. Since P is possible, the limit assumption has the result that there is a closest
P world. This then with the two true counterfactuals created the contradiction.
This reasoning is embodied in the following semantic validity proof:
Note the use of the limit assumption. It is the limit assumption we must use
when we need to know that there is a nearest φ-world, and we can’t get this
knowledge from other things in the proof.
For the last example of a countermodel, I’ll do a new kind of problem: one
with a counterfactual nested within another counterfactual. Let’s show that the
following formula is SC-invalid: [P 2→(Q2→R)]→[(P ∧Q)2→R] (this is the
formula corresponding to “importation”). We begin by making the formula
false in r, the actual world of the model. This means making the antecedent true
and the consequent false. Now, since the consequent is a false counterfactual,
we are forced to make there be a nearest P ∧Q world in which R is false:
/.1 1 1 -,
()P ∧Q R*+
0
a
O
no P ∧Q
/. -,
Now we’ve got to make the first subjunctive conditional true. We can’t make it
vacuously true, because we’ve already got a P -world in the model: a. So, we’ve
got to put in the nearest-to-r P world. Could it be farther away than a? No,
because a would be a closer P world. Could it be a? No, because we’ve got to
make Q2→R true in the closest P world, and since Q is true but R is false in a,
Q2→R is already false in a. So, we do it as follows:
/.1 1 1 -,
()P ∧Q R*+
0
a
O
/.1 -,
()P Q2→R*+
0 1
no P ∧Q b
O
no P
/. 0 1 -,
()[P 2→(Q2→ R)]→[(P ∧Q)2→ R]*+
0 0
r
Why did I make Q false at b? Well, because b is in the “no P ∧Q” zone, and P
is true at b, so Q had to be false there.
CHAPTER 8. COUNTERFACTUALS 208
no Q
/.1 -,
()P Q2→R*+
0 1
b
I made there be a nearest-to-b Q world, and made R true there. Notice that
I kept the old truth values of b from the other diagram. This is because this
new diagram is a diagram of the same worlds as the old diagram; the difference
is that the new diagram represents the b nearness relation, whereas the old
one represented a different relation: r . Now, this diagram isn’t finished. The
diagram is that of the b relation, and that relation relates all the worlds in
the model. So, worlds r and a have to show up somewhere here. But the
safest practice is to put them far away from b, so that there isn’t any possibility
of conflict with the “no Q” zone that has been established. Thus, the final
appearance of this part of the diagram is as follows:
/. -,
r () *+
/. -,
a () *+
/.1 -,
()Q R*+
1
c
O
no Q
/.1 -,
()P Q2→R*+
0 1
b
CHAPTER 8. COUNTERFACTUALS 209
The old truth values from r and a still are had (remember that this is another
diagram of the same model, but representing a different nearness relation),
but I left them out because of the fact that they’ve already been written on the
other part of the diagram.
Notice that the order of the worlds in the r-diagram does not in any way
affect the order of the worlds on the b diagram. The nearness relations in
the two diagrams are completely independent, because in our definition of ‘SC-
model’, we entered in no conditions constraining the relations between i and
j when i 6= j . This sometimes seems unintuitive. For example, we could have
two halves of a model looking as follows:
If this seems odd, remember that only some of the features of a diagram are
intended to be genuinely representative. For example, these diagrams are in ink;
this is not intended to convey the idea that the worlds in the model are made
of ink. This feature of the diagram isn’t intended to convey information about
the model. Analogously, the fact that in the left diagram, b is physically closer
to a than c is not intended to convey the information that, in the model, ba c.
In fact, the diagram of the view from r is only intended to convey information
about r ; it doesn’t carry any information about a , b , or c .
Back to the countermodel. That other part of the diagram, the view from r,
must be updated to include world c. The safest procedure is to put c far away
on the model to minimize possibility of conflict. Thus, the final picture of the
view from r is:
CHAPTER 8. COUNTERFACTUALS 210
/. -,
c () *+
/.1 1 1 -,
()P ∧Q R*+
0
a
O
/.1 -,
()P Q2→R*+
0 1
no P ∧Q b
O
no P
/. 0 1 -,
()[P 2→(Q2→ R)]→[(P ∧Q)2→ R]*+
0 0
r
Again, I haven’t re-written the truth values in world c, because they’re already
in the other diagram, but they are to be understood as carrying over. Now for
the official model:
W = {r, a, b, c}
V(P, a) = V(Q, a) = V(P, b) = V(Q, c) = V(R, c) = 1; all else (atom-
ics) 0 elsewhere
r = {〈b, a〉, 〈a, c〉 . . . }
b = {〈c, a〉, 〈a, r〉 . . . }
∼P 2 P 2→Q
Q 2 P 2→Q
Our semantics does indeed have this result, because the similarity metrics based
on different worlds can be very different. For example: consider a model with
worlds r and a, in which Q is true in the nearest-to-r P world, but in which Q
is false at the nearest-to-a P world. P 2→Q is true at r and false at a, whence
2(P 2→Q) is false at r.
8.6.3 No augmentation
Earlier we produced a model containing a world in which this conditional was
false:
[P 2→Q]→[(P ∧R)2→Q]
8.6.4 No contraposition
Let’s show that P 2→Q 2 ∼Q2→∼P :
CHAPTER 8. COUNTERFACTUALS 212
/.1 0 -,
()∼Q ∼P*+
0 1
a
O
/.1 -,
()P Q*+
1
no ∼Q b
O
no P
/. -,
()(P 2→Q) (∼Q2→∼P )*+
1 0
r
8.6.6 No exportation
We have shown that the SC-semantics reproduces the logical features of natural
language counterfactuals discussed in section 8.1. In the next few sections we
discuss some further logical features of the SC-semantics, and compare them
with the logical features of the →, the ⇒, and natural language counterfactuals.
The → obeys exportation:
(P ∧Q)→R
P →(Q→R)
But the ⇒ doesn’t in any system; (P ∧Q)⇒R 2S5 P ⇒(Q⇒R). Nor does
the 2→; it can be easily shown with a countermodel that (P ∧Q)2→R 2SC
P 2→(Q2→R).
Does the natural language counterfactual obey exportation? Here is an
argument that it does not. The following is true:
(L&H) If Bill had married Laura and Hillary, he would have been a
bigamist.
(L2→H) If Bill had married Laura, then it would have been the case
that if he had married Hillary, he would have been a bigamist.
Suppose Bill had married Laura. Would it then have been true that: if he had
married Hillary, he would have been a bigamist? Well, let’s ask for comparison:
what would the world have been like, had George W. Bush married Hillary
Rodham Clinton? Would Bush have been a bigamist? Here the natural answer
is no. George W. Bush is in fact married to Laura Bush; but when imagining him
married to Hillary Rodham Clinton, we don’t hold constant his actual marriage.
We imagine him being married to Hillary instead. If this is true for Bush, then
one might think it’s also true for Bill in the counterfactual circumstance in
which he’s married to Laura: it would then have been true of him that, if he
had married Hillary, he wouldn’t have still been married to Laura, and hence
would not have been a bigamist.
It’s unclear whether this is a good argument, though, since it assumes that
ordinary standards for evaluating unembedded counterfactuals (“If George had
married Hillary, he would have been a bigamist”) apply to counterfactuals em-
bedded within other counterfactuals (“If Bill had married Hillary, he would have
CHAPTER 8. COUNTERFACTUALS 214
8.6.7 No importation
Importation holds for →, and for ⇒ in T and stronger systems:
P →(Q→R) P ⇒(Q⇒R)
(P ∧Q)→R (P ∧Q)⇒R
but not for the 2→: above we produced an SC-model with a world in which
the conditional [P 2→(Q2→R)]→[(P ∧Q)2→R] was false.
The status of importation for natural language counterfactuals is similar
to the status of exportation. If (L2→H) is false (on at least one reading), then
presumably the following could be true (on at least one reading):
If Bill had married Laura, then it would have been the case that if
he had married Hillary, he would have been happy.
If Bill had married Laura and Hillary, he would have been happy
(if he had married both, his relatives would have hated him, he would have
become a public spectacle, etc.)
Q→R Q⇒R
P →Q P ⇒Q
P →R P ⇒R
CHAPTER 8. COUNTERFACTUALS 215
but not for 2→, as our proof above of the invalidity of [(P 2→Q)
∧(Q2→R)]→(P 2→R) shows.
Natural language counterfactuals also seem not to obey hypothetical syllo-
gism. I am the oldest child in my family; my brother Mike is the second-oldest.
So the following counterfactuals seem true:4
If my parents had never met, Mike would have been their oldest
child.
8.6.9 No transposition
Transposition governs the →:
P →(Q→R)
Q→(P →R)
but not the ⇒ (in any of our modal systems); P ⇒(Q⇒R) 2S5 Q⇒(P ⇒R). Nor
does it govern the 2→; it’s easy to show that P 2→(Q2→R) 2SC Q2→(P 2→R).
The status of transposition for natural language counterfactuals is sim-
ilar to that of importation and exportation. If we can ignore the effects of
embedding on the evaluation of counterfactuals, then we have the following
counterexample to transposition. It is true that:
4
Note that if these two statements are written in the reverse order, it seems far less clear
that they’re both true:
It’s natural in this case to interpret the second conditional by holding constant the antecedent
of the first conditional. This fact, together with what we observed about embedded counter-
factuals, suggests a systematic dependence of the interpretation of counterfactuals on their
immediate linguistic context. See von Fintel (2001) for a “dynamic” semantics for counterfac-
tuals, which more accurately models this feature of their use.
CHAPTER 8. COUNTERFACTUALS 216
Take the first one, for example. Suppose you gave up anti-symmetry, thereby
allowing ties. Then the following would be a countermodel for the law of
conditional excluded middle:
5
For an interesting response to Lewis, see Stalnaker (1981).
CHAPTER 8. COUNTERFACTUALS 217
/.1 -, /.1 -,
()P ∼Q*+
0
()P Q*+
0 1
a
O b
no Q
/. 0 0 0
-,
Remember that P 2→Q is true only if Q is true in all the nearest P worlds.
In this model, Q is true in one of the nearest P worlds, but not all, so that
counterfactual is false at r. Similarly for P 2→∼Q.
A similar model shows that distribution fails if the “no ties” assumption is
given up.
So, should we give up conditional excluded middle? As Lewis concedes,
the principle is initially plausible. An equivalent formulation of conditional
excluded middle is this:
But everyone agrees that (P 2→∼Q)→∼(P 2→Q) is always true, at least, when
P is possibly true. So, in cases where P is possibly true anyway, the question
of whether conditional excluded middle is valid is the question of whether
∼(P 2→Q) and P 2→∼Q are equivalent to each other. And it does indeed
seem that in ordinary usage, one expresses the negation of a counterfactual
by negating its consequent. To deny the counterfactual “if she had played,
she would have won", one says "no, she wouldn’t have!”, meaning “if she had
played, she would not have won”.
And take the other formula validated by Stalnaker’s theory, distribution. In
reply to: “if the coin had been flipped, it would have come out either heads or
tails”, one might ask: “which would it have been, heads or tails?”. The thinking
behind the reply is that “if the coin had been flipped, it would have come up
heads”, or “if the coin had been flipped, it would have come up tails” must be
true.
So there’s some plausibility to both these formulas. But Lewis says two
things. The first is metaphysical: if we’re going to accept the similarity analysis,
we’ve got to give them up, because ties just are possible. The second is purely
semantic: the intuitions aren’t completely compelling. About the coin-flipping
case, Lewis denies that if the coin had been flipped, it would have come up
CHAPTER 8. COUNTERFACTUALS 218
heads, and he also denies that if the coin had been flipped, it would have come
up tails. Rather, he says, if it had been flipped, it might have come up heads.
And if it had been flipped, it might have come up tails. But neither outcome is
such that it would have resulted, had the coin been flipped.
Concerning excluded middle, Lewis says:
It is not the case that if Bizet and Verdi were compatriots, Bizet
would be Italian; and it is not the case that if Bizet and Verdi were
compatriots, Bizet would not be Italian; nevertheless, if Bizet and
Verdi were compatriots, Bizet either would or would not be Italian.
(Counterfactuals, p. 80.)
Lewis can follow this up by noting that if Bizet and Verdi were compatriots,
Bizet might be Italian, but it’s not the case that if they were compatriots, he
would be Italian.
Here is a related complaint of Lewis’s about Stalnaker’s semantics. In the
last little bit, I’ve used English phrases of the form “if it were the case that φ,
then it might have been the case that ψ”. This conditional Lewis calls “the
‘might’ counterfactual”; he symbolizes it as φ3→ψ, and defines it thus:
Lewis criticizes Stalnaker’s system for the fact that this definition of 3→ doesn’t
work in Stalnaker’s system. Why not? Well, since internal negation is valid in
Stalnaker’s system, φ3→ψ would always imply φ2→ψ—not good, since the
might-conditional in English seems weaker than the would-conditional. So,
Lewis’s definition of 3→ doesn’t work in Stalnaker’s system. Moreover, there
doesn’t seem to be any other plausible definition. So, Stalnaker can’t define
3→.6
Lewis also objects to Stalnaker’s limit assumption. The following line is
less than one inch long:
If the line were more than 1 inch long, it would be over 100 miles
long.
6
Lewis (1973, p. 80).
CHAPTER 8. COUNTERFACTUALS 219
Seems false. But if we use Stalnaker’s truth conditions as truth conditions for
natural language counterfactuals, and take our intuitive judgments of similarity
seriously, we seem to get the result that it is true! The reason is that there
doesn’t seem to be a closest world in which the line is more than 1 inch long.
For every world in which the line is, say, 1+k inches long, there’s another world
in which the line has a length closer to its actual length but still more than 1
inch long: say, 1+ k2 inches. So there doesn’t seem to be any closest world in
which the line is over 1 inch long.
In light of these criticisms, Lewis proposes a new similarity-based seman-
tics for counterfactuals, which assumes neither anti-symmetry nor the limit
assumption. Let’s look at that system.
i) W is a nonempty set
ii) V is a function that assigns either 0 or 1 to each wff relative
to each member of W
7
See Lewis (1973, pp. 48-49). My formulation does away with the accessibility relation (in
Lewis’s terminology, Si , the set of worlds accessible from world i , is always W , the set of all
worlds in the model), so it is a bit simpler.
8
Actually, dropping the limit assumption doesn’t affect the class of valid formulas, which is
the same with or without the limit assumption. (See Lewis (1973, p. 121).
CHAPTER 8. COUNTERFACTUALS 220
It may be verified that every LC-valid wff is SC-valid (although not vice
versa).9
Comments on all this: First, notice that the limit and anti-symmetry con-
ditions are simply dropped. Second, the Base condition is modified; now it
says that no world is as close to a world as itself. Before, it said that each world
is at least as close to itself as any other. Stalnaker’s Base condition, plus anti-
symmetry, entails the present Base condition. But Lewis’s system doesn’t have
anti-symmetry, so the Base condition must be stated in the stronger form.10
Third, let’s think about what the truth condition for the 2→ says. First,
there’s the vacuous case: if φ is necessarily false then φ2→ψ comes out true.
But if φ is possibly true, then what the clause says is this: φ2→ψ is true at
w iff there’s some φ world where ψ is true, such that no matter how much
closer to w you go, you’ll never get a φ world where ψ is false. If there is a
nearest-to-w φ world, then this implies that φ2→ψ is true at w iff ψ is true in
all the nearest-to-w φ worlds.
So, thinking of these as truth-conditions for natural-langague counterfac-
tuals for a moment, recall the sentence:
9
Here’s a sketch of an argument. Let’s say that an LC model is “Stalnaker-acceptable” iff it
obeys the limit and anti-symmetry assumptions. Suppose that φ is LC-valid. Then it’s true in
all Stalnaker acceptable LC-models. Now, notice that in Stalnaker-acceptable models, Lewis’s
truth-conditions for 2→ are identical to Stalnaker’s. So, φ must be true in all SC-models.
10
Why do we want to prohibit worlds being just as close to w as w is to itself? So that P ∧Q
semantically implies P 2→Q. Otherwise P ∧Q could be true at w while P ∧∼Q was true at
some world as close to w as w is to itself, in which case P 2→Q would turn out false at w.
CHAPTER 8. COUNTERFACTUALS 221
If the line were over 1 inch long, it would be over 10 inches long.
There’s no nearest world in which the line is over 1 inch long, only an infinite
series of worlds where the line has lengths getting closer and closer to 1 inch
long. But this doesn’t make the counterfactual true. A counterfactual is true if
its antecedent is impossible, but that’s not true in this case. So the only way the
counterfactual could be true is if the second part of the definition is satisfied—if,
that is, there is some world, x, such that the antecedent is true at x, and the
material conditional (antecedent→ consequent) is true at every world at least
as similar to the actual world as is x. Since the “at least as similar as” relation is
reflexive, this can be rewritten thus:
for some world, x, the antecedent and consequent are both true at
x, and the material conditional (antecedent→ consequent) is true
at every world at least as similar to the actual world as is x
So, is there any such world, x? No. For let x be any world at which the
antecedent and consequent are both true—i.e., any world in which the line is
over 10 inches long. We can always find a world that is more similar to the actual
world than x in which the material conditional (antecedent→ consequent) is
false: just choose a world just like x but in which the line is only, say, 2 inches
long.
Let’s see how Lewis’s theory works in the case of a true counterfactual, for
instance:
If I were more than six feet tall, then I would be less than nine feet
tall
(I am, in fact, less than six feet tall.) The situation here is similar to the previous
example in that there is no nearest world in which the antecedent is true. But
now, we can find a world x, in which the antecedent and consequent are both
true, and such that the material conditional (antecedent→ consequent) is true
in every world at least as similar to the actual world as is x. Simply take x to
be a world just like the actual world but in which I am, say, six-feet-one. Any
world that is at least as similar to the actual world as this world must be one in
which I’m less than nine feet tall; so in any such world the material conditional
(I’m more than six feet tall→ I’m less than nine feet tall) is true.
Notice that the formulas representing Conditional Excluded Middle and
Distribution come out invalid now, because of the possibility of ties.
CHAPTER 8. COUNTERFACTUALS 222
Another thing: Lewis gives the following definition for the ‘might’-counter-
factual:
From this we may obtain a derived clause for the truth conditions of φ3→ψ:
That is, φ3→ψ is true at w iff φ is possible, and for any φ world, there’s a
world as close or closer to w in which φ and ψ are both true. In cases where
there is a nearest φ world, this means that ψ must be true in at least one of the
nearest φ worlds.
In general, one is entitled to conclude from “If P or Q had been the case,
then R would have been the case” that “if P had been the case, R would have
been the case”. If Butch Cassidy and the Sundance Kid could have survived by
11
For references, see the bibliography of Lewis (1977).
CHAPTER 8. COUNTERFACTUALS 223
surrendering, they certainly would not say to each other “If we had surrendered
or tried to run away, we would have been shot”.
Is this a problem for Lewis and Stalnaker? Some have argued this, but
others respond as follows. One must take great care in translating from natural
language into logic. For example,12 no one would want to criticize the law
∼∼P →P on the grounds that “There ain’t no way I’m doing that” doesn’t
imply that I might do that. And there are notorious peculiar things about the
behavior of ‘or’ in similar contexts. Consider:
One can argue that this does not have the form:
After all, suppose that you are permitted to stay, but not to go. If you stay, you
can’t help doing the following act: staying ∨ going. So, surely, you’re permitted
to do that. So, (2) is true. But (1) isn’t; if someone uttered (1) to you when you
were in jail, they’d be lying to you! (1) really means:
(3) You are permitted to stay, AND you are permitted to go.
Similarly, “If either P or Q were true then R would be true” seems usually to
mean “If P were true then R would be true, and if Q were true then R would be
true”. We can’t just expect natural language to translate directly into our logical
language—sometimes the surface structure of natural language is misleading.
12
The example is adapted from Loewer (1976).
Chapter 9
If we don’t have any quantifiers around, it seems that we have only one possible
symbolization: as 3P , where P stands for “Some rich person is poor”. But this
symbolization is incorrect, because in no possible world is the statement “Some
224
CHAPTER 9. QUANTIFIED MODAL LOGIC 225
rich person is poor” true! In contrast, (1) seems to have a reading on which it
is true. When we turn to symbolize it in QML, we have better luck. There
seem to be two possibilities:
(1a) 3∃x(Rx∧P x)
(1b) ∃x(Rx∧3P x)
The de dicto reading is true because it is true that in any possible world, anyone
that is in that world a bachelor is, in that world, male. The de re reading is
false because it makes the claim that if any object, u, is a bachelor in the actual
world, that object u is necessarily a bachelor—i.e., the object u is a bachelor in
all possible worlds. That claim is false: many objects are in fact bachelors, but
might have been married.
Definite descriptions also exhibit a de re/de dicto ambiguity. This may be
illustrated by using Russell’s theory of descriptions (section 5.3.3). Russell’s
method generates an ambiguity when a sentential operator such as “not” is
added to “is G”, for example:
Recall that our semantics for modal propositional logic assigned truth values
to sentence letters relative to possible worlds. We have something similar
here: we relativize the interpretation of predicates to possible worlds. The
interpretation of a two-place predicate, R, for example is a set of ordered
triples, two members of which are in the domain, and one member of which is
a possible world. When 〈u1 , u2 , w〉 is in the interpretation of R, that represents
R’s applying to u1 and u2 in possible world w. In a possible worlds setting, this
relativization makes intuitive sense: a predicate can apply to some objects in
one possible world but fail to apply to those same objects in some other possible
world.
Notice that the interpretations of constants are not relativized in any way to
possible worlds. The interpretation I assigns simply a member of the domain
to a name. This reflects the common belief that natural language proper
names—which constants are intended to represent—are rigid designators, i.e.,
terms that have the same denotation relative to every possible world (see Kripke
(1972).) We’ll discuss the significance of this feature of our semantics below.
On to the definition of the valuation function for an SQML-model. First,
we keep the definition of a variable assignment from nonmodal predicate logic
(section 4). Our variable assignments therefore assign members of the domain
to variables absolutely, rather than relative to worlds. (This is an appropriate
choice given our choice to assign constants absolute semantic values.) But the
valuation function will now relativize truth values to possible worlds (as well
as to variable assignments). After all, the sentence ‘F a’, if it represents “Ted is
tall”, should vary in truth value from world to world.
The derived clauses are what you’d expect, including the following one for
3:
Finally, we have:
account for this? Let’s look at an example: finding a countermodel for the
formula (3F a∧3Ga)→3(F a∧Ga).
We begin as follows:
∗
1 1 1 0 0
r
(3F a∧3Ga)→3(F a∧Ga)
∗ ∗
The understars make us create two new worlds:
∗
1 1 1 0 0
r
(3F a∧3Ga)→3(F a∧Ga)
∗ ∗
1 1
a b
Fa Ga
We must then discharge the overstar from the false diamond in each world
(since every world is accessible to every other world in our models):
∗
1 1 1 0 0 0 0
(3F a∧3Ga)→3(F a∧Ga)
r
∗ ∗ †
1 0 0 1 0 0
a b
Fa F a∧Ga Ga F a∧Ga
(I had to make either F a or Ga false in r—I chose F a arbitrarily.) Now, we’ve
indicated the truth-values that we want the atomics to have. How do we make
the atomics have the TVs we want in the picture?
We do this by introducing a domain for the model, and stipulating what the
names refer to and what objects are in the extensions of the predicates. Let’s
use letters like ‘u’ and ‘v’ as the members of the domain in our models. Now, if
we let the name ‘a’ refer to (the letter) u, and let the extension of F in world r
CHAPTER 9. QUANTIFIED MODAL LOGIC 230
be {} (the empty set), then the truth value of ‘F a’ in world r will be 0 (false),
since the denotation of a isn’t in the extension of F at world r. Likewise, we
need to put u in the extension of F (but not in the extension of G) in world
a, and put u in the extension of G ((but not in the extension of F ) in world b.
This all may be indicated on the diagram as follows:
a: u
∗
1 1 1 0 0 0 0
(3F a∧3Ga)→3(F a∧Ga)
r
∗ ∗ †
F :{}
1 0 0 1 0 0
Fa F a∧Ga Ga F a∧Ga
a b
F : {u} G : {} F : {} G : {u}
Within each world I’ve included a specification of the extension of each predi-
cate. But the specification of the referent of the name ‘a’ does not go within
any world; it was rather indicated (in boldface) at the top of model. This is
because names, unlike predicates, get assigned semantic values absolutely in a
model, not relative to worlds.
Time for the official model:
W = {r, a, b}
D = {u}
I (a) = u
I (F ) = {〈u, a〉}
I (G) = {〈u, b〉}
Let’s try one with quantifiers: 2∃xF x→∃x2F x:
∗ +
1 1 0 0
r
2 ∃ xF x→ ∃ x2F x
+
CHAPTER 9. QUANTIFIED MODAL LOGIC 231
The overstar above the 2 in the antecedent must be discharged in r itself, since,
remember, every world sees every world in these models. That gives us a true
existential. Now, a true existential is a bit like a true 3—the true ∃xF x means
that there must be some object u from the domain that’s in the extension of F
in r. I’ll put a + under true ∃s and false ∀s, to indicate a commitment to some
instance of some sort or other. Analogously, I’ll indicate a commitment to all
instances of a given type (which would arise from a true ∀ or a false ∃) with a +
above the connective in question.
OK, how do we make ∃xF x true in r? By making “F x” true for some
value of x. Let’s put the letter u in the domain, and make “F x” true when u is
assigned to x. We’ll indicate this by putting a 1 overtop of “F ux ” in the diagram.
Now, “F ux ” isn’t a formula of our language—what it indicates is that “F x” is to
be true when u is assigned to x. And to make this come true, we treat it as an
atomic—we put u in the extension of F at r:
∗ +
1 1 0 0 1
2 ∃ xF x→ ∃ x2F x F ux
+
r
F : {u}
Good. Now we’ve got to attend to the overplus, the + sign overtop the false
∃x2F x. Since it’s a false ∃, we’ve got to make 2F x false for every object in the
domain (otherwise—if there were something in the domain for which 2F x was
true—∃x2F x would be true after all). So far, we’ve got only one object in our
domain, u, so we’ve got to make 2F x false, when u is assigned to the variable
‘x’. We’ll indicate this on the diagram by putting a 0 overtop of “2F ux ”:
∗ +
1 1 0 0 1 0
2 ∃ xF x→ ∃ x2F x F u
x
2F ux
+
r
∗
F : {u}
Ok, now we have an understar, which means we should add a new world to
our model. When doing so, we’ll need to discharge the overstar from the
antecedent. We get:
CHAPTER 9. QUANTIFIED MODAL LOGIC 232
∗ +
1 1 0 0 1 0
2 ∃ xF x→ ∃ x2F x F u
x
2F ux
+
r
∗
F : {u}
0 1 1
F ux ∃ xF x F v
x
a +
F : {v}
This move requires some explanation. Why the v? Well, I was required to
make F x false, with u assigned to x. Well, that means keeping u out of the
extension of F at a. Easy enough, right? Just make F ’s extension {}? Well,
no—because of the true 2 in r, I’ve got to make ∃xF x true in a. But that means
that something’s got to be in F ’s extension in a! It can’t be u, so I’ll add a new
object, v, to the domain, and put it in F ’s extension in a.
But adding v to the domain of the model adds a complication. We had
an overplus in r—over the false ∃. That meant that, in r, for every member of
the domain, 2F x is false. So, 2F x is false in r when v is assigned to x. That
creates another understar, requiring the creation of a new world. The model
then looks as follows:
CHAPTER 9. QUANTIFIED MODAL LOGIC 233
∗ +
1 1 0 0 1 0 0
2 ∃ xF x→ ∃ x2F x F u
x
2F u
x
2F v
x
+
r
∗ ∗
F : {u}
0 1 1 0 1 1
F u
x
∃ xF x F v
x
F v
x
∃ xF x F ux
a + b +
F : {v} F : {u}
(Notice that we needn’t have made another world b—we could simply have
discharged the understar on r.)
Ok, here’s the official model:
W = {r, a, b}
D = {u, v}
I (F ) = {〈u, r〉, 〈u, b〉, 〈v, a〉}
Another example: I’ll do a validity proof for the following formula: 3∃x(x =
a∧2F x)→F a:
i) suppose for reductio that (for some model, some world r , and
some variable assignment g ,) Vg (3∃x(x=a∧2F x)→F a, r ) =
0. Thus Vg (3∃x(x=a∧2F x), r ) = 1 and …
ii) …Vg (F a, r ) = 0
iii) From i), for some w ∈ W , Vg (∃x(x=a∧2F x), w) = 1
iv) so for some u ∈ D, Vgu/x (x=a∧2F x, w) = 1)
v) Thus, Vgu/x (x=a, w) = 1 and …
vi) …Vgu/x (2F x, w) = 1
vii) from vi), Vgu/x (F x, r ) = 1
viii) Thus, 〈[x] gu/x , r 〉 ∈ I (F )—that is, 〈u, r 〉 ∈ I (F )
CHAPTER 9. QUANTIFIED MODAL LOGIC 234
Notice that in line xii) I inferred that Vg assigned “F a” truth at r . I could have
subscripted V with any variable assignment, since the truth condition for the
formula “F a” is the same, regardless of the variable assignment; I picked g
because that’s what I needed to get the contradiction.
∀x∀y(x=y→2(x=y))
When we try to make the formula false by putting a 0 over the initial ∀, we get
an under-plus. So we’ve got to make the inside part, ∀y(x=y→2x=y), false
for some value of x. We do this by putting some object u in the domain, and
letting that be the value of x for which ∀y(x=y→2x=y) is false. We get:
0 0
r ∀ x∀y(x=y→2x=y) ∀y( ux =y→2( ux =y))
+ +
Now we need to do the same thing for our new false universal: ∀y(x=y→2x=y).
For some value of y, the inside conditional has to be false. But that means that
the antecedent must be true. So the value for y has to be u again. We get:
CHAPTER 9. QUANTIFIED MODAL LOGIC 235
0 0 1 0 0
The understar now requires creation of a world in which x=y is false, when
both x and y are assigned u. But there cannot be any such world! An identity
sentence is true (at any world) if the denotations of the terms are identical.
Our attempt to find a countermodel has failed; we must do a validity proof.
Consider any QML model 〈W , D, I 〉, any r ∈ W , and any variable assignment
g ; we’ll show that Vg (∀x∀y(x=y→2x=y), r ) = 1:
Notice at the end how the particular world at which the identity sentence was
false didn’t matter. The truth condition for an identity sentence is simply that
the terms denote the same thing; it doesn’t matter what world this is evaluated
relative to.2
2
A note about variables. In validity proofs, I’m using italicized ‘u’ and ‘v’ as variables to
range over objects in the domain of the model I’m considering. So, a sentence like ‘u = v’
might be true, just as the sentence ‘x=y’ of our object language can be true. But when I’m
doing countermodels, I’m using the roman letters ‘u’ and ‘v’ as themselves being members of
the domain, not as variables ranging over members of the domain. Since the letters ‘u’ and ‘v’
are different letters, they are different members of the domain. Thus, in a countermodel with
letters in the domain, if the denotation of a name ‘a’ is the letter ‘u’, and the denotation of the
name ‘b ’ is the letter ‘v’, then the sentence ‘a=b ’ has got to be false, since ‘u’6=‘v’. If I were
using ‘u’ and ‘v’ as variables ranging over members of the domain, then the sentence ‘u = v’
might be true! This just goes to show that it’s important to distinguish between the following
two sentences:
CHAPTER 9. QUANTIFIED MODAL LOGIC 236
The first could be true, depending on what ‘u’ and ‘v’ currently refer to, but the second one is
just plain false, since ‘u’ and ‘v’ are different letters.
3
See Gibbard (1975).
CHAPTER 9. QUANTIFIED MODAL LOGIC 237
0 0
+ ∀ xF x F ux
1 0 0 +
r a
∀ x2F x→2∀xF x
∗
F : {}
When you have a choice between discharging over-things and under-things,
whether plusses or stars, always do the under things first. In this case, this
means discharging the understar and ignoring the over-plus for the moment.
So, discharging the understar gave us world a, in which we made a universal
false. This gave an underplus, and forced us to make an instance false. So I put
object u in our domain, and keep it out of the extension of F in a. This makes
F x false in a, when x is assigned u.
But now, I need to discharge the overplus in r. I must make 2F x true for
every member of the domain, including u, which is now in the domain. But
then this requires F x to be true, when u is assigned to x, in a:
+ ∗ 0 0 1
1 0 0 1 1 ∀ xF x F u
x
F ux
∀ x2F x→2∀xF x 2F u
x
+
r a
∗
F : {u} F : {?}
So, we fail to get a model. Time for a validity proof; let’s show that every
instance of the Barcan Schema is valid:
vi) thus, for every world, and so for w in particular, Vgu/α (φ, w) = 1.
Contradicts iv).
The validity of the Barcan formula in our semantics is infamous because the
Barcan formula seems, intuitively, to be invalid. To see why, we need to think a
bit about the intuitive significance of the relative order of quantifiers and modal
operators. Consider the difference between the following two sentences:
3∃xF x
∃x3F x
In general, a sentence of the form 3φ says that it’s possible for the component
sentence, φ, to be true. So the first of our two sentences, 3∃xF x, says that
it’s possible for “∃xF x” to be true. That is: it’s possible for there to exist
an F . What about the second sentence? In general, a sentence that begins
without a modal operator in front makes a statement about the actual world.
Thus, a statement that begins with “∃x . . . ” is saying that there exists, in the
actual world, an object x, such that…. Our second statement, then, says that
there actually exists an object, x, that is possibly F . We have seen that our two
statements have the following meanings:
It matters, therefore, whether the ∃ comes after or before the 3. If the ∃ comes
first, then the statement is saying that there actually exists a certain sort of
object (namely, an object that could have been a certain way.) But if it comes
second, after the 3, then the statement is merely saying that there could have
existed a certain sort of object. There is a similar contrast with the following
two statements:
2∀xF x
∀x2F x
The first says that it’s necessary that: everything is F . That is, in every possible
world, every object that exists in that world is F in that world. The objects
ranged over by the ∀, so to speak, are drawn from the worlds the 2 introduces,
because the ∀ occurs inside the scope of the 2. The second statement, by
CHAPTER 9. QUANTIFIED MODAL LOGIC 239
contrast, says that: every actual object is necessarily F . That is, every object that
exists in the actual world is F in every possible world. The second statement
concerns just actually existing objects because the ∀ occurs in the front of the
formula, not inside the scope of the 2.
With all this in mind, return to the Barcan formula, ∀x2F x→2∀xF x. It
says:
“If every actually existing thing is F in every possible world, then
in every world, every object in that world is F in that world”
Now we can see why this claim is questionable. Even if every actual thing is
necessarily F , there could still be worlds containing non-F things, so long as
those non-F things don’t exist in the actual world. Suppose, for instance, that
every object in the actual world is necessarily a material object. Then, letting
F stand for “is a material object”, ∀x2F x is true. Nevertheless, 2∀xF x seems
false—it would presumably be possible for there to exist an immaterial object:
a ghost, say. Possible worlds containing ghosts would simply need to contain
objects that do not exist in the actual world (since all the objects in the actual
world are necessarily material.)
This objection to the validity of the Barcan formula is obviously based on
the idea that what objects exist can vary from possible world to possible world.
But this sort of variation is not represented in the SQML definition of a model.
Each such model contains a single domain, D, rather than different domains
for different possible worlds. The truth condition we specified for a quantified
sentence ∀αφ, at a world w, was simply that φ is true at w of every member of
D—the quantifier ranges over the same domain, regardless of which possible
world is being described. That is why the Barcan formula turns out valid under
our definition.
This feature of SQML models is problematic for an even more direct reason:
the sentence ∀x2∃y y = x, i.e., “everything necessarily exists”, turns out valid!:
i) Suppose for reductio that Vg (∀x2∃y y=x, w) = 0.
ii) Then Vg (u/x) (2∃y y=x, w) = 0, for some u ∈ D
iii) So Vg (u/x) (∃y y=x, w 0 ) = 0, for some w 0 ∈ W
iv) So Vg (u/x u 0/y) (y=x, w 0 ) = 0, for every u 0 ∈ D
v) So, since u ∈ D, we have Vg (u/x u/y) (y=x, w 0 ) = 0. But that can’t
be, given the clause for ‘=’ in the definition of the valuation
function.
CHAPTER 9. QUANTIFIED MODAL LOGIC 240
It’s clear that this formula turns out valid for the same reason that the Barcan
formula turns out valid: SQML models have a single domain common to each
possible world.
The Barcan schema is just one of a number of interesting schemas concern-
ing how quantifiers and modal operators interact (for each schema I also list an
equivalent schema with 3 in place of 2):
∀α2φ→2∀αφ 3∃αφ→∃α3φ (Barcan)
2∀αφ→∀α2φ ∃α3φ→3∃αφ (Converse Barcan)
∃α2φ→2∃αφ 3∀αφ→∀α3φ
2∃αφ→∃α2φ ∀α3φ→3∀αφ
We have already discussed the Barcan schema. The third schema raises no philo-
sophical problems for SQML, since, quite properly, it has instances that turn out
invalid. As we saw above, there are SQML models in which 2∃xF x→∃x2F x
is false. Let’s look at the other two schemas.
First, the converse Barcan schema. Like the Barcan schema, each of its
instances is valid given the SQML semantics (I’ll leave this to the reader to
demonstrate), and like the Barcan schema, this verdict faces a philosophical
challenge. The antecedent says that in every world, everything that exists in
that world is φ. Existents are thus always φ. It might still be that some object
isn’t necessarily φ: perhaps some object that is φ in every world in which it
exists, fails to be φ in worlds in which it doesn’t exist. This talk of an object
being φ in a world in which it doesn’t exist may seem strange, but consider
the following instance of the converse Barcan schema, substituting “∃y y=x”
(think: “x exists”) for φ:
2∀x∃y y=x→∀x2∃y y=x
This formula seems to be false. Its antecedent is clearly true; but its consequent
says that every object in the actual world exists necessarily, and hence seems
intuitively to be false.
Each instance of the fourth schema, ∃α2φ→2∃αφ, is also validated by
the SQML semantics (again, an exercise for the reader); and again, this is
philosophically questionable. Let’s suppose that physical objects are necessarily
physical. Then, ∃x2P x seems true, letting P mean ‘is physical’. But 2∃xP x
seems false—it seems possible that there are no physical objects. This coun-
terexample requires that there be worlds with fewer objects than those that
actually exist, whereas the counterexample to the Barcan formula involved the
possibility that there be more objects than those that actually exist.
CHAPTER 9. QUANTIFIED MODAL LOGIC 241
presumably have no mass (nor any spatial location, nor any other physical
feature.) Ordinary quantification is restricted to normal things. So if we want
to translate an ordinary claim into the language of QML, we must introduce
a predicate for the normal things, “N ”, and use it to restrict quantifiers. But
now, consider the following ordinary English statement:
which says:
And this formula is not an instance of the Barcan schema, nor is it valid, as may
be shown by the following countermodel:
+
1 0 0 0 1 0
∀ x(N x→2F x)→2∀x(N x→F x) N u
x
→2F ux
r
∗ †
N : {}
0 1 0 0
∀ x(N x→F x) N u
x
→F ux
a +
N : {u} F : {}
So in a sense, the ordinary intuitions that were alleged to undermine the Barcan
schema are in fact consistent with Constancy.
CHAPTER 9. QUANTIFIED MODAL LOGIC 243
The defender of Constancy can defend the converse Barcan schema and the
fourth schema in similar fashion. The objection to the converse Barcan schema
assumed the falsity of ∀x2∃y y=x. “Sheer prejudice!”, according to the friend
of constancy. “And recall further that an ordinary utterance of ‘Everything exists
necessarily’ expresses, not ∀x2∃y y=x, but rather ∀x(N x→2∃y(N y∧y=x)),
(N for ‘normal’), the falsity of which is is perfectly compatible with Constancy.
It’s possible to fail to be normal; all that’s impossible is to utterly fail to exist.
Likewise for the fourth schema.”
This defense of SQML is hard to take. Let “G” stand for a kind of object
that, in fact, has no members, but which could have had members. Perhaps ghost
is such a kind. ∀x2∼Gx→2∀x∼Gx is an instance of the Barcan schema, and
so true according to the defender of Constancy. Since there could have existed
ghosts, the consequent of this conditional is false. Therefore, its antecedent
∀x2∼Gx must be false. That is, there exists something that could have been a
ghost. But this is a very surprising result. The alleged possible ghost couldn’t
be any material object, presumably, assuming it would be impossible for any
material object to be a ghost. The defender of Constancy, then, is committed to
the existence of objects which we wouldn’t otherwise have dreamed of accepting:
things that could have been ghosts, things that been dragons, things that could
have been gods, and so on.
The defender of Constancy might try to defend this conclusion by remind-
ing us that these “possible-ghosts”, “possible-dragons”, and so on, are not
normal objects. They aren’t in space and time, presumably, which explains why
no one has ever seen, heard, felt, or smelled one. He might even say that they
are non-actual, or even that they do not exist (though they are). We are quite cor-
rect, she might say, to scoff at the idea that some normal/actual/existing objects
are capable of being ghosts; but what’s the big deal about saying that some non-
normal/non-actual/non-existing objects have these capabilities? This move,
too, will be considered philosophically suspect by many. Many philosophers
regard the idea that there are some non-existent things, or some non-actual
things, as being anywhere from obviously false to conceptually incoherent, or
subversive, or worse.4 And how does it help to point out that the objects aren’t
normal? The postulation of non-normal objects—objects above and beyond
the objects that the rest of us believe in—was exactly what I was claiming is
philosophically suspect!
On the other hand, Constancy’s defenders can point to certain powerful
4
See Quine (1948); Lycan (1979).
CHAPTER 9. QUANTIFIED MODAL LOGIC 244
arguments in its favor. Here’s a quick sketch of one such argument. First, the
following seems to be a logical truth:
Ted = Ted
∃y y = Ted
This latter formula, too, is therefore a logical truth. But if φ is a logical truth
then so is 2φ (recall the rule of necessitation from chapter 6). So we may infer
that the following is a logical truth:
2∃y y = Ted
Next, notice that nothing in the argument for 2∃y y = Ted depended on any
special features of me. We may therefore conclude that the reasoning holds
good for every object; and so ∀x2∃y y = x is indeed a logical truth. Since,
therefore, every object exists necessarily, it should come as no surprise that
there are things that might have been ghosts, dragons, and so on—for if there
had been a ghost, it would have necessarily existed, and thus must actually exist.
This and other related arguments have apparently wild conclusions, but they
cannot be lightly dismissed, for it is no mean feat to say exactly where they go
wrong (if they go wrong at all!).5
The third one on the list was invalid before, and so is still invalid now. For
note: if M is a SQML model, then we can construct a corresponding VDQML
model with the same set of worlds, (super-) domain, and interpretation function,
in which every world is accessible from every other, and in which Q is a constant
function assigning the whole super-domain to each world. It is intuitively clear
that the same sentences are true in this corresponding model as are true in M .
Hence, whenever a sentence is SQML-invalid, it is VDQML-invalid.
As for the Barcan formula ∀x2F x→2∀xF x, consider the following VDQML-
model:
CHAPTER 9. QUANTIFIED MODAL LOGIC 247
r Dr : {u} F : {u}
0
a Da : {u, v} F : {u}
0
Official model:
W : {r, a}
R : {〈r, r〉, 〈r, a〉, 〈a, a〉}
D : {u, v}
Dr : {u} Da : {u, v}
I (F ) : {〈u, r〉, 〈u, a〉}
As for the fourth formula on the list, ∃x2F x→2∃xF x, we can construct a
countermodel as follows:
r Dr : {u, v} F : {u}
0
a Da : {v} F : {u}
0
W : {r, a}
R : {〈r, r〉, 〈r, a〉, 〈a, a〉}
D : {u, v}
Dr : {u, v} Da : {v}
I (F ) : {〈u, r〉, 〈u, a〉}
And for the converse Barcan formula, 2∀xF x→∀x2F x, we have the fol-
lowing countermodel:
r Dr : {u, v} F : {u, v}
0
a Da : {v} F : {v}
0
CHAPTER 9. QUANTIFIED MODAL LOGIC 248
W : {r, a}
R : {〈r, r〉, 〈r, a〉, 〈a, a〉}
D : {u, v}
Dr : {u, v} Da : {v}
I (F ) : {〈u, r〉, 〈v, r〉, 〈v, a〉}
The counterexample would then go away. Indeed, every instance of the Barcan
schema would then become VDSQML-valid, which may be proved as follows:
Note further that even after imposing the increasing domains requirement,
the Barcan formula remains VDQML-invalid; and after imposing the decreas-
ing domains requirement, the converse Barcan formula and also ∃x2F x→2∃xF x
remain VDQML-invalid; this can be seen by the original countermodels for
these formulas above, each of which obeys the requirements in question. How-
ever, it should also be noted that in systems in which the accessibility relation
is symmetric, this collapses: imposing either of these requirements results in
imposing the other. That is, in B or S5, imposing either the increasing or the
decreasing domains requirement results in imposing both, and hence results in
all three formulas being validated.
This argument goes by a little too quickly. The main worry of its proponent
is that our semantics requires a to exist necessarily, in order for 2F a to come
out true. But our semantics doesn’t require this. It does require F a to be
true in every world, in order for 2F a to be true; but it does not require a
to exist in every world in which F a is true. The clause in the definition of a
VDQML-model for the interpretation of predicates was this:
thus barring objects from having properties at worlds where they don’t ex-
ist. But some would argue that this goes too far. Given b0 ), the sentence
∀x2(F x→∃y y=x) turns out valid. “An object must exist in order to be F ”—
sounds clearly true if F stands for ‘is human’, but what if F stands for ‘is famous’?
Suppose John Lennon had never existed (it was all a communist plot); some
would argue that in such a circumstance Lennon still would have been famous.
The issues here are complex.8 But whether or not b) should be replaced
with b0 ), it looks as though there are some existence-entailing English predicates
π: predicates π such that nothing can be a π without existing. ‘Is human’ seems
to be such a predicate. So we’re back to our original worry about VDQML-
semantics: its truth condition for 2φ requires truth of φ at all worlds, which is
allegedly too strong, at least when φ is a sentence like πa, where π is existence-
entailing.
8
The question is that of so-called “serious actualism” (Plantinga, 1983).
CHAPTER 9. QUANTIFIED MODAL LOGIC 251
One could modify the clause for the 2 in the definition of the valuation
function, so that in order for 2F a to be true, a only needs to be F in worlds in
which it exists:
This would indeed have the result that 2F a gets to be true provided a is F in
every world in which it exists. But be careful what you wish for. Along with
this result comes the following: even if a doesn’t necessarily exist, the sentence
2∃x x=a comes out true. For according to d0 ), in order for 2∃x x=a to be
true, it must merely be the case that ∃x x=a is true in every world in which a
exists, and of course this is indeed the case.
If 2∃x x=a comes out true even if a doesn’t necessarily exist, then 2∃x x=a
doesn’t say that a necessarily exists. Indeed, it doesn’t look like we have any way
of saying that a necessarily exists, using the language of QML, if the 2 has the
meaning provided for it by d0 ).
A notion of necessity according to which “Necessarily φ” requires truth in
all possible worlds is sometimes called a notion of strong necessity. In contrast,
a notion of weak necessity is one according to which “Necessarily φ” requires
merely that φ be true in all worlds in which objects named within φ exist. d0 )
corresponds to weak necessity, whereas our original definition d) corresponds
to strong necessity.
As we saw, if the 2 expresses weak necessity, then one cannot even express
the idea that a thing necessarily exists. That’s because one needs strong necessity
to say that a thing necessarily exists: in order to necessarily exist, you need to
exist at all worlds, not just at all worlds at which you exist! So this is a serious
deficiency of having the 2 of QML express weak necessity. But if we allow the
2 to express strong necessity instead, there is no corresponding deficiency, for
one can still express weak necessity using the strong 2 and other connectives.
For example, to say that a is weakly necessarily F (that is, that a is F in every
world in which it exists), one can say: 2(∃x x=a→F a).
So it would seem that we should stick with our original truth condition d)
for the 2, and live with the fact that statements like 2F a turn out false if a fails
to be F at worlds in which it doesn’t exist. Those who think that “Necessarily,
9
I.e., not bound to any quantifier in φ.
CHAPTER 9. QUANTIFIED MODAL LOGIC 252
Ted is human” is true despite Ted’s possible nonexistence can always translate
this natural language sentence into the language of QML as 2(∃x x=a→F a)
(which requires a to be F only at worlds at which it exists) rather than as 2F a
(which requires a to be F at all worlds).
Chapter 10
10.1 Actuality
The word ‘actually’, in one of its senses anyway, can be thought of as a one-place
sentence operator: “Actually, φ.”
‘Actually’ might at first seem redundant. “Actually, snow is white” basically
amounts to: “snow is white”. But the actuality operator interacts with modal
operators in interesting ways. The following two sentences, for example, clearly
have different meanings:
The first sentence expresses the triviality that snow is white in any possible
world in which snow is white. But the second sentence makes the nontrivial
statement that if snow is white in any world, then snow is white in the actual
world.
So, ‘actually’ is nonredundant, and consequently, worth thinking about.
Let’s add a symbol to modal logic for it. “@φ” will symbolize “Actually, φ”.
We can now symbolize the pair of sentences above as 2(S→S) and 2(S→@S),
253
CHAPTER 10. TWO-DIMENSIONAL MODAL LOGIC 254
The old definition of validity for a system, recall, never employed the notion
of truth in a model; rather, it proceeded via the notion of validity in a frame.
The nice thing about the new definition is that it’s parallel to the way validity
is usually defined in model theory: one first defines truth in a model, and then
defines validity as truth in all models. But the new definition doesn’t differ in
any substantive way from the old definition, in that it yields exactly the same
class of valid formulas:
10.1.3 Examples
Example 1: Show that ∀x(F x∨2Gx)→2∀x(Gx∨@F x)
i) Suppose for reductio that this formula is not valid. Then for
some model and some variable assignment g , Vg (∀x(F x∨2Gx)→
2∀x(Gx∨@F x), w@ ) = 0.
ii) Then Vg (∀x(F x∨2Gx), w@ ) = 1 and…
iii) …Vg (2∀x(Gx∨@F x), w@ ) = 0
iv) Given the latter, there is some world, call it “a”, such that
R w@ a and Vg (∀x(Gx∨@F x), a) = 0. And so, there is some
object, call it “u”, in the model’s domain, D, such that
Vgu/x (Gx∨@F x, a) = 0
v) And so Vgu/x (Gx, a) = 0 and…
vi) …Vgu/x (@F x, a) = 0
vii) Given the latter, Vgu/x (F x, w@ ) = 0 (by the clause in the truth
definition for @)
viii) Given ii), for every object in D, and so for u in particular,
Vgu/x (F x∨2Gx, w@ ) = 1.
ix) And so, either Vgu/x (F x, w@ ) = 1 or Vgu/x (2Gx, w@ ) = 1
x) From ix and vii, Vgu/x (2Gx, w@ ) = 1
xi) And so, Vgu/x (Gx, a) = 1—contradicts v)
10.2 ×
Adding @ to the language of quantified modal logic is a step in the right
direction, since it allows us to express certain kinds of comparisons between
possible worlds that we couldn’t express otherwise. But it doesn’t go far enough;
we need a further addition.2 Consider this sentence:
It might have been the case that, if all those then rich might all
have been poor, then someone is happy
This is a bit like “It might have been that everyone who is actually rich is poor”;
in this new sentence the word ‘then’ plays a role a bit like the role ‘actually’
played in the earlier sentence. But the ‘then’ does not take us back to the actual
world of the model; it rather takes us back to the world, w, that is introduced
by the first possibility operator, ‘it might have been the case that’. We cannot,
therefore, symbolize our new sentence thus:
3(3∀x(@Rx→P x)→∃xH x)
for this has the truth condition that there is world w such that, if there’s a world
v such that (everyone who is rich in w@ is poor in v), then someone is happy in
w. The problem is that the @, as we’ve defined it, always takes us back to the
model’s designated world, whereas what we need to do is to “mark” a world,
and have @ take us back to the “marked” world:
3×(3∀x(@Rx→P x)→∃xH x)
VM , g (φ, w1 , w2 )
will mean that φ is true at world w2 , with reference world w1 . We define [α]M ,g ,
the denotation of term α relative to model M and variable assignment g , as
before. And VM ,g is defined as the function that assigns to each wff, relative to
each pair of worlds, either 0 or 1 subject to the following constraints:
Note what the × does: change the reference world. When evaluating a formula,
it says to forget about the old reference world, and make the new reference
world whatever the current world of evaluation happens to be.
We can define validity and consequence thus:
φ is 2D-valid (“2D φ”) iff for every model M , every world w in that
model, and every assignment g based on that model, VM ,g (φ, w, w) =
1
φ is a 2D-semantic consequence of Γ (“Γ 2D φ”) iff for every model
M , every assignment g based on that model, and every world w in
that model, if VM ,g (γ , w, w) = 1 for each γ ∈ Γ, then VM , g (φ, w, w) =
1
Valid formulas are thus defined as those that are true at every pair of worlds of
the form 〈w, w〉; semantic consequence is truth-preservation at every pair of
worlds 〈w, w〉.
Notice, however, that these aren’t the only notions of validity and con-
sequence that one could introduce. There is also the notion of truth, and
truth-preservation, at every pair of worlds:3
φ is generally 2D-valid (“G2D φ”) iff for every model M , any worlds
v and w in that model, and every assignment g based on that model,
VM ,g (φ, v, w) = 1
φ is a general 2D-semantic consequence of Γ (“Γ G2D φ”) iff for every
model M , every assignment g based on that model, and any worlds
v and w in that model, if VM ,g (γ , v, w) = 1 for each γ ∈ Γ, then
VM ,g (φ, v, w) = 1
Validity and general validity, and consequence and general consequence, come
apart in various ways, as we’ll see below.
As we saw, moving to this new language increases the flexibility of the @;
we can symbolize
It might have been the case that, if all those then rich might all
have been poor, then someone is happy
3
The term ‘general validity’ is from Davies and Humberstone (1980); the first definition of
validity corresponds to their “real-world validity”.
CHAPTER 10. TWO-DIMENSIONAL MODAL LOGIC 260
as
3×(3∀x(@Rx→P x)→∃xH x)
Moreover, it costs us nothing. For we can replace any sentence φ of the old
language with ×φ in the new language (i.e. we just put the × operator at the
front of the sentence.)4 For example, instead of symbolizing
10.2.2 Examples
Example 1: show that if φ then @φ:
Example 2: φ↔@φ, but 2 2(φ↔@φ). (Moral: any proof theory for this
logic had better not include the rule of necessitation!)
The truth condition for @ insures that for any world w in any model
(and any variable assignment), V(@φ, w, w) = 1 iff V(φ, w, w) = 1,
and so V(φ↔@φ, w, w) = 1. Thus, φ↔@φ.
But some instances of 2(φ↔@φ) aren’t valid. Let φ be ‘Fa’; here’s
a countermodel:
W = {c, d}
D = {u}
I (a) = u
I (F ) = {〈u, c〉}
4
This amounts to the same thing as the old symbolization in the following sense. Let
φ be any wff of the old language. Thus, φ may have some occurrences of @, but it has no
occurrences of ×. Then, for every QML-model M = 〈W , D, I 〉, and any v, w ∈ W , ×φ is
true at 〈v, w〉 in M iff φ is true in the designated-world QML model 〈W , w, D, I 〉.
CHAPTER 10. TWO-DIMENSIONAL MODAL LOGIC 261
Example 3: φ→2@φ
10.3 Fixedly
The two-dimensional approach to semantics—evaluating formulas at pairs of
worlds rather than single worlds—raises an intriguing possibility. The 2 is a
universal quantifier over the world of evaluation; we might, by analogy, follow
Davies and Humberstone (1980) and introduce an operator that is a universal
quantifier over the reference world. Davies and Humberstone call this operator
F, and read “Fφ” as “fixedly, φ”. Grammatically, F is a one-place sentential
operator. Its semantic clause is this:
But this isn’t significantly different from the present two-dimensional approach.
10.3.1 Examples
Example 1: for any wff, φ, 2D F@φ→φ
CHAPTER 10. TWO-DIMENSIONAL MODAL LOGIC 263
W = {c, d}
D = {u}
I (a) = u
I (G) = {〈u, c〉}
First, let φ be atomic. Then we’re trying to show that for any worlds
v, w, and any variable assignment g , Vg (Πα1 . . . αn →F Πα1 . . . αn , v, w) =
1. Suppose otherwise—suppose that (i) Vg (Πα1 . . . αn , v, w) = 1
and (ii) Vg (F Πα1 . . . αn , v, w) = 0. Given (ii), for some world, call
it v 0 , Vg (Πα1 . . . αn , v 0 , w) = 0, and so the ordered n-tuple of the
denotations of α1 . . . αn is not in the extension of Π at w, which
contradicts (i).
Now the inductive step. We must assume that φ and ψ obey our
statement, and show that complex formulas built from φ and ψ also
obey our statement. That is, we assume the inductive hypothesis:
and we must show that the following are also generally valid:
∼φ→F∼φ
(φ→ψ)→F(φ→ψ)
∀αφ→F∀αφ
2φ→F2φ
Fφ→FF φ
×φ→F ×φ
some u in the domain, Vgu/α (φ, v 0 , w) = 0. Given the former, Vgu/α (φ, v, w) =
1; given (ih) it follows that Vgu/α (Fφ, v, w) = 1, and so, Vgu/α (φ, v 0 , w) =
1. Contradiction.
2 : suppose (i) V(2φ, v, w) = 1 and (ii) V(F2φ, v, w) = 0, for some
v, w. From (ii), V(2φ, v 0 , w) = 0 for some v 0 , and so V(φ, v 0 , w 0 ) =
0 for some w 0 . Given (i), V(φ, v, w 0 ) = 1; and so, given (ih),
V(Fφ, v, w 0 ) = 1, and so V(φ, v 0 , w 0 ) = 1. Contradiction.
F: suppose V(Fφ, v, w) = 1 and V(FFφ, v, w) = 0, for some v, w.
From the latter, V(Fφ, v 0 , w) = 0 for some v 0 , and so V(φ, v 00 , w) =
0 for some v 00 , which contradicts the former.
×: suppose Vg (×φ, v, w) = 1 but Vg (F×φ, v, w) = 0, for some
v, w. Given the latter, Vg (×φ, v 0 , w) = 0 for some v 0 , and so
Vg (φ, w, w) = 0, which contradicts the former.
Hesperus = Phosphorus
B (the standard meter bar) is one meter long
Phosphorus are in fact one and the same entity—namely, the planet Venus—
there is no possible world in which Hesperus is different from Phosphorus, for
such a world would have to be a world in which Venus is distinct from itself.
Thus, the statement is necessary, despite its a posteriority: it took astronomical
evidence to learn that Hesperus and Phosphorus were identical; no amount
of pure rational reflection would have sufficed. As for the second statement,
Kripke argues that one can know its truth as soon as one knows that the phrase
‘one meter’ has its reference fixed by the description: “the length of bar B”.
Thus it is a priori. Nevertheless, he argues, it is contingent: bar B does not
have its length essentially, and thus could have been longer or shorter than one
meter.
On the face of it, the existence of necessary a priori or contingent a posteriori
statements is paradoxical. How can a statement that is true in all possible worlds
be in principle resistant to a priori investigation? Worse, how can a statement
that might have been false be known a priori?
The two-dimensional framework has been thought by some to shed light
on all this. Let’s consider the contingent a priori first. Let’s define the following
notion of contingency:
That was a relatively simple example; but one can give other examples that
are similar in spirit both to Kripke’s example of the meter bar, and to a related
example due to Gareth Evans (1979):
where bar B is the standard meter bar, and the “descriptive names” ‘one meter’
and ‘Julius’ are said to be “rigid designators” whose references are “fixed” by the
descriptions ‘the length of bar B’ and ‘the inventor of the zip’, respectively. Now,
whether or not these sentences, understood as sentences of everyday English,
are indeed genuinely contingent and a priori depends on delicate issues in the
philosophy of language concerning descriptive names, rigid designation, and
reference fixing. Rather than going into all that, let’s construct some examples
that are similar to Kripke’s and Evans’s. Let’s simply stipulate that ‘one meter’
and ‘Julius’ are to abbreviate “actualized descriptions”: ‘the actual length of bar
B’ and ‘the actual inventor of the zip’. With a little creative reconstruing in the
first case, the sentences then have the form: “the actual G is G”:
Now, these sentences are not quite a priori, since for all one knows, the G might
not exist—there might exist no unique length of bar B, no unique inventor of
the zip. So suppose we consider instead the following sentences:
If there is exactly one length of bar B, then the actual length of bar
B is a length of bar B
If there is exactly one inventor of the zip, then the actual inventor
of the zip invented the zip
Any sentence of this form is 2D-valid (though not generally 2D-valid), and is
superficially contingent. So we have further examples of the contingent a priori
in the neighborhood of the examples of Kripke and Evans.
Various philosophers want to concede that these sentences are contingent
in one sense—namely, in the sense of superficial contingency. But, they claim,
this is a relatively unimportant sense (hence the term ‘superficial contingency’,
which was coined by Evans). In another sense, they’re not contingent at all.
Evans calls the second sense of contingency “deep contingency”, and defines it
thus (1979, p. 185):
If a deeply contingent statement is true, there will exist some state of
affairs of which we can say both that had it not existed the statement
would not have been true, and that it might not have existed.
The intended meaning of ‘the statement would not have been true’ is that the
statement, as uttered with its actual meaning, would not have been true. The
idea is supposed to be that ‘Julius invented the zip’ is not deeply contingent,
because we can’t locate the required state of affairs, since in any situation in
which ‘Julius invented the zip’ is uttered with its actual meaning, it is uttered
truly. So the Julius example is not one of a deeply contingent a priori truth.
Evans’s notion of deep contingency is far from clear. One of the nice things
about the two-dimensional modal framework is that it allows us to give a
clear definition of deep contingency. Davies and Humberstone (1980) give a
definition of deep contingency which is parallel to the definition of superficial
contingency, but with F@ in place of 2:
Under this definition, the examples we have given are not deeply contingent.
To be sure, this definition is only as clear as the two-dimensional notions of
fixedness and actuality. The formal structure of the two-dimensional framework
is of course clear, but one can raise philosophical questions about how that
formalism is to be interpreted. But at least the formalism provides a clear
framework for the philosophical debate to occur.
Our discussion of the necessary a posteriori will be parallel to that of the
contingent a priori. Just as we defined superficial contingency as the falsity of
the 2, so we can define superficial necessity as the truth of the 2:
How shall we construe a posteriority? Let’s follow our earlier strategy, and take
the failure to be 2D-valid as our guide.
But here we must take a bit more care. It’s quite a trivial matter to construct
models in which 2D-invalid sentences are necessarily true; and we don’t need
the two-dimensional framework to do it. We clearly don’t want to say that
‘Everything is a lawyer” is an example of the necessary a posteriori. But let F
stand for ‘is a lawyer’; we can construct a model in which the predicate F is
true of every member of the domain at any world, ∀xF x is true, and so is
superficially necessary at every world, despite the fact that it is not 2D-valid.
But this is too cheap. We began by letting the predicate F stand for a predicate
of English, but then constructed our model without attending to the modal
fact that it’s simply not the case that it’s necessarily true that everything is a
lawyer. If F is indeed to stand for ‘is a lawyer’, we would need to include in any
realistic model—any model faithful to the modal facts—worlds in which not
everything is in the extension of F .
To provide nontrivial models of the necessary a posteriori, when we have
chosen to think of the nonlogical expressions of the language of QML as
standing for certain expressions of English, our strategy will be provide realistic
models—models that are faithful to the real modal facts in relevant respects,
given the choice of what the nonlogical expressions stand—in which 2D-invalid
sentences are necessarily true. Now, since the notion of a “realistic model”
has not been made precise, the argument here will be imprecise; but in the
circumstances this imprecision is inevitable.
So: consider now, as a schematic example of an a posteriori and superficially
necessary sentence:
If the actual F and the actual G exist, then they are identical
[∃x(@F x∧∀y(@F y→x=y)) ∧ ∃z(@Gz∧∀y(@Gy→z=y)] →
∃x[@F x∧∀y(@F y→x=y) ∧ ∃z(@Gz∧∀y(@Gy→z=y) ∧ z=x)]
be short for ‘the actual F ’, and that Phosphorus is to be short for ‘the actual
G’. And let’s think of F as standing for ‘the first heavenly body visible in the
evening’, and G for ‘the last heavenly body visible in the morning’. Then
has the form ‘If the actual F and the actual G exist then they are identical’,
which was discussed in the previous paragraph. We may then construct a
realistic model in which F and G each have a single object in their extension in
some world, w, but in which they have different objects in their extensions in
other worlds. In such a model, the sentence
(2HP) 2(If Hesperus and Phosphorus exist then they are identical)
is true at 〈w, w〉, and so we again have our desired example: (HP) is superficially
necessary, despite the fact that it is a posteriori (2D-invalid).
Isn’t it strange that (HP) is both a posteriori and necessary? The two-
dimensional response is: no, it’s not, since although it is superficially necessary,
it isn’t deeply necessary in the following sense:
It isn’t deeply necessary because in any realistic model (given what F and G
currently stand for), there must be worlds and objects other than c and u, that
are configured as they are in the model below:
W = {c, d}
D = {u, v}
I (F ) = {〈u, c〉, 〈u, d〉}
I (G) = {〈u, c〉, 〈v, d〉}
In this model, even though (2HP) is false at 〈c, c〉, still, F@(HP), i.e.:
is false at 〈c, c〉 (and indeed, at every pair of worlds), since (HP) is false at 〈d, d〉.
And so, (HP) is not deeply necessary in this model.
One might try to take this two-dimensional line further, and claim that in
every case of the necessary a posteriori (or the contingent a priori), the necessity
(contingency) is merely superficial. But defending this stronger line would
require more than we have in place so far. To take one example, return again
to ‘Hesperus = Phosphorus’, but now, instead of thinking of ‘Hesperus’ and
‘Phosphorus’ as abbreviations for actualized descriptions, let us represent them
by names in the logical sense (i.e., the expressions called “names” in the defini-
tion of well-formed formulas, which are assigned denotations by interpretation
functions in models). Thus, ‘Hesperus=Phosphorus’ is now represented as:
a=b . Consider the following model:
W = {c, d}
D = {u, v}
I (a) = u
I (b ) = u
The model is apparently realistic; it falsifies no relevant modal facts. But the
sentence a=b is deeply necessary (at any world in the model). And yet it is a
posteriori (2D-invalid).
Bibliography
— (1998). Logic, Logic, and Logic. Cambridge, MA: Harvard University Press.
Boolos, George and Richard Jeffrey (1989). Computability and Logic. 3rd edition.
Cambridge: Cambridge University Press.
Chalmers, David (1996). The Conscious Mind. Oxford: Oxford University Press.
Cresswell, M.J. and G.E. Hughes (1996). A New Introduction to Modal Logic.
London: Routledge.
272
BIBLIOGRAPHY 273
Enderton, Herbert (1977). Elements of Set Theory. New York: Academic Press.
Evans, Gareth (1979). “Reference and Contingency.” The Monist 62: 161–189.
Reprinted in Evans 1985.
— (1991b). Logic, Language, and Meaning, Volume 2: Intensional Logic and Logical
Grammar. Chicago: University of Chicago Press.
Harper, William L., Robert Stalnaker and Glenn Pearce (eds.) (1981). Ifs: Con-
ditionals, Belief, Decision, Chance, and Time. Dordrecht: D. Reidel Publishing
Company.
Hodes, Harold (1984a). “On Modal Logics Which Enrich First-order S5.”
Journal of Philosophical Logic 13: 423–454.
BIBLIOGRAPHY 274
Kripke, Saul (1972). “Naming and Necessity.” In Donald Davidson and Gilbert
Harman (eds.), Semantics of Natural Language, 253–355, 763–769. Dordrecht:
Reidel. Revised edition published in 1980 as Naming and Necessity (Cambridge,
MA: Harvard University Press).
Linsky, Bernard and Edward N. Zalta (1994). “In Defense of the Simplest
Quantified Modal Logic.” In James Tomberlin (ed.), Philosophical Perspectives
8: Logic and Language, 431–458. Atascadero: Ridgeview.
Prior, A. N. (1967). Past, Present, and Future. Oxford: Oxford University Press.
— (1956). Logic and Knowledge. Ed. Robert Charles Marsh. New York: G.P.
Putnam’s Sons.
Soames, Scott (2004). Reference and Description: The Case against Two-
Dimensionalism. Princeton: Princeton University Press.
— (1978). “Assertion.” In Peter Cole and Jerry Morgan (eds.), Syntax and Se-
mantics, Volume 9: Pragmatics, 315–332. New York: Academic Press. Reprinted
in Stalnaker 1999: 78–95.